Add Batch 0a4a2b80-9072-433a-8a67-c3f068517452 data

.gitattributes

@@ -3577,3 +3577,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2023/PreSTU_[[:space:]]Pre-Training[[:space:]]for[[:space:]]Scene-Text[[:space:]]Understanding/c5c18218-53d8-4a71-9148-4d0d43178fa9_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Predict[[:space:]]to[[:space:]]Detect_[[:space:]]Prediction-guided[[:space:]]3D[[:space:]]Object[[:space:]]Detection[[:space:]]using[[:space:]]Sequential[[:space:]]Images/fcbf30f9-f195-4f61-9083-e1dbafc732e8_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Preface_[[:space:]]A[[:space:]]Data-driven[[:space:]]Volumetric[[:space:]]Prior[[:space:]]for[[:space:]]Few-shot[[:space:]]Ultra[[:space:]]High-resolution[[:space:]]Face[[:space:]]Synthesis/904af390-9b66-44f1-b757-c369bcb42d27_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Preparing[[:space:]]the[[:space:]]Future[[:space:]]for[[:space:]]Continual[[:space:]]Semantic[[:space:]]Segmentation/72d5919b-3823-475f-9939-cac5c29b48a5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Preserve[[:space:]]Your[[:space:]]Own[[:space:]]Correlation_[[:space:]]A[[:space:]]Noise[[:space:]]Prior[[:space:]]for[[:space:]]Video[[:space:]]Diffusion[[:space:]]Models/12c6fb66-a455-4547-8b5a-4a378b5ee51f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Preserving[[:space:]]Modality[[:space:]]Structure[[:space:]]Improves[[:space:]]Multi-Modal[[:space:]]Learning/9786f414-b2fe-4822-86c7-2868d21a76fe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Preserving[[:space:]]Tumor[[:space:]]Volumes[[:space:]]for[[:space:]]Unsupervised[[:space:]]Medical[[:space:]]Image[[:space:]]Registration/dbd41759-9a48-41b9-9206-f4ea354bc2a2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Pretrained[[:space:]]Language[[:space:]]Models[[:space:]]as[[:space:]]Visual[[:space:]]Planners[[:space:]]for[[:space:]]Human[[:space:]]Assistance/057af23b-bb0a-449d-acad-c434ed40a938_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Preventing[[:space:]]Zero-Shot[[:space:]]Transfer[[:space:]]Degradation[[:space:]]in[[:space:]]Continual[[:space:]]Learning[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models/438697ca-0d24-4663-a1e4-f11b5f684ada_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Prior-guided[[:space:]]Source-free[[:space:]]Domain[[:space:]]Adaptation[[:space:]]for[[:space:]]Human[[:space:]]Pose[[:space:]]Estimation/67d12eb1-f1e4-4b90-aba2-4771e3bf7c39_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Priority-Centric[[:space:]]Human[[:space:]]Motion[[:space:]]Generation[[:space:]]in[[:space:]]Discrete[[:space:]]Latent[[:space:]]Space/58c7e478-f288-424e-a702-051324564633_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Privacy[[:space:]]Preserving[[:space:]]Localization[[:space:]]via[[:space:]]Coordinate[[:space:]]Permutations/f992eeb4-2e60-42f9-9e28-abc98f428b2a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Privacy-Preserving[[:space:]]Face[[:space:]]Recognition[[:space:]]Using[[:space:]]Random[[:space:]]Frequency[[:space:]]Components/3da3fc63-806e-4e3e-99b2-e2f33c132bad_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ProPainter_[[:space:]]Improving[[:space:]]Propagation[[:space:]]and[[:space:]]Transformer[[:space:]]for[[:space:]]Video[[:space:]]Inpainting/447df786-8aae-4bc4-b7d5-1e90caa96cc1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ProbVLM_[[:space:]]Probabilistic[[:space:]]Adapter[[:space:]]for[[:space:]]Frozen[[:space:]]Vison-Language[[:space:]]Models/52935457-f9bc-4c99-866c-671f6012153b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Probabilistic[[:space:]]Human[[:space:]]Mesh[[:space:]]Recovery[[:space:]]in[[:space:]]3D[[:space:]]Scenes[[:space:]]from[[:space:]]Egocentric[[:space:]]Views/c490ef67-0a1e-4595-8f34-5035e971ee89_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Probabilistic[[:space:]]Modeling[[:space:]]of[[:space:]]Inter-[[:space:]]and[[:space:]]Intra-observer[[:space:]]Variability[[:space:]]in[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation/bf4cff8f-5588-469f-97d0-102dd2317c9c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Probabilistic[[:space:]]Precision[[:space:]]and[[:space:]]Recall[[:space:]]Towards[[:space:]]Reliable[[:space:]]Evaluation[[:space:]]of[[:space:]]Generative[[:space:]]Models/dd6377f8-01ac-488f-b3b2-5ae1ebd477f1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Probabilistic[[:space:]]Triangulation[[:space:]]for[[:space:]]Uncalibrated[[:space:]]Multi-View[[:space:]]3D[[:space:]]Human[[:space:]]Pose[[:space:]]Estimation/b6b4e04e-10fb-49dc-be18-fbc2e2b53da9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Progressive[[:space:]]Spatio-Temporal[[:space:]]Prototype[[:space:]]Matching[[:space:]]for[[:space:]]Text-Video[[:space:]]Retrieval/480eca60-b42e-4764-a2ee-88684bf2c389_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Prompt[[:space:]]Switch_[[:space:]]Efficient[[:space:]]CLIP[[:space:]]Adaptation[[:space:]]for[[:space:]]Text-Video[[:space:]]Retrieval/408eb595-fd0b-42c6-8600-a2efb5916c5c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Prompt[[:space:]]Tuning[[:space:]]Inversion[[:space:]]for[[:space:]]Text-driven[[:space:]]Image[[:space:]]Editing[[:space:]]Using[[:space:]]Diffusion[[:space:]]Models/f6450b64-da9d-48f8-bcd6-4de4863bfc2b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Prompt-aligned[[:space:]]Gradient[[:space:]]for[[:space:]]Prompt[[:space:]]Tuning/7786570f-2ec2-4f85-a04c-e553c7832bc6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PromptCap_[[:space:]]Prompt-Guided[[:space:]]Image[[:space:]]Captioning[[:space:]]for[[:space:]]VQA[[:space:]]with[[:space:]]GPT-3/4ffd89d6-8734-4748-beec-1543acfdd6d0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PromptStyler_[[:space:]]Prompt-driven[[:space:]]Style[[:space:]]Generation[[:space:]]for[[:space:]]Source-free[[:space:]]Domain[[:space:]]Generalization/f8d2a45d-4c84-4d88-913c-50e0d9fbe4a8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ProtoFL_[[:space:]]Unsupervised[[:space:]]Federated[[:space:]]Learning[[:space:]]via[[:space:]]Prototypical[[:space:]]Distillation/c82e3c6e-4735-400a-b460-81642c25b787_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ProtoTransfer_[[:space:]]Cross-Modal[[:space:]]Prototype[[:space:]]Transfer[[:space:]]for[[:space:]]Point[[:space:]]Cloud[[:space:]]Segmentation/4c052c14-482f-4c6c-bea8-5d4ef790aea8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Prototype[[:space:]]Reminiscence[[:space:]]and[[:space:]]Augmented[[:space:]]Asymmetric[[:space:]]Knowledge[[:space:]]Aggregation[[:space:]]for[[:space:]]Non-Exemplar[[:space:]]Class-Incremental[[:space:]]Learning/f06e28ba-39c2-463d-9177-75781355ca62_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Prototype-based[[:space:]]Dataset[[:space:]]Comparison/02a62e7f-ba00-4044-92de-55f8e976a093_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Prototypes-oriented[[:space:]]Transductive[[:space:]]Few-shot[[:space:]]Learning[[:space:]]with[[:space:]]Conditional[[:space:]]Transport/e3d7c832-dffc-42b7-b29a-faccf7386804_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Prototypical[[:space:]]Kernel[[:space:]]Learning[[:space:]]and[[:space:]]Open-set[[:space:]]Foreground[[:space:]]Perception[[:space:]]for[[:space:]]Generalized[[:space:]]Few-shot[[:space:]]Semantic[[:space:]]Segmentation/ae5d3fad-adfb-4c23-912c-f33a2a374a40_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Prototypical[[:space:]]Mixing[[:space:]]and[[:space:]]Retrieval-Based[[:space:]]Refinement[[:space:]]for[[:space:]]Label[[:space:]]Noise-Resistant[[:space:]]Image[[:space:]]Retrieval/ec71d51a-9735-4817-bbda-a4f4fe8a6bfa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Proxy[[:space:]]Anchor-based[[:space:]]Unsupervised[[:space:]]Learning[[:space:]]for[[:space:]]Continuous[[:space:]]Generalized[[:space:]]Category[[:space:]]Discovery/82f1fc34-be8f-46cf-8cc9-6e1e6fe0df48_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Prune[[:space:]]Spatio-temporal[[:space:]]Tokens[[:space:]]by[[:space:]]Semantic-aware[[:space:]]Temporal[[:space:]]Accumulation/6db1efd1-b765-4bc8-8a5a-d2a425ae5e95_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Pseudo[[:space:]]Flow[[:space:]]Consistency[[:space:]]for[[:space:]]Self-Supervised[[:space:]]6D[[:space:]]Object[[:space:]]Pose[[:space:]]Estimation/cf29abfd-328a-4679-9b7c-7340065b4fa2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Pseudo-label[[:space:]]Alignment[[:space:]]for[[:space:]]Semi-supervised[[:space:]]Instance[[:space:]]Segmentation/42aab451-44d8-431c-bd3b-859749553f79_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Pyramid[[:space:]]Dual[[:space:]]Domain[[:space:]]Injection[[:space:]]Network[[:space:]]for[[:space:]]Pan-sharpening/aa199a66-751b-4158-bd1f-1783173c94fd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Q-Diffusion_[[:space:]]Quantizing[[:space:]]Diffusion[[:space:]]Models/64ad5dff-59b7-42a2-a4ac-ca7c9541d350_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/QD-BEV[[:space:]]_[[:space:]]Quantization-aware[[:space:]]View-guided[[:space:]]Distillation[[:space:]]for[[:space:]]Multi-view[[:space:]]3D[[:space:]]Object[[:space:]]Detection/ce2686ce-9d14-4ea8-b9f7-4395d3561b65_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Quality[[:space:]]Diversity[[:space:]]for[[:space:]]Visual[[:space:]]Pre-Training/3f69c090-f906-4d45-8423-292cfc679eb9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Quality-Agnostic[[:space:]]Deepfake[[:space:]]Detection[[:space:]]with[[:space:]]Intra-model[[:space:]]Collaborative[[:space:]]Learning/77142358-08da-4d47-9173-9c11400a028e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Query[[:space:]]Refinement[[:space:]]Transformer[[:space:]]for[[:space:]]3D[[:space:]]Instance[[:space:]]Segmentation/8435ea40-79e7-4103-b1c6-be175c97058d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Query6DoF_[[:space:]]Learning[[:space:]]Sparse[[:space:]]Queries[[:space:]]as[[:space:]]Implicit[[:space:]]Shape[[:space:]]Prior[[:space:]]for[[:space:]]Category-Level[[:space:]]6DoF[[:space:]]Pose[[:space:]]Estimation/4d1aafad-a31c-4773-aac8-62fd2260bfdc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/R-Pred_[[:space:]]Two-Stage[[:space:]]Motion[[:space:]]Prediction[[:space:]]Via[[:space:]]Tube-Query[[:space:]]Attention-Based[[:space:]]Trajectory[[:space:]]Refinement/cbd955af-2686-4f7c-843a-a3576e04bc6e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/R3D3_[[:space:]]Dense[[:space:]]3D[[:space:]]Reconstruction[[:space:]]of[[:space:]]Dynamic[[:space:]]Scenes[[:space:]]from[[:space:]]Multiple[[:space:]]Cameras/df3f55bd-3906-48f7-856c-09ad31d9f51b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RANA_[[:space:]]Relightable[[:space:]]Articulated[[:space:]]Neural[[:space:]]Avatars/ae3b789d-a192-4696-b97f-daa6528a52ba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RCA-NOC_[[:space:]]Relative[[:space:]]Contrastive[[:space:]]Alignment[[:space:]]for[[:space:]]Novel[[:space:]]Object[[:space:]]Captioning/cbdb5e43-66b1-4704-93d9-a919fc74c296_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/REAP_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Realistic[[:space:]]Adversarial[[:space:]]Patch[[:space:]]Benchmark/451492ac-a33d-4658-b61d-707c2258ba9b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RED-PSM_[[:space:]]Regularization[[:space:]]by[[:space:]]Denoising[[:space:]]of[[:space:]]Partially[[:space:]]Separable[[:space:]]Models[[:space:]]for[[:space:]]Dynamic[[:space:]]Imaging/ffd9f980-64f4-43d0-a9f4-e03ede33c976_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RFD-ECNet_[[:space:]]Extreme[[:space:]]Underwater[[:space:]]Image[[:space:]]Compression[[:space:]]with[[:space:]]Reference[[:space:]]to[[:space:]]Feature[[:space:]]Dictionary/cdd3e31c-4266-47f7-a8c7-755cca493958_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RFLA_[[:space:]]A[[:space:]]Stealthy[[:space:]]Reflected[[:space:]]Light[[:space:]]Adversarial[[:space:]]Attack[[:space:]]in[[:space:]]the[[:space:]]Physical[[:space:]]World/bdd5612c-711f-444e-bf67-fd043897ebe9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RICO_[[:space:]]Regularizing[[:space:]]the[[:space:]]Unobservable[[:space:]]for[[:space:]]Indoor[[:space:]]Compositional[[:space:]]Reconstruction/9a248d4c-688f-46d5-8671-0975c0bef03f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RIGID_[[:space:]]Recurrent[[:space:]]GAN[[:space:]]Inversion[[:space:]]and[[:space:]]Editing[[:space:]]of[[:space:]]Real[[:space:]]Face[[:space:]]Videos/0ac9c66e-6779-4caf-9449-892ceb58d43e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RLIPv2_[[:space:]]Fast[[:space:]]Scaling[[:space:]]of[[:space:]]Relational[[:space:]]Language-Image[[:space:]]Pre-Training/dedbf300-120b-4912-a26e-f96e152258fe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RLSAC_[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]Enhanced[[:space:]]Sample[[:space:]]Consensus[[:space:]]for[[:space:]]End-to-End[[:space:]]Robust[[:space:]]Estimation/253f2fc2-5243-41a5-8404-4467c420f19c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RMP-Loss_[[:space:]]Regularizing[[:space:]]Membrane[[:space:]]Potential[[:space:]]Distribution[[:space:]]for[[:space:]]Spiking[[:space:]]Neural[[:space:]]Networks/8dd5399b-d34f-43a1-82f0-9e644d0103b9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ROME_[[:space:]]Robustifying[[:space:]]Memory-Efficient[[:space:]]NAS[[:space:]]via[[:space:]]Topology[[:space:]]Disentanglement[[:space:]]and[[:space:]]Gradient[[:space:]]Accumulation/6ade4227-5686-42c9-830e-da85efbc5dca_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RPEFlow_[[:space:]]Multimodal[[:space:]]Fusion[[:space:]]of[[:space:]]RGB-PointCloud-Event[[:space:]]for[[:space:]]Joint[[:space:]]Optical[[:space:]]Flow[[:space:]]and[[:space:]]Scene[[:space:]]Flow[[:space:]]Estimation/d027f499-898c-464c-acf6-a1eba174840a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RPG-Palm_[[:space:]]Realistic[[:space:]]Pseudo-data[[:space:]]Generation[[:space:]]for[[:space:]]Palmprint[[:space:]]Recognition/6f1b3792-7545-4d6e-8992-4404f77414a4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RSFNet_[[:space:]]A[[:space:]]White-Box[[:space:]]Image[[:space:]]Retouching[[:space:]]Approach[[:space:]]using[[:space:]]Region-Specific[[:space:]]Color[[:space:]]Filters/a1a06f38-ffb3-49f3-88a3-88bb8d7af4de_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Random[[:space:]]Boxes[[:space:]]Are[[:space:]]Open-world[[:space:]]Object[[:space:]]Detectors/99919d69-2dc0-454c-a4b9-63b911115db6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Random[[:space:]]Sub-Samples[[:space:]]Generation[[:space:]]for[[:space:]]Self-Supervised[[:space:]]Real[[:space:]]Image[[:space:]]Denoising/0ac3ef7d-9ebc-4369-9f12-6e7b6e92e896_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Randomized[[:space:]]Quantization_[[:space:]]A[[:space:]]Generic[[:space:]]Augmentation[[:space:]]for[[:space:]]Data[[:space:]]Agnostic[[:space:]]Self-supervised[[:space:]]Learning/35e2ad13-e7ee-4623-a3bb-087851f2da1d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RankMatch_[[:space:]]Fostering[[:space:]]Confidence[[:space:]]and[[:space:]]Consistency[[:space:]]in[[:space:]]Learning[[:space:]]with[[:space:]]Noisy[[:space:]]Labels/0d715541-80cd-458e-81f4-423a1c46d83e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/RankMixup_[[:space:]]Ranking-Based[[:space:]]Mixup[[:space:]]Training[[:space:]]for[[:space:]]Network[[:space:]]Calibration/4c024ec3-bba1-4470-bd07-d0e7529be938_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Rapid[[:space:]]Adaptation[[:space:]]in[[:space:]]Online[[:space:]]Continual[[:space:]]Learning_[[:space:]]Are[[:space:]]We[[:space:]]Evaluating[[:space:]]It[[:space:]]Right_/165500ab-b2c3-4db0-ba04-4570a2ad4282_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Rapid[[:space:]]Network[[:space:]]Adaptation_[[:space:]]Learning[[:space:]]to[[:space:]]Adapt[[:space:]]Neural[[:space:]]Networks[[:space:]]Using[[:space:]]Test-Time[[:space:]]Feedback/32896e91-9d8e-4427-a165-f57c45c9afec_origin.pdf filter=lfs diff=lfs merge=lfs -text

2023/Preparing the Future for Continual Semantic Segmentation/72d5919b-3823-475f-9939-cac5c29b48a5_content_list.json

@@ -0,0 +1,1538 @@
+[
+    {
+        "type": "text",
+        "text": "Preparing the Future for Continual Semantic Segmentation",
+        "text_level": 1,
+        "bbox": [
+            184,
+            130,
+            785,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zihan Lin<sup>1</sup> Zilei Wang<sup>*1</sup> Yixin Zhang<sup>1,2</sup>",
+        "bbox": [
+            305,
+            179,
+            663,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup> University of Science and Technology of China",
+        "bbox": [
+            292,
+            198,
+            678,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$  Institute of Artificial Intelligence, Hefei Comprehensive National Science Center",
+        "bbox": [
+            158,
+            215,
+            808,
+            234
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "myustc@mail.ustc.edu.cn, {zlwang, zhyx12}@ustc.edu.cn",
+        "bbox": [
+            250,
+            236,
+            718,
+            251
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            286,
+            313,
+            301
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this study, we focus on Continual Semantic Segmentation (CSS) and present a novel approach to tackle the issue of existing methods struggling to learn new classes. The primary challenge of CSS is to learn new knowledge while retaining old knowledge, which is commonly known as the rigidity-plasticity dilemma. Existing approaches strive to address this by carefully balancing the learning of new and old classes during training on new data. Differently, this work aims to avoid this dilemma fundamentally rather than handling the difficulties involved in it. Specifically, we reveal that this dilemma mainly arises from the greater fluctuation of knowledge for new classes because they have never been learned before the current step. Additionally, the data available in incremental steps are usually inadequate, which can impede the model's ability to learn discriminative features for both new and old classes. To address these challenges, we introduce a novel concept of pre-learning for future knowledge. Our approach entails optimizing the feature space and output space for unlabeled data, which thus enables the model to acquire knowledge for future classes. With this approach, updating the model for new classes becomes as smooth as for old classes, effectively avoiding the rigidity-plasticity dilemma. We conducted extensive experiments and the results demonstrate a significant improvement in the learning of new classes compared to previous state-of-the-art methods.",
+        "bbox": [
+            75,
+            316,
+            472,
+            710
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            739,
+            209,
+            753
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Deep neural networks have demonstrated their superiority in many computer vision tasks [26, 31, 14]. Conventionally, they are trained in an offline manner with all data collected beforehand. When novel classes need to be handled, the networks need to be re-trained using either updated training set with both new and previous data, or directly finetuned on new data. The former has higher costs and is even impracticable sometimes like, when privacy raises concerns. The latter is prone to forget previ",
+        "bbox": [
+            75,
+            763,
+            468,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/301e91d65d39335f139ddfa2115135c85334d75e9536194ad13a2b6c5bd7c493.jpg",
+        "image_caption": [
+            "Figure 1. On VOC2012 benchmark, our method boosts the performance of new classes by a significant margin, while also achieving promising performance for old classes, compared to recent state-of-the-art methods."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            292,
+            870,
+            462
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ously learned knowledge, which is called catastrophic forgetting (CF) [25]. To better fit neural networks into real-world scenarios, we want them to have the ability to learn new concepts while preserving the old knowledge (or even improve it using new knowledge). This is known as the continual learning (CL) problem and receives increasing attention recently. It has been explored first in the field of image classification [28, 42, 17, 11, 57, 44, 43], and some methods are proposed recently targeting semantic segmentation [5, 10, 6, 54, 51, 37]. Generally, it is often achieved by applying forgetting-preventing constraints between current and previous networks to prevent CF, which raises the rigidity (i.e., the ability to preserve old states) but harms the plasticity (i.e., the ability to learn new knowledge). It is known as the rigidity-plasticity (R-P) dilemma. A key aspect in designing constraints is to balance between rigidity and plasticity, which is what most works focus on.",
+        "bbox": [
+            496,
+            549,
+            893,
+            806
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite that several new constraints have been proposed for CSS to better tackle the dilemma, existing works still struggle to learn new classes. Taking the VOC2012 benchmark as an example, most methods can only achieve a performance around  $70\\%$  of that of the upperbound for 15-5 setup and around  $50\\%$  for 19-1 setup, which is even worse",
+        "bbox": [
+            496,
+            809,
+            893,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "11910",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "for harder setups. To get a clear picture of why this happens, we first analyze the intermediate features of the network and find it suffers serious underfitting for new classes, which indicates a large amount of knowledge is not learned. We then examine the CSS training pipeline and find two main obstacles causing this phenomenon. (1) R-P Dilemma: the network has no knowledge w.r.t. new classes until they are introduced. This generally requires a rather large update to incorporate the new knowledge, which will be significantly hindered by the forgetting-preventing constraints punishing the network from drifting away from its previous states. (2) Inadequate data for training: the amount of data available in incremental steps is usually limited, and this problem becomes more severe when a single class is added, resulting in a tiny training set with only a few negative samples. This can significantly limit the network to properly discriminate between old and new classes. This problem has never been explicitly considered before.",
+        "bbox": [
+            76,
+            90,
+            472,
+            362
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Unlike existing works trying to tackle the R-P dilemma directly, we are the first to argue that this obstacle can be elegantly avoided by exploring future knowledge in an unsupervised way. Assume that the network has obtained partial knowledge of future classes in advance, then the network needs fewer updates to learn the upcoming knowledge, which consequently lowers the impacts of the R-P dilemma. The more future knowledge it learns in advance, the fewer the impacts will be caused. Considering an extreme situation where all the future knowledge is learned beforehand, the network does not need any update, and thus the influence of constraints could be totally avoided. On top of this, it makes us the first attempt to explicitly tackle the second obstacle by utilizing the abundant data in previous steps to optimize discriminability between old and new classes. Based on these understandings, we design a novel framework where unsupervised contrastive learning is performed on intermediate features to optimize future classes with both visual similarity and feature affinity as supervisions. One step further, an auxiliary classifier is trained using pseudo labels of future classes generated via clustering, which is used to initialize the actual classifier for future steps. Our method surpasses existing works by a huge margin as shown in Figure 1 and the effectiveness is further demonstrated with extensive experiments. Our contributions are summarized as follows.",
+        "bbox": [
+            76,
+            363,
+            472,
+            755
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose to pre-learn future knowledge for CSS with the aim of improving the performance especially for new classes, which can simultaneously tackle the R-P dilemma and the obstacle of inadequate data.",
+            "- We design a novel framework that can pre-learn future knowledge in an unsupervised way, where both the feature space and output space are handled.",
+            "- We have validated the effectiveness of our proposed"
+        ],
+        "bbox": [
+            76,
+            762,
+            468,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "approach through extensive experiments. The results demonstrate significant improvement on standard CSS benchmarks.",
+        "bbox": [
+            511,
+            90,
+            890,
+            137
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            152,
+            640,
+            167
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Continual Learning (CL). Continual learning has been studied for years in image classification. Existing techniques can be roughly divided into regularization methods, rehearsal methods, and architectural methods. Regularization methods apply consistency constraints between current network and previous network to avoid drifts from previous states. They can be applied on network weights [53, 25, 1], the intermediate features [9, 11, 45, 13, 22] or the output logits [28, 42, 17, 48]. Parameter regularization is rarely seen recently due to the lack of plasticity, while the other two are mostly based on knowledge distillation (KD) [16]. To better solve the forgetting problem, rehearsal methods replay previous data in the incremental steps, including image replay [42, 41, 3], generative replay [49, 39], feature replay [20, 57], and memory management [30, 29]. Architectural methods use a dynamic network to handle incoming knowledge, like sub-network [32, 33], complementary network [47, 40] or expandable network [27, 50, 52].",
+        "bbox": [
+            496,
+            178,
+            890,
+            450
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Continual Semantic Segmentation (CSS). Some works extended the aforementioned techniques to semantic segmentation tasks. For the KD-based methods, MiB [5] formalized and targeted background shift problem with output space distillation. PLOP [10], SDR [36], and RCIL [54] use feature space distillation to further mitigate forgetting. For the rehearsal-based methods, SSUL [6] and RECALL [34] utilize memory or external data to overcome forgetting and work well in multiple-step learning. For the architectural methods, RCIL [54] proposes a structural reparameterization design to better learn new knowledge.",
+        "bbox": [
+            496,
+            450,
+            890,
+            618
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Unsupervised Semantic Segmentation. These works focus on learning to cluster semantically related pixels without human labeling. Existing works can be roughly divided into three types. PiCIE [8] iteratively performs clustering and uses it to refine network features. This relies on the architectural prior and is prone to degenerate solutions. IIC [21] maximizes the mutual information between two augmentations of the same image which enforces spatial stability on them. But this solution has trouble dealing with complex scenes. The last type uses visual similarity. These works [19, 55, 46, 23] use handcrafted priors, like boundary, superpixels, or saliency detectors to first group pixels by visual similarity. This is then used to learn pixel-level representations using contrastive learning.",
+        "bbox": [
+            496,
+            618,
+            890,
+            829
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3. Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            500,
+            844,
+            638,
+            859
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we first explain the definition of CSS (Sec. 3.1), and present the underfitting phenomenon of new",
+        "bbox": [
+            500,
+            869,
+            890,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "11911",
+        "bbox": [
+            480,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/7c464a504f500bccef080c7376cf9f4d8e5a117b9e37411e79d55f94d75c2aa6.jpg",
+        "image_caption": [
+            "Figure 2. Feature fitness analysis of different methods using L1 norm (Left), activation strength (Mid), and cosine similarity (Right) on VOC2012 15-5 setup. It is demonstrated that existing works suffer underfitting on both intra-class level (left and mid) and inter-class level (right) for new classes. Our method effectively addresses this problem."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            102,
+            344,
+            214
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/196888f1c316e672db90607467f10607b1ec82a221c00885053b765c22d1160e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            359,
+            102,
+            609,
+            215
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/4f84dd7ca2e2af2a43674bb3157b812ed37d730691a0015ce1dc4f2aa1c259c0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            625,
+            102,
+            872,
+            215
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "classes (Sec. 3.2). After that, we analyze how the performance is impacted in CSS (Sec. 3.3). Finally, we discuss the solutions to boost new class performance (Sec. 3.4).",
+        "bbox": [
+            75,
+            284,
+            470,
+            330
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. CSS Definition",
+        "text_level": 1,
+        "bbox": [
+            76,
+            340,
+            227,
+            357
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In CSS, the training procedure is composed of a series of  $T$  learning steps. At each step  $t$  where  $t \\in \\{1, \\dots, T\\}$ , a set of new categories  $\\mathcal{C}_t$  are introduced with a training set  $D_t = \\{(x_n, y_n)\\}$  where  $x_n$  is an image and  $y_n$  is the corresponding label. Only  $\\mathcal{C}_t$  is labeled in  $y_n$  and all the other classes are labeled as background. It is assumed that the categories learned at each step are disjoint, i.e.,  $\\mathcal{C}_i \\cap \\mathcal{C}_j = \\emptyset$ . The network is trained sequentially on each step, where it learns the knowledge w.r.t.  $\\mathcal{C}_t$  and expands its output space to  $\\mathcal{C}_{1:t} = \\mathcal{C}_{1:t-1} \\cup \\mathcal{C}_t$ . After all learning steps, the network is expected to perform well on all classes.",
+        "bbox": [
+            75,
+            364,
+            470,
+            532
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Analysis of Feature Fitness",
+        "text_level": 1,
+        "bbox": [
+            76,
+            542,
+            320,
+            559
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To get a clear view of why new classes perform poorly in existing methods, we analyze the intermediate features of two works: MiB [5] based on conventional logits distillation, and a more recent work, RCIL [54], based on feature distillation and structural re-parameterization aiming to better tackle the R-P dilemma. To simplify the situation, we opt for the two-step 15-5 setup on VOC2012 to conduct the analysis. The results are reported in Figure 2 (Detailed information about incremental setups and validation set can be found in Section 5.). The results of joint training (i.e., performance upperbound) and initial state (i.e., backbone is initialized with ImageNet pre-trained weights while the other parts are randomly initialized) are also provided for reference. Ideally, the results are expected to be close to that of the upperbound in order to achieve good performance.",
+        "bbox": [
+            75,
+            566,
+            468,
+            794
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Intra-class Fitness. We first evaluate how well the network fits to each individual class. In neural networks, each channel in the feature vectors could represent a certain pattern [4] which composes the learned classes [58]. When the network fits a class well, the output feature vector should have specific channels being activated while the others deactivated, resulting in sparsity in the feature vector. The",
+        "bbox": [
+            75,
+            795,
+            470,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "activation strength further shows how strong the network's response is to each pattern. Both of them can serve as clues indicating how well the network fits a class. We use the average L1 norm of L2-normalized feature vectors to measure the sparsity, and the average top activation values to measure the response strength. As can be seen in Figure 2 (left and mid), the new classes of both MiB and RCIL have high L1 norm and low activation strength, indicating a serious underfitting problem.",
+        "bbox": [
+            496,
+            284,
+            893,
+            419
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Inter-class Fitness. We further evaluate how well the network learns to distinguish different classes, which is measured by the cosine similarity between the feature vectors of different classes. The better the network can distinguish each class, the lower the cosine similarities will be. We compute the mean cosine distance of a class to any other classes, which is then averaged based on class group to get the final results, as shown in Figure 2 (right). It can be seen that the network has good discriminative power on old classes, but the new classes again present poor results. It indicates the network cannot distinguish a new class between other classes well.",
+        "bbox": [
+            496,
+            421,
+            892,
+            602
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Discussion. From the above analyses, it is evident that new classes suffer serious underfitting on both intra-class and inter-class aspects. This exactly explains why most methods show low performance on new classes, even in simple setups with only a single incremental step.",
+        "bbox": [
+            496,
+            604,
+            893,
+            681
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3. Understanding Training Pipeline",
+        "text_level": 1,
+        "bbox": [
+            500,
+            693,
+            790,
+            709
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this part, we dive deeper to investigate why the underfitting problem happens for new classes in the conventional training pipeline. We use Figure 3 to give a clear view of how the two obstacles impact the performance in a conventional pipeline (green background). Without loss of generality, a two-step situation is adopted as an example.",
+        "bbox": [
+            496,
+            718,
+            893,
+            808
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Starting from step 1 where  $\\mathcal{C}_1$  is labeled in the given data  $D_{1}$  while others are marked as background, the network acquires most knowledge of  $\\mathcal{C}_1$  (i.e., distinguish between classes in  $\\mathcal{C}_1$ ) after finishing this step. Considering that it only accesses partial data at this step compared to joint training, the knowledge of  $\\mathcal{C}_1$  remains to be updated",
+        "bbox": [
+            496,
+            810,
+            893,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "11912",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/a284cd579c0221d99adab4bb6e6c37940eaaaffa4cfdeabfc8de1a28bfe70af4.jpg",
+        "image_caption": [
+            "Figure 3. Illustration of the difference between knowledge learning in a conventional training pipeline (green background) and our solution (blue background). We jointly optimize new classes together with old classes in an unsupervised way before new classes are introduced."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            89,
+            460,
+            265
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "in the next step. However, even though the future classes  $\\mathcal{C}_2$  do exist in  $D_{1}$ , they do not receive explicit supervision, thus the network obtains almost no knowledge of  $\\mathcal{C}_2$  after finishing step 1.",
+        "bbox": [
+            75,
+            363,
+            468,
+            425
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Moving to step 2, a constraint is usually applied to avoid forgetting previous knowledge by punishing the network from drifting away from its previous states. In this step, the network needs to update its knowledge of  $\\mathcal{C}_1$  on  $D_2$ . Given that  $\\mathcal{C}_1$  is already properly optimized at step 1, the network only needs a small adjustment to complete its knowledge of  $\\mathcal{C}_1$  and receives little punishment from the constraint. However, the network is also intended to learn the knowledge w.r.t.  $\\mathcal{C}_2$  (i.e., distinguish the classes in  $\\mathcal{C}_2$  and distinguish between  $\\mathcal{C}_2$  and  $\\mathcal{C}_1$ ) from the start, which requires a rather large update. Based on the fact that the constraints make more punishment for greater changes, this update will be suppressed more severely compared to that of  $\\mathcal{C}_1$ , making more knowledge w.r.t.  $\\mathcal{C}_2$  cannot be learned. This is why the R-P dilemma often leads to low performance for new classes. To make matters worse, the scale of the training set is usually limited in incremental steps, especially for the situation when a single class is added with only a few samples of  $\\mathcal{C}_1$ . Such limited data could impede the network from successfully distinguishing between  $\\mathcal{C}_1$  and  $\\mathcal{C}_2$ , which is the inadequate data obstacle. Both of these factors result in the inability to properly learn knowledge w.r.t.  $\\mathcal{C}_2$  and contribute to the aforementioned underfitting phenomenon.",
+        "bbox": [
+            75,
+            425,
+            468,
+            773
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4. Discussions for Solutions",
+        "text_level": 1,
+        "bbox": [
+            76,
+            785,
+            303,
+            800
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Pre-learning future knowledge. Based on the analysis, if we modify the training pipeline by adding explicit supervision to  $\\mathcal{C}_2$  before it is introduced (blue background in Figure 3). The knowledge w.r.t.  $\\mathcal{C}_2$  could be learned in advance. By doing so, we first avoid the R-P dilemma by reducing the great update to learning the knowledge w.r.t.  $\\mathcal{C}_2$  at the next",
+        "bbox": [
+            75,
+            809,
+            470,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "step. Meanwhile, we can use the abundant samples of  $\\mathcal{C}_1$  at step 1 to make the network better learn how to distinguish between  $\\mathcal{C}_2$  and  $\\mathcal{C}_1$ , mitigating the influence of inadequate data of  $\\mathcal{C}_1$  at next step. Consequently, both obstacles are tackled in a uniform and elegant way. The main challenge falls to how we can learn knowledge w.r.t.  $\\mathcal{C}_2$  as accurately as possible in an unsupervised way in order to reduce the extent of updates required in the future.",
+        "bbox": [
+            496,
+            90,
+            890,
+            210
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Memory. Memory is a widely used technique in CL and has been adapted to CSS recently [6, 34]. Based on our analysis above, here we clearly explain the effect of memory in CSS. Memory is applied only in incremental steps. With the help of it, new classes are possible to be optimized with partial previous data, which can also alleviate the impact of inadequate data. Meanwhile, memory helps to ease the forgetting problem, which makes it possible to relax the constraints without forgetting. As a result, new knowledge can be more easily learned. We would like to point out that memory can be used to solve the two obstacles at a different aspect. It could further boost the performance in conjunction with our solution.",
+        "bbox": [
+            496,
+            210,
+            892,
+            407
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4. Method",
+        "text_level": 1,
+        "bbox": [
+            500,
+            420,
+            589,
+            435
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce a novel framework based on pre-learning future knowledge as depicted in Figure 4. It takes into account future knowledge in an unsupervised way, both in the feature space (green background) and the output space (blue background).",
+        "bbox": [
+            496,
+            446,
+            890,
+            522
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1. Learning Future Knowledge in Feature Space",
+        "text_level": 1,
+        "bbox": [
+            498,
+            530,
+            885,
+            546
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Contrastive learning. Contrastive learning is a widely used technique in unsupervised learning [7, 15]. In our work, we resort to it to optimize future classes in feature space. In semantic segmentation, directly using feature vectors at pixel level to perform pixel-pixel contrastive is redundant and results in high computational cost. Therefore, we follow [24, 46, 18] to perform contrast in a pixel-segment manner. Formally, supposing an image  $I$  is partitioned into segments  $S$  by a given prior. Let  $\\pmb{f}_i$  be the unit-length feature vector of pixel  $i$ , we define the (normalized) prototype  $\\pmb{p}$  of a segment  $s \\in S$  by the average of feature vectors within this segment.",
+        "bbox": [
+            496,
+            553,
+            890,
+            734
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\boldsymbol {p} _ {s} = \\frac {1}{| s |} \\sum_ {i \\in s} \\boldsymbol {f} _ {i}, \\| \\boldsymbol {p} _ {s} \\| = 1 \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            599,
+            744,
+            890,
+            779
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We adopt the InfoNCE [38] loss and formalize our loss as",
+        "bbox": [
+            500,
+            785,
+            880,
+            800
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} \\left(\\boldsymbol {f} _ {i}, S\\right) = - \\log \\frac {\\exp \\left(\\boldsymbol {f} _ {i} \\cdot \\boldsymbol {p} _ {s _ {+}} / \\tau\\right)}{\\exp \\left(\\boldsymbol {f} _ {i} \\cdot \\boldsymbol {p} _ {s _ {+}} / \\tau\\right) + \\sum_ {s _ {-}} \\exp \\left(\\boldsymbol {f} _ {i} \\cdot \\boldsymbol {p} _ {s _ {-}} / \\tau\\right)} \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            503,
+            821,
+            890,
+            869
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\tau$  is a temperature hyper-parameter. Ideally, it pulls  $f_{i}$  close to the positive segment  $s_{+}$  and pushes it away from",
+        "bbox": [
+            500,
+            869,
+            890,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "11913",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/578b344b9dfb493d19462b36cf82484cdfcd0bea0835ebb175e0fdbc86fb417a.jpg",
+        "image_caption": [
+            "Figure 4. Overall framework. Given an image at step  $t$ , a projection head maps the features from feature extractor  $F_{t}$  into an embedding space where  $\\mathcal{L}_{c}$  is used to optimize future classes (green block). Next, the features from  $F_{t}$  are also used to train an auxiliary classifier  $G_{aux}$  to initialize the classifier for future classes (blue block). Last, the pseudo label technique together with a knowledge distillation loss are adopted to mitigate forgetting (yellow block)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            88,
+            459,
+            297
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "all the negative segments  $s_{-}$ . For a given pixel, the segment it belongs to serves as  $s_{+}$ , whereas other segments, including those from other images, serve as  $s_{-}$ .",
+        "bbox": [
+            75,
+            441,
+            468,
+            488
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Contrast by Visual Similarity. In pixel-segment contrastive learning, the partition scheme determines how the network is optimized. We first resort to visual boundaries as a prior to partition the images into visual coherent segments  $S_{v}$  by OWT-UCM [2], which is a decent choice in unsupervised segmentation [19, 55, 23]. By utilizing visual similarity as supervision, we can bootstrap the representation learning and ground the features to respect low-level visual cues. But using this supervision alone makes the network hard to capture semantic-level information of future classes, and still needs further updates to properly learn them. This contradicts our goal to make the necessary updates as few as possible in the future.",
+        "bbox": [
+            75,
+            489,
+            468,
+            686
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Contrast by Feature Affinity. We also need a partition method that can match actual future classes accurately to serve as another supervision. Due to the bootstrap effect of the previous contrast, we can resort to clustering methods to explore feature affinity as semantic-level supervision without relying on architectural prior like [8]. However, applying clustering directly cannot generate satisfying results. In practice, we find that by hierarchically combining an over-clustering result, we could obtain much more accurate results. More in detail, starting from an over-clustering generated using intermediate features, we merge the two most similar segments each time till all segments are merged as one. The similarity between segments is measured by the cosine similarity between their prototypes. This procedure",
+        "bbox": [
+            75,
+            688,
+            470,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/3b3e192f3dd6e2b2518246bf9e82b12904482820528cf8133922f17ae7952bc6.jpg",
+        "image_caption": [
+            "Figure 5. Segments obtained using visual similarity and feature affinity respectively. Top: Segments generated by OWT-UCM [2] using boundaries. Bottom: Segments that are likely to match the actual classes are selected from a hierarchical combination of fine-grained clustering results."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            90,
+            885,
+            224
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "generates consistent segments across all granularities. But different classes naturally have different levels of granularities, directly selecting segments from a single granularity is often suboptimal. We further propose a confidence score to evaluate how likely a segment matches a class accurately to select segments across granularities. The design of the confidence score is based on the following observations. In the combination procedure where the most similar segments are merged each time, the longer a segment can remain not merged, the more different it is to all other segments, and is more likely to match a class. What's more, given that this procedure starts from an over-clustering, the segments in later steps have a higher chance to match a class and should be assigned higher weights. A possible choice is to use the cosine distance between the segments to be merged at each step as a weight factor, which monotonously increases by step. Formally, assuming segments  $a$  and  $b$  are merged into a new segment  $c$  at step  $n$  in this procedure, set  $S_{n}$  indicates all segments that are merged from start till now (include  $a$  and  $b$ ). The confidence score  $u$  for a segment  $s$  at step  $n$  is defined recursively as follows.",
+        "bbox": [
+            496,
+            320,
+            893,
+            638
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nu _ {s} ^ {n} = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} s = c, \\\\ u _ {s} ^ {n - 1} & \\text {e l i f} s \\in S _ {n}, \\\\ u _ {s} ^ {n - 1} + \\left(1 - \\boldsymbol {p} _ {a} \\cdot \\boldsymbol {p} _ {b}\\right) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            547,
+            654,
+            890,
+            710
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After finishing this procedure, we use the final score of each segment to select those with scores above the threshold to serve as high-confidence segments  $S_{f}$  and participate in contrastive learning. If two selected segments are partially overlapped, we only reserve the one with a higher score. This procedure is performed for each training sample before each epoch to select high-confidence segments as the network updates itself, except for the first epoch since the feature cannot be clustered at that moment.",
+        "bbox": [
+            496,
+            719,
+            893,
+            853
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Overall Contrastive Loss. In our contrastive learning, the pixels of  $\\mathcal{C}_t$  with ground truth annotations and  $\\mathcal{C}_{1:t-1}$  with pseudo labels are excluded. Meanwhile, they can naturally",
+        "bbox": [
+            496,
+            854,
+            893,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "11914",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "serve as high-confidence negative segments in contrastive learning. Since the batch size is limited in segmentation while the number of negative samples is crucial in contrastive learning [7, 15], we use a memory bank to cache the features of recent four batches to increase the number of negative samples. Given the segments  $S_{v}$  generate by visual similarity and segments  $S_{f}$  generated by feature affinity, our contrastive loss is formalized as",
+        "bbox": [
+            75,
+            90,
+            472,
+            210
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal{L}_{c} = \\frac{1}{|S_{v}|}\\sum_{\\substack{i\\in S_{v}\\\\ i\\notin \\mathcal{C}_{1:t}}}\\mathcal{L}(\\boldsymbol{f}_{i},S_{v}) + \\frac{1}{|S_{f}|}\\sum_{\\substack{i\\in S_{f}\\\\ i\\notin \\mathcal{C}_{1:t}}}\\mathcal{L}(\\boldsymbol{f}_{i},S_{f}) \\tag{4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            94,
+            228,
+            468,
+            276
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $|S|$  means the pixel number in all segments within  $S$ .",
+        "bbox": [
+            76,
+            281,
+            467,
+            297
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Learning Future Knowledge in Output Space",
+        "text_level": 1,
+        "bbox": [
+            76,
+            303,
+            460,
+            320
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Previous work [5] has claimed that a proper initialization of classifier weight at step  $t > 1$  could also benefit the performance and is widely adopted by other works [10, 54]. This motivates us to further introduce future knowledge to the output space. Existing practice eases the misalignment between features and classifier weights by using the weight of the background to initialize that of the new classes. But in the update process, the weights of new classes should be updated to separate them from the background. Making the classifier weights to be identical to that of the background has the risk of increasing the necessary update, which has a negative effect based on previous analysis.",
+        "bbox": [
+            75,
+            325,
+            468,
+            507
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To address this issue, we want the classifier to be initialized as close to the final optimal value as possible. When optimizing feature space, we manage to obtain high-confidence segments of future classes, and the intermediate features are also optimized accordingly. We could naturally utilize them to train an auxiliary classifier that can classify future classes in advance and use its weights as initialization. More in detail, we first perform clustering on the prototypes of segments to assign an auxiliary semantic label to each segment, which is then used to optimize the auxiliary classifier using cross-entropy loss. Note that the network receives no gradient from this classifier. In the next step when initialization is needed, we compute the prototype of each new class and select the auxiliary classifier weight which is closest to the prototype to initialize the actual classifier weight of it. On the one hand, we can make the classifier converge more easily and avoid the R-P dilemma. On the other hand, it solves the inadequate data obstacle in output space by utilizing the data of current step to serve as the negative samples of future classes, which can further boost the discriminability of the classifier compared to training it directly on the future steps.",
+        "bbox": [
+            75,
+            508,
+            470,
+            840
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Overall Framework",
+        "text_level": 1,
+        "bbox": [
+            76,
+            847,
+            266,
+            861
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Our proposed framework is depicted in Figure 4.  $\\mathcal{L}_{ce}$  and  $\\mathcal{L}_{aux}$  are both cross-entropy loss to learn new classes and to",
+        "bbox": [
+            76,
+            869,
+            468,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "train the auxiliary classifier, respectively. To complete our framework, we still need a KD loss to overcome forgetting. Since our focus is not on this part, we adopt the CKD loss in MiB [5] as  $\\mathcal{L}_{kd}$  to assist our training, with the addition of a temperature parameter as in original KD loss [16] to enhance its ability in knowledge preserving. This makes the overall loss function of our framework as follows.",
+        "bbox": [
+            496,
+            90,
+            890,
+            196
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c e} + \\alpha \\mathcal {L} _ {k d} + \\beta \\mathcal {L} _ {c} + \\mathcal {L} _ {\\text {a u x}}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            583,
+            208,
+            890,
+            224
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\alpha$  and  $\\beta$  are hyper-parameters.",
+        "bbox": [
+            500,
+            234,
+            746,
+            250
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            263,
+            633,
+            280
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. Implement Details",
+        "text_level": 1,
+        "bbox": [
+            500,
+            287,
+            678,
+            304
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Datasets. Following previous works, we evaluate our method on VOC2012 [12] and ADE20K [56], which are the standard CSS benchmarks. We report the performance on the official validation set of each dataset. We also exclude  $20\\%$  of the training set as a validation set to tune hyperparameters and perform all analyses.",
+        "bbox": [
+            496,
+            311,
+            890,
+            402
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Protocols. There are two settings in CSS benchmarks: disjoint and overlapped. The disjoint setting limits the given images to contain pixels only belong to  $\\mathcal{C}_{1:t}$ , while the latter does not have this restriction, which means the pixels can also belong to the future classes, i.e.,  $\\mathcal{C}_{t+1:T}$ . Note that the overlapped setting is considered to be more realistic and more challenging by all existing works [5, 10, 6, 37], and the disjoint setting is even ignored in some works [6, 37]. In design, our work is fully based on the overlapped setting. Evaluation. The CSS benchmark setups are denoted as  $N_{o} - N_{n}$  where  $N_{o}$  and  $N_{n}$  are the number of classes to be learned at the initial step and each incremental step, respectively. For example, the 15-1 setup of VOC2012 gives 15 classes at the initial step and adds one class per incremental step, resulting in a total of 6 steps. On VOC2012, we conduct experiments on four setups: 19-1 (2 steps), 15-5 (2 steps), 15-1 (6 steps), and 10-1 (11 steps). On ADE20K, we use three setups: 100-50 (2 tasks), 50-50 (3 tasks), and 100-10 (6 tasks). The network is evaluated after finishing all learning steps using mean Intersection over Union (mIoU) as metric. The performance of initial classes  $\\mathcal{C}_{1}$ , incremental classes  $\\mathcal{C}_{2:T}$ , and all classes  $\\mathcal{C}_{1:T}$  are denoted as old, new, and all, respectively.",
+        "bbox": [
+            496,
+            402,
+            892,
+            750
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Training. Following previous works [5, 10], we adopt DeepLab-V3 with an output stride of 16. ResNet-101 serves as the backbone, which is pre-trained on ImageNet. We use the SGD optimizer with momentum set to 0.9. The learning rate is set to 1e-2 for the initial learning step and 1e-3 for the following learning step. The learning rate is adjusted by the polynomial decay schedule with power set to 0.9. The network is trained on VOC2012 for 30 epochs, and 60 epochs on ADE20K, both with a batch size of 16. The image is cropped to  $512 \\times 512$  for both training and validation, with",
+        "bbox": [
+            496,
+            750,
+            890,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "11915",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/38b65846c664a259995b1d4175be99b52341fc5acbf7b347db38c31a8af6a499.jpg",
+        "table_caption": [
+            "Table 1. Results of different methods on VOC2012. The best is in bold while the second best is in underline. Results are averaged over 3 runs with standard deviation provided."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">19-1 (2 tasks)</td><td colspan=\"3\">15-5 (2 tasks)</td><td colspan=\"3\">15-1 (6 tasks)</td><td colspan=\"3\">10-1 (11 tasks)</td></tr><tr><td>old</td><td>new</td><td>all</td><td>old</td><td>new</td><td>all</td><td>old</td><td>new</td><td>all</td><td>old</td><td>new</td><td>all</td></tr><tr><td>ILT [35]</td><td>67.1</td><td>12.3</td><td>64.4</td><td>66.3</td><td>40.6</td><td>59.9</td><td>4.9</td><td>7.8</td><td>5.7</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MiB [5]</td><td>70.2</td><td>22.1</td><td>67.8</td><td>75.5</td><td>49.4</td><td>69.0</td><td>35.1</td><td>13.5</td><td>29.7</td><td>12.2</td><td>13.0</td><td>12.6</td></tr><tr><td>PLOP [10]</td><td>75.3</td><td>37.3</td><td>73.5</td><td>75.7</td><td>51.7</td><td>70.0</td><td>65.1</td><td>21.1</td><td>54.6</td><td>44.0</td><td>15.5</td><td>22.9</td></tr><tr><td>SDR [36]</td><td>69.1</td><td>32.6</td><td>67.4</td><td>75.4</td><td>52.6</td><td>69.9</td><td>44.7</td><td>21.8</td><td>39.2</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SSUL [6]</td><td>77.7</td><td>29.6</td><td>75.4</td><td>77.8</td><td>50.1</td><td>71.2</td><td>77.3</td><td>36.5</td><td>67.6</td><td>71.3</td><td>45.9</td><td>59.2</td></tr><tr><td>RCIL [54]</td><td>-</td><td>-</td><td>-</td><td>78.8</td><td>52.0</td><td>72.4</td><td>70.6</td><td>23.7</td><td>59.4</td><td>55.4</td><td>15.1</td><td>34.3</td></tr><tr><td>ALIFE [37]</td><td>76.6</td><td>49.3</td><td>75.3</td><td>77.1</td><td>52.5</td><td>71.3</td><td>64.4</td><td>34.9</td><td>57.4</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Ours</td><td>77.7 ±0.22</td><td>59.8 ±1.38</td><td>76.8 ±0.35</td><td>79.1 ±0.21</td><td>67.7 ±0.46</td><td>76.4 ±0.29</td><td>75.2 ±0.75</td><td>52.2 ±1.33</td><td>69.7 ±1.09</td><td>68.1 ±0.96</td><td>54.5 ±1.57</td><td>61.6 ±1.13</td></tr><tr><td>RECALL [34]</td><td>68.1</td><td>55.3</td><td>68.6</td><td>67.7</td><td>54.3</td><td>65.6</td><td>67.8</td><td>50.9</td><td>64.8</td><td>65.0</td><td>53.7</td><td>60.7</td></tr><tr><td>SSUL-M [6]</td><td>77.8</td><td>49.7</td><td>76.4</td><td>78.4</td><td>55.8</td><td>73.0</td><td>78.3</td><td>49.0</td><td>71.3</td><td>74.0</td><td>53.2</td><td>64.1</td></tr><tr><td>ALIFE-M [37]</td><td>76.7</td><td>52.2</td><td>75.5</td><td>77.6</td><td>55.2</td><td>72.3</td><td>66.0</td><td>38.8</td><td>59.5</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Ours-M</td><td>77.9 ±0.19</td><td>67.7 ±0.74</td><td>77.4 ±0.32</td><td>79.3 ±0.22</td><td>70.2 ±0.29</td><td>77.1 ±0.24</td><td>77.1 ±0.65</td><td>60.4 ±0.73</td><td>73.1 ±0.68</td><td>69.5 ±0.60</td><td>63.2 ±1.25</td><td>66.5 ±0.78</td></tr><tr><td>Joint</td><td>77.6</td><td>77.7</td><td>77.6</td><td>79.5</td><td>71.5</td><td>77.6</td><td>79.5</td><td>71.5</td><td>77.6</td><td>78.5</td><td>76.6</td><td>77.6</td></tr></table>",
+        "bbox": [
+            127,
+            114,
+            841,
+            329
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/2b878ddc6981c3a914540097085f7266c34ce9cbd1eb7f2d42df5bd25ce58653.jpg",
+        "table_caption": [
+            "Table 2. Results of different methods on ADE20K. The best is in bold while the second best is in underline. Results are averaged over 3 runs with standard deviation provided."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">100-50 (2 tasks)</td><td colspan=\"3\">50-50 (3 tasks)</td><td colspan=\"3\">100-10 (6 tasks).</td></tr><tr><td>old</td><td>new</td><td>all</td><td>old</td><td>new</td><td>all</td><td>old</td><td>new</td><td>all</td></tr><tr><td>ILT [35]</td><td>18.2</td><td>14.4</td><td>17.0</td><td>3.5</td><td>12.8</td><td>9.7</td><td>0.1</td><td>1.3</td><td>0.5</td></tr><tr><td>MiB [5]</td><td>40.5</td><td>17.1</td><td>32.7</td><td>45.5</td><td>21.0</td><td>29.3</td><td>38.2</td><td>11.1</td><td>29.2</td></tr><tr><td>PLOP [10]</td><td>41.8</td><td>14.8</td><td>32.9</td><td>48.8</td><td>20.9</td><td>30.4</td><td>40.4</td><td>13.6</td><td>31.5</td></tr><tr><td>SSUL [6]</td><td>41.2</td><td>18.0</td><td>33.5</td><td>48.3</td><td>20.1</td><td>29.5</td><td>40.2</td><td>18.7</td><td>33.1</td></tr><tr><td>RCIL [54]</td><td>42.3</td><td>18.8</td><td>34.5</td><td>48.3</td><td>25.0</td><td>32.9</td><td>39.3</td><td>17.6</td><td>32.1</td></tr><tr><td>ALIFE [37]</td><td>42.1</td><td>23.0</td><td>35.8</td><td>48.9</td><td>25.6</td><td>33.5</td><td>41.0</td><td>22.7</td><td>34.9</td></tr><tr><td>Ours</td><td>43.1 ±0.17</td><td>26.0 ±0.36</td><td>37.4 ±0.25</td><td>49.1 ±0.24</td><td>28.3 ±0.43</td><td>35.2 ±0.33</td><td>41.4 ±0.28</td><td>25.5 ±0.45</td><td>36.1 ±0.35</td></tr><tr><td>SSUL-M [6]</td><td>42.7</td><td>17.5</td><td>34.3</td><td>49.1</td><td>20.1</td><td>29.7</td><td>42.8</td><td>17.6</td><td>34.4</td></tr><tr><td>ALIFE-M [37]</td><td>42.2</td><td>23.5</td><td>36.0</td><td>49.0</td><td>26.1</td><td>33.8</td><td>41.1</td><td>23.0</td><td>35.1</td></tr><tr><td>Ours-M</td><td>43.1 ±0.15</td><td>26.2 ±0.29</td><td>37.5 ±0.21</td><td>49.2 ±0.12</td><td>28.5 ±0.27</td><td>35.4 ±0.18</td><td>41.6 ±0.21</td><td>25.8 ±0.26</td><td>36.3 ±0.22</td></tr><tr><td>Joint</td><td>43.5</td><td>29.4</td><td>38.8</td><td>50.3</td><td>32.7</td><td>38.8</td><td>43.5</td><td>29.4</td><td>38.8</td></tr></table>",
+        "bbox": [
+            202,
+            371,
+            764,
+            563
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/8066eebb0a9e448826a25a517bbc4d23e5e8c44466d09326524826b0c4261577.jpg",
+        "image_caption": [
+            "Figure 6. mIoU comparison between CSS methods over steps on VOC 15-1 setups. The performance of initial classes and each new class added per step are plotted separately from left to right. Our method shows great ability in learning new classes."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            580,
+            212,
+            667
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/1456f2f508fda48f3552742c5238f4220be55d929a74f03258c36cdbf422fa04.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            223,
+            580,
+            346,
+            667
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/dac3a1bc85afb87d35a21c63297a342ca4d3799976f2363f7fe10305dd3f7269.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            357,
+            580,
+            478,
+            667
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d9b160afa6126415c6d98b25328ea13e8e956f48c714c7043319d87f48c4b57a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            488,
+            580,
+            609,
+            667
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f72d05b7f88f7d937ef29b727e0e21a3d6c265214cd0beee4f49af21cc17dcb3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            622,
+            580,
+            743,
+            667
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/8c7e91b67ba41d4d2802679391173cca990126900e950f96ecc2612a4a149e96.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            754,
+            580,
+            875,
+            667
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "a random scaling and a random left-right flip additionally applied to training.",
+        "bbox": [
+            75,
+            724,
+            467,
+            755
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2. Results",
+        "text_level": 1,
+        "bbox": [
+            76,
+            768,
+            169,
+            784
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We provide the results on VOC2012 and ADE20K in Table 1 and Table 2, respectively. Some works [6, 34] exploit memory or external data to assist training. We also report the results obtained using memory (denoted by -M) following the implementation of [6] to fairly compare with these methods. The result of joint training is also provided as the upperbound for reference.",
+        "bbox": [
+            75,
+            794,
+            468,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "For all setups on VOC2012, our method surpasses all existing methods by a large margin on new classes. This demonstrates the strong ability of our method to learn new classes. Note that our method also shows promising results on old classes even with a simple approach to preserve old knowledge. This is due to that a higher performance on new classes indicates that there also might be less confusion between old and new classes, which also benefits the mIoU of old classes. On the more challenging setups with multiple steps (15-1, 10-1), the new classes cannot be properly learned in the first place for most methods. To",
+        "bbox": [
+            496,
+            724,
+            890,
+            891
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "11916",
+        "bbox": [
+            478,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "make matters worse, these poorly optimized new classes will suffer forgetting in the following steps which further reduces their performance. Even in these challenging circumstances, our method still shows great performance in learning new classes, as shown in Figure 6. We can also see that even though SSUL shows higher performance on old classes on long-term setups, it relies on completely freezing the feature extractor to prevent forgetting, which does help a lot for old classes on these steps, but also harms its flexibility greatly. We believe that if equipped with a modern knowledge-preserving constraint, our method could offer another boost in performance.",
+        "bbox": [
+            75,
+            90,
+            472,
+            272
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "On the more challenging ADE20K benchmarks, we obtain similar results as that of VOC2012. Our method shows some improvement on old classes and is significantly better on new classes. On top of the results of VOC2012 benchmark, this further demonstrates that our approach works well on more complex datasets, and has the ability to handle both scenes and objects centric datasets.",
+        "bbox": [
+            75,
+            276,
+            468,
+            381
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "When memory is applied on both benchmarks, our method receives further improvements and holds the leading position. This also indicates that our proposed training method has a non-overlapping effect with the memory technique. They can be used as two complementary techniques as pointed out in Section 3.4",
+        "bbox": [
+            75,
+            386,
+            468,
+            476
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            76,
+            497,
+            228,
+            513
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We investigate the effects of the proposed components in our framework using both 15-1 and 15-5 setups on VOC2012. Starting from a baseline composed of  $L_{ce}$ ,  $L_{kd}$  and pseudo label, we add the components one by one and show the ablation results in Table 3. As can be seen, the performance increase steadily as each proposed component is gradually added. Only applying contrast by visual similarity does not receive much improvement since it cannot instruct the network to capture semantic-level information. After applying contrast by feature affinity, the performance receives the highest gain, indicating the necessity to add semantic-level supervision and make the network learn more accurate knowledge. On top of that, the performance receives another boost from our classifier initialization technique which introduces future knowledge to output space.",
+        "bbox": [
+            75,
+            525,
+            470,
+            752
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/07afd2a5f13f14635242afd93d61c36a9d2ecd7cbdc8724f34625d556a36271a.jpg",
+        "table_caption": [
+            "Table 3. Ablation study on VOC2012 15-1 and 15-5 setup."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">15-1 (6 tasks)</td><td colspan=\"3\">15-5 (2 tasks)</td></tr><tr><td>old</td><td>new</td><td>all</td><td>old</td><td>new</td><td>all</td></tr><tr><td>baseline</td><td>65.3</td><td>18.8</td><td>54.2</td><td>76.9</td><td>51.3</td><td>70.8</td></tr><tr><td>+ visual similarity contrast</td><td>67.1</td><td>26.3</td><td>57.4</td><td>77.2</td><td>54.5</td><td>71.8</td></tr><tr><td>+ feature affinity contrast</td><td>74.5</td><td>45.1</td><td>67.5</td><td>78.7</td><td>65.4</td><td>75.5</td></tr><tr><td>+ classifier initialization</td><td>75.2</td><td>52.2</td><td>69.7</td><td>79.1</td><td>67.7</td><td>76.4</td></tr></table>",
+        "bbox": [
+            78,
+            786,
+            470,
+            873
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/40581768f3806ea880e27c8aa7a621e968f63154814aa1b47d483e2c1282af1f.jpg",
+        "image_caption": [
+            "Figure 7. Visualization of clustering results of direct clustering and our practice. The high-confidence segments selected by our criterion are also presented on the right."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            88,
+            888,
+            258
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/8218f054600d24222cc7444f31ca5dd8a5fb4bc3003c0c7e2acc0718c776b4d1.jpg",
+        "image_caption": [
+            "Figure 8. Performance comparison between different classifier initialization over training epochs on VOC2012 19-1 setup."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            550,
+            319,
+            839,
+            470
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4. Further Discussion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            527,
+            681,
+            542
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Feature Quality Analysis. To further demonstrate the effect of our method to solve the underfitting problem pointed out in Section 3.2, we conduct the same analysis on our method and incorporate the findings into Figure 2 (denoted by Ours). As can be seen, pre-learning future classes can significantly improve both intra-class and inter-class fitness.",
+        "bbox": [
+            496,
+            551,
+            890,
+            643
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Segment Generation. We provide a qualitative comparison between direct clustering and our proposed hierarchical combination at different granularities to validate our design in Section 4.1. The results are generated from the same intermediate features and are plotted in Figure 7. As can be seen, our practice consistently generates better results across different granularities. We also present the high-confidence segments selected by our criterion for each image, which match the actual classes with high precision.",
+        "bbox": [
+            496,
+            643,
+            890,
+            779
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Classifier Initialization. We give a comparison of performance over training epochs for different classifier initialization methods in Figure 8. As can be seen, the initialization technique proposed by MiB [5] converges much slower due to larger update paces. Random initialization converges fast but the performance stops increasing shortly afterward. Among all of them, our method stands out with its fast convergence speed and superior performance.",
+        "bbox": [
+            496,
+            779,
+            890,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "11917",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Limitation and Conclusion",
+        "text_level": 1,
+        "bbox": [
+            76,
+            90,
+            326,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our method has two limitations. Firstly, it relies on the natural attribute of CSS that unlabeled future classes exist in current training data. So it may not perform well when there are only a few unlabeled future classes in the current training data. However, in such cases, we can still leverage external unlabeled data, such as web-crawled data as used in RECALL [34]. Secondly, we point out that even though our proposed framework achieves great performance, it as a practicable solution might not be able to unleash the full potential of the training method, which requires further exploration in future works.",
+        "bbox": [
+            75,
+            116,
+            468,
+            282
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we present an in-depth analysis of the reason why existing methods suffer low performance on new classes. Building on this analysis, we introduce a novel training pipeline for CSS that optimizes potential future classes in the current training step. This offers a new perspective for enhancing new class performance in contrast to existing works that focus on addressing the R-P dilemma. We propose a framework based on this pipeline and it shows a significant improvement for new classes, while also benefiting the performance of old classes. We believe that this work can provide valuable insights for the research community on CSS.",
+        "bbox": [
+            75,
+            282,
+            472,
+            464
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7. Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            76,
+            481,
+            264,
+            500
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work is supported by the National Natural Science Foundation of China under Grant 62176246 and Grant 61836008. This work is also supported by Anhui Province Key Research and Development Plan (202304a05020045), Anhui Provincial Natural Science Foundation 2208085UD17 and the Fundamental Research Funds for the Central Universities (WK3490000006).",
+        "bbox": [
+            75,
+            507,
+            468,
+            613
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            76,
+            630,
+            174,
+            645
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, 2018.",
+            "[2] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 33(5):898-916, 2010.",
+            "[3] Jihwan Bang, Heesu Kim, YoungJoon Yoo, Jung-Woo Ha, and Jonghyun Choi. Rainbow memory: Continual learning with a memory of diverse samples. In CVPR, 2021.",
+            "[4] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017.",
+            "[5] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background"
+        ],
+        "bbox": [
+            84,
+            655,
+            470,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "for incremental learning in semantic segmentation. In ICCV, 2020.",
+            "[6] Sungmin Cha, YoungJoon Yoo, Taesup Moon, et al. Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. NIPS, 2021.",
+            "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020.",
+            "[8] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In CVPR, 2021.",
+            "[9] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019.",
+            "[10] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In CVPR, 2021.",
+            "[11] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020.",
+            "[12] Mark Everingham and John Winn. The pascal visual object classes challenge 2012 (voc2012) development kit. Pattern Analysis, Statistical Modelling and Computational Learning, Tech. Rep, 2011.",
+            "[13] Qiankun Gao, Chen Zhao, Bernard Ghanem, and Jian Zhang. R-dfcil: Relation-guided representation learning for data-free class incremental learning. In ECCV, 2018.",
+            "[14] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Region-based convolutional networks for accurate object detection and segmentation. IEEE transactions on pattern analysis and machine intelligence, 38(1):142-158, 2015.",
+            "[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020.",
+            "[16] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015.",
+            "[17] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In ICCV, 2019.",
+            "[18] Hanzhe Hu, Jinshi Cui, and Liwei Wang. Region-aware contrastive learning for semantic segmentation. In ICCV, 2021.",
+            "[19] Jyh-Jing Hwang, Stella X Yu, Jianbo Shi, Maxwell D Collins, Tien-Ju Yang, Xiao Zhang, and Liang-Chieh Chen. Segsort: Segmentation by discriminative sorting of segments. In ICCV, 2019.",
+            "[20] Ahmet Iscen, Jeffrey Zhang, Svetlana Lazebnik, and Cordelia Schmid. Memory-efficient incremental learning through feature adaptation. In ECCV, 2020.",
+            "[21] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, 2019.",
+            "[22] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In CVPR, 2022."
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "11918",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[23] Tsung-Wei Ke, Jyh-Jing Hwang, Yunhui Guo, Xudong Wang, and Stella X Yu. Unsupervised hierarchical semantic segmentation with multiview cosegmentation and clustering transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2571-2581, 2022.",
+            "[24] Tsung-Wei Ke, Jyh-Jing Hwang, and Stella Yu. Universal weakly supervised segmentation by pixel-to-segment contrastive learning. In ICLR, 2021.",
+            "[25] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017.",
+            "[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017.",
+            "[27] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In ICML, 2019.",
+            "[28] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017.",
+            "[29] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Rmm: Reinforced memory management for class-incremental learning. NIPS, 2021.",
+            "[30] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In CVPR, 2020.",
+            "[31] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015.",
+            "[32] Arun Mallya, Dillon Davis, and Svetlana Lazebnik. Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In ECCV, 2018.",
+            "[33] Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning. In CVPR, 2018.",
+            "[34] Andrea Maracani, Umberto Michieli, Marco Toldo, and Pietro Zanuttigh. Recall: Replay-based continual learning in semantic segmentation. In ICCV, 2021.",
+            "[35] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In ICCVW, 2019.",
+            "[36] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In CVPR, 2021.",
+            "[37] Youngmin Oh, Donghyeon Baek, and Bumsub Ham. Alife: Adaptive logit regularizer and feature replay for incremental semantic segmentation. NIPS, 2022.",
+            "[38] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018.",
+            "[39] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Johannichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In CVPR, 2019."
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[40] Quang Pham, Chenghao Liu, and Steven Hoi. Dualnet: Continual learning, fast and slow. NIPS, 2021.",
+            "[41] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In ECCV, 2020.",
+            "[42] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017.",
+            "[43] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip HS Torr, Song Bai, and Vincent YF Tan. Mimicking the oracle: an initial phase decorrelation approach for class incremental learning. In CVPR, 2022.",
+            "[44] Pravendra Singh, Pratik Mazumder, Piyush Rai, and Vinay P Namboodiri. Rectification-based knowledge retention for continual learning. In CVPR, 2021.",
+            "[45] Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 254-270. Springer, 2020.",
+            "[46] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In ICCV, 2021.",
+            "[47] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In ECCV, 2022.",
+            "[48] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019.",
+            "[49] Ye Xiang, Ying Fu, Pan Ji, and Hua Huang. Incremental learning using conditional adversarial networks. In ICCV, 2019.",
+            "[50] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, 2021.",
+            "[51] Guanglei Yang, Enrico Fini, Dan Xu, Paolo Rota, Mingli Ding, Moin Nabi, Xavier Alameda-Pineda, and Elisa Ricci. Uncertainty-aware contrastive distillation for incremental semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022.",
+            "[52] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. In International Conference on Learning Representations, 2018.",
+            "[53] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, pages 3987-3995, 2017.",
+            "[54] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In CVPR, 2022.",
+            "[55] Xiao Zhang and Michael Maire. Self-supervised visual representation learning from hierarchical grouping. NIPS, 2020.",
+            "[56] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017."
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "11919",
+        "bbox": [
+            480,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[57] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and Cheng-Lin Liu. Prototype augmentation and self-supervision for incremental learning. In CVPR, 2021.",
+            "[58] Yixiong Zou, Shanghang Zhang, Ke Chen, Yonghong Tian, Yaowei Wang, and José MF Moura. Compositional few-shot recognition with primitive discovery and enhancing. In Proceedings of the 28th ACM International Conference on Multimedia, pages 156-164, 2020."
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            204
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "11920",
+        "bbox": [
+            480,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2023/Preparing the Future for Continual Semantic Segmentation/72d5919b-3823-475f-9939-cac5c29b48a5_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d26cf624ccd36fff4dff89997a02f719a75552b6ffc11286af1b3258c5617d41
+size 3408325

2023/Preparing the Future for Continual Semantic Segmentation/full.md

@@ -0,0 +1,303 @@
+# Preparing the Future for Continual Semantic Segmentation
+
+Zihan Lin<sup>1</sup> Zilei Wang<sup>*1</sup> Yixin Zhang<sup>1,2</sup>
+
+<sup>1</sup> University of Science and Technology of China
+
+$^{2}$  Institute of Artificial Intelligence, Hefei Comprehensive National Science Center
+
+myustc@mail.ustc.edu.cn, {zlwang, zhyx12}@ustc.edu.cn
+
+# Abstract
+
+In this study, we focus on Continual Semantic Segmentation (CSS) and present a novel approach to tackle the issue of existing methods struggling to learn new classes. The primary challenge of CSS is to learn new knowledge while retaining old knowledge, which is commonly known as the rigidity-plasticity dilemma. Existing approaches strive to address this by carefully balancing the learning of new and old classes during training on new data. Differently, this work aims to avoid this dilemma fundamentally rather than handling the difficulties involved in it. Specifically, we reveal that this dilemma mainly arises from the greater fluctuation of knowledge for new classes because they have never been learned before the current step. Additionally, the data available in incremental steps are usually inadequate, which can impede the model's ability to learn discriminative features for both new and old classes. To address these challenges, we introduce a novel concept of pre-learning for future knowledge. Our approach entails optimizing the feature space and output space for unlabeled data, which thus enables the model to acquire knowledge for future classes. With this approach, updating the model for new classes becomes as smooth as for old classes, effectively avoiding the rigidity-plasticity dilemma. We conducted extensive experiments and the results demonstrate a significant improvement in the learning of new classes compared to previous state-of-the-art methods.
+
+# 1. Introduction
+
+Deep neural networks have demonstrated their superiority in many computer vision tasks [26, 31, 14]. Conventionally, they are trained in an offline manner with all data collected beforehand. When novel classes need to be handled, the networks need to be re-trained using either updated training set with both new and previous data, or directly finetuned on new data. The former has higher costs and is even impracticable sometimes like, when privacy raises concerns. The latter is prone to forget previ
+
+![](images/301e91d65d39335f139ddfa2115135c85334d75e9536194ad13a2b6c5bd7c493.jpg)  
+Figure 1. On VOC2012 benchmark, our method boosts the performance of new classes by a significant margin, while also achieving promising performance for old classes, compared to recent state-of-the-art methods.
+
+ously learned knowledge, which is called catastrophic forgetting (CF) [25]. To better fit neural networks into real-world scenarios, we want them to have the ability to learn new concepts while preserving the old knowledge (or even improve it using new knowledge). This is known as the continual learning (CL) problem and receives increasing attention recently. It has been explored first in the field of image classification [28, 42, 17, 11, 57, 44, 43], and some methods are proposed recently targeting semantic segmentation [5, 10, 6, 54, 51, 37]. Generally, it is often achieved by applying forgetting-preventing constraints between current and previous networks to prevent CF, which raises the rigidity (i.e., the ability to preserve old states) but harms the plasticity (i.e., the ability to learn new knowledge). It is known as the rigidity-plasticity (R-P) dilemma. A key aspect in designing constraints is to balance between rigidity and plasticity, which is what most works focus on.
+
+Despite that several new constraints have been proposed for CSS to better tackle the dilemma, existing works still struggle to learn new classes. Taking the VOC2012 benchmark as an example, most methods can only achieve a performance around  $70\%$  of that of the upperbound for 15-5 setup and around  $50\%$  for 19-1 setup, which is even worse
+
+for harder setups. To get a clear picture of why this happens, we first analyze the intermediate features of the network and find it suffers serious underfitting for new classes, which indicates a large amount of knowledge is not learned. We then examine the CSS training pipeline and find two main obstacles causing this phenomenon. (1) R-P Dilemma: the network has no knowledge w.r.t. new classes until they are introduced. This generally requires a rather large update to incorporate the new knowledge, which will be significantly hindered by the forgetting-preventing constraints punishing the network from drifting away from its previous states. (2) Inadequate data for training: the amount of data available in incremental steps is usually limited, and this problem becomes more severe when a single class is added, resulting in a tiny training set with only a few negative samples. This can significantly limit the network to properly discriminate between old and new classes. This problem has never been explicitly considered before.
+
+Unlike existing works trying to tackle the R-P dilemma directly, we are the first to argue that this obstacle can be elegantly avoided by exploring future knowledge in an unsupervised way. Assume that the network has obtained partial knowledge of future classes in advance, then the network needs fewer updates to learn the upcoming knowledge, which consequently lowers the impacts of the R-P dilemma. The more future knowledge it learns in advance, the fewer the impacts will be caused. Considering an extreme situation where all the future knowledge is learned beforehand, the network does not need any update, and thus the influence of constraints could be totally avoided. On top of this, it makes us the first attempt to explicitly tackle the second obstacle by utilizing the abundant data in previous steps to optimize discriminability between old and new classes. Based on these understandings, we design a novel framework where unsupervised contrastive learning is performed on intermediate features to optimize future classes with both visual similarity and feature affinity as supervisions. One step further, an auxiliary classifier is trained using pseudo labels of future classes generated via clustering, which is used to initialize the actual classifier for future steps. Our method surpasses existing works by a huge margin as shown in Figure 1 and the effectiveness is further demonstrated with extensive experiments. Our contributions are summarized as follows.
+
+- We propose to pre-learn future knowledge for CSS with the aim of improving the performance especially for new classes, which can simultaneously tackle the R-P dilemma and the obstacle of inadequate data.  
+- We design a novel framework that can pre-learn future knowledge in an unsupervised way, where both the feature space and output space are handled.  
+- We have validated the effectiveness of our proposed
+
+approach through extensive experiments. The results demonstrate significant improvement on standard CSS benchmarks.
+
+# 2. Related Work
+
+Continual Learning (CL). Continual learning has been studied for years in image classification. Existing techniques can be roughly divided into regularization methods, rehearsal methods, and architectural methods. Regularization methods apply consistency constraints between current network and previous network to avoid drifts from previous states. They can be applied on network weights [53, 25, 1], the intermediate features [9, 11, 45, 13, 22] or the output logits [28, 42, 17, 48]. Parameter regularization is rarely seen recently due to the lack of plasticity, while the other two are mostly based on knowledge distillation (KD) [16]. To better solve the forgetting problem, rehearsal methods replay previous data in the incremental steps, including image replay [42, 41, 3], generative replay [49, 39], feature replay [20, 57], and memory management [30, 29]. Architectural methods use a dynamic network to handle incoming knowledge, like sub-network [32, 33], complementary network [47, 40] or expandable network [27, 50, 52].
+
+Continual Semantic Segmentation (CSS). Some works extended the aforementioned techniques to semantic segmentation tasks. For the KD-based methods, MiB [5] formalized and targeted background shift problem with output space distillation. PLOP [10], SDR [36], and RCIL [54] use feature space distillation to further mitigate forgetting. For the rehearsal-based methods, SSUL [6] and RECALL [34] utilize memory or external data to overcome forgetting and work well in multiple-step learning. For the architectural methods, RCIL [54] proposes a structural reparameterization design to better learn new knowledge.
+
+Unsupervised Semantic Segmentation. These works focus on learning to cluster semantically related pixels without human labeling. Existing works can be roughly divided into three types. PiCIE [8] iteratively performs clustering and uses it to refine network features. This relies on the architectural prior and is prone to degenerate solutions. IIC [21] maximizes the mutual information between two augmentations of the same image which enforces spatial stability on them. But this solution has trouble dealing with complex scenes. The last type uses visual similarity. These works [19, 55, 46, 23] use handcrafted priors, like boundary, superpixels, or saliency detectors to first group pixels by visual similarity. This is then used to learn pixel-level representations using contrastive learning.
+
+# 3. Preliminaries
+
+In this section, we first explain the definition of CSS (Sec. 3.1), and present the underfitting phenomenon of new
+
+![](images/7c464a504f500bccef080c7376cf9f4d8e5a117b9e37411e79d55f94d75c2aa6.jpg)  
+Figure 2. Feature fitness analysis of different methods using L1 norm (Left), activation strength (Mid), and cosine similarity (Right) on VOC2012 15-5 setup. It is demonstrated that existing works suffer underfitting on both intra-class level (left and mid) and inter-class level (right) for new classes. Our method effectively addresses this problem.
+
+![](images/196888f1c316e672db90607467f10607b1ec82a221c00885053b765c22d1160e.jpg)
+
+![](images/4f84dd7ca2e2af2a43674bb3157b812ed37d730691a0015ce1dc4f2aa1c259c0.jpg)
+
+classes (Sec. 3.2). After that, we analyze how the performance is impacted in CSS (Sec. 3.3). Finally, we discuss the solutions to boost new class performance (Sec. 3.4).
+
+# 3.1. CSS Definition
+
+In CSS, the training procedure is composed of a series of  $T$  learning steps. At each step  $t$  where  $t \in \{1, \dots, T\}$ , a set of new categories  $\mathcal{C}_t$  are introduced with a training set  $D_t = \{(x_n, y_n)\}$  where  $x_n$  is an image and  $y_n$  is the corresponding label. Only  $\mathcal{C}_t$  is labeled in  $y_n$  and all the other classes are labeled as background. It is assumed that the categories learned at each step are disjoint, i.e.,  $\mathcal{C}_i \cap \mathcal{C}_j = \emptyset$ . The network is trained sequentially on each step, where it learns the knowledge w.r.t.  $\mathcal{C}_t$  and expands its output space to  $\mathcal{C}_{1:t} = \mathcal{C}_{1:t-1} \cup \mathcal{C}_t$ . After all learning steps, the network is expected to perform well on all classes.
+
+# 3.2. Analysis of Feature Fitness
+
+To get a clear view of why new classes perform poorly in existing methods, we analyze the intermediate features of two works: MiB [5] based on conventional logits distillation, and a more recent work, RCIL [54], based on feature distillation and structural re-parameterization aiming to better tackle the R-P dilemma. To simplify the situation, we opt for the two-step 15-5 setup on VOC2012 to conduct the analysis. The results are reported in Figure 2 (Detailed information about incremental setups and validation set can be found in Section 5.). The results of joint training (i.e., performance upperbound) and initial state (i.e., backbone is initialized with ImageNet pre-trained weights while the other parts are randomly initialized) are also provided for reference. Ideally, the results are expected to be close to that of the upperbound in order to achieve good performance.
+
+Intra-class Fitness. We first evaluate how well the network fits to each individual class. In neural networks, each channel in the feature vectors could represent a certain pattern [4] which composes the learned classes [58]. When the network fits a class well, the output feature vector should have specific channels being activated while the others deactivated, resulting in sparsity in the feature vector. The
+
+activation strength further shows how strong the network's response is to each pattern. Both of them can serve as clues indicating how well the network fits a class. We use the average L1 norm of L2-normalized feature vectors to measure the sparsity, and the average top activation values to measure the response strength. As can be seen in Figure 2 (left and mid), the new classes of both MiB and RCIL have high L1 norm and low activation strength, indicating a serious underfitting problem.
+
+Inter-class Fitness. We further evaluate how well the network learns to distinguish different classes, which is measured by the cosine similarity between the feature vectors of different classes. The better the network can distinguish each class, the lower the cosine similarities will be. We compute the mean cosine distance of a class to any other classes, which is then averaged based on class group to get the final results, as shown in Figure 2 (right). It can be seen that the network has good discriminative power on old classes, but the new classes again present poor results. It indicates the network cannot distinguish a new class between other classes well.
+
+Discussion. From the above analyses, it is evident that new classes suffer serious underfitting on both intra-class and inter-class aspects. This exactly explains why most methods show low performance on new classes, even in simple setups with only a single incremental step.
+
+# 3.3. Understanding Training Pipeline
+
+In this part, we dive deeper to investigate why the underfitting problem happens for new classes in the conventional training pipeline. We use Figure 3 to give a clear view of how the two obstacles impact the performance in a conventional pipeline (green background). Without loss of generality, a two-step situation is adopted as an example.
+
+Starting from step 1 where  $\mathcal{C}_1$  is labeled in the given data  $D_{1}$  while others are marked as background, the network acquires most knowledge of  $\mathcal{C}_1$  (i.e., distinguish between classes in  $\mathcal{C}_1$ ) after finishing this step. Considering that it only accesses partial data at this step compared to joint training, the knowledge of  $\mathcal{C}_1$  remains to be updated
+
+![](images/a284cd579c0221d99adab4bb6e6c37940eaaaffa4cfdeabfc8de1a28bfe70af4.jpg)  
+Figure 3. Illustration of the difference between knowledge learning in a conventional training pipeline (green background) and our solution (blue background). We jointly optimize new classes together with old classes in an unsupervised way before new classes are introduced.
+
+in the next step. However, even though the future classes  $\mathcal{C}_2$  do exist in  $D_{1}$ , they do not receive explicit supervision, thus the network obtains almost no knowledge of  $\mathcal{C}_2$  after finishing step 1.
+
+Moving to step 2, a constraint is usually applied to avoid forgetting previous knowledge by punishing the network from drifting away from its previous states. In this step, the network needs to update its knowledge of  $\mathcal{C}_1$  on  $D_2$ . Given that  $\mathcal{C}_1$  is already properly optimized at step 1, the network only needs a small adjustment to complete its knowledge of  $\mathcal{C}_1$  and receives little punishment from the constraint. However, the network is also intended to learn the knowledge w.r.t.  $\mathcal{C}_2$  (i.e., distinguish the classes in  $\mathcal{C}_2$  and distinguish between  $\mathcal{C}_2$  and  $\mathcal{C}_1$ ) from the start, which requires a rather large update. Based on the fact that the constraints make more punishment for greater changes, this update will be suppressed more severely compared to that of  $\mathcal{C}_1$ , making more knowledge w.r.t.  $\mathcal{C}_2$  cannot be learned. This is why the R-P dilemma often leads to low performance for new classes. To make matters worse, the scale of the training set is usually limited in incremental steps, especially for the situation when a single class is added with only a few samples of  $\mathcal{C}_1$ . Such limited data could impede the network from successfully distinguishing between  $\mathcal{C}_1$  and  $\mathcal{C}_2$ , which is the inadequate data obstacle. Both of these factors result in the inability to properly learn knowledge w.r.t.  $\mathcal{C}_2$  and contribute to the aforementioned underfitting phenomenon.
+
+# 3.4. Discussions for Solutions
+
+Pre-learning future knowledge. Based on the analysis, if we modify the training pipeline by adding explicit supervision to  $\mathcal{C}_2$  before it is introduced (blue background in Figure 3). The knowledge w.r.t.  $\mathcal{C}_2$  could be learned in advance. By doing so, we first avoid the R-P dilemma by reducing the great update to learning the knowledge w.r.t.  $\mathcal{C}_2$  at the next
+
+step. Meanwhile, we can use the abundant samples of  $\mathcal{C}_1$  at step 1 to make the network better learn how to distinguish between  $\mathcal{C}_2$  and  $\mathcal{C}_1$ , mitigating the influence of inadequate data of  $\mathcal{C}_1$  at next step. Consequently, both obstacles are tackled in a uniform and elegant way. The main challenge falls to how we can learn knowledge w.r.t.  $\mathcal{C}_2$  as accurately as possible in an unsupervised way in order to reduce the extent of updates required in the future.
+
+Memory. Memory is a widely used technique in CL and has been adapted to CSS recently [6, 34]. Based on our analysis above, here we clearly explain the effect of memory in CSS. Memory is applied only in incremental steps. With the help of it, new classes are possible to be optimized with partial previous data, which can also alleviate the impact of inadequate data. Meanwhile, memory helps to ease the forgetting problem, which makes it possible to relax the constraints without forgetting. As a result, new knowledge can be more easily learned. We would like to point out that memory can be used to solve the two obstacles at a different aspect. It could further boost the performance in conjunction with our solution.
+
+# 4. Method
+
+In this section, we introduce a novel framework based on pre-learning future knowledge as depicted in Figure 4. It takes into account future knowledge in an unsupervised way, both in the feature space (green background) and the output space (blue background).
+
+# 4.1. Learning Future Knowledge in Feature Space
+
+Contrastive learning. Contrastive learning is a widely used technique in unsupervised learning [7, 15]. In our work, we resort to it to optimize future classes in feature space. In semantic segmentation, directly using feature vectors at pixel level to perform pixel-pixel contrastive is redundant and results in high computational cost. Therefore, we follow [24, 46, 18] to perform contrast in a pixel-segment manner. Formally, supposing an image  $I$  is partitioned into segments  $S$  by a given prior. Let  $\pmb{f}_i$  be the unit-length feature vector of pixel  $i$ , we define the (normalized) prototype  $\pmb{p}$  of a segment  $s \in S$  by the average of feature vectors within this segment.
+
+$$
+\boldsymbol {p} _ {s} = \frac {1}{| s |} \sum_ {i \in s} \boldsymbol {f} _ {i}, \| \boldsymbol {p} _ {s} \| = 1 \tag {1}
+$$
+
+We adopt the InfoNCE [38] loss and formalize our loss as
+
+$$
+\mathcal {L} \left(\boldsymbol {f} _ {i}, S\right) = - \log \frac {\exp \left(\boldsymbol {f} _ {i} \cdot \boldsymbol {p} _ {s _ {+}} / \tau\right)}{\exp \left(\boldsymbol {f} _ {i} \cdot \boldsymbol {p} _ {s _ {+}} / \tau\right) + \sum_ {s _ {-}} \exp \left(\boldsymbol {f} _ {i} \cdot \boldsymbol {p} _ {s _ {-}} / \tau\right)} \tag {2}
+$$
+
+where  $\tau$  is a temperature hyper-parameter. Ideally, it pulls  $f_{i}$  close to the positive segment  $s_{+}$  and pushes it away from
+
+![](images/578b344b9dfb493d19462b36cf82484cdfcd0bea0835ebb175e0fdbc86fb417a.jpg)  
+Figure 4. Overall framework. Given an image at step  $t$ , a projection head maps the features from feature extractor  $F_{t}$  into an embedding space where  $\mathcal{L}_{c}$  is used to optimize future classes (green block). Next, the features from  $F_{t}$  are also used to train an auxiliary classifier  $G_{aux}$  to initialize the classifier for future classes (blue block). Last, the pseudo label technique together with a knowledge distillation loss are adopted to mitigate forgetting (yellow block).
+
+all the negative segments  $s_{-}$ . For a given pixel, the segment it belongs to serves as  $s_{+}$ , whereas other segments, including those from other images, serve as  $s_{-}$ .
+
+Contrast by Visual Similarity. In pixel-segment contrastive learning, the partition scheme determines how the network is optimized. We first resort to visual boundaries as a prior to partition the images into visual coherent segments  $S_{v}$  by OWT-UCM [2], which is a decent choice in unsupervised segmentation [19, 55, 23]. By utilizing visual similarity as supervision, we can bootstrap the representation learning and ground the features to respect low-level visual cues. But using this supervision alone makes the network hard to capture semantic-level information of future classes, and still needs further updates to properly learn them. This contradicts our goal to make the necessary updates as few as possible in the future.
+
+Contrast by Feature Affinity. We also need a partition method that can match actual future classes accurately to serve as another supervision. Due to the bootstrap effect of the previous contrast, we can resort to clustering methods to explore feature affinity as semantic-level supervision without relying on architectural prior like [8]. However, applying clustering directly cannot generate satisfying results. In practice, we find that by hierarchically combining an over-clustering result, we could obtain much more accurate results. More in detail, starting from an over-clustering generated using intermediate features, we merge the two most similar segments each time till all segments are merged as one. The similarity between segments is measured by the cosine similarity between their prototypes. This procedure
+
+![](images/3b3e192f3dd6e2b2518246bf9e82b12904482820528cf8133922f17ae7952bc6.jpg)  
+Figure 5. Segments obtained using visual similarity and feature affinity respectively. Top: Segments generated by OWT-UCM [2] using boundaries. Bottom: Segments that are likely to match the actual classes are selected from a hierarchical combination of fine-grained clustering results.
+
+generates consistent segments across all granularities. But different classes naturally have different levels of granularities, directly selecting segments from a single granularity is often suboptimal. We further propose a confidence score to evaluate how likely a segment matches a class accurately to select segments across granularities. The design of the confidence score is based on the following observations. In the combination procedure where the most similar segments are merged each time, the longer a segment can remain not merged, the more different it is to all other segments, and is more likely to match a class. What's more, given that this procedure starts from an over-clustering, the segments in later steps have a higher chance to match a class and should be assigned higher weights. A possible choice is to use the cosine distance between the segments to be merged at each step as a weight factor, which monotonously increases by step. Formally, assuming segments  $a$  and  $b$  are merged into a new segment  $c$  at step  $n$  in this procedure, set  $S_{n}$  indicates all segments that are merged from start till now (include  $a$  and  $b$ ). The confidence score  $u$  for a segment  $s$  at step  $n$  is defined recursively as follows.
+
+$$
+u _ {s} ^ {n} = \left\{ \begin{array}{l l} 0 & \text {i f} s = c, \\ u _ {s} ^ {n - 1} & \text {e l i f} s \in S _ {n}, \\ u _ {s} ^ {n - 1} + \left(1 - \boldsymbol {p} _ {a} \cdot \boldsymbol {p} _ {b}\right) & \text {o t h e r w i s e .} \end{array} \right. \tag {3}
+$$
+
+After finishing this procedure, we use the final score of each segment to select those with scores above the threshold to serve as high-confidence segments  $S_{f}$  and participate in contrastive learning. If two selected segments are partially overlapped, we only reserve the one with a higher score. This procedure is performed for each training sample before each epoch to select high-confidence segments as the network updates itself, except for the first epoch since the feature cannot be clustered at that moment.
+
+Overall Contrastive Loss. In our contrastive learning, the pixels of  $\mathcal{C}_t$  with ground truth annotations and  $\mathcal{C}_{1:t-1}$  with pseudo labels are excluded. Meanwhile, they can naturally
+
+serve as high-confidence negative segments in contrastive learning. Since the batch size is limited in segmentation while the number of negative samples is crucial in contrastive learning [7, 15], we use a memory bank to cache the features of recent four batches to increase the number of negative samples. Given the segments  $S_{v}$  generate by visual similarity and segments  $S_{f}$  generated by feature affinity, our contrastive loss is formalized as
+
+$$
+\mathcal{L}_{c} = \frac{1}{|S_{v}|}\sum_{\substack{i\in S_{v}\\ i\notin \mathcal{C}_{1:t}}}\mathcal{L}(\boldsymbol{f}_{i},S_{v}) + \frac{1}{|S_{f}|}\sum_{\substack{i\in S_{f}\\ i\notin \mathcal{C}_{1:t}}}\mathcal{L}(\boldsymbol{f}_{i},S_{f}) \tag{4}
+$$
+
+where  $|S|$  means the pixel number in all segments within  $S$ .
+
+# 4.2. Learning Future Knowledge in Output Space
+
+Previous work [5] has claimed that a proper initialization of classifier weight at step  $t > 1$  could also benefit the performance and is widely adopted by other works [10, 54]. This motivates us to further introduce future knowledge to the output space. Existing practice eases the misalignment between features and classifier weights by using the weight of the background to initialize that of the new classes. But in the update process, the weights of new classes should be updated to separate them from the background. Making the classifier weights to be identical to that of the background has the risk of increasing the necessary update, which has a negative effect based on previous analysis.
+
+To address this issue, we want the classifier to be initialized as close to the final optimal value as possible. When optimizing feature space, we manage to obtain high-confidence segments of future classes, and the intermediate features are also optimized accordingly. We could naturally utilize them to train an auxiliary classifier that can classify future classes in advance and use its weights as initialization. More in detail, we first perform clustering on the prototypes of segments to assign an auxiliary semantic label to each segment, which is then used to optimize the auxiliary classifier using cross-entropy loss. Note that the network receives no gradient from this classifier. In the next step when initialization is needed, we compute the prototype of each new class and select the auxiliary classifier weight which is closest to the prototype to initialize the actual classifier weight of it. On the one hand, we can make the classifier converge more easily and avoid the R-P dilemma. On the other hand, it solves the inadequate data obstacle in output space by utilizing the data of current step to serve as the negative samples of future classes, which can further boost the discriminability of the classifier compared to training it directly on the future steps.
+
+# 4.3. Overall Framework
+
+Our proposed framework is depicted in Figure 4.  $\mathcal{L}_{ce}$  and  $\mathcal{L}_{aux}$  are both cross-entropy loss to learn new classes and to
+
+train the auxiliary classifier, respectively. To complete our framework, we still need a KD loss to overcome forgetting. Since our focus is not on this part, we adopt the CKD loss in MiB [5] as  $\mathcal{L}_{kd}$  to assist our training, with the addition of a temperature parameter as in original KD loss [16] to enhance its ability in knowledge preserving. This makes the overall loss function of our framework as follows.
+
+$$
+\mathcal {L} = \mathcal {L} _ {c e} + \alpha \mathcal {L} _ {k d} + \beta \mathcal {L} _ {c} + \mathcal {L} _ {\text {a u x}}, \tag {5}
+$$
+
+where  $\alpha$  and  $\beta$  are hyper-parameters.
+
+# 5. Experiments
+
+# 5.1. Implement Details
+
+Datasets. Following previous works, we evaluate our method on VOC2012 [12] and ADE20K [56], which are the standard CSS benchmarks. We report the performance on the official validation set of each dataset. We also exclude  $20\%$  of the training set as a validation set to tune hyperparameters and perform all analyses.
+
+Protocols. There are two settings in CSS benchmarks: disjoint and overlapped. The disjoint setting limits the given images to contain pixels only belong to  $\mathcal{C}_{1:t}$ , while the latter does not have this restriction, which means the pixels can also belong to the future classes, i.e.,  $\mathcal{C}_{t+1:T}$ . Note that the overlapped setting is considered to be more realistic and more challenging by all existing works [5, 10, 6, 37], and the disjoint setting is even ignored in some works [6, 37]. In design, our work is fully based on the overlapped setting. Evaluation. The CSS benchmark setups are denoted as  $N_{o} - N_{n}$  where  $N_{o}$  and  $N_{n}$  are the number of classes to be learned at the initial step and each incremental step, respectively. For example, the 15-1 setup of VOC2012 gives 15 classes at the initial step and adds one class per incremental step, resulting in a total of 6 steps. On VOC2012, we conduct experiments on four setups: 19-1 (2 steps), 15-5 (2 steps), 15-1 (6 steps), and 10-1 (11 steps). On ADE20K, we use three setups: 100-50 (2 tasks), 50-50 (3 tasks), and 100-10 (6 tasks). The network is evaluated after finishing all learning steps using mean Intersection over Union (mIoU) as metric. The performance of initial classes  $\mathcal{C}_{1}$ , incremental classes  $\mathcal{C}_{2:T}$ , and all classes  $\mathcal{C}_{1:T}$  are denoted as old, new, and all, respectively.
+
+Training. Following previous works [5, 10], we adopt DeepLab-V3 with an output stride of 16. ResNet-101 serves as the backbone, which is pre-trained on ImageNet. We use the SGD optimizer with momentum set to 0.9. The learning rate is set to 1e-2 for the initial learning step and 1e-3 for the following learning step. The learning rate is adjusted by the polynomial decay schedule with power set to 0.9. The network is trained on VOC2012 for 30 epochs, and 60 epochs on ADE20K, both with a batch size of 16. The image is cropped to  $512 \times 512$  for both training and validation, with
+
+Table 1. Results of different methods on VOC2012. The best is in bold while the second best is in underline. Results are averaged over 3 runs with standard deviation provided.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">19-1 (2 tasks)</td><td colspan="3">15-5 (2 tasks)</td><td colspan="3">15-1 (6 tasks)</td><td colspan="3">10-1 (11 tasks)</td></tr><tr><td>old</td><td>new</td><td>all</td><td>old</td><td>new</td><td>all</td><td>old</td><td>new</td><td>all</td><td>old</td><td>new</td><td>all</td></tr><tr><td>ILT [35]</td><td>67.1</td><td>12.3</td><td>64.4</td><td>66.3</td><td>40.6</td><td>59.9</td><td>4.9</td><td>7.8</td><td>5.7</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MiB [5]</td><td>70.2</td><td>22.1</td><td>67.8</td><td>75.5</td><td>49.4</td><td>69.0</td><td>35.1</td><td>13.5</td><td>29.7</td><td>12.2</td><td>13.0</td><td>12.6</td></tr><tr><td>PLOP [10]</td><td>75.3</td><td>37.3</td><td>73.5</td><td>75.7</td><td>51.7</td><td>70.0</td><td>65.1</td><td>21.1</td><td>54.6</td><td>44.0</td><td>15.5</td><td>22.9</td></tr><tr><td>SDR [36]</td><td>69.1</td><td>32.6</td><td>67.4</td><td>75.4</td><td>52.6</td><td>69.9</td><td>44.7</td><td>21.8</td><td>39.2</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SSUL [6]</td><td>77.7</td><td>29.6</td><td>75.4</td><td>77.8</td><td>50.1</td><td>71.2</td><td>77.3</td><td>36.5</td><td>67.6</td><td>71.3</td><td>45.9</td><td>59.2</td></tr><tr><td>RCIL [54]</td><td>-</td><td>-</td><td>-</td><td>78.8</td><td>52.0</td><td>72.4</td><td>70.6</td><td>23.7</td><td>59.4</td><td>55.4</td><td>15.1</td><td>34.3</td></tr><tr><td>ALIFE [37]</td><td>76.6</td><td>49.3</td><td>75.3</td><td>77.1</td><td>52.5</td><td>71.3</td><td>64.4</td><td>34.9</td><td>57.4</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Ours</td><td>77.7 ±0.22</td><td>59.8 ±1.38</td><td>76.8 ±0.35</td><td>79.1 ±0.21</td><td>67.7 ±0.46</td><td>76.4 ±0.29</td><td>75.2 ±0.75</td><td>52.2 ±1.33</td><td>69.7 ±1.09</td><td>68.1 ±0.96</td><td>54.5 ±1.57</td><td>61.6 ±1.13</td></tr><tr><td>RECALL [34]</td><td>68.1</td><td>55.3</td><td>68.6</td><td>67.7</td><td>54.3</td><td>65.6</td><td>67.8</td><td>50.9</td><td>64.8</td><td>65.0</td><td>53.7</td><td>60.7</td></tr><tr><td>SSUL-M [6]</td><td>77.8</td><td>49.7</td><td>76.4</td><td>78.4</td><td>55.8</td><td>73.0</td><td>78.3</td><td>49.0</td><td>71.3</td><td>74.0</td><td>53.2</td><td>64.1</td></tr><tr><td>ALIFE-M [37]</td><td>76.7</td><td>52.2</td><td>75.5</td><td>77.6</td><td>55.2</td><td>72.3</td><td>66.0</td><td>38.8</td><td>59.5</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Ours-M</td><td>77.9 ±0.19</td><td>67.7 ±0.74</td><td>77.4 ±0.32</td><td>79.3 ±0.22</td><td>70.2 ±0.29</td><td>77.1 ±0.24</td><td>77.1 ±0.65</td><td>60.4 ±0.73</td><td>73.1 ±0.68</td><td>69.5 ±0.60</td><td>63.2 ±1.25</td><td>66.5 ±0.78</td></tr><tr><td>Joint</td><td>77.6</td><td>77.7</td><td>77.6</td><td>79.5</td><td>71.5</td><td>77.6</td><td>79.5</td><td>71.5</td><td>77.6</td><td>78.5</td><td>76.6</td><td>77.6</td></tr></table>
+
+Table 2. Results of different methods on ADE20K. The best is in bold while the second best is in underline. Results are averaged over 3 runs with standard deviation provided.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">100-50 (2 tasks)</td><td colspan="3">50-50 (3 tasks)</td><td colspan="3">100-10 (6 tasks).</td></tr><tr><td>old</td><td>new</td><td>all</td><td>old</td><td>new</td><td>all</td><td>old</td><td>new</td><td>all</td></tr><tr><td>ILT [35]</td><td>18.2</td><td>14.4</td><td>17.0</td><td>3.5</td><td>12.8</td><td>9.7</td><td>0.1</td><td>1.3</td><td>0.5</td></tr><tr><td>MiB [5]</td><td>40.5</td><td>17.1</td><td>32.7</td><td>45.5</td><td>21.0</td><td>29.3</td><td>38.2</td><td>11.1</td><td>29.2</td></tr><tr><td>PLOP [10]</td><td>41.8</td><td>14.8</td><td>32.9</td><td>48.8</td><td>20.9</td><td>30.4</td><td>40.4</td><td>13.6</td><td>31.5</td></tr><tr><td>SSUL [6]</td><td>41.2</td><td>18.0</td><td>33.5</td><td>48.3</td><td>20.1</td><td>29.5</td><td>40.2</td><td>18.7</td><td>33.1</td></tr><tr><td>RCIL [54]</td><td>42.3</td><td>18.8</td><td>34.5</td><td>48.3</td><td>25.0</td><td>32.9</td><td>39.3</td><td>17.6</td><td>32.1</td></tr><tr><td>ALIFE [37]</td><td>42.1</td><td>23.0</td><td>35.8</td><td>48.9</td><td>25.6</td><td>33.5</td><td>41.0</td><td>22.7</td><td>34.9</td></tr><tr><td>Ours</td><td>43.1 ±0.17</td><td>26.0 ±0.36</td><td>37.4 ±0.25</td><td>49.1 ±0.24</td><td>28.3 ±0.43</td><td>35.2 ±0.33</td><td>41.4 ±0.28</td><td>25.5 ±0.45</td><td>36.1 ±0.35</td></tr><tr><td>SSUL-M [6]</td><td>42.7</td><td>17.5</td><td>34.3</td><td>49.1</td><td>20.1</td><td>29.7</td><td>42.8</td><td>17.6</td><td>34.4</td></tr><tr><td>ALIFE-M [37]</td><td>42.2</td><td>23.5</td><td>36.0</td><td>49.0</td><td>26.1</td><td>33.8</td><td>41.1</td><td>23.0</td><td>35.1</td></tr><tr><td>Ours-M</td><td>43.1 ±0.15</td><td>26.2 ±0.29</td><td>37.5 ±0.21</td><td>49.2 ±0.12</td><td>28.5 ±0.27</td><td>35.4 ±0.18</td><td>41.6 ±0.21</td><td>25.8 ±0.26</td><td>36.3 ±0.22</td></tr><tr><td>Joint</td><td>43.5</td><td>29.4</td><td>38.8</td><td>50.3</td><td>32.7</td><td>38.8</td><td>43.5</td><td>29.4</td><td>38.8</td></tr></table>
+
+![](images/8066eebb0a9e448826a25a517bbc4d23e5e8c44466d09326524826b0c4261577.jpg)  
+Figure 6. mIoU comparison between CSS methods over steps on VOC 15-1 setups. The performance of initial classes and each new class added per step are plotted separately from left to right. Our method shows great ability in learning new classes.
+
+![](images/1456f2f508fda48f3552742c5238f4220be55d929a74f03258c36cdbf422fa04.jpg)
+
+![](images/dac3a1bc85afb87d35a21c63297a342ca4d3799976f2363f7fe10305dd3f7269.jpg)
+
+![](images/d9b160afa6126415c6d98b25328ea13e8e956f48c714c7043319d87f48c4b57a.jpg)
+
+![](images/f72d05b7f88f7d937ef29b727e0e21a3d6c265214cd0beee4f49af21cc17dcb3.jpg)
+
+![](images/8c7e91b67ba41d4d2802679391173cca990126900e950f96ecc2612a4a149e96.jpg)
+
+a random scaling and a random left-right flip additionally applied to training.
+
+# 5.2. Results
+
+We provide the results on VOC2012 and ADE20K in Table 1 and Table 2, respectively. Some works [6, 34] exploit memory or external data to assist training. We also report the results obtained using memory (denoted by -M) following the implementation of [6] to fairly compare with these methods. The result of joint training is also provided as the upperbound for reference.
+
+For all setups on VOC2012, our method surpasses all existing methods by a large margin on new classes. This demonstrates the strong ability of our method to learn new classes. Note that our method also shows promising results on old classes even with a simple approach to preserve old knowledge. This is due to that a higher performance on new classes indicates that there also might be less confusion between old and new classes, which also benefits the mIoU of old classes. On the more challenging setups with multiple steps (15-1, 10-1), the new classes cannot be properly learned in the first place for most methods. To
+
+make matters worse, these poorly optimized new classes will suffer forgetting in the following steps which further reduces their performance. Even in these challenging circumstances, our method still shows great performance in learning new classes, as shown in Figure 6. We can also see that even though SSUL shows higher performance on old classes on long-term setups, it relies on completely freezing the feature extractor to prevent forgetting, which does help a lot for old classes on these steps, but also harms its flexibility greatly. We believe that if equipped with a modern knowledge-preserving constraint, our method could offer another boost in performance.
+
+On the more challenging ADE20K benchmarks, we obtain similar results as that of VOC2012. Our method shows some improvement on old classes and is significantly better on new classes. On top of the results of VOC2012 benchmark, this further demonstrates that our approach works well on more complex datasets, and has the ability to handle both scenes and objects centric datasets.
+
+When memory is applied on both benchmarks, our method receives further improvements and holds the leading position. This also indicates that our proposed training method has a non-overlapping effect with the memory technique. They can be used as two complementary techniques as pointed out in Section 3.4
+
+# 5.3. Ablation Study
+
+We investigate the effects of the proposed components in our framework using both 15-1 and 15-5 setups on VOC2012. Starting from a baseline composed of  $L_{ce}$ ,  $L_{kd}$  and pseudo label, we add the components one by one and show the ablation results in Table 3. As can be seen, the performance increase steadily as each proposed component is gradually added. Only applying contrast by visual similarity does not receive much improvement since it cannot instruct the network to capture semantic-level information. After applying contrast by feature affinity, the performance receives the highest gain, indicating the necessity to add semantic-level supervision and make the network learn more accurate knowledge. On top of that, the performance receives another boost from our classifier initialization technique which introduces future knowledge to output space.
+
+Table 3. Ablation study on VOC2012 15-1 and 15-5 setup.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">15-1 (6 tasks)</td><td colspan="3">15-5 (2 tasks)</td></tr><tr><td>old</td><td>new</td><td>all</td><td>old</td><td>new</td><td>all</td></tr><tr><td>baseline</td><td>65.3</td><td>18.8</td><td>54.2</td><td>76.9</td><td>51.3</td><td>70.8</td></tr><tr><td>+ visual similarity contrast</td><td>67.1</td><td>26.3</td><td>57.4</td><td>77.2</td><td>54.5</td><td>71.8</td></tr><tr><td>+ feature affinity contrast</td><td>74.5</td><td>45.1</td><td>67.5</td><td>78.7</td><td>65.4</td><td>75.5</td></tr><tr><td>+ classifier initialization</td><td>75.2</td><td>52.2</td><td>69.7</td><td>79.1</td><td>67.7</td><td>76.4</td></tr></table>
+
+![](images/40581768f3806ea880e27c8aa7a621e968f63154814aa1b47d483e2c1282af1f.jpg)  
+Figure 7. Visualization of clustering results of direct clustering and our practice. The high-confidence segments selected by our criterion are also presented on the right.
+
+![](images/8218f054600d24222cc7444f31ca5dd8a5fb4bc3003c0c7e2acc0718c776b4d1.jpg)  
+Figure 8. Performance comparison between different classifier initialization over training epochs on VOC2012 19-1 setup.
+
+# 5.4. Further Discussion
+
+Feature Quality Analysis. To further demonstrate the effect of our method to solve the underfitting problem pointed out in Section 3.2, we conduct the same analysis on our method and incorporate the findings into Figure 2 (denoted by Ours). As can be seen, pre-learning future classes can significantly improve both intra-class and inter-class fitness.
+
+Segment Generation. We provide a qualitative comparison between direct clustering and our proposed hierarchical combination at different granularities to validate our design in Section 4.1. The results are generated from the same intermediate features and are plotted in Figure 7. As can be seen, our practice consistently generates better results across different granularities. We also present the high-confidence segments selected by our criterion for each image, which match the actual classes with high precision.
+
+Classifier Initialization. We give a comparison of performance over training epochs for different classifier initialization methods in Figure 8. As can be seen, the initialization technique proposed by MiB [5] converges much slower due to larger update paces. Random initialization converges fast but the performance stops increasing shortly afterward. Among all of them, our method stands out with its fast convergence speed and superior performance.
+
+# 6. Limitation and Conclusion
+
+Our method has two limitations. Firstly, it relies on the natural attribute of CSS that unlabeled future classes exist in current training data. So it may not perform well when there are only a few unlabeled future classes in the current training data. However, in such cases, we can still leverage external unlabeled data, such as web-crawled data as used in RECALL [34]. Secondly, we point out that even though our proposed framework achieves great performance, it as a practicable solution might not be able to unleash the full potential of the training method, which requires further exploration in future works.
+
+In this work, we present an in-depth analysis of the reason why existing methods suffer low performance on new classes. Building on this analysis, we introduce a novel training pipeline for CSS that optimizes potential future classes in the current training step. This offers a new perspective for enhancing new class performance in contrast to existing works that focus on addressing the R-P dilemma. We propose a framework based on this pipeline and it shows a significant improvement for new classes, while also benefiting the performance of old classes. We believe that this work can provide valuable insights for the research community on CSS.
+
+# 7. Acknowledgements
+
+This work is supported by the National Natural Science Foundation of China under Grant 62176246 and Grant 61836008. This work is also supported by Anhui Province Key Research and Development Plan (202304a05020045), Anhui Provincial Natural Science Foundation 2208085UD17 and the Fundamental Research Funds for the Central Universities (WK3490000006).
+
+# References
+
+[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, 2018.  
+[2] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 33(5):898-916, 2010.  
+[3] Jihwan Bang, Heesu Kim, YoungJoon Yoo, Jung-Woo Ha, and Jonghyun Choi. Rainbow memory: Continual learning with a memory of diverse samples. In CVPR, 2021.  
+[4] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017.  
+[5] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background
+
+for incremental learning in semantic segmentation. In ICCV, 2020.  
+[6] Sungmin Cha, YoungJoon Yoo, Taesup Moon, et al. Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. NIPS, 2021.  
+[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020.  
+[8] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In CVPR, 2021.  
+[9] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019.  
+[10] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In CVPR, 2021.  
+[11] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020.  
+[12] Mark Everingham and John Winn. The pascal visual object classes challenge 2012 (voc2012) development kit. Pattern Analysis, Statistical Modelling and Computational Learning, Tech. Rep, 2011.  
+[13] Qiankun Gao, Chen Zhao, Bernard Ghanem, and Jian Zhang. R-dfcil: Relation-guided representation learning for data-free class incremental learning. In ECCV, 2018.  
+[14] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Region-based convolutional networks for accurate object detection and segmentation. IEEE transactions on pattern analysis and machine intelligence, 38(1):142-158, 2015.  
+[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020.  
+[16] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015.  
+[17] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In ICCV, 2019.  
+[18] Hanzhe Hu, Jinshi Cui, and Liwei Wang. Region-aware contrastive learning for semantic segmentation. In ICCV, 2021.  
+[19] Jyh-Jing Hwang, Stella X Yu, Jianbo Shi, Maxwell D Collins, Tien-Ju Yang, Xiao Zhang, and Liang-Chieh Chen. Segsort: Segmentation by discriminative sorting of segments. In ICCV, 2019.  
+[20] Ahmet Iscen, Jeffrey Zhang, Svetlana Lazebnik, and Cordelia Schmid. Memory-efficient incremental learning through feature adaptation. In ECCV, 2020.  
+[21] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, 2019.  
+[22] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In CVPR, 2022.
+
+[23] Tsung-Wei Ke, Jyh-Jing Hwang, Yunhui Guo, Xudong Wang, and Stella X Yu. Unsupervised hierarchical semantic segmentation with multiview cosegmentation and clustering transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2571-2581, 2022.  
+[24] Tsung-Wei Ke, Jyh-Jing Hwang, and Stella Yu. Universal weakly supervised segmentation by pixel-to-segment contrastive learning. In ICLR, 2021.  
+[25] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017.  
+[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017.  
+[27] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In ICML, 2019.  
+[28] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017.  
+[29] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Rmm: Reinforced memory management for class-incremental learning. NIPS, 2021.  
+[30] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In CVPR, 2020.  
+[31] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015.  
+[32] Arun Mallya, Dillon Davis, and Svetlana Lazebnik. Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In ECCV, 2018.  
+[33] Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning. In CVPR, 2018.  
+[34] Andrea Maracani, Umberto Michieli, Marco Toldo, and Pietro Zanuttigh. Recall: Replay-based continual learning in semantic segmentation. In ICCV, 2021.  
+[35] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In ICCVW, 2019.  
+[36] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In CVPR, 2021.  
+[37] Youngmin Oh, Donghyeon Baek, and Bumsub Ham. Alife: Adaptive logit regularizer and feature replay for incremental semantic segmentation. NIPS, 2022.  
+[38] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018.  
+[39] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Johannichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In CVPR, 2019.
+
+[40] Quang Pham, Chenghao Liu, and Steven Hoi. Dualnet: Continual learning, fast and slow. NIPS, 2021.  
+[41] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In ECCV, 2020.  
+[42] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017.  
+[43] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip HS Torr, Song Bai, and Vincent YF Tan. Mimicking the oracle: an initial phase decorrelation approach for class incremental learning. In CVPR, 2022.  
+[44] Pravendra Singh, Pratik Mazumder, Piyush Rai, and Vinay P Namboodiri. Rectification-based knowledge retention for continual learning. In CVPR, 2021.  
+[45] Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 254-270. Springer, 2020.  
+[46] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In ICCV, 2021.  
+[47] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In ECCV, 2022.  
+[48] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019.  
+[49] Ye Xiang, Ying Fu, Pan Ji, and Hua Huang. Incremental learning using conditional adversarial networks. In ICCV, 2019.  
+[50] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, 2021.  
+[51] Guanglei Yang, Enrico Fini, Dan Xu, Paolo Rota, Mingli Ding, Moin Nabi, Xavier Alameda-Pineda, and Elisa Ricci. Uncertainty-aware contrastive distillation for incremental semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022.  
+[52] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. In International Conference on Learning Representations, 2018.  
+[53] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, pages 3987-3995, 2017.  
+[54] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In CVPR, 2022.  
+[55] Xiao Zhang and Michael Maire. Self-supervised visual representation learning from hierarchical grouping. NIPS, 2020.  
+[56] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017.
+
+[57] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and Cheng-Lin Liu. Prototype augmentation and self-supervision for incremental learning. In CVPR, 2021.  
+[58] Yixiong Zou, Shanghang Zhang, Ke Chen, Yonghong Tian, Yaowei Wang, and José MF Moura. Compositional few-shot recognition with primitive discovery and enhancing. In Proceedings of the 28th ACM International Conference on Multimedia, pages 156-164, 2020.

2023/Preparing the Future for Continual Semantic Segmentation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d24bb8eddcad4bb943db4d6bcac98256a77d9b8c2ac9445736f6173240254d46
+size 432379

2023/Preserve Your Own Correlation_ A Noise Prior for Video Diffusion Models/12c6fb66-a455-4547-8b5a-4a378b5ee51f_content_list.json

@@ -0,0 +1,1835 @@
+[
+    {
+        "type": "text",
+        "text": "Preserve Your Own Correlation: A Noise Prior for Video Diffusion Models",
+        "text_level": 1,
+        "bbox": [
+            109,
+            130,
+            861,
+            151
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Songwei Ge*",
+        "bbox": [
+            137,
+            181,
+            243,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Maryland",
+        "bbox": [
+            116,
+            200,
+            258,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Seungjun Nah  \nNVIDIA",
+        "bbox": [
+            297,
+            181,
+            413,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Guilin Liu NVIDIA",
+        "bbox": [
+            454,
+            181,
+            539,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Tyler Poon",
+        "bbox": [
+            602,
+            181,
+            691,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Chicago",
+        "bbox": [
+            580,
+            200,
+            712,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Andrew Tao NVIDIA",
+        "bbox": [
+            753,
+            181,
+            852,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Bryan Catanzaro",
+        "bbox": [
+            94,
+            242,
+            228,
+            258
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "NVIDIA",
+        "bbox": [
+            133,
+            262,
+            189,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "David Jacobs",
+        "bbox": [
+            282,
+            242,
+            388,
+            257
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Maryland",
+        "bbox": [
+            264,
+            262,
+            406,
+            276
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jia-Bin Huang",
+        "bbox": [
+            454,
+            242,
+            571,
+            258
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Maryland",
+        "bbox": [
+            444,
+            262,
+            584,
+            276
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ming-Yu Liu",
+        "bbox": [
+            620,
+            242,
+            725,
+            258
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "NVIDIA",
+        "bbox": [
+            645,
+            262,
+            700,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yogesh Balaji",
+        "bbox": [
+            764,
+            242,
+            875,
+            258
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "NVIDIA",
+        "bbox": [
+            792,
+            262,
+            846,
+            273
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "A very happy fuzzy panda dressed as a chef eating pizza in the New York street food truck.",
+        "bbox": [
+            76,
+            512,
+            341,
+            555
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The supernova explosion of a white dwarf in the universe, photo realistic.",
+        "bbox": [
+            348,
+            520,
+            609,
+            549
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "A high-quality 3D render of hyperrealist, super strong, multicolor stripped, and fluffy bear with wings, highly detailed.",
+        "bbox": [
+            617,
+            513,
+            883,
+            556
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Figure 1: Given a text description, our approach can faithfully generate videos that are consistent with the input text while being photorealistic and temporally consistent. Best viewed with Acrobat Reader. Click the images to play the video clips.",
+        "bbox": [
+            75,
+            566,
+            892,
+            598
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            616,
+            313,
+            632
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite tremendous progress in generating high-quality images using diffusion models, synthesizing a sequence of animated frames that are both photorealistic and temporally coherent is still in its infancy. While off-the-shelf billion-scale datasets for image generation are available, collecting similar video data of the same scale is still challenging. Also, training a video diffusion model is computationally much more expensive than its image counterpart. In this work, we explore finetuning a pretrained image diffusion model with video data as a practical solution for the video synthesis task. We find that naively extending the image noise prior to video noise prior in video diffusion leads to sub-optimal performance. Our carefully designed video noise prior leads to substantially better performance. Extensive experimental validation shows that our",
+        "bbox": [
+            73,
+            648,
+            473,
+            875
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "model, Preserve Your Own COrelation (PYoCo), attains SOTA zero-shot text-to-video results on the UCF-101 and MSR-VTT benchmarks. It also achieves SOTA video generation quality on the small-scale UCF-101 benchmark with a  $10 \\times$  smaller model using significantly less computation than the prior art. The project page is available at https://research.nvidia.com/labs/dir/pyoco/.",
+        "bbox": [
+            500,
+            617,
+            893,
+            724
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            500,
+            753,
+            630,
+            768
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large-scale diffusion-based text-to-image models [38, 42, 2] have demonstrated impressive capabilities in turning complex text descriptions into photorealistic images. They can generate images with novel concepts unseen during training. Sophisticated image editing and processing tasks can easily be accomplished through guidance control and embedding techniques. Due to the immense success in several applications [30, 67, 5], these models are established as pow-",
+        "bbox": [
+            496,
+            779,
+            895,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Work done during an internship at NVIDIA.",
+        "bbox": [
+            94,
+            886,
+            336,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "22930",
+        "bbox": [
+            478,
+            924,
+            519,
+            936
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/3ea06ec868739f34b3665eb70207e13f0537015064d7770050b4c180314a5265.jpg",
+        "image_caption": [
+            "(a)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            78,
+            87,
+            524,
+            273
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/e6d7c286fd69acb9c2a133e2940d19045711e89bb8826830f27c7352a116b27c.jpg",
+        "image_caption": [
+            "(b)",
+            "Figure 2: Visualizing the noise map correlations. (a) visualizes the t-SNE plot of the noise maps corresponding to input frames randomly sampled from videos. These noise maps are obtained by running a diffusion ODE [49, 48] on the input frames using a trained text-to-image model, but in the opposite direction of image synthesis ( $\\sigma : 0 \\to \\sigma_{\\max}$ ). The green dots in the background denote the reference noise maps sampled from an i.i.d. Gaussian distribution. The red dots and yellow dots are noise maps corresponding to input frames coming from different videos. We found they are spread out and share no correlation. On the other hand, the noise maps corresponding to the frames coming from the same video (shown in blue dots) are clustered together. (b) Using an i.i.d. noise model (orange dots) for finetuning text-to-image models for video synthesis is not ideal since temporal correlations between frames are not modeled. To remedy this, we propose a progressive noise model in which the correlation between different noise maps is injected along the temporal axis. Our progressive noise model (blue dots) aptly models the correlations present in the video noise maps."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            588,
+            93,
+            879,
+            271
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "erful image synthesis tools for content generation. As image synthesis is largely democratized with the success of these text-to-image models, it is natural to ask whether we can repeat the same success in video synthesis with large-scale diffusion-based text-to-video models.",
+        "bbox": [
+            75,
+            483,
+            470,
+            558
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Multiple attempts have been made to build large-scale video diffusion models. Ho et al. [17] proposed a UNet-based architecture for the video synthesis task that is trained using joint image-video denoising losses.Imagen video [14] extends the cascaded text-to-image generation architecture ofImagen [42] for video generation. In both works, the authors directly train a video generation model from scratch. While these approaches achieve great success and produce high-quality videos, they are inherently expensive to train, requiring hundreds of high-end GPUs or TPU's and several weeks of training. After all, video generators not only need to learn to form individual images but should also learn to synthesize coherent temporal dynamics, which makes the video generation task much more challenging. While the formation of individual frames is a shared component in an image and video synthesis, these works disregard the existence of powerful pretrained text-to-image diffusion models and train their video generators from scratch.",
+        "bbox": [
+            75,
+            563,
+            470,
+            835
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We explore a different avenue for building large-scale text-to-video diffusion models by starting with a pretrained text-to-image diffusion model. Our motivation is that most of the components learned for the image synthesis task can",
+        "bbox": [
+            75,
+            839,
+            468,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "effectively be reused for video generation, leading to knowledge transfer and efficient training. A similar idea is adopted by several recent works [46, 69, 4]. Without exception, when finetuning, they naively extend the image diffusion noise prior (i.i.d. noise) used in the text-to-image model to a video diffusion noise prior by adding an extra dimension to the 2D noise map. We argue that this approach is not ideal as it does not utilize the natural correlations in videos that are already learned by the image models. This is illustrated in Figure 2, where we visualize the t-SNE plot of noise maps corresponding to different input frames as obtained from a pretrained text-to-image diffusion model. The noise maps corresponding to different frames coming from the same video (blue dots in Figure 2a are clustered together, exhibiting a high degree of correlation. The use of i.i.d. noise prior does not model this correlation, which would impede the finetuning process. Our careful analysis of the video diffusion noise prior leads us to a noise prior that is better tailored for finetuning an image synthesis model to the video generation task. As illustrated in Figure 2b, our proposed noise prior (shown in blue dots) aptly captures the correlations in noise maps corresponding to video frames.",
+        "bbox": [
+            496,
+            483,
+            893,
+            816
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We then proceed to build a large-scale diffusion-based text-to-video model. We leverage several design choices from the prior works, including the use of temporal attention [17], joint image-video finetuning [17], a cascaded generation architecture [14], and an ensemble of expert denois-",
+        "bbox": [
+            496,
+            824,
+            895,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "22931",
+        "bbox": [
+            478,
+            924,
+            517,
+            936
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "ers [2]. Together with these techniques and the proposed video noise prior, our model establishes a new state-of-the-art for video generation outperforming competing methods on several benchmark datasets. Figure 1 shows our model can achieve high-quality zero-shot video synthesis capability with SOTA photorealism and temporal consistency.",
+        "bbox": [
+            75,
+            90,
+            468,
+            181
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In short, our work makes the following key contributions.",
+        "bbox": [
+            96,
+            181,
+            468,
+            196
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. We propose a video diffusion noise tailored for finetuning text-to-image diffusion models for text-to-video.",
+            "2. We conduct extensive experimental validation and verify the effectiveness of the proposed noise prior.",
+            "3. We build a large-scale text-to-video diffusion model by finetuning a pretrained eDiff-I model with our noise prior and achieve state-of-the-art results on several benchmarks."
+        ],
+        "bbox": [
+            70,
+            205,
+            468,
+            319
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            330,
+            218,
+            345
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Diffusion-based text-to-image models: Diffusion models have significantly advanced the progress of text-based photorealistic, compositional image generation [38, 42]. Given the nature of the iterative denoising process that requires massive numbers of score function evaluations, earlier diffusion models focused on generating low-resolution images, e.g.,  $64 \\times 64$  [15, 48]. To generate high-resolution images, two common approaches have been used. The first approach applies cascaded super-resolution models in the RGB space [32, 16, 42, 38], while the second approach leverages a decoder to exploit latent space [40, 11]. Based on these models, advanced image and video editing have been achieved through finetuning the model [41, 67, 5, 23, 61, 29] or controlling the inference process [30, 13, 34, 10, 35, 7, 31, 3]. Here, we study the problem of using large-scale diffusion models for text-to-video generation.",
+        "bbox": [
+            75,
+            357,
+            468,
+            598
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Video generation models: Generating realistic and novel videos have long been an attractive and essential research direction [58, 39, 65]. Previously studies have resorted to different types of generative models such as GANs [58, 43, 54, 52, 45], Autoregressive models [51, 63, 25, 9, 18], and implicit neural representations [47, 66]. Recently, driven by the tremendous success of applying the diffusion model to image synthesis, multiple works have proposed to explore diffusion models for conditional and unconditional video synthesis [57, 12, 69, 61, 4, 22, 19, 57, 64, 33, 28, 1, 59]. For example, Singer et al. extend the unCLIP framework [38] to text-to-video generation, which allows training without video captions [46]. Ho et al. [17] extend the Imagen framework [42] by repeatedly up-scaling low-resolution small-fps videos in both spatial and temporal directions with multiple models [14]. Our work also falls into this line of work which uses a diffusion model. We focus on augmenting an image diffusion model for video and study the design choice of the diffusion noise priors for such an image-to-video finetuning task.",
+        "bbox": [
+            75,
+            598,
+            470,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Leverage knowledge from images for text-to-video generation: Like text-to-image models, text-to-video models require massive amounts of data to learn caption-relatedness, frame photorealism, and temporal dynamics. But in contrast to the abundant image data resource, video data are more limited in style, volume, and quality. To resolve such scarcity issue of text-video data, previous works have resorted to different strategies to leverage knowledge from image data for text-to-video generation, including joint training on the text-image data from scratch [17, 14, 56, 60], first training a text-to-image model and then finetuning partially [18, 4, 61, 29] or entirely [46, 8] on the video dataset, and using CLIP image features as the conditional information [46, 69]. In this paper, we propose a new video diffusion noise prior that is tailored for finetuning a pretrained diffusion-based image generation model for the video generation task. We reuse several design choices in the prior work by finetuning jointly on text-image and text-video datasets. As a result, we can build a text-to-video generation system that achieves state-of-the-art zero-shot performances.",
+        "bbox": [
+            496,
+            90,
+            893,
+            393
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            500,
+            409,
+            637,
+            425
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Diffusion models generate data by iteratively denoising samples drawn from a noise distribution. In the case of text-to-video models, text embeddings obtained from a pre-trained text encoder are used as additional inputs in the denoising process. Formally, let  $D(\\mathbf{x}, \\mathbf{e}, \\sigma)$  denote a denoising network that operates on the noisy input video  $\\mathbf{x} \\in \\mathbb{R}^{b \\times n_s \\times 3 \\times h \\times w}$  where  $\\mathbf{e}$  is the text embedding, and  $\\sigma$  is the noise level. Here  $n_s$  is the sequence length of the input video,  $b$  is the batch size, and  $h \\times w$  is the spatial resolution. The model  $D$  is trained to denoise the input  $\\mathbf{x}$ .",
+        "bbox": [
+            496,
+            435,
+            893,
+            585
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Training We follow the EDM formulation of Karras et al. [21] to optimize the denoiser  $D$  using the following objective",
+        "bbox": [
+            500,
+            609,
+            893,
+            640
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathbb {E} _ {p _ {\\text {d a t a}} (\\mathbf {x} _ {\\text {c l e a n}}, \\mathbf {e}), p (\\epsilon), p (\\sigma)} \\left[ \\lambda (\\sigma) \\| D \\left(\\mathbf {x} _ {\\text {n o i s e}}; \\mathbf {e}, \\sigma\\right) - \\mathbf {x} _ {\\text {c l e a n}} \\| _ {2} ^ {2} \\right] \\tag {1} \\\\ w h e r e \\quad \\mathbf {x} _ {\\text {n o i s e}} = \\mathbf {x} _ {\\text {c l e a n}} + \\sigma \\epsilon \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            514,
+            652,
+            890,
+            704
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Here,  $\\mathbf{x}_{\\mathrm{noise}}$  is the noisy sample obtained by corrupting the clean video  $\\mathbf{x}$  with noise  $\\sigma \\epsilon$ , where  $p(\\epsilon) = \\mathcal{N}(\\mathbf{0},\\mathbf{I})$  and  $\\sigma$  is a scalar for the noise level drawn from  $p(\\sigma)$ . The loss weight,  $\\lambda (\\sigma)$ , is a function of  $\\sigma$  given by  $\\lambda (\\sigma) = (\\sigma^2 +\\sigma_{\\mathrm{data}}^2) / (\\sigma \\cdot \\sigma_{\\mathrm{data}})^2$ . Eq. (1) is a simple denoising objective in which the denoiser  $D$  is trained to estimate the clean video  $\\mathbf{x}_{\\mathrm{clean}}$  from the noisy input  $\\mathbf{x}_{\\mathrm{noise}}$ . Following EDM, we use a log-normal distribution for  $\\sigma$  i.e.,  $\\ln (p(\\sigma)) = \\mathcal{N}(P_{\\mathrm{mean}},P_{\\mathrm{std}}^2)$  with  $P_{\\mathrm{mean}} = -1.2$  and  $P_{\\mathrm{std}} = 1.2$ .",
+        "bbox": [
+            496,
+            718,
+            893,
+            853
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To train the denoising model, EDM uses preconditioning terms in its objective function to properly scale the inputs and output of the denoiser model  $D$ . More specifically, the",
+        "bbox": [
+            496,
+            854,
+            890,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "22932",
+        "bbox": [
+            478,
+            924,
+            517,
+            936
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "denoising model  $D$  is written as",
+        "bbox": [
+            76,
+            90,
+            292,
+            104
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nD (\\mathbf {x}; \\mathbf {e}, \\sigma) := \\left(\\frac {\\sigma_ {\\mathrm {d a t a}}}{\\sigma^ {*}}\\right) ^ {2} \\mathbf {x} + \\frac {\\sigma \\cdot \\sigma_ {\\mathrm {d a t a}}}{\\sigma^ {*}} F _ {\\theta} \\left(\\frac {\\mathbf {x}}{\\sigma^ {*}}; \\mathbf {e}, \\frac {\\ln (\\sigma)}{4}\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            94,
+            116,
+            450,
+            148
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Here,  $F_{\\theta}$  is a neural network with parameters  $\\theta$  and  $\\sigma^{*} = \\sqrt{\\sigma^{2} + \\sigma_{\\mathrm{data}}^{2}}$ . We use  $\\sigma_{\\mathrm{data}} = 0.5$ .",
+        "bbox": [
+            76,
+            156,
+            468,
+            188
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Sampling Once the denoising model is trained, sampling can be performed by solving the following ODE [21]",
+        "bbox": [
+            76,
+            205,
+            468,
+            234
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\frac {d \\mathbf {x}}{d \\sigma} = - \\sigma \\nabla_ {\\mathbf {x}} \\log p (\\mathbf {x} | \\mathbf {e}, \\sigma) = \\frac {\\mathbf {x} - D (\\mathbf {x} ; \\mathbf {e} , \\sigma)}{\\sigma} \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            119,
+            244,
+            468,
+            276
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "for  $\\sigma$  flowing backwards from  $\\sigma = \\sigma_{\\mathrm{max}}$  to  $\\sigma = 0$ . The initial value for  $\\mathbf{x}$  is obtained by sampling from the prior distribution  $\\mathbf{x} \\sim \\mathcal{N}(\\mathbf{0}, \\sigma_{\\mathrm{max}}^2 \\mathbf{I})$ . Over the recent years, several samplers have been proposed for sampling from the trained diffusion models [68, 48, 26, 27, 15]. In this paper, we use DEIS [68] and its stochastic variant [21] for synthesizing samples from our model.",
+        "bbox": [
+            76,
+            285,
+            470,
+            390
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4. Method",
+        "text_level": 1,
+        "bbox": [
+            76,
+            402,
+            168,
+            419
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Training text-to-video models is much more challenging than training text-to-image diffusion models due to practical difficulties in collecting billion-scale video datasets and securing enough computational resources. Additionally, generating videos is much more challenging since individual frames need to be both photorealistic and temporally coherent. Prior works leverage large-scale image datasets to mitigate these difficulties by either joint training on the image datasets [60, 17, 14] or finetuning a text-to-image model on the video datasets [18, 46]. Here, we are interested in finetuning text-to-image diffusion models jointly on image and video datasets. We postulate that naively extending the image noise prior to video diffusion is not ideal. We carefully explore the design space of noise priors and propose one that is well suited for our video finetuning task, which leads to significant performance gains.",
+        "bbox": [
+            75,
+            429,
+            472,
+            672
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Correlated noise model An image diffusion model is trained to denoise independent noise from a perturbed image. The noise vector  $\\epsilon$  in the denoising objective (1) is sampled from an i.i.d. Gaussian distribution  $\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})$ . However, after training the image diffusion model and applying it to reverse real frames from a video into the noise space in a per-frame manner, we find that the noise maps corresponding to different frames are highly correlated. This is illustrated in Figure 2, where the t-SNE plot of noise maps corresponding to different video frames are plotted. When the input frames come from the same video (shown in blue dots in Figure 2a, noise maps are clustered. The use of i.i.d. sampling (shown in orange dots in Figure 2b does not capture these correlations. This is also depicted quantitatively",
+        "bbox": [
+            75,
+            688,
+            472,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Cosine similarity of the reversed noise. The noise maps corresponding to the frames sampled from the same videos have a higher similarity than those sampled from different videos.",
+        "bbox": [
+            496,
+            88,
+            893,
+            148
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/4383742137095222d6e0d6fa310f41451b12d67c8f340416c502a161b125485a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Cosine Similarity</td></tr><tr><td>(a) Same video noise</td><td>0.206±0.156</td></tr><tr><td>(b) Different video noise</td><td>0.001±0.009</td></tr></table>",
+        "bbox": [
+            529,
+            154,
+            862,
+            215
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "in Table 1 where we compute the average pairwise cosine similarity between noise corresponding to (a) same video and (b) different video. (a) is much higher than (b). As a result, the video diffusion model trained with i.i.d. noise is coerced to forget such correlation among the noise between different frames, making it difficult to preserve knowledge from the image diffusion model. Motivated by this observation, we propose to modify the noise process to preserve the correlation between different frames. To this end, we investigate two noisng strategies - mixed and progressive noising.",
+        "bbox": [
+            496,
+            237,
+            893,
+            402
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Mixed noise model: Let  $\\epsilon^1, \\epsilon^2, \\ldots, \\epsilon^{n_s}$  denote the noise corresponding to individual video frames i.e.,  $\\epsilon^i$  corresponds to the  $i^{th}$  element of the noise tensor  $\\epsilon$ . In the mixed noise model, we generate two noise vectors  $\\epsilon_{\\mathrm{shared}}$  and  $\\epsilon_{\\mathrm{ind}}$ .  $\\epsilon_{\\mathrm{shared}}$  is a common noise vector shared among all video frames, while  $\\epsilon_{\\mathrm{ind}}$  is the individual noise per frame. The linear combination of both these vectors is used as the final noise.",
+        "bbox": [
+            496,
+            402,
+            893,
+            508
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\epsilon_ {\\text {s h a r e d}} \\sim \\mathcal {N} \\left(\\mathbf {0}, \\frac {\\alpha^ {2}}{1 + \\alpha^ {2}} \\mathbf {I}\\right), \\epsilon_ {\\text {i n d}} ^ {i} \\sim \\mathcal {N} \\left(\\mathbf {0}, \\frac {1}{1 + \\alpha^ {2}} \\mathbf {I}\\right) \\tag {3} \\\\ \\epsilon^ {i} = \\epsilon_ {\\mathrm {s h a r e d}} + \\epsilon_ {\\mathrm {i n d}} ^ {i} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            511,
+            512,
+            893,
+            566
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Progressive noise model: In the progressive noise model, the noise for each frame is generated in an autoregressive fashion in which the noise at frame  $i$  is generated by perturbing the noise at frame  $i - 1$ . Let  $\\epsilon_{\\mathrm{ind}}^i$  denote the independent noise generated for frame  $i$ . Then, progressive noising can be formulated as",
+        "bbox": [
+            496,
+            571,
+            893,
+            662
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\epsilon^ {0} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}) \\quad \\epsilon_ {\\mathrm {i n d}} ^ {i} \\sim \\mathcal {N} (\\mathbf {0}, \\frac {1}{1 + \\alpha^ {2}} \\mathbf {I}) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            568,
+            667,
+            893,
+            699
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\epsilon^ {i} = \\frac {\\alpha}{\\sqrt {1 + \\alpha^ {2}}} \\epsilon^ {i - 1} + \\epsilon_ {\\mathrm {i n d}} ^ {i}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            629,
+            700,
+            803,
+            729
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In both these models,  $\\alpha$  controls how much noise is shared among different video frames. The higher the value of  $\\alpha$ , the more correlation exists among the noise maps corresponding to different frames. As  $\\alpha \\rightarrow \\infty$ , all frames would have the same noise which results in generating a frozen video. On the other hand,  $\\alpha = 0$  corresponds to i.i.d. noise.",
+        "bbox": [
+            496,
+            734,
+            893,
+            824
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 2b, the use of progressive noise sampling (blue dots) better models the correlations between different noise maps by obtaining similar clustering patterns to the noise maps of real video frames embedded by a pretrained text-to-image model in Figure 2a (blue dots).",
+        "bbox": [
+            496,
+            825,
+            893,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "footer",
+        "text": "22933",
+        "bbox": [
+            478,
+            924,
+            519,
+            936
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/8d6b3d4c8df87f34a0ce9703c0a7f610b6687418e672edb194d3268825477a99.jpg",
+        "image_caption": [
+            "Figure 3: Model architecture. Our pipeline consists of a cascade of four networks — a base model and three upsampling models. All four models take inputs as the text embeddings obtained from the T5 encoder and the CLIP text encoder. The base model produces 16 video frames of spatial resolution  $64 \\times 64$  with a frameskip of 5. The first upsampling model performs a temporal interpolation, resulting in videos of size  $76 \\times 64 \\times 64$  while the subsequent two super-resolution models perform spatial super-resolution to produce videos of sizes  $76 \\times 256 \\times 256$  and  $76 \\times 1024 \\times 1024$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            88,
+            893,
+            296
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Model architecture As visualized in Figure 3, our model consists of a cascade of four networks — a base network and three upsampling stacks. The base network generates an output video of dimension  $16 \\times 64 \\times 64$  with a frameskip of 5. It generates the frames  $\\{1,6,11,\\ldots,76\\}$ . The first upsampling network performs a temporal interpolation to produce a video of size  $76 \\times 64 \\times 64$ . The second and the third super-resolution network performs spatial upsampling to produce the outputs of sizes  $76 \\times 256 \\times 256$  and  $76 \\times 1024 \\times 1024$ . We utilize eDiff-I [2], a state-of-the-art text-to-image diffusion model, to initialize our base and spatial super-resolution models. Similar to prior works [17, 46], we adapt the image-based U-Net model for the video synthesis task by making the following changes: (1) Transforming 2D convolutions to 3D by adding a dimension 1 to temporal axis and (2) Adding temporal attention layers. Please refer to the supplementary material for more details.",
+        "bbox": [
+            75,
+            409,
+            472,
+            666
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Similar to Ho et al. [17], we jointly finetune the model on video and image datasets by concatenating videos and images in the temporal axis and applying our temporal modules only on the video part. Similarly to eDiff-I, our model uses both T5 text embeddings [37] and CLIP text embeddings [36]. We drop each of the embeddings independently at random during training, as in eDiff-I.",
+        "bbox": [
+            75,
+            666,
+            470,
+            772
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            784,
+            209,
+            801
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we evaluate our proposed strategy of training diffusion models for video synthesis on two sets of experiments. We first comprehensively analyze our proposed noise model on the small-scale UCF-101 dataset. We then scale up our experiments to the challenging large-scale text-to-video synthesis task.",
+        "bbox": [
+            75,
+            810,
+            472,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1. Experimental Setups",
+        "text_level": 1,
+        "bbox": [
+            500,
+            409,
+            696,
+            425
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We conduct ablation experiments in a small-scale unconditional video generation setting and pick the best configuration for our large-scale text-to-video generation run.",
+        "bbox": [
+            498,
+            431,
+            895,
+            478
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Datasets We train our model on the UCF-101 dataset [50] for the small-scale experiments, where we follow the protocol defined in Ho et al. [17] to generate videos of size  $16 \\times 64 \\times 64$ . UCF-101 dataset contains 13,320 videos. We randomly sample frames from these videos to train our image synthesis model. For our large-scale experiments, we use a combination of public and proprietary datasets for text-to-image and text-to-video finetuning. Most of the videos are of 2K resolution with 16:9 aspect ratio. All data was filtered using a preset CLIP and aesthetic scores* to ensure high quality. Our final image dataset contains around 1.2 billion text-image pairs and 22.5 million text-video pairs.",
+        "bbox": [
+            496,
+            497,
+            893,
+            679
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training details In the unconditional generation experiment on the UCF-101 dataset, to do an ablation study on the model size, we design 3 models where each model has 69M, 112M, and 253M parameters, respectively. As a comparison, the baseline Video Diffusion Model (VDM) [17] contains 1.2B parameters. In the large-scale text-to-video experiment, our base and temporal interpolation models contain 1.08B parameters. Our super-resolution model adapted from the efficient U-Net [42] architecture with temporal convolution layers [14, 46] contains 313M parameters. Please refer to the supplementary material for more training details.",
+        "bbox": [
+            496,
+            699,
+            893,
+            866
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "\\*https://github.com/christophschuhmann/ improved-aesthetic-predictor",
+        "bbox": [
+            500,
+            875,
+            823,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "22934",
+        "bbox": [
+            478,
+            924,
+            517,
+            936
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/be0d32f6c5261750250a55f4cff915a053e9259ad86489ddce09cc007e0763ed.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            89,
+            277,
+            241
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/4b5f2ccad8a2a714b4777005a783956b00b8fe07cd922ddc9476fa3b96ba75c6.jpg",
+        "image_caption": [
+            "An epic tornado attacking above a glowing city at night, the tornado is made of smoke, highly detailed."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            277,
+            89,
+            473,
+            241
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/0b71e5432321661882bb8e9d07f88930204323a4e84990b34323694870a32f4e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            473,
+            89,
+            671,
+            241
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/7497cd615041a3013314d3ae02fc31c19b21ba4245c7efeae2c13dcc7ecf9e74.jpg",
+        "image_caption": [
+            "A golden retriever puppy holding a green sign that says \"NVIDIA ROCKS\". Background is a classroom."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            671,
+            89,
+            870,
+            241
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/31f24d40b980ff0fc7922ccd7cfcbfe5b0d5a54f03b55e39df6481d82290aaca.jpg",
+        "image_caption": [
+            "A cute corgi wearing a red robe holding a sign that says \"Merry Christmas\". There is a Christmas tree in the background."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            308,
+            276,
+            460
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/181156bf69a3ac777b674dfabb451cac0bf126ed68722d49f6b63b74c54d733f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            277,
+            308,
+            473,
+            460
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/8f9a5a97c5cd51d09128de896bf9ded108cfd243c94a7f05c4a9a840b15e6ebf.jpg",
+        "image_caption": [
+            "Small boat sailing in the ocean, giant Cthulhu monster coming out a dense mist in the background, giant waves attacking."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            473,
+            308,
+            671,
+            460
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/a8e97061fcc21c77cd0c887851bf593e0b2b7dfd929be30c5adb3933145bc9a8.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            671,
+            308,
+            870,
+            460
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/8962c3cab1893100cc256b523b53fa2604524a275958c2bb078790ce3e85c544.jpg",
+        "image_caption": [
+            "A cute funny robot dancing, centered, award winning watercolor pen illustration.",
+            "A skull burning while being held up by a skeletal hand.",
+            "Figure 4: Sample generations. Please check our project website to view the videos."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            512,
+            277,
+            662
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/e7fd793e68cfa900bbc13500147bbcb8cac48b979f63a288ea6c2e2090c78b53.jpg",
+        "image_caption": [
+            "A cartoon white wolf is giving puppy-dog eyes, detailed fur, very cute kid's film character.",
+            "A huge dinosaur skeleton walking in a golden wheat field on a bright sunny day."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            277,
+            511,
+            473,
+            662
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/8f943694a9daad4968571f69c1d1f69ba8a562ee9e68557a56bed7b52081d8ff.jpg",
+        "image_caption": [
+            "A lightning striking atop of eiffel tower, dark clouds in the sky, slow motion.",
+            "A cute rabbit is eating grass, wildlife photography."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            473,
+            511,
+            671,
+            662
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/de933d40c5692d4eb338a50bcf8966472a661d2ba6bdafedbe195396e20a70fa.jpg",
+        "image_caption": [
+            "An anime girl looks at the beautiful nature through the window of a moving train, well rendered.",
+            "Tomato sauce pouring over fries."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            671,
+            511,
+            870,
+            662
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation For the small-scale experiments on UCF-101 dataset, we follow the protocol defined in the prior approaches [52, 47, 17] and report the Inception Score (IS) [44] calculated by a trained C3D model [53] and Fréchet Video Distance (FVD) [55] by a trained I3D model [6]. For the large-scale text-to-video experiments, we perform the zero-shot evaluation of the video generation quality on the UCF-101 and MSR-VTT datasets following Make-A-Video [46]. We carefully discuss the evaluation process below.",
+        "bbox": [
+            75,
+            761,
+            473,
+            898
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "UCF-101 experiment We use IS and FVD for evaluation in our small-scale experiments. UCF-101 is a categorical video dataset designed for action recognition. When sampling from the text-to-video model, we devise a set of prompts for each class name to be used as the conditional input. This is necessary as some class names (such as jump rope) are not descriptive. We list all the prompts we use in the supplementary material. We sample 20 videos for each prompt to compute the IS metric. For FVD, we follow the",
+        "bbox": [
+            496,
+            761,
+            893,
+            898
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "22935",
+        "bbox": [
+            478,
+            924,
+            519,
+            936
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/de325ac2e516572d7c1e70adc0f690a8cdaed82b24974216a2bf2eb5007c177f.jpg",
+        "table_caption": [
+            "Table 2: Zero-shot text to video generation on UCF-101. Our approach gives significant performance gains compared to the prior baselines both in inception score and FVD metrics."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>IS (↑)</td><td>FVD (↓)</td></tr><tr><td>CogVideo [18] (Chinese)</td><td>23.55</td><td>751.34</td></tr><tr><td>CogVideo [18] (English)</td><td>25.27</td><td>701.59</td></tr><tr><td>Make-A-Video [46]</td><td>33.00</td><td>367.23</td></tr><tr><td>MagicVideo [69]</td><td>-</td><td>655.00</td></tr><tr><td>Video LDM [4]</td><td>33.45</td><td>550.61</td></tr><tr><td>VideoFactory [59]</td><td>-</td><td>410.00</td></tr><tr><td>PYoCo</td><td>47.76</td><td>355.19</td></tr></table>",
+        "bbox": [
+            120,
+            138,
+            426,
+            276
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/98ef80e11a92ad26e3ddf3b139388c1b0c92e278f338771783a4c35a635f056f.jpg",
+        "table_caption": [
+            "Table 3: Text conditional zero-shot generation on MSRVTT. Our approach with the base config achieves the best results, and using an ensemble further improves the FIDs."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>CLIP-FID (↓)</td><td>FID (↓)</td></tr><tr><td>NUWA [60] (Chinese)</td><td>47.68</td><td>-</td></tr><tr><td>CogVideo [18] (Chinese)</td><td>24.78</td><td>-</td></tr><tr><td>CogVideo [18] (English)</td><td>23.59</td><td>-</td></tr><tr><td>Make-A-Video [46]</td><td>13.17</td><td>-</td></tr><tr><td>MagicVideo [69]</td><td>-</td><td>36.50</td></tr><tr><td>Latent-Shift [1]</td><td>15.23</td><td>-</td></tr><tr><td>PYoCo (Config-A)</td><td>10.21</td><td>25.39</td></tr><tr><td>PYoCo (Config-B)</td><td>9.95</td><td>24.28</td></tr><tr><td>PYoCo (Config-C)</td><td>9.91</td><td>24.54</td></tr><tr><td>PYoCo (Config-D)</td><td>9.73</td><td>22.14</td></tr></table>",
+        "bbox": [
+            98,
+            347,
+            450,
+            530
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "prior work [25, 52] and sample 2, 048 videos for evaluation.",
+        "bbox": [
+            75,
+            556,
+            470,
+            571
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "MSR-VTT experiment MSR-VTT [62] test set contains 2,990 videos as well as 59,794 captions. All the videos have the same resolution of  $320 \\times 240$ . We generate a  $76 \\times 256 \\times 256$  video for each 59,794 caption and save the videos in an mp4 format with a high bit rate. To compare with Make-A-Video, we compute FID using a ViT-B/32 model [24]. We also report a more common FID metric computed by an Inception-V3 model. We also examine the idea of ensemble denoiser [2] by finetuning the level-1 experts of each model. We denote Config-A as the configuration of using only baseline models and Config-B to Config-D as incrementally changing super-resolution model, temporal interpolation model, and base model with the corresponding ensemble models.",
+        "bbox": [
+            75,
+            593,
+            470,
+            803
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2. Main Results",
+        "text_level": 1,
+        "bbox": [
+            76,
+            816,
+            215,
+            830
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Large-scale text-to-video synthesis We quantitatively compare our method against Make-A-Video [46], NUWA [60], CogVideo [18], and several concurrent works [4, 69, 4, 59, 1]. Table 2 shows that our method",
+        "bbox": [
+            75,
+            839,
+            470,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/f62d1eeed007f3585f1e09ece8f1a1362f075440aaf56762c2188274607b354d.jpg",
+        "table_caption": [
+            "Table 4: Unconditional UCF-101 generation results. Our approach achieves the state-of-the-art inception score and FVD, while having considerably smaller parameter count compared to other diffusion-based approaches such as VDM (1B parameters)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>IS (↑)</td><td>FVD (↓)</td></tr><tr><td>TGAN [43]</td><td>15.83±.18</td><td>-</td></tr><tr><td>LDVD-GAN [20]</td><td>22.91±.19</td><td>-</td></tr><tr><td>VideoGPT [63]</td><td>24.69±.30</td><td>-</td></tr><tr><td>MoCoGAN-HD [52]</td><td>32.36</td><td>838</td></tr><tr><td>DIGAN [66]</td><td>29.71±.53</td><td>655±22</td></tr><tr><td>CCVS [25]</td><td>24.47±.13</td><td>386±15</td></tr><tr><td>StyleGAN-V [47]</td><td>23.94±.73</td><td>-</td></tr><tr><td>VDM [17]</td><td>57.00±.62</td><td>-</td></tr><tr><td>TATS [9]</td><td>57.63±.73</td><td>430 ±18</td></tr><tr><td>PYoCo (112M)</td><td>57.93±.24</td><td>332 ±13</td></tr><tr><td>PYoCo (253M)</td><td>60.01±.51</td><td>310 ±13</td></tr></table>",
+        "bbox": [
+            526,
+            169,
+            864,
+            367
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "outperforms all the baselines on the UCF-101 dataset and improves the zero-shot Inception Score from 33.45 to 47.76. In Table 3, we show that our baseline model achieves a new state-of-the-art CLIP-FID score [24] of 10.21, while using ensemble models further improves both CLIP-FID and FID scores. In Figure 4, we qualitatively visualize the synthesis capability of our approach. Our model achieves high-quality zero-shot video synthesis capability with good photorealism and temporal coherency. We also provide a qualitative comparison with Make-A-Video [46] andImagen Video [14] in Figure 5. We observe that our model is able to produce videos with better details than both approaches, as shown in the animal videos. We also produce better-stylized videos thanImagen Video.",
+        "bbox": [
+            496,
+            383,
+            893,
+            595
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Small-scale unconditional video synthesis We report IS and FVD scores on UCF-101 dataset in Table 4 and compare our model with multiple unconditional video generation baselines. Note that using class labels as conditional information could lead to sizeable improvement in IS and FVD scores [9], which we do not consider as the comparison. Our method attains state-of-the-art unconditional video generation quality. Compared with previous diffusion-based unconditional generation model [17], our model is  $\\sim 10\\times$  smaller and has  $\\sim 14\\times$  less training time (75 GPU-days vs. 925 GPU-days).",
+        "bbox": [
+            496,
+            628,
+            893,
+            796
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            500,
+            813,
+            653,
+            830
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We quantitatively compare several training strategies for video diffusion models. Then, we perform ablation on the correlation ratio in the Equations 3 and 4, a key hyperparameter in our approach.",
+        "bbox": [
+            496,
+            839,
+            893,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "22936",
+        "bbox": [
+            478,
+            924,
+            519,
+            936
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f6a0d29dd22ed069c738d632fc980df02d7577656b6e6bd19cd3796a56121db7.jpg",
+        "image_caption": [
+            "Make-A-Video [46]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            104,
+            251,
+            237
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5df5207698abb05a05e378f80ce72886b6456c216ebda1f9d3b371c0e5609342.jpg",
+        "image_caption": [
+            "PYoCo"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            251,
+            104,
+            421,
+            237
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/83d81abbfb238a9855453e259f275ebda3921ee15a564c7b3fb9f7db156e2345.jpg",
+        "image_caption": [
+            "Imagen Video [14]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            426,
+            104,
+            709,
+            237
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/9ff683812bea89fc21a52045226a7b707e14228bab302290c9f1d4b390656de3.jpg",
+        "image_caption": [
+            "PYoCo"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            709,
+            104,
+            880,
+            237
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/948ed773c9da2f0cccb1b0ded2fb2dbac6a3addc8835810020572224731d24f0.jpg",
+        "image_caption": [
+            "A confused grizzly bear in calculus class."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            257,
+            250,
+            388
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/250a4b80c2e0ad72e8ae4221f419e71914963a8b29b642054accdce2d7f86fd6.jpg",
+        "image_caption": [
+            "Sailboat sailing on a sunny day in a mountainlake."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            250,
+            257,
+            421,
+            388
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/2856d764a26c0dceff776c469c325eecabc2d37157d43553670cc7bc4a9a8161.jpg",
+        "image_caption": [
+            "A sheep to the right of a wineglass.",
+            "A cat eating food out of a bowl, in style of Van Gogh."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            426,
+            257,
+            709,
+            388
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/05aba182271478008d147821594306e7c1e14e854c9c1fa92d32d525b402c841.jpg",
+        "image_caption": [
+            "Figure 5: Qualitative comparison with baseline approaches. The two panels on the left show the comparison of our approach with Make-A-Video [46], while those on the right show the comparison withImagen Video [14]. PYoCo achieves better photorealism compared to the two approaches."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            709,
+            257,
+            880,
+            388
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/b7e8a40bf3b433a6bb279667cea21e9ce3bc29e8fdbe343d7f1c35b135b3875f.jpg",
+        "table_caption": [
+            "Table 5: Quantitative results of different training strategies on UCF-101 dataset."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>IS(↑)</td><td>FVD (↓)</td><td>FID (↓)</td></tr><tr><td>Image Diffusion (ID)</td><td>-</td><td>-</td><td>30.05</td></tr><tr><td>Training from scratch</td><td>28.25</td><td>903.37</td><td>124.75</td></tr><tr><td>Finetuning from ID</td><td>41.25</td><td>566.67</td><td>56.43</td></tr><tr><td>+ Mixed Noise</td><td>52.71</td><td>337.40</td><td>31.57</td></tr><tr><td>+ Progressive Noise</td><td>53.52</td><td>339.67</td><td>31.88</td></tr></table>",
+        "bbox": [
+            86,
+            526,
+            457,
+            633
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Training strategies We compare training from scratch, a simple finetuning baseline, finetuning with mixed noising, and progressive noising using IS, FVD, and averaged frame FID metrics on the UCF-101 dataset in Table 5. We first find that finetuning from an image diffusion model is much more effective than training from scratch. For finetuning from the image model, the correlated noise model produces better video generation quality than the independent noise model. In addition, we notice that the correlated noise better preserves the image quality learned by the pretrained image model and produces a lower frame FID. This is particularly desired in large-scale text-to-video training to fulfill the goal of inheriting the knowledge from the image model missing in the video datasets. Specifically, most videos contain realistic scenes captured by cameras and have infrequent media types like paintings, illustrations, sketches, etc. Moreover, the",
+        "bbox": [
+            75,
+            659,
+            472,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "video data is much smaller in volume, and the scenes are less diverse than image datasets. As shown in Figure 4, our model can preserve properties learned from image datasets that are not presented in our video dataset, such as the artistic styles, and generate faithful motion on them.",
+        "bbox": [
+            496,
+            493,
+            892,
+            569
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Correlation ratio The hyperparameter  $\\alpha$  in the Equations 3 and 4 controls the correlation between the noise of different frames. A larger  $\\alpha$  injects more correlation into the noise. The correlation disappears when  $\\alpha \\rightarrow 0$ , and the mixed and progressive noise models reproduce the vanilla noise model. To find optimal  $\\alpha$ , we train our UCF-small model (69M parameters) using  $\\alpha \\in \\{0, 0.1, 0.2, 0.5, 1, 1, 2, 5, 10, \\infty\\}$  and report FVD in Figure 7. For each  $\\alpha$  value, we repeat the experiment 3 times and report the mean. Note that  $\\alpha = 0$  indicates finetuning with the independent frame noise, and  $\\alpha = \\infty$  indicates using identical noise maps for all the frames, which produces frozen videos during the inference time. Finetuning an image diffusion model almost consistently outperforms the training-from-scratch baseline with different  $\\alpha$ s. Using  $\\alpha = 1$  for mixed noisng and  $\\alpha = 2$  for progressive noisng produces similar best results. We also show qualitative results for models trained with  $\\alpha = 0, 1, 10$  in Figure 6. When  $\\alpha$  is too small, we notice a degradation in visual quality in the generated video frames and a reduced video diversity. For example, we notice many repeated samples and black borders in almost every video generated with",
+        "bbox": [
+            496,
+            583,
+            893,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "22937",
+        "bbox": [
+            478,
+            924,
+            517,
+            936
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/84043d9b5f0034b76c6835e2d4e087432971d2d22535d277eb4e1c20e2e5608a.jpg",
+        "image_caption": [
+            "$\\alpha = 0$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            88,
+            344,
+            292
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/999c099386c0fd284da43b5fb62b6ec5de5670e01424725f8869ea74c5d45ab6.jpg",
+        "image_caption": [
+            "$\\alpha = 1$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            344,
+            88,
+            609,
+            292
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/26e5b0e3adcaeb96a05393aada9def68b3f50bb8173fd43ac1a8b5ee7ec07265.jpg",
+        "image_caption": [
+            "$\\alpha = 10$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            609,
+            89,
+            874,
+            292
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/2e6035bc99550b824a6ea57ef45e3d24c183214050feafa286f4413540cd2415.jpg",
+        "image_caption": [
+            "Figure 6: Visual ablation on  $\\alpha$ . Small  $\\alpha = 0$  reduces video quality and diversity and large  $\\alpha = 10$  yields motion artifacts."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            358,
+            465,
+            506
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "$\\alpha = 0$ . On the other hand, when  $\\alpha$  is too large, the model has difficulty generating proper motions.",
+        "bbox": [
+            75,
+            623,
+            468,
+            655
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Model size We pick the best  $\\alpha$  for the mixed and progressive noise models and compare them with the model trained from scratch on models with different numbers of parameters, 69M, 112M, and 253M. Figure 8 shows that our mixed and progressive models outperform the baseline consistently by a large margin in terms of FVD. Overall, mixed and progressive noising provide similar performance. In our large large-scale experiments, we choose progressive noising with  $\\alpha = 2$  due to its autoregressive nature.",
+        "bbox": [
+            75,
+            676,
+            470,
+            813
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            76,
+            829,
+            194,
+            844
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We proposed a new efficient way of training text-to-video generation models. By observing that the noise maps generating the frames of a video are clustered together, we study",
+        "bbox": [
+            75,
+            854,
+            470,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/0230cc56489705efdce7ccceeafb558d19fccf3188e8d60146c865c67f51f050.jpg",
+        "image_caption": [
+            "Figure 7: Quantitative ablation on hyperparameter  $\\alpha$ . Finetuning with temporally correlated prior improves over training from scratch. Using a too-large or too-small  $\\alpha$  leads to inferior results.  $\\alpha = 1$ ,  $\\alpha = 2$  each works the best for mixed and progressive noising, respectively.",
+            "Figure 8: Ablation on model size. Larger models consistently improve the performance of both finetuning and training from scratch. Finetuning from image model consistently outperforms training from scratch."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            358,
+            887,
+            506
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "mixed and progressive noise priors well-suited for sequential video frame generation. We apply our progressive noise prior to finetuning a state-of-the-art diffusion-based text-to-image model to achieve a state-of-the-art large-scale text-to-video model. The high quality of the generated videos and the state-of-the-art Inception and FID scores demonstrate the strength of our approach.",
+        "bbox": [
+            496,
+            611,
+            890,
+            717
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgment. We would like to thank Amanda Moran, John Dickinson, Sivakumar Arayandi Thottakara, David Page, Ranjitha Prasanna, Venkata Karri, and others in NGC and PBSS team for the computing infrastructure support. We also give thanks to Qinsheng Zhang, Zekun Hao, Tsung-Yi Lin, Ajay Jain, and Chen-Hsuan Lin for useful discussions and feedback. Thanks also go to Yin Xi and Thomas Hayes for clarifying the evaluation protocol of Make-A-Video and to Ming Ding for the details of CogVideo. We thank the ICCV PCs, ACs, and reviewers for their service and valuable feedback. This work is partly supported by NSF grants No. IIS-1910132 and IIS-2213335.",
+        "bbox": [
+            496,
+            719,
+            893,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "22938",
+        "bbox": [
+            478,
+            924,
+            517,
+            936
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            174,
+            104
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 3, 7",
+            "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. eDiff-I: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1, 3, 5, 7",
+            "[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, pages 707-723. Springer, 2022. 3",
+            "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 3, 7",
+            "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. CVPR, 2023. 1, 3",
+            "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 6",
+            "[7] Duygu Ceylan, Chun-Hao Huang, and Niloy J. Mitra. Pix2video: Video editing using image diffusion. arXiv:2303.12688, 2023. 3",
+            "[8] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 3",
+            "[9] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. arXiv preprint arXiv:2204.03638, 2022. 3, 7",
+            "[10] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. arXiv preprint arXiv:2304.06720, 2023. 3",
+            "[11] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, pages 10696-10706, 2022. 3",
+            "[12] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 3",
+            "[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3",
+            "[14] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3, 4, 5, 7, 8",
+            "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 3, 4"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[16] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 23:47-1, 2022. 3",
+            "[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 3, 4, 5, 6, 7",
+            "[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3, 4, 7",
+            "[19] Tobias Hübpe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. arXiv preprint arXiv:2206.07696, 2022. 3",
+            "[20] Emmanuel Kahembwe and Subramanian Ramamoorthy. Lower dimensional kernels for video discriminators. Neural Networks, 2020. 7",
+            "[21] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 3, 4",
+            "[22] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 3",
+            "[23] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. arXiv preprint arXiv:2212.04488, 2022. 3",
+            "[24] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fréchet inception distance. In ICLR, 2023. 7",
+            "[25] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Ccvs: Context-aware controllable video synthesis. NeurIPS, 2021. 3, 7",
+            "[26] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 4",
+            "[27] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 4",
+            "[28] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 3",
+            "[29] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Ying Shan, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186, 2023. 3",
+            "[30] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 1, 3",
+            "[31] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen."
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "22939",
+        "bbox": [
+            478,
+            924,
+            519,
+            936
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 3",
+            "[32] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3",
+            "[33] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 3",
+            "[34] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. arXiv preprint arXiv:2302.03027, 2023. 3",
+            "[35] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 3",
+            "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 5",
+            "[37] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 5",
+            "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 3",
+            "[39] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv preprint arXiv:1412.6604, 2014. 3",
+            "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3",
+            "[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3",
+            "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2, 3, 5",
+            "[43] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In ICCV, 2017. 3, 7",
+            "[44] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan. IJCV, 2020. 6"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            900
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[45] Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Mostganv: Video generation with temporal motion styles. In CVPR, 2023. 3",
+            "[46] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 3, 4, 5, 6, 7, 8",
+            "[47] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. arXiv preprint arXiv:2112.14683, 2021. 3, 6, 7",
+            "[48] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 4",
+            "[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2",
+            "[50] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5",
+            "[51] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In ICML, 2015. 3",
+            "[52] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. In ICLR, 2021. 3, 6, 7",
+            "[53] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 6",
+            "[54] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, June 2018. 3",
+            "[55] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. *ICLR*, 2019. 6",
+            "[56] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 3",
+            "[57] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Masked conditional video diffusion for prediction, generation, and interpolation. arXiv preprint arXiv:2205.09853, 2022. 3",
+            "[58] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. NIPS, 2016. 3",
+            "[59] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. arXiv preprint arXiv:2305.10874, 2023. 3, 7",
+            "[60] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In ECCV, pages 720-736. Springer, 2022. 3, 4, 7"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            900
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "22940",
+        "bbox": [
+            478,
+            924,
+            519,
+            936
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[61] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022.3",
+            "[62] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 7",
+            "[63] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3, 7",
+            "[64] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481, 2022. 3",
+            "[65] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. arXiv preprint arXiv:2212.05199, 2022.3",
+            "[66] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. In ICLR, 2021. 3, 7",
+            "[67] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 1, 3",
+            "[68] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. In ICLR, 2023. 4",
+            "[69] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2, 3, 7"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            558
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "22941",
+        "bbox": [
+            478,
+            924,
+            517,
+            936
+        ],
+        "page_idx": 11
+    }
+]

2023/Preserve Your Own Correlation_ A Noise Prior for Video Diffusion Models/12c6fb66-a455-4547-8b5a-4a378b5ee51f_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:174988f7e8f3a9f4fa76bbc83a87d352f180f80bbccd2dbb9eec4f72a8508dd0
+size 9514040

2023/Preserve Your Own Correlation_ A Noise Prior for Video Diffusion Models/full.md

@@ -0,0 +1,377 @@
+# Preserve Your Own Correlation: A Noise Prior for Video Diffusion Models
+
+Songwei Ge*
+
+University of Maryland
+
+Seungjun Nah  
+NVIDIA
+
+Guilin Liu NVIDIA
+
+Tyler Poon
+
+University of Chicago
+
+Andrew Tao NVIDIA
+
+Bryan Catanzaro
+
+NVIDIA
+
+David Jacobs
+
+University of Maryland
+
+Jia-Bin Huang
+
+University of Maryland
+
+Ming-Yu Liu
+
+NVIDIA
+
+Yogesh Balaji
+
+NVIDIA
+
+A very happy fuzzy panda dressed as a chef eating pizza in the New York street food truck.
+
+The supernova explosion of a white dwarf in the universe, photo realistic.
+
+A high-quality 3D render of hyperrealist, super strong, multicolor stripped, and fluffy bear with wings, highly detailed.
+
+Figure 1: Given a text description, our approach can faithfully generate videos that are consistent with the input text while being photorealistic and temporally consistent. Best viewed with Acrobat Reader. Click the images to play the video clips.
+
+# Abstract
+
+Despite tremendous progress in generating high-quality images using diffusion models, synthesizing a sequence of animated frames that are both photorealistic and temporally coherent is still in its infancy. While off-the-shelf billion-scale datasets for image generation are available, collecting similar video data of the same scale is still challenging. Also, training a video diffusion model is computationally much more expensive than its image counterpart. In this work, we explore finetuning a pretrained image diffusion model with video data as a practical solution for the video synthesis task. We find that naively extending the image noise prior to video noise prior in video diffusion leads to sub-optimal performance. Our carefully designed video noise prior leads to substantially better performance. Extensive experimental validation shows that our
+
+model, Preserve Your Own COrelation (PYoCo), attains SOTA zero-shot text-to-video results on the UCF-101 and MSR-VTT benchmarks. It also achieves SOTA video generation quality on the small-scale UCF-101 benchmark with a  $10 \times$  smaller model using significantly less computation than the prior art. The project page is available at https://research.nvidia.com/labs/dir/pyoco/.
+
+# 1. Introduction
+
+Large-scale diffusion-based text-to-image models [38, 42, 2] have demonstrated impressive capabilities in turning complex text descriptions into photorealistic images. They can generate images with novel concepts unseen during training. Sophisticated image editing and processing tasks can easily be accomplished through guidance control and embedding techniques. Due to the immense success in several applications [30, 67, 5], these models are established as pow-
+
+![](images/3ea06ec868739f34b3665eb70207e13f0537015064d7770050b4c180314a5265.jpg)  
+(a)
+
+![](images/e6d7c286fd69acb9c2a133e2940d19045711e89bb8826830f27c7352a116b27c.jpg)  
+(b)  
+Figure 2: Visualizing the noise map correlations. (a) visualizes the t-SNE plot of the noise maps corresponding to input frames randomly sampled from videos. These noise maps are obtained by running a diffusion ODE [49, 48] on the input frames using a trained text-to-image model, but in the opposite direction of image synthesis ( $\sigma : 0 \to \sigma_{\max}$ ). The green dots in the background denote the reference noise maps sampled from an i.i.d. Gaussian distribution. The red dots and yellow dots are noise maps corresponding to input frames coming from different videos. We found they are spread out and share no correlation. On the other hand, the noise maps corresponding to the frames coming from the same video (shown in blue dots) are clustered together. (b) Using an i.i.d. noise model (orange dots) for finetuning text-to-image models for video synthesis is not ideal since temporal correlations between frames are not modeled. To remedy this, we propose a progressive noise model in which the correlation between different noise maps is injected along the temporal axis. Our progressive noise model (blue dots) aptly models the correlations present in the video noise maps.
+
+erful image synthesis tools for content generation. As image synthesis is largely democratized with the success of these text-to-image models, it is natural to ask whether we can repeat the same success in video synthesis with large-scale diffusion-based text-to-video models.
+
+Multiple attempts have been made to build large-scale video diffusion models. Ho et al. [17] proposed a UNet-based architecture for the video synthesis task that is trained using joint image-video denoising losses.Imagen video [14] extends the cascaded text-to-image generation architecture ofImagen [42] for video generation. In both works, the authors directly train a video generation model from scratch. While these approaches achieve great success and produce high-quality videos, they are inherently expensive to train, requiring hundreds of high-end GPUs or TPU's and several weeks of training. After all, video generators not only need to learn to form individual images but should also learn to synthesize coherent temporal dynamics, which makes the video generation task much more challenging. While the formation of individual frames is a shared component in an image and video synthesis, these works disregard the existence of powerful pretrained text-to-image diffusion models and train their video generators from scratch.
+
+We explore a different avenue for building large-scale text-to-video diffusion models by starting with a pretrained text-to-image diffusion model. Our motivation is that most of the components learned for the image synthesis task can
+
+effectively be reused for video generation, leading to knowledge transfer and efficient training. A similar idea is adopted by several recent works [46, 69, 4]. Without exception, when finetuning, they naively extend the image diffusion noise prior (i.i.d. noise) used in the text-to-image model to a video diffusion noise prior by adding an extra dimension to the 2D noise map. We argue that this approach is not ideal as it does not utilize the natural correlations in videos that are already learned by the image models. This is illustrated in Figure 2, where we visualize the t-SNE plot of noise maps corresponding to different input frames as obtained from a pretrained text-to-image diffusion model. The noise maps corresponding to different frames coming from the same video (blue dots in Figure 2a are clustered together, exhibiting a high degree of correlation. The use of i.i.d. noise prior does not model this correlation, which would impede the finetuning process. Our careful analysis of the video diffusion noise prior leads us to a noise prior that is better tailored for finetuning an image synthesis model to the video generation task. As illustrated in Figure 2b, our proposed noise prior (shown in blue dots) aptly captures the correlations in noise maps corresponding to video frames.
+
+We then proceed to build a large-scale diffusion-based text-to-video model. We leverage several design choices from the prior works, including the use of temporal attention [17], joint image-video finetuning [17], a cascaded generation architecture [14], and an ensemble of expert denois-
+
+ers [2]. Together with these techniques and the proposed video noise prior, our model establishes a new state-of-the-art for video generation outperforming competing methods on several benchmark datasets. Figure 1 shows our model can achieve high-quality zero-shot video synthesis capability with SOTA photorealism and temporal consistency.
+
+In short, our work makes the following key contributions.
+
+1. We propose a video diffusion noise tailored for finetuning text-to-image diffusion models for text-to-video.  
+2. We conduct extensive experimental validation and verify the effectiveness of the proposed noise prior.  
+3. We build a large-scale text-to-video diffusion model by finetuning a pretrained eDiff-I model with our noise prior and achieve state-of-the-art results on several benchmarks.
+
+# 2. Related Work
+
+Diffusion-based text-to-image models: Diffusion models have significantly advanced the progress of text-based photorealistic, compositional image generation [38, 42]. Given the nature of the iterative denoising process that requires massive numbers of score function evaluations, earlier diffusion models focused on generating low-resolution images, e.g.,  $64 \times 64$  [15, 48]. To generate high-resolution images, two common approaches have been used. The first approach applies cascaded super-resolution models in the RGB space [32, 16, 42, 38], while the second approach leverages a decoder to exploit latent space [40, 11]. Based on these models, advanced image and video editing have been achieved through finetuning the model [41, 67, 5, 23, 61, 29] or controlling the inference process [30, 13, 34, 10, 35, 7, 31, 3]. Here, we study the problem of using large-scale diffusion models for text-to-video generation.
+
+Video generation models: Generating realistic and novel videos have long been an attractive and essential research direction [58, 39, 65]. Previously studies have resorted to different types of generative models such as GANs [58, 43, 54, 52, 45], Autoregressive models [51, 63, 25, 9, 18], and implicit neural representations [47, 66]. Recently, driven by the tremendous success of applying the diffusion model to image synthesis, multiple works have proposed to explore diffusion models for conditional and unconditional video synthesis [57, 12, 69, 61, 4, 22, 19, 57, 64, 33, 28, 1, 59]. For example, Singer et al. extend the unCLIP framework [38] to text-to-video generation, which allows training without video captions [46]. Ho et al. [17] extend the Imagen framework [42] by repeatedly up-scaling low-resolution small-fps videos in both spatial and temporal directions with multiple models [14]. Our work also falls into this line of work which uses a diffusion model. We focus on augmenting an image diffusion model for video and study the design choice of the diffusion noise priors for such an image-to-video finetuning task.
+
+Leverage knowledge from images for text-to-video generation: Like text-to-image models, text-to-video models require massive amounts of data to learn caption-relatedness, frame photorealism, and temporal dynamics. But in contrast to the abundant image data resource, video data are more limited in style, volume, and quality. To resolve such scarcity issue of text-video data, previous works have resorted to different strategies to leverage knowledge from image data for text-to-video generation, including joint training on the text-image data from scratch [17, 14, 56, 60], first training a text-to-image model and then finetuning partially [18, 4, 61, 29] or entirely [46, 8] on the video dataset, and using CLIP image features as the conditional information [46, 69]. In this paper, we propose a new video diffusion noise prior that is tailored for finetuning a pretrained diffusion-based image generation model for the video generation task. We reuse several design choices in the prior work by finetuning jointly on text-image and text-video datasets. As a result, we can build a text-to-video generation system that achieves state-of-the-art zero-shot performances.
+
+# 3. Preliminaries
+
+Diffusion models generate data by iteratively denoising samples drawn from a noise distribution. In the case of text-to-video models, text embeddings obtained from a pre-trained text encoder are used as additional inputs in the denoising process. Formally, let  $D(\mathbf{x}, \mathbf{e}, \sigma)$  denote a denoising network that operates on the noisy input video  $\mathbf{x} \in \mathbb{R}^{b \times n_s \times 3 \times h \times w}$  where  $\mathbf{e}$  is the text embedding, and  $\sigma$  is the noise level. Here  $n_s$  is the sequence length of the input video,  $b$  is the batch size, and  $h \times w$  is the spatial resolution. The model  $D$  is trained to denoise the input  $\mathbf{x}$ .
+
+Training We follow the EDM formulation of Karras et al. [21] to optimize the denoiser  $D$  using the following objective
+
+$$
+\begin{array}{l} \mathbb {E} _ {p _ {\text {d a t a}} (\mathbf {x} _ {\text {c l e a n}}, \mathbf {e}), p (\epsilon), p (\sigma)} \left[ \lambda (\sigma) \| D \left(\mathbf {x} _ {\text {n o i s e}}; \mathbf {e}, \sigma\right) - \mathbf {x} _ {\text {c l e a n}} \| _ {2} ^ {2} \right] \tag {1} \\ w h e r e \quad \mathbf {x} _ {\text {n o i s e}} = \mathbf {x} _ {\text {c l e a n}} + \sigma \epsilon \\ \end{array}
+$$
+
+Here,  $\mathbf{x}_{\mathrm{noise}}$  is the noisy sample obtained by corrupting the clean video  $\mathbf{x}$  with noise  $\sigma \epsilon$ , where  $p(\epsilon) = \mathcal{N}(\mathbf{0},\mathbf{I})$  and  $\sigma$  is a scalar for the noise level drawn from  $p(\sigma)$ . The loss weight,  $\lambda (\sigma)$ , is a function of  $\sigma$  given by  $\lambda (\sigma) = (\sigma^2 +\sigma_{\mathrm{data}}^2) / (\sigma \cdot \sigma_{\mathrm{data}})^2$ . Eq. (1) is a simple denoising objective in which the denoiser  $D$  is trained to estimate the clean video  $\mathbf{x}_{\mathrm{clean}}$  from the noisy input  $\mathbf{x}_{\mathrm{noise}}$ . Following EDM, we use a log-normal distribution for  $\sigma$  i.e.,  $\ln (p(\sigma)) = \mathcal{N}(P_{\mathrm{mean}},P_{\mathrm{std}}^2)$  with  $P_{\mathrm{mean}} = -1.2$  and  $P_{\mathrm{std}} = 1.2$ .
+
+To train the denoising model, EDM uses preconditioning terms in its objective function to properly scale the inputs and output of the denoiser model  $D$ . More specifically, the
+
+denoising model  $D$  is written as
+
+$$
+D (\mathbf {x}; \mathbf {e}, \sigma) := \left(\frac {\sigma_ {\mathrm {d a t a}}}{\sigma^ {*}}\right) ^ {2} \mathbf {x} + \frac {\sigma \cdot \sigma_ {\mathrm {d a t a}}}{\sigma^ {*}} F _ {\theta} \left(\frac {\mathbf {x}}{\sigma^ {*}}; \mathbf {e}, \frac {\ln (\sigma)}{4}\right)
+$$
+
+Here,  $F_{\theta}$  is a neural network with parameters  $\theta$  and  $\sigma^{*} = \sqrt{\sigma^{2} + \sigma_{\mathrm{data}}^{2}}$ . We use  $\sigma_{\mathrm{data}} = 0.5$ .
+
+Sampling Once the denoising model is trained, sampling can be performed by solving the following ODE [21]
+
+$$
+\frac {d \mathbf {x}}{d \sigma} = - \sigma \nabla_ {\mathbf {x}} \log p (\mathbf {x} | \mathbf {e}, \sigma) = \frac {\mathbf {x} - D (\mathbf {x} ; \mathbf {e} , \sigma)}{\sigma} \tag {2}
+$$
+
+for  $\sigma$  flowing backwards from  $\sigma = \sigma_{\mathrm{max}}$  to  $\sigma = 0$ . The initial value for  $\mathbf{x}$  is obtained by sampling from the prior distribution  $\mathbf{x} \sim \mathcal{N}(\mathbf{0}, \sigma_{\mathrm{max}}^2 \mathbf{I})$ . Over the recent years, several samplers have been proposed for sampling from the trained diffusion models [68, 48, 26, 27, 15]. In this paper, we use DEIS [68] and its stochastic variant [21] for synthesizing samples from our model.
+
+# 4. Method
+
+Training text-to-video models is much more challenging than training text-to-image diffusion models due to practical difficulties in collecting billion-scale video datasets and securing enough computational resources. Additionally, generating videos is much more challenging since individual frames need to be both photorealistic and temporally coherent. Prior works leverage large-scale image datasets to mitigate these difficulties by either joint training on the image datasets [60, 17, 14] or finetuning a text-to-image model on the video datasets [18, 46]. Here, we are interested in finetuning text-to-image diffusion models jointly on image and video datasets. We postulate that naively extending the image noise prior to video diffusion is not ideal. We carefully explore the design space of noise priors and propose one that is well suited for our video finetuning task, which leads to significant performance gains.
+
+Correlated noise model An image diffusion model is trained to denoise independent noise from a perturbed image. The noise vector  $\epsilon$  in the denoising objective (1) is sampled from an i.i.d. Gaussian distribution  $\epsilon \sim \mathcal{N}(0, \mathbf{I})$ . However, after training the image diffusion model and applying it to reverse real frames from a video into the noise space in a per-frame manner, we find that the noise maps corresponding to different frames are highly correlated. This is illustrated in Figure 2, where the t-SNE plot of noise maps corresponding to different video frames are plotted. When the input frames come from the same video (shown in blue dots in Figure 2a, noise maps are clustered. The use of i.i.d. sampling (shown in orange dots in Figure 2b does not capture these correlations. This is also depicted quantitatively
+
+Table 1: Cosine similarity of the reversed noise. The noise maps corresponding to the frames sampled from the same videos have a higher similarity than those sampled from different videos.
+
+<table><tr><td></td><td>Cosine Similarity</td></tr><tr><td>(a) Same video noise</td><td>0.206±0.156</td></tr><tr><td>(b) Different video noise</td><td>0.001±0.009</td></tr></table>
+
+in Table 1 where we compute the average pairwise cosine similarity between noise corresponding to (a) same video and (b) different video. (a) is much higher than (b). As a result, the video diffusion model trained with i.i.d. noise is coerced to forget such correlation among the noise between different frames, making it difficult to preserve knowledge from the image diffusion model. Motivated by this observation, we propose to modify the noise process to preserve the correlation between different frames. To this end, we investigate two noisng strategies - mixed and progressive noising.
+
+Mixed noise model: Let  $\epsilon^1, \epsilon^2, \ldots, \epsilon^{n_s}$  denote the noise corresponding to individual video frames i.e.,  $\epsilon^i$  corresponds to the  $i^{th}$  element of the noise tensor  $\epsilon$ . In the mixed noise model, we generate two noise vectors  $\epsilon_{\mathrm{shared}}$  and  $\epsilon_{\mathrm{ind}}$ .  $\epsilon_{\mathrm{shared}}$  is a common noise vector shared among all video frames, while  $\epsilon_{\mathrm{ind}}$  is the individual noise per frame. The linear combination of both these vectors is used as the final noise.
+
+$$
+\begin{array}{l} \epsilon_ {\text {s h a r e d}} \sim \mathcal {N} \left(\mathbf {0}, \frac {\alpha^ {2}}{1 + \alpha^ {2}} \mathbf {I}\right), \epsilon_ {\text {i n d}} ^ {i} \sim \mathcal {N} \left(\mathbf {0}, \frac {1}{1 + \alpha^ {2}} \mathbf {I}\right) \tag {3} \\ \epsilon^ {i} = \epsilon_ {\mathrm {s h a r e d}} + \epsilon_ {\mathrm {i n d}} ^ {i} \\ \end{array}
+$$
+
+Progressive noise model: In the progressive noise model, the noise for each frame is generated in an autoregressive fashion in which the noise at frame  $i$  is generated by perturbing the noise at frame  $i - 1$ . Let  $\epsilon_{\mathrm{ind}}^i$  denote the independent noise generated for frame  $i$ . Then, progressive noising can be formulated as
+
+$$
+\epsilon^ {0} \sim \mathcal {N} (\mathbf {0}, \mathbf {I}) \quad \epsilon_ {\mathrm {i n d}} ^ {i} \sim \mathcal {N} (\mathbf {0}, \frac {1}{1 + \alpha^ {2}} \mathbf {I}) \tag {4}
+$$
+
+$$
+\epsilon^ {i} = \frac {\alpha}{\sqrt {1 + \alpha^ {2}}} \epsilon^ {i - 1} + \epsilon_ {\mathrm {i n d}} ^ {i}
+$$
+
+In both these models,  $\alpha$  controls how much noise is shared among different video frames. The higher the value of  $\alpha$ , the more correlation exists among the noise maps corresponding to different frames. As  $\alpha \rightarrow \infty$ , all frames would have the same noise which results in generating a frozen video. On the other hand,  $\alpha = 0$  corresponds to i.i.d. noise.
+
+As shown in Figure 2b, the use of progressive noise sampling (blue dots) better models the correlations between different noise maps by obtaining similar clustering patterns to the noise maps of real video frames embedded by a pretrained text-to-image model in Figure 2a (blue dots).
+
+![](images/8d6b3d4c8df87f34a0ce9703c0a7f610b6687418e672edb194d3268825477a99.jpg)  
+Figure 3: Model architecture. Our pipeline consists of a cascade of four networks — a base model and three upsampling models. All four models take inputs as the text embeddings obtained from the T5 encoder and the CLIP text encoder. The base model produces 16 video frames of spatial resolution  $64 \times 64$  with a frameskip of 5. The first upsampling model performs a temporal interpolation, resulting in videos of size  $76 \times 64 \times 64$  while the subsequent two super-resolution models perform spatial super-resolution to produce videos of sizes  $76 \times 256 \times 256$  and  $76 \times 1024 \times 1024$ .
+
+Model architecture As visualized in Figure 3, our model consists of a cascade of four networks — a base network and three upsampling stacks. The base network generates an output video of dimension  $16 \times 64 \times 64$  with a frameskip of 5. It generates the frames  $\{1,6,11,\ldots,76\}$ . The first upsampling network performs a temporal interpolation to produce a video of size  $76 \times 64 \times 64$ . The second and the third super-resolution network performs spatial upsampling to produce the outputs of sizes  $76 \times 256 \times 256$  and  $76 \times 1024 \times 1024$ . We utilize eDiff-I [2], a state-of-the-art text-to-image diffusion model, to initialize our base and spatial super-resolution models. Similar to prior works [17, 46], we adapt the image-based U-Net model for the video synthesis task by making the following changes: (1) Transforming 2D convolutions to 3D by adding a dimension 1 to temporal axis and (2) Adding temporal attention layers. Please refer to the supplementary material for more details.
+
+Similar to Ho et al. [17], we jointly finetune the model on video and image datasets by concatenating videos and images in the temporal axis and applying our temporal modules only on the video part. Similarly to eDiff-I, our model uses both T5 text embeddings [37] and CLIP text embeddings [36]. We drop each of the embeddings independently at random during training, as in eDiff-I.
+
+# 5. Experiments
+
+In this section, we evaluate our proposed strategy of training diffusion models for video synthesis on two sets of experiments. We first comprehensively analyze our proposed noise model on the small-scale UCF-101 dataset. We then scale up our experiments to the challenging large-scale text-to-video synthesis task.
+
+# 5.1. Experimental Setups
+
+We conduct ablation experiments in a small-scale unconditional video generation setting and pick the best configuration for our large-scale text-to-video generation run.
+
+Datasets We train our model on the UCF-101 dataset [50] for the small-scale experiments, where we follow the protocol defined in Ho et al. [17] to generate videos of size  $16 \times 64 \times 64$ . UCF-101 dataset contains 13,320 videos. We randomly sample frames from these videos to train our image synthesis model. For our large-scale experiments, we use a combination of public and proprietary datasets for text-to-image and text-to-video finetuning. Most of the videos are of 2K resolution with 16:9 aspect ratio. All data was filtered using a preset CLIP and aesthetic scores* to ensure high quality. Our final image dataset contains around 1.2 billion text-image pairs and 22.5 million text-video pairs.
+
+Training details In the unconditional generation experiment on the UCF-101 dataset, to do an ablation study on the model size, we design 3 models where each model has 69M, 112M, and 253M parameters, respectively. As a comparison, the baseline Video Diffusion Model (VDM) [17] contains 1.2B parameters. In the large-scale text-to-video experiment, our base and temporal interpolation models contain 1.08B parameters. Our super-resolution model adapted from the efficient U-Net [42] architecture with temporal convolution layers [14, 46] contains 313M parameters. Please refer to the supplementary material for more training details.
+
+![](images/be0d32f6c5261750250a55f4cff915a053e9259ad86489ddce09cc007e0763ed.jpg)
+
+![](images/4b5f2ccad8a2a714b4777005a783956b00b8fe07cd922ddc9476fa3b96ba75c6.jpg)  
+An epic tornado attacking above a glowing city at night, the tornado is made of smoke, highly detailed.
+
+![](images/0b71e5432321661882bb8e9d07f88930204323a4e84990b34323694870a32f4e.jpg)
+
+![](images/7497cd615041a3013314d3ae02fc31c19b21ba4245c7efeae2c13dcc7ecf9e74.jpg)  
+A golden retriever puppy holding a green sign that says "NVIDIA ROCKS". Background is a classroom.
+
+![](images/31f24d40b980ff0fc7922ccd7cfcbfe5b0d5a54f03b55e39df6481d82290aaca.jpg)  
+A cute corgi wearing a red robe holding a sign that says "Merry Christmas". There is a Christmas tree in the background.
+
+![](images/181156bf69a3ac777b674dfabb451cac0bf126ed68722d49f6b63b74c54d733f.jpg)
+
+![](images/8f9a5a97c5cd51d09128de896bf9ded108cfd243c94a7f05c4a9a840b15e6ebf.jpg)  
+Small boat sailing in the ocean, giant Cthulhu monster coming out a dense mist in the background, giant waves attacking.
+
+![](images/a8e97061fcc21c77cd0c887851bf593e0b2b7dfd929be30c5adb3933145bc9a8.jpg)
+
+![](images/8962c3cab1893100cc256b523b53fa2604524a275958c2bb078790ce3e85c544.jpg)  
+A cute funny robot dancing, centered, award winning watercolor pen illustration.  
+A skull burning while being held up by a skeletal hand.  
+Figure 4: Sample generations. Please check our project website to view the videos.
+
+![](images/e7fd793e68cfa900bbc13500147bbcb8cac48b979f63a288ea6c2e2090c78b53.jpg)  
+A cartoon white wolf is giving puppy-dog eyes, detailed fur, very cute kid's film character.  
+A huge dinosaur skeleton walking in a golden wheat field on a bright sunny day.
+
+![](images/8f943694a9daad4968571f69c1d1f69ba8a562ee9e68557a56bed7b52081d8ff.jpg)  
+A lightning striking atop of eiffel tower, dark clouds in the sky, slow motion.  
+A cute rabbit is eating grass, wildlife photography.
+
+![](images/de933d40c5692d4eb338a50bcf8966472a661d2ba6bdafedbe195396e20a70fa.jpg)  
+An anime girl looks at the beautiful nature through the window of a moving train, well rendered.  
+Tomato sauce pouring over fries.
+
+Evaluation For the small-scale experiments on UCF-101 dataset, we follow the protocol defined in the prior approaches [52, 47, 17] and report the Inception Score (IS) [44] calculated by a trained C3D model [53] and Fréchet Video Distance (FVD) [55] by a trained I3D model [6]. For the large-scale text-to-video experiments, we perform the zero-shot evaluation of the video generation quality on the UCF-101 and MSR-VTT datasets following Make-A-Video [46]. We carefully discuss the evaluation process below.
+
+UCF-101 experiment We use IS and FVD for evaluation in our small-scale experiments. UCF-101 is a categorical video dataset designed for action recognition. When sampling from the text-to-video model, we devise a set of prompts for each class name to be used as the conditional input. This is necessary as some class names (such as jump rope) are not descriptive. We list all the prompts we use in the supplementary material. We sample 20 videos for each prompt to compute the IS metric. For FVD, we follow the
+
+Table 2: Zero-shot text to video generation on UCF-101. Our approach gives significant performance gains compared to the prior baselines both in inception score and FVD metrics.  
+
+<table><tr><td>Method</td><td>IS (↑)</td><td>FVD (↓)</td></tr><tr><td>CogVideo [18] (Chinese)</td><td>23.55</td><td>751.34</td></tr><tr><td>CogVideo [18] (English)</td><td>25.27</td><td>701.59</td></tr><tr><td>Make-A-Video [46]</td><td>33.00</td><td>367.23</td></tr><tr><td>MagicVideo [69]</td><td>-</td><td>655.00</td></tr><tr><td>Video LDM [4]</td><td>33.45</td><td>550.61</td></tr><tr><td>VideoFactory [59]</td><td>-</td><td>410.00</td></tr><tr><td>PYoCo</td><td>47.76</td><td>355.19</td></tr></table>
+
+Table 3: Text conditional zero-shot generation on MSRVTT. Our approach with the base config achieves the best results, and using an ensemble further improves the FIDs.  
+
+<table><tr><td>Method</td><td>CLIP-FID (↓)</td><td>FID (↓)</td></tr><tr><td>NUWA [60] (Chinese)</td><td>47.68</td><td>-</td></tr><tr><td>CogVideo [18] (Chinese)</td><td>24.78</td><td>-</td></tr><tr><td>CogVideo [18] (English)</td><td>23.59</td><td>-</td></tr><tr><td>Make-A-Video [46]</td><td>13.17</td><td>-</td></tr><tr><td>MagicVideo [69]</td><td>-</td><td>36.50</td></tr><tr><td>Latent-Shift [1]</td><td>15.23</td><td>-</td></tr><tr><td>PYoCo (Config-A)</td><td>10.21</td><td>25.39</td></tr><tr><td>PYoCo (Config-B)</td><td>9.95</td><td>24.28</td></tr><tr><td>PYoCo (Config-C)</td><td>9.91</td><td>24.54</td></tr><tr><td>PYoCo (Config-D)</td><td>9.73</td><td>22.14</td></tr></table>
+
+prior work [25, 52] and sample 2, 048 videos for evaluation.
+
+MSR-VTT experiment MSR-VTT [62] test set contains 2,990 videos as well as 59,794 captions. All the videos have the same resolution of  $320 \times 240$ . We generate a  $76 \times 256 \times 256$  video for each 59,794 caption and save the videos in an mp4 format with a high bit rate. To compare with Make-A-Video, we compute FID using a ViT-B/32 model [24]. We also report a more common FID metric computed by an Inception-V3 model. We also examine the idea of ensemble denoiser [2] by finetuning the level-1 experts of each model. We denote Config-A as the configuration of using only baseline models and Config-B to Config-D as incrementally changing super-resolution model, temporal interpolation model, and base model with the corresponding ensemble models.
+
+# 5.2. Main Results
+
+Large-scale text-to-video synthesis We quantitatively compare our method against Make-A-Video [46], NUWA [60], CogVideo [18], and several concurrent works [4, 69, 4, 59, 1]. Table 2 shows that our method
+
+Table 4: Unconditional UCF-101 generation results. Our approach achieves the state-of-the-art inception score and FVD, while having considerably smaller parameter count compared to other diffusion-based approaches such as VDM (1B parameters).  
+
+<table><tr><td>Method</td><td>IS (↑)</td><td>FVD (↓)</td></tr><tr><td>TGAN [43]</td><td>15.83±.18</td><td>-</td></tr><tr><td>LDVD-GAN [20]</td><td>22.91±.19</td><td>-</td></tr><tr><td>VideoGPT [63]</td><td>24.69±.30</td><td>-</td></tr><tr><td>MoCoGAN-HD [52]</td><td>32.36</td><td>838</td></tr><tr><td>DIGAN [66]</td><td>29.71±.53</td><td>655±22</td></tr><tr><td>CCVS [25]</td><td>24.47±.13</td><td>386±15</td></tr><tr><td>StyleGAN-V [47]</td><td>23.94±.73</td><td>-</td></tr><tr><td>VDM [17]</td><td>57.00±.62</td><td>-</td></tr><tr><td>TATS [9]</td><td>57.63±.73</td><td>430 ±18</td></tr><tr><td>PYoCo (112M)</td><td>57.93±.24</td><td>332 ±13</td></tr><tr><td>PYoCo (253M)</td><td>60.01±.51</td><td>310 ±13</td></tr></table>
+
+outperforms all the baselines on the UCF-101 dataset and improves the zero-shot Inception Score from 33.45 to 47.76. In Table 3, we show that our baseline model achieves a new state-of-the-art CLIP-FID score [24] of 10.21, while using ensemble models further improves both CLIP-FID and FID scores. In Figure 4, we qualitatively visualize the synthesis capability of our approach. Our model achieves high-quality zero-shot video synthesis capability with good photorealism and temporal coherency. We also provide a qualitative comparison with Make-A-Video [46] andImagen Video [14] in Figure 5. We observe that our model is able to produce videos with better details than both approaches, as shown in the animal videos. We also produce better-stylized videos thanImagen Video.
+
+Small-scale unconditional video synthesis We report IS and FVD scores on UCF-101 dataset in Table 4 and compare our model with multiple unconditional video generation baselines. Note that using class labels as conditional information could lead to sizeable improvement in IS and FVD scores [9], which we do not consider as the comparison. Our method attains state-of-the-art unconditional video generation quality. Compared with previous diffusion-based unconditional generation model [17], our model is  $\sim 10\times$  smaller and has  $\sim 14\times$  less training time (75 GPU-days vs. 925 GPU-days).
+
+# 5.3. Ablation Study
+
+We quantitatively compare several training strategies for video diffusion models. Then, we perform ablation on the correlation ratio in the Equations 3 and 4, a key hyperparameter in our approach.
+
+![](images/f6a0d29dd22ed069c738d632fc980df02d7577656b6e6bd19cd3796a56121db7.jpg)  
+Make-A-Video [46]
+
+![](images/5df5207698abb05a05e378f80ce72886b6456c216ebda1f9d3b371c0e5609342.jpg)  
+PYoCo
+
+![](images/83d81abbfb238a9855453e259f275ebda3921ee15a564c7b3fb9f7db156e2345.jpg)  
+Imagen Video [14]
+
+![](images/9ff683812bea89fc21a52045226a7b707e14228bab302290c9f1d4b390656de3.jpg)  
+PYoCo
+
+![](images/948ed773c9da2f0cccb1b0ded2fb2dbac6a3addc8835810020572224731d24f0.jpg)  
+A confused grizzly bear in calculus class.
+
+![](images/250a4b80c2e0ad72e8ae4221f419e71914963a8b29b642054accdce2d7f86fd6.jpg)  
+Sailboat sailing on a sunny day in a mountainlake.
+
+![](images/2856d764a26c0dceff776c469c325eecabc2d37157d43553670cc7bc4a9a8161.jpg)  
+A sheep to the right of a wineglass.  
+A cat eating food out of a bowl, in style of Van Gogh.
+
+![](images/05aba182271478008d147821594306e7c1e14e854c9c1fa92d32d525b402c841.jpg)  
+Figure 5: Qualitative comparison with baseline approaches. The two panels on the left show the comparison of our approach with Make-A-Video [46], while those on the right show the comparison withImagen Video [14]. PYoCo achieves better photorealism compared to the two approaches.
+
+Table 5: Quantitative results of different training strategies on UCF-101 dataset.  
+
+<table><tr><td></td><td>IS(↑)</td><td>FVD (↓)</td><td>FID (↓)</td></tr><tr><td>Image Diffusion (ID)</td><td>-</td><td>-</td><td>30.05</td></tr><tr><td>Training from scratch</td><td>28.25</td><td>903.37</td><td>124.75</td></tr><tr><td>Finetuning from ID</td><td>41.25</td><td>566.67</td><td>56.43</td></tr><tr><td>+ Mixed Noise</td><td>52.71</td><td>337.40</td><td>31.57</td></tr><tr><td>+ Progressive Noise</td><td>53.52</td><td>339.67</td><td>31.88</td></tr></table>
+
+Training strategies We compare training from scratch, a simple finetuning baseline, finetuning with mixed noising, and progressive noising using IS, FVD, and averaged frame FID metrics on the UCF-101 dataset in Table 5. We first find that finetuning from an image diffusion model is much more effective than training from scratch. For finetuning from the image model, the correlated noise model produces better video generation quality than the independent noise model. In addition, we notice that the correlated noise better preserves the image quality learned by the pretrained image model and produces a lower frame FID. This is particularly desired in large-scale text-to-video training to fulfill the goal of inheriting the knowledge from the image model missing in the video datasets. Specifically, most videos contain realistic scenes captured by cameras and have infrequent media types like paintings, illustrations, sketches, etc. Moreover, the
+
+video data is much smaller in volume, and the scenes are less diverse than image datasets. As shown in Figure 4, our model can preserve properties learned from image datasets that are not presented in our video dataset, such as the artistic styles, and generate faithful motion on them.
+
+Correlation ratio The hyperparameter  $\alpha$  in the Equations 3 and 4 controls the correlation between the noise of different frames. A larger  $\alpha$  injects more correlation into the noise. The correlation disappears when  $\alpha \rightarrow 0$ , and the mixed and progressive noise models reproduce the vanilla noise model. To find optimal  $\alpha$ , we train our UCF-small model (69M parameters) using  $\alpha \in \{0, 0.1, 0.2, 0.5, 1, 1, 2, 5, 10, \infty\}$  and report FVD in Figure 7. For each  $\alpha$  value, we repeat the experiment 3 times and report the mean. Note that  $\alpha = 0$  indicates finetuning with the independent frame noise, and  $\alpha = \infty$  indicates using identical noise maps for all the frames, which produces frozen videos during the inference time. Finetuning an image diffusion model almost consistently outperforms the training-from-scratch baseline with different  $\alpha$ s. Using  $\alpha = 1$  for mixed noisng and  $\alpha = 2$  for progressive noisng produces similar best results. We also show qualitative results for models trained with  $\alpha = 0, 1, 10$  in Figure 6. When  $\alpha$  is too small, we notice a degradation in visual quality in the generated video frames and a reduced video diversity. For example, we notice many repeated samples and black borders in almost every video generated with
+
+![](images/84043d9b5f0034b76c6835e2d4e087432971d2d22535d277eb4e1c20e2e5608a.jpg)  
+$\alpha = 0$
+
+![](images/999c099386c0fd284da43b5fb62b6ec5de5670e01424725f8869ea74c5d45ab6.jpg)  
+$\alpha = 1$
+
+![](images/26e5b0e3adcaeb96a05393aada9def68b3f50bb8173fd43ac1a8b5ee7ec07265.jpg)  
+$\alpha = 10$
+
+![](images/2e6035bc99550b824a6ea57ef45e3d24c183214050feafa286f4413540cd2415.jpg)  
+Figure 6: Visual ablation on  $\alpha$ . Small  $\alpha = 0$  reduces video quality and diversity and large  $\alpha = 10$  yields motion artifacts.
+
+$\alpha = 0$ . On the other hand, when  $\alpha$  is too large, the model has difficulty generating proper motions.
+
+Model size We pick the best  $\alpha$  for the mixed and progressive noise models and compare them with the model trained from scratch on models with different numbers of parameters, 69M, 112M, and 253M. Figure 8 shows that our mixed and progressive models outperform the baseline consistently by a large margin in terms of FVD. Overall, mixed and progressive noising provide similar performance. In our large large-scale experiments, we choose progressive noising with  $\alpha = 2$  due to its autoregressive nature.
+
+# 6. Conclusion
+
+We proposed a new efficient way of training text-to-video generation models. By observing that the noise maps generating the frames of a video are clustered together, we study
+
+![](images/0230cc56489705efdce7ccceeafb558d19fccf3188e8d60146c865c67f51f050.jpg)  
+Figure 7: Quantitative ablation on hyperparameter  $\alpha$ . Finetuning with temporally correlated prior improves over training from scratch. Using a too-large or too-small  $\alpha$  leads to inferior results.  $\alpha = 1$ ,  $\alpha = 2$  each works the best for mixed and progressive noising, respectively.  
+Figure 8: Ablation on model size. Larger models consistently improve the performance of both finetuning and training from scratch. Finetuning from image model consistently outperforms training from scratch.
+
+mixed and progressive noise priors well-suited for sequential video frame generation. We apply our progressive noise prior to finetuning a state-of-the-art diffusion-based text-to-image model to achieve a state-of-the-art large-scale text-to-video model. The high quality of the generated videos and the state-of-the-art Inception and FID scores demonstrate the strength of our approach.
+
+Acknowledgment. We would like to thank Amanda Moran, John Dickinson, Sivakumar Arayandi Thottakara, David Page, Ranjitha Prasanna, Venkata Karri, and others in NGC and PBSS team for the computing infrastructure support. We also give thanks to Qinsheng Zhang, Zekun Hao, Tsung-Yi Lin, Ajay Jain, and Chen-Hsuan Lin for useful discussions and feedback. Thanks also go to Yin Xi and Thomas Hayes for clarifying the evaluation protocol of Make-A-Video and to Ming Ding for the details of CogVideo. We thank the ICCV PCs, ACs, and reviewers for their service and valuable feedback. This work is partly supported by NSF grants No. IIS-1910132 and IIS-2213335.
+
+# References
+
+[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 3, 7  
+[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. eDiff-I: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1, 3, 5, 7  
+[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, pages 707-723. Springer, 2022. 3  
+[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 3, 7  
+[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. CVPR, 2023. 1, 3  
+[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 6  
+[7] Duygu Ceylan, Chun-Hao Huang, and Niloy J. Mitra. Pix2video: Video editing using image diffusion. arXiv:2303.12688, 2023. 3  
+[8] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 3  
+[9] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. arXiv preprint arXiv:2204.03638, 2022. 3, 7  
+[10] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. arXiv preprint arXiv:2304.06720, 2023. 3  
+[11] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, pages 10696-10706, 2022. 3  
+[12] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 3  
+[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3  
+[14] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3, 4, 5, 7, 8  
+[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 3, 4
+
+[16] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 23:47-1, 2022. 3  
+[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 3, 4, 5, 6, 7  
+[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3, 4, 7  
+[19] Tobias Hübpe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. arXiv preprint arXiv:2206.07696, 2022. 3  
+[20] Emmanuel Kahembwe and Subramanian Ramamoorthy. Lower dimensional kernels for video discriminators. Neural Networks, 2020. 7  
+[21] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 3, 4  
+[22] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 3  
+[23] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. arXiv preprint arXiv:2212.04488, 2022. 3  
+[24] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fréchet inception distance. In ICLR, 2023. 7  
+[25] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Ccvs: Context-aware controllable video synthesis. NeurIPS, 2021. 3, 7  
+[26] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 4  
+[27] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 4  
+[28] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 3  
+[29] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Ying Shan, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186, 2023. 3  
+[30] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 1, 3  
+[31] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen.
+
+Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 3  
+[32] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3  
+[33] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 3  
+[34] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. arXiv preprint arXiv:2302.03027, 2023. 3  
+[35] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 3  
+[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 5  
+[37] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 5  
+[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 3  
+[39] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv preprint arXiv:1412.6604, 2014. 3  
+[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3  
+[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3  
+[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2, 3, 5  
+[43] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In ICCV, 2017. 3, 7  
+[44] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan. IJCV, 2020. 6
+
+[45] Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Mostganv: Video generation with temporal motion styles. In CVPR, 2023. 3  
+[46] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 3, 4, 5, 6, 7, 8  
+[47] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. arXiv preprint arXiv:2112.14683, 2021. 3, 6, 7  
+[48] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 4  
+[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2  
+[50] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5  
+[51] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In ICML, 2015. 3  
+[52] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. In ICLR, 2021. 3, 6, 7  
+[53] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 6  
+[54] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, June 2018. 3  
+[55] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. *ICLR*, 2019. 6  
+[56] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 3  
+[57] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Masked conditional video diffusion for prediction, generation, and interpolation. arXiv preprint arXiv:2205.09853, 2022. 3  
+[58] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. NIPS, 2016. 3  
+[59] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. arXiv preprint arXiv:2305.10874, 2023. 3, 7  
+[60] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In ECCV, pages 720-736. Springer, 2022. 3, 4, 7
+
+[61] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022.3  
+[62] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 7  
+[63] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3, 7  
+[64] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481, 2022. 3  
+[65] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. arXiv preprint arXiv:2212.05199, 2022.3  
+[66] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. In ICLR, 2021. 3, 7  
+[67] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 1, 3  
+[68] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. In ICLR, 2023. 4  
+[69] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2, 3, 7

2023/Preserve Your Own Correlation_ A Noise Prior for Video Diffusion Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f87ee3736d81d23612453005a114883d4b131cf617b4c047d9eedc77326e3696
+size 827552

2023/Preserving Modality Structure Improves Multi-Modal Learning/9786f414-b2fe-4822-86c7-2868d21a76fe_content_list.json

@@ -0,0 +1,1886 @@
+[
+    {
+        "type": "text",
+        "text": "Preserving Modality Structure Improves Multi-Modal Learning",
+        "text_level": 1,
+        "bbox": [
+            163,
+            130,
+            808,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sirnam Swetha<sup>1</sup>, Mamshad Nayeem Rizve<sup>1</sup>, Nina Shvetsova<sup>2,3</sup>, Hilde Kuehne<sup>2,3,4</sup>, Mubarak Shah<sup>1</sup>",
+        "bbox": [
+            96,
+            191,
+            870,
+            212
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ CRCV, University of Central Florida,  $^{2}$ Goethe University Frankfurt Germany,  $^{3}$ University of Bonn Germany,  $^{4}$ MIT-IBM Watson AI Lab",
+        "bbox": [
+            168,
+            222,
+            800,
+            258
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            304,
+            313,
+            320
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Self-supervised learning on large-scale multi-modal datasets allows learning semantically meaningful embeddings in a joint multi-modal representation space without relying on human annotations. These joint embeddings enable zero-shot cross-modal tasks like retrieval and classification. However, these methods often struggle to generalize well on out-of-domain data as they ignore the semantic structure present in modality-specific embeddings. In this context, we propose a novel Semantic-Structure-Preserving Consistency approach to improve generalizability by preserving the modality-specific relationships in the joint embedding space. To capture modality-specific semantic relationships between samples, we propose to learn multiple anchors and represent the multifaceted relationship between samples with respect to their relationship with these anchors. To assign multiple anchors to each sample, we propose a novel Multi-Assignment Sinkhorn-Knopp algorithm. Our experimentation demonstrates that our proposed approach learns semantically meaningful anchors in a self-supervised manner. Furthermore, our evaluation on MSR-VTT and YouCook2 datasets demonstrates that our proposed multi-anchor assignment based solution achieves state-of-the-art performance and generalizes to both in-and out-of-domain datasets. Code: https://github.com/Swetha5/Multi_Sinkhorn_Knopp",
+        "bbox": [
+            76,
+            338,
+            472,
+            717
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            752,
+            209,
+            767
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Humans often rely on multiple sensory inputs to have a better understanding of everyday events. Most commonly, we utilize vision, audio, and language to perceive an event as they provide complementary information for robust reasoning. The closest approximation of this setup is video data as it provides both visual and audio information along with a text description as a caption. Recently, researchers have started to explore learning meaningful representations",
+        "bbox": [
+            75,
+            780,
+            470,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "by leveraging multiple modalities to train efficient models at scale [5, 11, 42]. Such systems focus on representation learning that either improves features for each modality separately [5] or learns a joint multi-modal embedding [11, 42] space that enables various zero-shot tasks like retrieval or classification. However, given the inherent differences across the modalities, it is challenging to learn effective joint embeddings. Furthermore, the real-world data presents additional challenges like misalignment between modalities, leading to weak supervision.",
+        "bbox": [
+            496,
+            305,
+            893,
+            457
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Current pre-training approaches in this area usually employ a contrastive objective [33] to learn the joint embeddings that pulls the cross-modal embeddings of a sample from the same temporal instance closer and pushes embeddings of other samples farther. Despite promising performances, these methods struggle with generalizability. This is particularly evident in previous approaches trained on HT100M [11, 42], which do well on the closely related downstream dataset YouCook2 but struggle to improve on the MSR-VTT dataset, which exhibits a relatively larger domain shift with respect to HT100M [42]. This is due to the contrastive objective's emphasis on strict alignment between modalities in the joint embedding space while ignoring the inherent weak alignment between different modalities [45], as well as the underlying semantic structure across samples [43, 50]. Recent works have tackled these issues, either by using joint multi-modal clustering [11] to preserve the semantic structure in the joint embedding space or by incorporating a reconstruction objective [11, 25] to retain modality-specific features in the joint embedding space, allowing for weak multi-modal alignment. However, the usual reconstruction objective trivially tries to retain most modality-specific features in the joint space, thus preventing the learning of optimal features for cross-modal tasks. And the multi-modal clustering approaches perform hard-clustering making it less flexible. Therefore, the limitations of the contrastive objective cannot be adequately addressed even after combining these independent objectives.",
+        "bbox": [
+            496,
+            460,
+            895,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address this, we propose a semantic-structure",
+        "bbox": [
+            519,
+            885,
+            890,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            805,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "21993",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "preserving consistency loss (SSPC) to only retain information that is beneficial for both cross-modal embedding learning and retaining modality-specific semantic structure. In particular, for SSPC loss we consider each sample (e.g., a video clip) to be composed of multiple concepts: scene or objects involved in the downstream task. Therefore, the relationship between samples is multifaceted, representing both shared and unique concepts across samples. To capture this multifaceted relationship in a flexible manner, we propose to learn anchors (latent codes) and model the relationship between samples with respect to their relationships with these anchors. Therefore, these anchors act as a proxy to represent the modality-specific relationships between samples (semantic structure) which can be preserved using the proposed SSPC loss. Since we have no supervision to learn these anchors, we formulate this anchor learning problem as a many-to-many assignment problem, as modeling this multifaceted relationship simultaneously involves assigning multiple anchors to one sample and multiple samples to one anchor. Although there is a vast literature on solving the many-to-one assignment problem [9, 5, 38, 48], there is no efficient way to solve this many-to-many assignment problem.",
+        "bbox": [
+            76,
+            90,
+            472,
+            436
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To this end, we propose a novel Multi-Assignment Sinkhorn-Knopp (Multi-SK) algorithm that iteratively optimizes the many-to-many anchor assignments for both the modality-specific embeddings (in input space) and modality-agnostic multi-modal embeddings (in joint embedding space). To allow for weak alignment between modalities, we select the dominant anchors for each sample to represent the relationship between different samples. Our proposed SSPC loss enforces consistency between the dominant anchor assignments at the input and joint embedding spaces to preserve the modality-specific semantic structure. To demonstrate the effectiveness of our proposed solution, we train our model on HT100M dataset and test on 6 zero-shot tasks on multiple downstream datasets and observe that our approach leads to state-of-the-art results in all settings.",
+        "bbox": [
+            76,
+            439,
+            472,
+            665
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In summary, we make the following contributions: (i) We propose a flexible modality-specific semantic-structure-preserving approach to improve the generalizability of cross-modal features. (ii) We introduce Multi-Assignment Sinkhorn-Knopp, a novel algorithm to enable multiple assignments for flexible sample relationship modeling. (iii) Our proposed method outperforms the current state-of-the-art for multi-modal self-supervised representation learning on both in- and out-of domain datasets.",
+        "bbox": [
+            76,
+            665,
+            472,
+            800
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            814,
+            218,
+            830
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Multi-Modal Learning",
+        "text_level": 1,
+        "bbox": [
+            76,
+            835,
+            287,
+            852
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "With the availability of large-scale multi-modal datasets [31, 41, 7], multi-modal learning research has received a lot of attention. It comprises of vision-language",
+        "bbox": [
+            76,
+            854,
+            470,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "learning [35, 51], vision-audio learning [2, 4, 6, 12, 46], video-audio-language learning [39, 11, 42], zero-shot learning [21, 28], cross-modal generation [37, 53, 27] and multi-modal multi-task learning [22]. Miech et al. [31] proposed a large-scale multi-modal dataset consisting of video, audio and text by collecting instructional videos from YouTube without requiring any human annotations. Note that the text is generated from audio using Automatic Speech Recognition (ASR) and has noisy alignment between the text and video. They also proposed a multimodal system to demonstrate the potential for learning video-text embedding via contrastive loss. To handle the noise in the dataset, Amrani et al. [3] proposed a noise estimation for multi-modal data via multi-modal density estimation. A noise-contrastive estimation approach in a multi-instance learning framework was proposed by Miech et al. [29]. XDC [2] performs clustering on audio-video for learning better features for each modality separately. These works utilize only two modalities for multi-modal learning, while others have explored utilizing audio, video, and text together for multi-modal learning. Multi-Modal versatile networks [1] was proposed to learn different embedding spaces for each combination of modalities. AVLNet [39] proposed to learn a shared embedding that maps all modalities to a single joint embedding space. Following this, MCN [11] proposed to perform joint clustering and reconstruction to learn joint embedding space. Note that, [11] performs multi-modal K-means clustering to learn hard semantic clusters. Unlike strict assignment in [11], we propose flexible learning with multiple assignments and separately for each modality. More recently, EAO [42] utilizes transformers and combinatorial fusion of modalities to learn the joint embedding with contrastive loss.",
+        "bbox": [
+            496,
+            90,
+            893,
+            590
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Most of these works, utilize contrastive or clustering loss over fused multi-modal representation to learn the joint embedding space. By doing so, these models do not retain the modality-specific semantic structuring between samples encoded by the pre-trained modality specific backbones, hurting the generalization ability of the model. Additionally, recent works have reported that large-scale contrastive multimodal models (e.g., CLIP [36]) are somewhat robust to distributional shifts mainly due to diverse large-scale training data and prompt-engineering [15]. Therefore, our work focuses on making the pre-training objective robust to distributional shifts. In this context, we propose a novel approach to preserve the modality-specific semantic relationships in the joint embedding space by modeling the relationship between samples w.r.t learnable anchors. To enable flexible relationship modeling between samples, we learn multiple anchor assignments per sample, where anchors shared across samples model the commonality between them, and the distinct anchors between samples highlight the uniqueness of the samples.",
+        "bbox": [
+            496,
+            598,
+            895,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "21994",
+        "bbox": [
+            478,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/36b1353f0d482ab61ec574a70a19b3d49c9e6d43f912d0b1a2da048bb98354b6.jpg",
+        "image_caption": [
+            "Figure 1: Overview of the proposed model. Given weakly aligned text, video, and audio, we first extract features using frozen modality-specific backbones. These features are then passed through a token projection layer to obtain input features  $(\\mathbf{v},\\mathbf{t},\\mathbf{a})$  specific to each modality. Next, the modality-specific transformer models project the input features into a joint multi-modal representation space  $(\\hat{\\mathbf{v}},\\hat{\\mathbf{t}},\\hat{\\mathbf{a}})$ . The similarity between the input features  $(\\mathbf{v},\\mathbf{t},\\mathbf{a})$  and input space anchors  $(\\mathbf{Z}_{\\mathbf{v}},\\mathbf{Z}_{\\mathbf{t}},\\mathbf{Z}_{\\mathbf{a}})$  (projected features  $(\\hat{\\mathbf{v}},\\hat{\\mathbf{t}},\\hat{\\mathbf{a}})$  and projected space anchors  $(\\hat{\\mathbf{Z}}_{\\mathbf{v}},\\hat{\\mathbf{Z}}_{\\mathbf{t}},\\hat{\\mathbf{Z}}_{\\mathbf{a}}))$  is computed, and our Multi-SK algorithm  $(M(.))$  is used to optimize multiple anchor assignments per sample, as shown in the Input Space Assignments (Projected Space Assignments). We do this for each modality and enforce the respective consistency losses, but in this figure we only show modality-specific consistency loss for text anchors and cross-modal consistency between text and video modalities for brevity. LN, MHSA represents LayerNorm and Multi-Head Self-Attention."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            76,
+            88,
+            895,
+            410
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2. Sinkhorn-Knopp",
+        "text_level": 1,
+        "bbox": [
+            76,
+            551,
+            246,
+            569
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Recently, Sinkhorn-Knopp algorithm [44] has drawn huge attention because of its effectiveness in solving optimal-transport problems [23, 8]. Specifically, [13] proposed an entropic relaxation of the optimal transport problem which can be efficiently solved using Sinkhorn's matrix scaling algorithm. Many following works have since successfully utilized the Sinkhorn-Knopp algorithm to solve different label assignment problems framed as an optimal transport problems. For instance, SeLa [48] cast the unsupervised clustering problem as a pseudo-label assignment problem and used the Sinkhorn-Knopp algorithm to solve it. SeLaVi [5] extended this idea to self-supervised representation learning for multi-modal data where the cluster assignments between different modalities are swapped to encourage modality invariant representation learning. Similarly, SwAV [9] used the Sinkhorn-Knopp algorithm for self-supervised representation learning and proposed to swap the pseudo-labels for differently augmented versions of a sample and use soft assignments instead of hard pseudolabels. In contrast to these works, SuperGlue [40] used the Sinkhorn-Knopp algorithm to solve the correspondence",
+        "bbox": [
+            75,
+            583,
+            472,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "problem between two sets of local features. Moreover, Sinkhorn-Knopp has been used in detection problems [17], where it was used to match the anchors with ground truths. Recently, UNO [16], and TRSSL [38] have successfully used the Sinkhorn-Knopp algorithm in solving novel class discovery and open-world semi-supervised learning problems, respectively. One key limitation of the traditional Sinkhorn-Knopp algorithm is that it cannot be directly utilized to compute multiple assignments necessary to perform multi-anchor based learning i.e. many-to-many assignments.",
+        "bbox": [
+            496,
+            553,
+            890,
+            718
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Some prior works [18, 26] have attempted to solve many-to-many assignments in indirect ways. While authors of [18] use an intermediate graph to match groups of vertices from source to target graph, which can only perform group-to-group assignments and is inadequate for our problem that requires true many-to-many matching. [26] modifies the Sinkhorn-Knopp row and column constraints to obtain many-to-many assignment to model dense correspondences, however, we have found this approach to be inferior in solving the multiple anchor assignment problem. The modified Sinkhorn-Knopp constraints approach yields suboptimal results, as discussed in Sec. 4.4. To address these",
+        "bbox": [
+            496,
+            719,
+            893,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "21995",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "limitations, we propose a novel algorithm Multi-SK, which outperforms the modified Sinkhorn-Knopp constraints approach to get true many-to-many assignments.",
+        "bbox": [
+            75,
+            90,
+            468,
+            137
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            76,
+            145,
+            168,
+            160
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Given a set of multi-modal inputs  $\\{\\mathbf{t}^{(i)},\\mathbf{v}^{(i)},\\mathbf{a}^{(i)}\\}_{i = 1}^{N}$  from N video clips, we learn modality-specific projection functions  $f_{t},f_{v},f_{a}$ , that transform t,v,a into a d-dimensional joint embedding space,  $\\mathbb{R}^d$  to obtain  $\\hat{\\mathbf{t}},\\hat{\\mathbf{v}},\\hat{\\mathbf{a}}$  respectively. Our goal is to optimize the parameters of  $f_{t},f_{v},f_{a}$  in such a way that they maintain the semantic structure amongst samples from a particular modality in the joint embedding space, as discussed in Sec 3.1, and simultaneously brings the semantically related cross-modal inputs closer. In the following, first, we formulate our approach to modeling the relationship between samples using anchors in Sec 3.1, then in Sec 3.2, we discuss our novel Multi-SK algorithm to learn these anchors for representing sample relationships, and finally, we present the overall training objective to train the model in Sec 3.3.",
+        "bbox": [
+            75,
+            169,
+            472,
+            398
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1. Modeling Sample Relationships with Anchors",
+        "text_level": 1,
+        "bbox": [
+            76,
+            406,
+            465,
+            422
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our work aims to preserve the relationship between samples i.e.  $\\mathbf{x}^{(i)},\\mathbf{x}^{(j)}$  from a particular modality, to better generalize on the unseen data. To this end, we propose to model the relationship between samples using anchors, where the similarity of each sample w.r.t. anchors encodes the semantic structure of the feature space. Unlike clustering approaches, which involve hard-assignment to a specific cluster, our approach offers flexibility i.e., there can be shared and unique anchors across samples that define the relationship between them. To be particular, for each sample, we learn  $K$  anchors, and the similarity of assignments over these anchors represents the relationship (which we want to preserve) between samples from a particular modality.",
+        "bbox": [
+            75,
+            431,
+            468,
+            626
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As the pre-trained features extracted from modality-specific backbones encode the semantic structure between samples within that particular modality, preserving such modality-specific relationships in the joint embedding space would boost the generalizability of the model. To achieve this, we define two sets of  $K$  learnable anchors  $\\mathbf{z} = \\{\\mathbf{z}^{(i)}\\}_{i=1}^{K}$ , and  $\\hat{\\mathbf{z}} = \\{\\hat{\\mathbf{z}}^{(i)}\\}_{i=1}^{K}$  to model the sample relationships before and after projecting them to the joint embedding space respectively. We repeat this for all the modalities,  $t, v, a$ . We propose to preserve the modality-specific semantic structure at the joint embedding space by enforcing consistency in anchor assignments before and after performing feature projections (Eq. 3).",
+        "bbox": [
+            75,
+            628,
+            468,
+            824
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We still have one remaining challenge, i.e. how to learn/discover these anchors in an unsupervised manner. To this end, we cast the anchor discovery as a label assignment task with a uniform prior, i.e. each anchor will have an equal number of sample assignments. Additionally, to",
+        "bbox": [
+            75,
+            825,
+            468,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "encourage flexible modeling, we enforce multiple anchor assignments per sample. However, it is difficult to estimate the exact number of anchors for each sample without additional prior information, thus we select the top  $K'$  anchors for each sample to effectively model the sample relationships. Even though this optimization task may seem like a difficult combinatorial problem, we model this combinatorial optimization task as an optimal transport problem. Following recent works [5, 9], one might assume the Sinkhorn-Knopp algorithm is a natural choice to solve this problem, however, the vanilla Sinkhorn-Knopp algorithm cannot handle multiple assignments per sample i.e. many-to-many assignment. To address this limitation, we propose Multi-Assignment Sinkhorn-Knopp algorithm, presented in the following.",
+        "bbox": [
+            496,
+            90,
+            893,
+            318
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Multi-Assignment Sinkhorn-Knopp",
+        "text_level": 1,
+        "bbox": [
+            498,
+            329,
+            808,
+            345
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Given a sample matrix  $\\mathbf{B}$  s.t.  $\\mathbf{B} \\in \\mathbb{R}^{N \\times d}$ , representing  $N$  samples and an anchor matrix  $\\mathbf{Z}$  s.t.  $\\mathbf{Z} \\in \\mathbb{R}^{K \\times d}$ , of  $K$  anchor vectors, we obtain the similarity matrix  $\\mathbf{S}$  s.t.  $\\mathbf{S} = \\mathbf{B}\\mathbf{Z}^{\\intercal}$  and  $\\mathbf{S} \\in \\mathbb{R}^{N \\times K}$ , where,  $\\mathbf{S}_{ij}$  is the probability of assigning  $j$ th anchor to the  $i$ th sample vector. The goal is to find an anchor assignment matrix,  $\\mathbf{Q}$ , such that it satisfies the following constraints: (i) a sample should be assigned to exactly  $K'$  anchors, to learn top  $K'$  anchor assignments per sample. (ii) anchor assignments must be equally partitioned i.e each anchor must be selected exactly  $N \\times K' / K$  times, for uniform anchor assignment.",
+        "bbox": [
+            496,
+            352,
+            893,
+            518
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To obtain such multiple anchor assignments per sample, we propose to create a substitute 3D assignment matrix  $\\mathbf{Q}'$  s.t.  $\\mathbf{Q}' \\in \\mathbb{R}^{K \\times N \\times K}$ . We also generate a 3D similarity matrix  $\\mathbf{S}'$  from the 2D similarity matrix  $\\mathbf{S}$  by introducing  $K$  channels (depth dimension) to get  $K \\times N \\times K$  matrix such that each channel is a scaled matrix of  $S$  with a predefined ranking between the channels enabling top  $K'$  anchor selection. Since we are only interested in selecting the top  $K'$  anchors, we set the first  $K'$  channels of  $\\mathbf{S}'$  to be the same as  $\\mathbf{S}$ . The remaining  $K - K'$  channels are set to  $\\mu \\mathbf{S}$ . Here,  $\\mu$  is a damping factor s.t.  $0 < \\mu < 1$ , to help select the top  $K'$  anchors for each sample. We discuss alternate designs for  $\\mathbf{S}'$  generation in the Supplementary Sec. 2.",
+        "bbox": [
+            496,
+            520,
+            893,
+            715
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The optimization objective of Multi-Assignment Sinkhorn-Knopp is to find an assignment matrix  $\\mathbf{Q}'$  such that it satisfies our multi-anchor assignment constraints while maximizing similarity with the initial assignment/similarity matrix,  $\\mathbf{S}'$ . This optimization problem is defined as:",
+        "bbox": [
+            496,
+            717,
+            893,
+            805
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathbf {Q} ^ {*}: \\max  _ {\\mathbf {Q} ^ {\\prime}} <   \\mathbf {Q} ^ {\\prime}, \\mathbf {S} ^ {\\prime} > + \\epsilon H (\\mathbf {Q} ^ {\\prime}) \\\\ H \\left(\\mathbf {Q} ^ {*}\\right) = - \\sum \\mathbf {Q} _ {i j k} ^ {\\prime} \\log \\mathbf {Q} _ {i j k} ^ {\\prime} \\tag {1} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            803,
+            890,
+            852
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "$\\mathbf{Q}^*$  needs to satisfy the following constraints to be a valid solution for our multi-anchor assignment problem.",
+        "bbox": [
+            496,
+            851,
+            890,
+            881
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "-  $\\mathbf{Q}^*$  Row Constraint: Within a channel, the sum of all el",
+        "bbox": [
+            500,
+            885,
+            890,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "21996",
+        "bbox": [
+            478,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ements in a particular row must be equal to one. This is because we only want one anchor assignment for a sample in a particular channel.  $\\forall i,k\\sum_{j}\\mathbf{Q}^{*}_{ijk} = 1$",
+        "bbox": [
+            89,
+            90,
+            468,
+            138
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "-  $\\mathbf{Q}^*$  Column Constraint: In a channel, the sum of all elements in a column should be equal to  $N / K$ . This constraint enforces equal partitioning of anchor assignments.  $\\forall j, k \\sum_{i} \\mathbf{Q}^{*}_{ijk} = N / K$ .",
+            "-  $\\mathbf{Q}^*$  Depth constraint: Depth-wise sum should be equal to one for every sample and anchor combination. This constraint prevents selecting the same anchor across different channels.  $\\forall i, j \\sum_{k} \\mathbf{Q}^{*}_{ijk} = 1$ ."
+        ],
+        "bbox": [
+            76,
+            142,
+            468,
+            271
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The traditional Sinkhorn-Knopp method uses an iterative matrix scaling algorithm that scales the rows and columns alternatively (with desired constraints) till the desired assignment matrix is obtained. We employ a similar scheme and extend the iterative scaling to the depth dimension for estimating  $\\mathbf{Q}^*$ . We iteratively scale the rows, columns, and channels (depth dimension) till convergence. The final 2D assignment matrix  $\\mathbf{Q}$  is computed by performing a depth-wise sum on the top  $K^{\\prime}$  channels. We provide Pytorch-style pseudo-code for our Multi-Assignment Sinkhorn-Knop algorithm in the Supplementary Sec 3.",
+        "bbox": [
+            75,
+            272,
+            470,
+            438
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3. Training Objective",
+        "text_level": 1,
+        "bbox": [
+            76,
+            445,
+            259,
+            462
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Semantic Structure Preserving Consistency Loss. To preserve the semantic structure of each modality,  $t, v, a$ , we apply consistency loss to enforce similar anchor assignments between input and joint embedding space. As the cross-modal contrastive loss in the joint embedding space tries to bring different modalities together, features in the joint embedding space from a particular modality should preserve the common anchors that exist in corresponding features from the other modalities. Therefore, we also apply cross-modal anchor consistency across all modalities as shown in Eq. 3. Since we are dealing with 3 input modalities, this cross-modal consistency results in 9 consistency constraints.",
+        "bbox": [
+            75,
+            468,
+            468,
+            664
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Let's denote  $\\mathcal{L}(\\mathbf{t},\\hat{\\mathbf{v}},\\mathbf{z}_t,\\hat{\\mathbf{z}}_t)$  as the consistency loss between text anchor assignment in the input space and the textual anchor assignment of the corresponding video features at the joint embedding space as shown below in Eq. 2:",
+        "bbox": [
+            75,
+            665,
+            468,
+            726
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\mathbf {t}, \\hat {\\mathbf {v}}, \\mathbf {z} _ {t}, \\hat {\\mathbf {z}} _ {t}) = \\alpha_ {\\mathbf {t}, \\hat {\\mathbf {v}}} g \\left(\\operatorname {s i m} \\left(\\mathbf {t}, \\mathbf {z} _ {t}\\right), M \\left(\\operatorname {s i m} \\left(\\hat {\\mathbf {v}}, \\hat {\\mathbf {z}} _ {t}\\right)\\right)\\right) (2) \\\\ + \\beta_ {\\mathbf {t}, \\hat {\\mathbf {v}}} g (\\operatorname {s i m} (\\hat {\\mathbf {v}}, \\hat {\\mathbf {z}} _ {t})), M (\\operatorname {s i m} (\\mathbf {t}, \\mathbf {z} _ {t}))), (2) \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            96,
+            731,
+            468,
+            766
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Here,  $\\mathbf{z}_t, \\hat{\\mathbf{z}}_t$  respectively represent input and output learnable anchor vectors for the text modality.  $g(.)$ , and  $M(.)$  respectively represent binary cross-entropy-with-logits loss and Multi-Assignment Sinkhorn-Knopp (discussed in Sec. 3.2),  $\\alpha$  and  $\\beta$  represents loss coefficients, and  $\\mathrm{sim}(\\mathbf{a},\\mathbf{b}) = \\exp (\\mathbf{a.b} / \\tau \\| \\mathbf{a}\\| \\| \\mathbf{b}\\|)$ ,  $\\tau$  is the temperature hyperparameter of the similarity metric.",
+        "bbox": [
+            75,
+            766,
+            468,
+            869
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Overall semantic structure preserving consistency loss for all the modalities is defined as:",
+        "bbox": [
+            75,
+            869,
+            468,
+            898
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {s s p c} = \\sum_ {m \\in \\{t, v, a \\}} \\sum_ {n \\in \\{t, v, a \\}} \\mathcal {L} (\\mathbf {m}, \\hat {\\mathbf {n}}, \\mathbf {z} _ {m}, \\hat {\\mathbf {z}} _ {m}). \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            542,
+            102,
+            890,
+            136
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Contrastive Loss. Following [11, 42], we also use contrastive loss to bring cross-modal embeddings of the same sample closer while pushing away embeddings from other sample. For this, we use 3 pairwise single-modality constrastive losses,  $\\mathcal{L}_{nce\\_ tv}$ ,  $\\mathcal{L}_{nce\\_ ta}$ ,  $\\mathcal{L}_{nce\\_ va}$  between  $(t,v)$ ,  $(t,a)$ ,  $(v,a)$  respectively. Specifically, we use Noise Contrastive Estimation [33] with temperature  $\\kappa$  as shown in Eq. 4.",
+        "bbox": [
+            496,
+            135,
+            893,
+            256
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {n c e - x y} = - \\log \\frac {\\exp \\left(\\mathbf {x} ^ {\\intercal} \\mathbf {y} / \\kappa\\right)}{\\sum_ {i = 1} ^ {N} \\exp \\left(\\mathbf {x} ^ {(i) \\intercal} \\mathbf {y} ^ {(i)} / \\kappa\\right)} \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            557,
+            268,
+            890,
+            306
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The overall contrastive loss for all modalities is defined as  $\\mathcal{L}_{nce} = \\lambda_{tv}\\mathcal{L}_{nce\\_tv} + \\lambda_{ta}\\mathcal{L}_{nce\\_ta} + \\lambda_{va}\\mathcal{L}_{nce\\_va}$",
+        "bbox": [
+            498,
+            311,
+            890,
+            343
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Overall Loss",
+        "text_level": 1,
+        "bbox": [
+            500,
+            357,
+            611,
+            372
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The overall training objective is a combination of SSPC loss (Eq. 3) and contrastive loss (4):  $\\mathcal{L}_f = \\lambda_{sspc}\\mathcal{L}_{sspc} + \\lambda_{nce}\\mathcal{L}_{nce}$ , where,  $\\lambda_{sspc}$  and  $\\lambda_{nce}$  are loss coefficients. By combining both losses, the model learns a more generic joint embedding space which preserves the modality-specific semantic structure by enforcing the anchor assignment similarity before and after feature projection and also brings representations of different modalities together by utilizing contrastive loss.",
+        "bbox": [
+            496,
+            382,
+            890,
+            518
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            534,
+            633,
+            550
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            500,
+            559,
+            689,
+            574
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Backbones. For comparability, we follow the same setup of previous works [31, 39, 11, 42]. As visual backbone, we use a combination of 2D features from ResNet-152 [20] pretrained onImagenet [14], and 3D features from ResNeXt101 [19] pretrained on Kinetics [10]. The text backbone is GoogleNews pretrained Word2vec model [32]. These backbones are fixed and not finetuned during training. Following [11, 42], we use a trainable CNN with residual layers as an audio backbone. We provide additional details in the Supplementary Sec. 5.",
+        "bbox": [
+            496,
+            582,
+            890,
+            733
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Data Sampling. We use a batch of 216 videos and randomly sample ten 8-second clips per video. If the sampled clip contains narration (95% clips), we use ASR time stamps to select clip borders. To disentangle high text-audio correlation in HT100M, we shift the audio clip randomly by 4 seconds with respect to the video and text boundaries.",
+        "bbox": [
+            496,
+            734,
+            890,
+            824
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Projections. Following [42, 11, 39], we use a gated linear projection [30] to project features into common token space, as well as to project resulting tokens into shared embedding space. We set the dimension of the common token space to 4096 and of the shared embedding space to 6144.",
+        "bbox": [
+            496,
+            825,
+            890,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "21997",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/ed572caec3374f9dd60945e2e137049d9279ef7f7644f808f777eba12bd1a223.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td>Retrieval</td><td>Train</td><td>Visual</td><td colspan=\"3\">Trainable BB</td><td colspan=\"4\">MSR-VTT</td><td colspan=\"3\">YouCook2</td><td></td></tr><tr><td></td><td>Dataset</td><td>BB</td><td>t</td><td>v</td><td>a</td><td>R@5↑</td><td>R@10↑</td><td>MedR↓</td><td>MeanR↓</td><td>R@5↑</td><td>R@10↑</td><td>MedR↓</td><td>MeanR↓</td></tr><tr><td>ActBERT [54]</td><td>t → v</td><td>HT100M</td><td>Res3D+Faster R-CNN</td><td></td><td></td><td></td><td>23.4</td><td>33.1</td><td>36</td><td>-</td><td>26.7</td><td>38.0</td><td>19</td><td>-</td></tr><tr><td>SupportSet [34]</td><td>t → v</td><td>HT100M</td><td>R152 + R(2+1)D-34</td><td>✓</td><td></td><td></td><td>23.0</td><td>31.1</td><td>31</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>HT100M [31]</td><td>t → v</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td></td><td>21.2</td><td>29.6</td><td>38</td><td>-</td><td>17.3</td><td>24.8</td><td>46</td><td>-</td></tr><tr><td>AVLNet [39]</td><td>t → v</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>24.7</td><td>34.2</td><td>-</td><td>-</td><td>21.1</td><td>29.6</td><td>-</td><td>-</td></tr><tr><td>EAO [42]</td><td>t → v</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>24.6</td><td>35.3</td><td>25</td><td>90.4</td><td>27.9</td><td>38.9</td><td>19</td><td>119.6</td></tr><tr><td>Ours</td><td>t → v</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>26.4</td><td>35.1</td><td>23</td><td>92.2</td><td>29.4</td><td>40.7</td><td>18</td><td>111.8</td></tr><tr><td>AVLNet [39]</td><td>v → t</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>27.2</td><td>35.7</td><td>25</td><td>86.5</td><td>22.8</td><td>32.9</td><td>30</td><td>142.2</td></tr><tr><td>EAO [42]</td><td>v → t</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>27.6</td><td>36.6</td><td>25</td><td>85</td><td>31.8</td><td>70.5</td><td>15</td><td>91.9</td></tr><tr><td>Ours</td><td>v → t</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>27.2</td><td>37.1</td><td>23</td><td>84.5</td><td>32</td><td>72</td><td>15</td><td>85.2</td></tr><tr><td>AVLNet [39]</td><td>t → v + a</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>19.2</td><td>27.4</td><td>47</td><td>-</td><td>36.1</td><td>44.3</td><td>16</td><td>-</td></tr><tr><td>MCN [11]</td><td>t → v + a</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>25.2</td><td>33.8</td><td>-</td><td>-</td><td>35.5</td><td>45.2</td><td>-</td><td>-</td></tr><tr><td>EAO [42]</td><td>t → v + a</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>23.3</td><td>33.2</td><td>29</td><td>94.8</td><td>38.5</td><td>49.2</td><td>11</td><td>82.7</td></tr><tr><td>Ours</td><td>t → v + a</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>25.1</td><td>34.5</td><td>26</td><td>91.8</td><td>39.4</td><td>50.1</td><td>10</td><td>83.3</td></tr><tr><td>AVLNet [39]</td><td>v + a → t</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>19</td><td>26.3</td><td>44</td><td>128.1</td><td>48.8</td><td>58.4</td><td>6</td><td>67.1</td></tr><tr><td>EAO [42]</td><td>v + a → t</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>21.8</td><td>31.4</td><td>28.5</td><td>98.9</td><td>49</td><td>60.9</td><td>6</td><td>43.8</td></tr><tr><td>Ours</td><td>v + a → t</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>24</td><td>32</td><td>27</td><td>95.9</td><td>48.8</td><td>61.3</td><td>6</td><td>43.5</td></tr></table>",
+        "bbox": [
+            78,
+            88,
+            895,
+            315
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Zero-shot Retrieval results on MSR-VTT/YouCook2. For fair comparison, we compare with models trained on text, video and audio. Retrieval column represents the evaluation task. BB=Backbone. Bold, underline represent highest and second-highest scores.",
+        "bbox": [
+            75,
+            320,
+            893,
+            349
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We use a single transformer block with hidden size of 4096 with 64 heads and an MLP size of 4096. We set the number of anchors  $K$  as 64 and  $K'$  as 32 with damping factor  $\\mu$  as 0.25. We train all models for 15 epochs using an Adam optimizer [24] with a learning rate of 5e-5, exponential decay of 0.9 and the temperature of cosine similarity  $(\\tau)$  as 0.1. We maintain a memory-bank of size 5500 while performing Multi-SK. Following [1, 42], we set higher weight for loss terms that involve text-video in Eq. 3, i.e., for all text-video terms the weight is set to 1.0 and the rest of the weights are set to 0.1.",
+        "bbox": [
+            75,
+            364,
+            472,
+            531
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Datasets, Tasks, Metrics",
+        "text_level": 1,
+        "bbox": [
+            76,
+            542,
+            299,
+            559
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Pretraining Dataset. We train our model on the HT100M dataset [31], which contains over 1 million instructional videos with automatically generated text narrations. The text narrations can be assumed to be noisy and to not always describe the video scene.",
+        "bbox": [
+            75,
+            566,
+            468,
+            643
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Zero-shot Retrieval. We use MSR-VTT [47] and YouCook2 [52] datasets to evaluate the zero-shot retrieval capability of our model. We report performance on 4 retrieval tasks: (i) Text-to-Video retrieval, (ii) Video-to-Text retrieval, (iii) Text-to-Video-Audio retrieval, (iv) Video-Audio-to-Text retrieval. The YouCook2 dataset contains cooking videos from YouTube with human-annotated clips ( $\\sim$  2 - 200 secs). For evaluation we use at maximum first 48 seconds of clip, since most clips are shorter than that. The MSR-VTT dataset contains human-annotated clips ( $\\sim$  10 - 30 secs) on various topics and provides captions with natural language sentences. Following [31, 39, 11, 42], to evaluate our model on MSR-VTT, we use the 1k set of test clips [49], and for YouCook2, we use 3,350 validation clips [31]. To perform  $(t\\rightarrow v + a)$  retrieval, we compute similarities by dot product between a text query  $t$  and all videos in the dataset using a averaged  $v + a$  representa",
+        "bbox": [
+            75,
+            643,
+            472,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "tion for each video. We report standard recall metrics  $\\mathbf{R}@\\mathbf{5}$ ,  $\\mathbf{R}@\\mathbf{10}$ , median rank (MedR) and the mean rank(MeanR). Further, we also evaluate our model using CLIP backbone and report results in Sec. 1 of the Supplementary.",
+        "bbox": [
+            496,
+            364,
+            890,
+            425
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Zero-Shot Full-Video Retrieval. Following [11], we evaluate Zero-Shot Full Video Retrieval from a set of captions on YouCook2 dataset. We report recall metrics following Caption averaging method [11] that finds maximal prediction over all the clips of video for each caption and averaging over set of captions in query leading to a single prediction for full video.",
+        "bbox": [
+            496,
+            426,
+            890,
+            530
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Text-to-video Retrieval after Fine-tuning. We additionally evaluate the retrieval performance of the models fine-tuned on downstream tasks. We use 6783 clips from MSR-VTT (which contain audio) and 9586 clips from YouCook2 train datasets to fine-tune the model as proposed by [39].",
+        "bbox": [
+            496,
+            532,
+            890,
+            608
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We also evaluate our model for Zero-shot Classification and report additional results in Sec. 1 of the Supplementary.",
+        "bbox": [
+            496,
+            608,
+            890,
+            640
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Comparison with State-of-the-art",
+        "text_level": 1,
+        "bbox": [
+            498,
+            650,
+            794,
+            666
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Zero-shot Retrieval Tasks. In Tab. 1, we report the performance of the learned multi-modal representations on four zero-shot retrieval tasks, (i) Text-to-Video, (ii) Video-to-Text, (iii) Text-to-Video-Audio, (iv) Video-Audio-to-Text., on MSR-VTT and YouCook2 datasets. For a fair comparison, we only compare with models trained on all three modalities i.e. text, video and audio. In summary, our proposed method outperforms the current state-of-the-art methods by a noticeable margin on both the datasets. The results on the MSR-VTT dataset are particularly interesting since it demonstrates the generalizability of our model. In particular, HT100M consists of instructional videos and the textual descriptions are generated from audio using ASR. The text narration has noisy alignment with video and typically describes the steps in the video. YouCook2 shares",
+        "bbox": [
+            496,
+            674,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "21998",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/9771281867bec15a8ba481b4691adaa823ae740ed92b82472ea88a3ea6b3067b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td>Retrieval</td><td>Train</td><td>Visual</td><td colspan=\"2\">Trainable BB</td><td colspan=\"4\">MSR-VTT</td><td colspan=\"4\">YouCook2</td><td></td></tr><tr><td></td><td>Dataset</td><td>BB</td><td>t</td><td>v</td><td>a</td><td>R@5↑</td><td>R@10↑</td><td>MedR↓</td><td>MeanR↓</td><td>R@5↑</td><td>R@10↑</td><td>MedR↓</td><td>MeanR↓</td></tr><tr><td>AVLNet [39]</td><td>t → v</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>42.2</td><td>56.2</td><td>7</td><td>35.1</td><td>23.7</td><td>32.7</td><td>28</td><td>122.6</td></tr><tr><td>EAO [42]</td><td>t → v</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>47.7</td><td>59.3</td><td>6</td><td>35.6</td><td>33.9</td><td>45.8</td><td>13</td><td>70.7</td></tr><tr><td>Ours</td><td>t → v</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>48.7</td><td>60.6</td><td>5</td><td>33.1</td><td>35.6</td><td>48.1</td><td>12</td><td>64.9</td></tr></table>",
+        "bbox": [
+            78,
+            88,
+            893,
+            165
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/e0dcb005520344e71c5cd8360820be3a47f7dd7e4e2773a2fa5d75082e01241d.jpg",
+        "table_caption": [
+            "Table 2: Text-to-video retrieval results on MSR-VTT/YouCook2 in the fine-tune setting. For fair comparison, we compare with models trained on all 3 modalities i.e. text, video and audio. Bold, underline represent best and second-best scores."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Aggregation</td><td>R@1↑</td><td>R@5↑</td><td>R@10↑</td></tr><tr><td>Random</td><td>-</td><td>0.23</td><td>1.15</td><td>2.32</td></tr><tr><td>HT100M [31]</td><td>Caption Avg.</td><td>43.1</td><td>68.6</td><td>79.1</td></tr><tr><td>MIL-NCE [29]</td><td>Caption Avg.</td><td>46.6</td><td>74.3</td><td>83.7</td></tr><tr><td>MCN [11]</td><td>Caption Avg.</td><td>53.4</td><td>75.0</td><td>81.4</td></tr><tr><td>EAO [42]</td><td>Caption Avg.</td><td>62.9</td><td>80.5</td><td>86.7</td></tr><tr><td>Ours</td><td>Caption Avg.</td><td>65.1</td><td>83.2</td><td>87.6</td></tr></table>",
+        "bbox": [
+            78,
+            214,
+            468,
+            327
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "domain similarity with HT100M as its videos are instructional format, and the text descriptions correspond to specific steps in the recipe. In contrast, MSR-VTT is not restricted to instructional videos and the text is typically a single sentence caption describing the whole video. As a result, there is a distributional shift between HT100M and MSR-VTT. Therefore, zero-shot retrieval on MSR-VTT has to overcome distributional shift, which is a crucial requirement for practical deployment. Our proposed method shows relatively higher improvement on this task validating the effectiveness of anchor-based learning.",
+        "bbox": [
+            75,
+            378,
+            468,
+            545
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In case of text-to-video-audio retrieval, our method improves the Median and Mean rank of the baseline [42] by  $3\\%$  on MSR-VTT along with gains on recall metrics  $R@5$  and  $R@10$ . For video-audio-to-text retrieval, our methods perform very well on MSR-VTT with  $3\\%$  improvement on MeanR,  $2.2\\%$  improvement on  $R@5$ . We also observe similar improvements on the YouCook2 dataset. Similarly, our proposed method outperforms the current state-of-the-art on most of the metrics for text-video-retrieval video-text-retrieval. Additionally, to compare our approach with text-video only model, we train our model on video and text to compare the performance with state-of-the-art in Sec 1 of the Supplementary.",
+        "bbox": [
+            75,
+            547,
+            470,
+            744
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Zero-shot Full Video Retrieval. In Tab. 3, we report results for text-to-full-video retrieval task. Our approach outperforms prior works by  $2.2\\%$ ,  $2.7\\%$  on R@1 and R@5 respectively.",
+        "bbox": [
+            75,
+            746,
+            468,
+            808
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Retrieval after Fine-tuning. We also evaluate the retrieval performance of our model after fine-tuning on the downstream datasets as shown in Tab. 2. For a fair comparison, we only report the baselines that use the same training split. We outperform state-of-the-art consistently on all the metrics and on both MSR-VTT & YouCook2 datasets",
+        "bbox": [
+            75,
+            810,
+            470,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "as shown in Tab. 2.",
+        "bbox": [
+            500,
+            217,
+            629,
+            231
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4. Ablation Studies",
+        "text_level": 1,
+        "bbox": [
+            500,
+            239,
+            663,
+            253
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "First we analyze the impact of the proposed components and report results in Tab. 4, followed by the effect of number of anchors  $(K)$  &  $(K^{\\prime})$  in Tab. 5.",
+        "bbox": [
+            498,
+            263,
+            893,
+            309
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/bf6e4abd95019bf4aba296720631dbbb6302db5fe0ddfe7c23dcb464d71a72a9.jpg",
+        "table_caption": [
+            "Table 3: Zero-Shot Text-to-Full Video retrieval on YouCook2. Averaging across captions is used to obtain video-level predictions."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">MSR-VTT</td><td colspan=\"2\">YouCook2</td></tr><tr><td>R@5↑</td><td>R@10↑</td><td>R@5↑</td><td>R@10↑</td></tr><tr><td>Recon. + CL</td><td>23.1</td><td>32.4</td><td>37.8</td><td>48.7</td></tr><tr><td>w/o CM SSPCL</td><td>22.7</td><td>31.8</td><td>36.9</td><td>48.3</td></tr><tr><td>w/o SSPCL</td><td>23.8</td><td>33.3</td><td>37.9</td><td>48.8</td></tr><tr><td>Modified SK</td><td>23.4</td><td>31.3</td><td>37.9</td><td>48.3</td></tr><tr><td>Ours</td><td>25.1</td><td>34.5</td><td>39.4</td><td>50.1</td></tr></table>",
+        "bbox": [
+            514,
+            319,
+            875,
+            438
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Ablation studies showing the impact of various components for zero-shot retrieval task. Recon.=Reconstruction Loss, CM SSPCL=Cross-Modal SSPC Loss, SK=Sinkhorn-Knopp, CL=Contrastive Loss.",
+        "bbox": [
+            498,
+            444,
+            892,
+            503
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Effect of Proposed Components. We report the results for this ablation in Tab. 4. In first row, we report results using reconstruction loss and contrastive loss, we notice that using reconstruction loss reduces the performance by  $2\\%$  on all metrics indicating the effectiveness of the proposed SSPC loss. In the second row, we report results without our proposed cross-modal SSPC loss ('w/o CM SSPC'). We employ this loss to bring the cross-modal representations closer in the joint embedding space to obtain better performance in zero-shot cross-modal tasks. We notice that removing the cross-modal SSPC loss drastically decreases the zero-shot retrieval performance on both MSR-VTT and YouCook2 datasets,  $2.4\\%$  and  $2.5\\%$  drop in R@5 performance respectively. This empirically validates the effectiveness of the cross-modal SSPC loss in obtaining better cross-modal representations.",
+        "bbox": [
+            496,
+            537,
+            890,
+            777
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In the third row, we analyze the effect of the proposed SSPC loss ('w/o SSPCL'). We apply the SSPC loss to retain modality-specific semantic structure from the pretrained models in the joint embedding space. To investigate its impact, we remove the anchor consistency between the modality-specific and joint embedding spaces. Instead, we enforce the anchor assignments before Multi-SK to be consistent with the Multi-SK optimized anchor as",
+        "bbox": [
+            496,
+            779,
+            892,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "21999",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/b2cc957fe4de615732ae3dfbc2a05de5dbf12c811084fd877125e2262a3ada0f.jpg",
+        "image_caption": [
+            "(a) UMeazzCYlps (Holidays and Traditions)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            88,
+            197,
+            141
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/8263a7ab1f21a9d5ba54fa0df65dff181159df1cd90f2150283e72300b5f1d53.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            202,
+            88,
+            321,
+            141
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/61fe488eb9ef28b0a46f636b11a59fdb546ec4494b06ce9f624e88b67cc4cfe5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            326,
+            88,
+            446,
+            141
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/2deac40873fcf07f5e27e410e2c90f9c34d5c2d674c8dbd2d806a878162499c0.jpg",
+        "image_caption": [
+            "(b) fdUgKk-G5Tw (Food and Entertaining)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            88,
+            643,
+            141
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/96296052ffce8bfcf13cced27a5f7ebc8c2ffa957ebf7b27539e93f161741a84.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            647,
+            88,
+            766,
+            141
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6a69f6250264733e7faaad8f4cccfab0da2a185e5af60020327526c5625df11b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            88,
+            890,
+            141
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/f9c1f037b7e0082a484beb440766b9e3808933f28d379e5e1d01b1bd205ba1be.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            84,
+            172,
+            883,
+            202
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/57f91f9de5423015637a8942a535d6095aeb60cc0d73e585210e8e50ea094708.jpg",
+        "image_caption": [
+            "(d) A0hURqF7h-c (Personal Care and Style)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            78,
+            226,
+            197,
+            280
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/24c730676e7520321e775bf58e7956c85dc832f12e26efc1912c06d46d44371c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            202,
+            226,
+            321,
+            279
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b4b48feb9d8852948609348aa4acd1ffb68a3512ae000e6de3bd7ed5e96af616.jpg",
+        "image_caption": [
+            "(c) Video Anchor Assignments"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            326,
+            226,
+            446,
+            279
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5eb9eb018f1a5cdd6437b2ab14e1e16f6076521c5f5a8a8b1778559db2862949.jpg",
+        "image_caption": [
+            "Figure 2: Anchor assignments illustrated in this figure demonstrate the visual similarity between video samples from two related categories i.e. (a) Holidays and Tradition and (b) Food and Entertaining; and two different categories i.e. (d) Personal Care ans Style and (e) Finance and Business within the HowTo100M dataset. The visual similarity is also reflected in video anchor assignments (c) as most assigned anchors are similar with minor differences in assignments, thereby showcasing the flexibility and effectiveness of our approach. (d), (e) samples look very different and therefore the anchor assignments are also very different as shown in (f). Green cell  $\\rightarrow$  Anchor assigned, Yellow  $\\rightarrow$  Anchor not assigned. Difference in anchor assignments indicated in red."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            226,
+            642,
+            279
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/51d8ddbe79ea5b9f265920a633c0e65b28d235cdf409ee39c8ee4e284455718d.jpg",
+        "image_caption": [
+            "(e) _Y51yX3KJoE (Finance and Business)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            645,
+            226,
+            766,
+            279
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/90effec5408bbe895c6bb2857e4cf0a5e5319eef736e15b533f5ccddcbf34d57.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            226,
+            890,
+            279
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6199cb8143c1a217676b9c629e752f1e33284306eea87dafd469e07455f56cd4.jpg",
+        "image_caption": [
+            "(f) Video anchors Assignments"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            84,
+            310,
+            883,
+            339
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/3809ccd243e6e9c7791b0aa9a5648baa913ca6551d8f7aec7e1c65d43add8e20.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">K</td><td rowspan=\"2\">K&#x27;</td><td colspan=\"2\">MSR-VTT</td><td colspan=\"2\">YouCook2</td></tr><tr><td>R@5↑</td><td>R@10↑</td><td>R@5↑</td><td>R@10↑</td></tr><tr><td>16</td><td>8</td><td>23.1</td><td>32</td><td>37.1</td><td>47.5</td></tr><tr><td>32</td><td>16</td><td>23.2</td><td>32.1</td><td>36.1</td><td>47.6</td></tr><tr><td>64</td><td>16</td><td>23.3</td><td>31.8</td><td>36.7</td><td>47.2</td></tr><tr><td>64</td><td>48</td><td>23.7</td><td>32.1</td><td>36.7</td><td>47.7</td></tr><tr><td>64(Ours)</td><td>32</td><td>25.1</td><td>34.5</td><td>39.4</td><td>50.1</td></tr></table>",
+        "bbox": [
+            89,
+            489,
+            457,
+            616
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Effect of different # of anchors on zero-shot retrieval.  $K \\rightarrow$  # of anchors and  $K' \\rightarrow$  # of selected anchors, respectively.",
+        "bbox": [
+            76,
+            627,
+            462,
+            657
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "signments from the same embedding space. We also apply the cross-modal SSPC loss to only isolate the impact of the proposed SSPC loss across feature projections. We observe that removing the anchor consistency between the modality-specific and joint embedding spaces reduces the zero-shot retrieval performance by a significant margin ( $\\sim 1.5\\%$  drop in R@5 performance), indicating the importance of SSPC loss in maintaining modality-specific semantic structure in the joint embedding space for better performance.",
+        "bbox": [
+            75,
+            686,
+            467,
+            823
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Finally, in the fourth row of Tab. 4, we analyse the effectiveness of the proposed Multi-Assignment Sinkhorn-Knoppalgorithm. To this end, we modify the vanilla Sinkhorn-Knopp [13] to obtain multiple anchor assignments per sample (refer to Sec. 5 of the Supplementary",
+        "bbox": [
+            75,
+            824,
+            467,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "for details). We notice that the modified Sinkhorn-Knopp does not perform well with a drop in performance ( $\\sim 1.6\\%$  drop in R@5) on both MSR-VTT and YouCook2 datasets on all the metrics indicating the significance of our proposed MULTI-SK algorithm.",
+        "bbox": [
+            496,
+            493,
+            890,
+            568
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of Number of Anchors. To analyze the effect of the number of anchors, we conduct experiments with different number of anchors,  $K$ , and different number of selected anchors,  $K'$ . We report the results in Tab. 5. We conduct experiments on both MSR-VTT and YouCook2 datasets. From the top half of Tab. 5, we observe that the performance of our proposed method improves with the increasing number of anchors. This is to be expected since a higher number of anchors have a higher representation learning capacity. We also observe that the method performs reasonably well even with a very small number of anchors showing the general effectiveness of our proposed solution. We notice that the performance of the proposed method improves as we select more anchors as shown in bottom half of Tab. 5. However, selecting a very large number of anchors (48 out of 64 for this experiment) adds more constraints leading to poor performance.",
+        "bbox": [
+            496,
+            575,
+            892,
+            833
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.5. Qualitative Analysis",
+        "text_level": 1,
+        "bbox": [
+            500,
+            845,
+            691,
+            862
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Here, first we present a fine-grained visual analysis of the learned anchors, followed by qualitative retrieval results.",
+        "bbox": [
+            500,
+            869,
+            890,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "22000",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "\"Animated comic scene of guy cutting up food for dinner\"",
+        "bbox": [
+            76,
+            104,
+            236,
+            143
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "\"a woman holding a ribbon\"",
+        "bbox": [
+            76,
+            175,
+            233,
+            200
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "\"mix ingredients refrigerate\"",
+        "bbox": [
+            76,
+            219,
+            233,
+            244
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "\"add mutton pan\"",
+        "bbox": [
+            102,
+            276,
+            210,
+            291
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/f7e7e48075b15f9cebdea991dca2cd685f88583ac000e5ef483866cdd5b9ec21.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            251,
+            87,
+            354,
+            146
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/6fef8ff10153ebdf7ac5c77cc855c27afb983fdf4313f22ded1e74fa85fa38ab.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            251,
+            146,
+            354,
+            205
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/512d1d94ddde5cf43663e4a2e81964b05b5a7cdddeecacf992d07d761e849b06.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            251,
+            205,
+            354,
+            247
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/ab8c55517bc30b4c7669d599c469af248a5b5e51ff5f07ba41c1f57877828df4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            251,
+            247,
+            354,
+            292
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/d0e9f4499376fb190e25ebf377003bdc013496e0898fcf285a4aee1788216100.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            354,
+            88,
+            455,
+            146
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/2606ee2d6bb550f4e50b4634782c64cb43f4270c6676bdfc28678fc1ae9cab97.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            354,
+            146,
+            455,
+            204
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/338f294da97650e8d905b2ca23441cd769228c343f180bf1a2f58a7b7a4e2240.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            354,
+            204,
+            455,
+            247
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/29b37b8697b4b205debcf43945958609e3896f11874ae2e6902d41eb45c53fd2.jpg",
+        "image_caption": [
+            "Ours",
+            "Figure 3: Examples of Zero-Shot Text-to-Video Retrieval on MSR-VTT and YouCook2 datasets. Each row consists of Textual Query (left), and top-3 retrieved videos for our method (center) and the state-of-the-art method EAO [42](right). Match indicates correct video for the query."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            354,
+            247,
+            455,
+            292
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/0cb0d4848fc921cc9b6e19727e77b9028c0e698a01a9a16bd45e24f76aeee2b9.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            457,
+            88,
+            558,
+            204
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/ed979fe8cc4337ba93053f95e71750f1bc7a26534d33199d5f8127cca553dc1b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            457,
+            204,
+            558,
+            292
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/a658fa31a34597a89d52f8422fe597aad42e5d996fd7433d7224243c46167342.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            580,
+            88,
+            681,
+            205
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/599738612238d178becabd032ef3bf12f1be5ee833e3f691923c2f63c7c539b4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            580,
+            205,
+            681,
+            247
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/e1af4cc6a71efce0bb02d25715f16f879b7a83012168aedf932c43e3be5d1b70.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            580,
+            247,
+            681,
+            292
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/87c0cfcde940184f1dd6aeb55faae211d18f326ea0d359136eb4099f0ebfeeea.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            683,
+            88,
+            784,
+            204
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/c867ea629bf90de482f897293f68428df213fce06ca51f8f1c8aeb0a598b9dc4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            683,
+            204,
+            784,
+            247
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/951cc2b16efba37ab8c6d6d9d3dc1b77bea082db667bd05ce5e81007f9f65907.jpg",
+        "image_caption": [
+            "EAO [42]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            683,
+            247,
+            784,
+            292
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/87ae398794fc54d31a8d14024f3f4bfe97471dba08d74bf625b58a5b64e3c8c9.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            785,
+            88,
+            885,
+            204
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/314daf3ecc2444e258048b4629fd33d4ecdd430322a99058b0db9be5655dede1.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            785,
+            204,
+            885,
+            247
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/5adc6043e40230bd6ac85dbe81b708b0840445aa16714abd37e90bff2a8abdf3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            785,
+            247,
+            885,
+            292
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We show fine-grained analysis of the learned anchors in Fig. 2. For the purpose of this analysis, we visualize the anchor assignments as binary assignments. However, during training we use soft anchor assignments. In Fig. 2(c), we compare the anchor assignments for samples from similar categories. It can be seen that the videos are visually similar even though they belong to different categories and the anchor assignments for these two examples are able to capture the sample similarity. In Fig. 2(f), we compare the anchor assignments for videos from different categories and it can be seen that the anchor assignments are very different as expected. This further validates our claim that our proposed method can assign semantically meaningful anchors without any explicit supervision. Further, we show qualitative retrieval comparison with EAO [42] in Fig. 3. We present more qualitative analysis in Supplementary Sec. 4.",
+        "bbox": [
+            75,
+            396,
+            472,
+            640
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            76,
+            660,
+            194,
+            676
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We proposed a novel approach that preserves the modality-specific semantic relationship between samples in the joint multi-modal embedding space. To this end, we propose a flexible sample relationship modeling approach by assigning multiple anchors to each sample, which captures both shared and unique aspects of samples. To obtain these assignments, we develop a novel Multi-Assignment Sinkhorn-Knopp (Multi-SK) algorithm, and also utilize the proposed anchor consistency loss to learn these anchors. Our qualitative results demonstrate that our learnt anchors correspond to meaningful semantic concepts. Our extensive experimentation demonstrates that the proposed approach improves generalizability by outperforming state-of-the-art methods on both in- and out-of-domain datasets. We also",
+        "bbox": [
+            75,
+            689,
+            470,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "show that our method achieves state-of-the-art performance on multiple zero-shot tasks, and also outperforms when fine-tuned on downstream datasets.",
+        "bbox": [
+            498,
+            396,
+            892,
+            441
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            500,
+            457,
+            666,
+            474
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Research was sponsored by the Army Research Office and was accomplished under Grant Number W911NF-19-1-0356. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Office or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. Nina Shvetsova is supported by German Federal Ministry of Education and Research (BMBF) project STCL - 01IS22067.",
+        "bbox": [
+            498,
+            482,
+            892,
+            650
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            500,
+            664,
+            596,
+            679
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Jean-Baptiste Alayrac, Adria Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. Advances in Neural Information Processing Systems, 33:25-37, 2020. 2, 6",
+            "[2] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. Advances in Neural Information Processing Systems, 33:9758-9770, 2020. 2",
+            "[3] Elad Amrani, Rami Ben-Ari, Daniel Rotman, and Alex Bronstein. Noise estimation using density estimation for self-supervised multimodal learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 6644-6652, 2021. 2"
+        ],
+        "bbox": [
+            506,
+            688,
+            893,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "22001",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[4] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In Proceedings of the IEEE International Conference on Computer Vision, pages 609-617, 2017. 2",
+            "[5] Yuki Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. Advances in Neural Information Processing Systems, 33:4660-4671, 2020. 1, 2, 3, 4",
+            "[6] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Sound-net: Learning sound representations from unlabeled video. Advances in neural information processing systems, 29, 2016. 2",
+            "[7] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 2",
+            "[8] Yann Brenier. D'écomposition polaire et r'arrangement monotone des champs de vecteurs. CR Acad. Sci. Paris Sér. I Math., 305:805-808, 1987. 3",
+            "[9] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 2, 3, 4",
+            "[10] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 5",
+            "[11] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie Boggust, Rameswar Panda, Brian Kingsbury, Rogerio Feris, David Harwath, et al. Multimodal clustering networks for self-supervised learning from unlabeled videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8012-8021, 2021. 1, 2, 5, 6, 7",
+            "[12] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 2",
+            "[13] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 3, 8",
+            "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5",
+            "[15] Alex Fang, Gabriel Ilharco, Mitchell Wortsman, Yuhao Wan, Vaishaal Shankar, Achal Dave, and Ludwig Schmidt. Data determines distributional robustness in contrastive language image pre-training (clip). In International Conference on Machine Learning. PMLR, 2022. 2",
+            "[16] Enrico Fini, Enver Sangineto, Stéphane Lathuilière, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3"
+        ],
+        "bbox": [
+            78,
+            92,
+            470,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[17] Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 303-312, 2021. 3",
+            "[18] Anca-Ioana Grapa, Laure Blanc-Feraud, Ellen van Obberghen-Schilling, and Xavier Descombes. Optimal transport vs many-to-many assignment for graph matching. In GRETSI 2019-XXVIIème Colloque francophone de traitement du signal et des images, 2019. 3",
+            "[19] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imageNet? In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6546-6555, 2018. 5",
+            "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. cvpr. 2016. arXiv preprint arXiv:1512.03385, 2016. 5",
+            "[21] Dat Huynh and Ehsan Elhamifar. A shared multi-attention framework for multi-label zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8776–8786, 2020. 2",
+            "[22] Lukasz Kaiser, Aidan N Gomez, Noam Shazeer, Ashish Vaswani, Niki Parmar, Llion Jones, and Jakob Uszkoreit. One model to learn them all. arXiv preprint arXiv:1706.05137, 2017. 2",
+            "[23] Leonid Kantorovich. On translation of mass. In Dokl. AN SSSR, volume 37, page 20, 1942. 3",
+            "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6",
+            "[25] Mingchao Li, Xiaoming Shi, Haitao Leng, Wei Zhoi, Haitao Zheng, and Kuncai Zhang. Learning semantic alignment with global modality reconstruction for video-language pretraining towards retrieval. Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 1",
+            "[26] Yanbin Liu, Linchao Zhu, Makoto Yamada, and Yi Yang. Semantic correspondence as an optimal transport problem. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4463-4472, 2020. 3",
+            "[27] Shuang Ma, Daniel McDuff, and Yale Song. Unpaired image-to-speech synthesis with multimodal information bottleneck. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 2",
+            "[28] Massimiliano Mancini, Muhammad Ferjad Naeem, Yongqin Xian, and Zeynep Akata. Open world compositional zero-shot learning. 2021. 2",
+            "[29] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 2, 7",
+            "[30] Antoine Miech, Ivan Laptev, and Josef Sivic. Learnable pooling with context gating for video classification. arXiv preprint arXiv:1706.06905, 2017. 5",
+            "[31] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "22002",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 2, 5, 6, 7",
+            "[32] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013. 5",
+            "[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 5",
+            "[34] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, Joao Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. arXiv preprint arXiv:2010.02824, 2020. 6",
+            "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2",
+            "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning. PMLR, 2021. 2",
+            "[37] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning. PMLR, 2016. 2",
+            "[38] Mamshad Nayeem Rizve, Navig Kardan, and Mubarak Shah. Towards realistic semi-supervised learning. In European Conference on Computer Vision, pages 437-455. Springer, 2022. 2, 3",
+            "[39] Andrew Rouditchenko, Angie Boggust, David Harwath, Brian Chen, Dhiraj Joshi, Samuel Thomas, Kartik Audhkhasi, Hilde Kuehne, Rameswar Panda, Rogerio Feris, et al. Avlnet: Learning audio-visual language representations from instructional videos. arXiv preprint arXiv:2006.09199, 2020. 2, 5, 6, 7",
+            "[40] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 3",
+            "[41] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 2",
+            "[42] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. 1, 2, 5, 6, 7, 9"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            897
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[43] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15638-15650, June 2022. 1",
+            "[44] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967. 3",
+            "[45] Zineng Tang, Jie Lei, and Mohit Bansal. DeCEMBERT: Learning from noisy instructional videos via dense captions and entropy minimization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Association for Computational Linguistics, 2021. 1",
+            "[46] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020. 2",
+            "[47] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 6",
+            "[48] Asano YM., Rupprecht C., and Vedaldi A. Self-labelling via simultaneous clustering and representation learning. In International Conference on Learning Representations, 2020. 2, 3",
+            "[49] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 6",
+            "[50] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634-23651, 2021. 1",
+            "[51] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 2",
+            "[52] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. 6",
+            "[53] Yipin Zhou, Zhaowen Wang, Chen Fang, Trung Bui, and Tamara L Berg. Visual to sound: Generating natural sound for videos in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 2",
+            "[54] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 6"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            852
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "22003",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2023/Preserving Modality Structure Improves Multi-Modal Learning/9786f414-b2fe-4822-86c7-2868d21a76fe_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ca36fdf7d4f3d75f15605e468d7b24b9570f6409ea9d6d75dbd6a6420c089a8
+size 3647939

2023/Preserving Modality Structure Improves Multi-Modal Learning/full.md

@@ -0,0 +1,353 @@
+# Preserving Modality Structure Improves Multi-Modal Learning
+
+Sirnam Swetha<sup>1</sup>, Mamshad Nayeem Rizve<sup>1</sup>, Nina Shvetsova<sup>2,3</sup>, Hilde Kuehne<sup>2,3,4</sup>, Mubarak Shah<sup>1</sup>
+
+$^{1}$ CRCV, University of Central Florida,  $^{2}$ Goethe University Frankfurt Germany,  $^{3}$ University of Bonn Germany,  $^{4}$ MIT-IBM Watson AI Lab
+
+# Abstract
+
+Self-supervised learning on large-scale multi-modal datasets allows learning semantically meaningful embeddings in a joint multi-modal representation space without relying on human annotations. These joint embeddings enable zero-shot cross-modal tasks like retrieval and classification. However, these methods often struggle to generalize well on out-of-domain data as they ignore the semantic structure present in modality-specific embeddings. In this context, we propose a novel Semantic-Structure-Preserving Consistency approach to improve generalizability by preserving the modality-specific relationships in the joint embedding space. To capture modality-specific semantic relationships between samples, we propose to learn multiple anchors and represent the multifaceted relationship between samples with respect to their relationship with these anchors. To assign multiple anchors to each sample, we propose a novel Multi-Assignment Sinkhorn-Knopp algorithm. Our experimentation demonstrates that our proposed approach learns semantically meaningful anchors in a self-supervised manner. Furthermore, our evaluation on MSR-VTT and YouCook2 datasets demonstrates that our proposed multi-anchor assignment based solution achieves state-of-the-art performance and generalizes to both in-and out-of-domain datasets. Code: https://github.com/Swetha5/Multi_Sinkhorn_Knopp
+
+# 1. Introduction
+
+Humans often rely on multiple sensory inputs to have a better understanding of everyday events. Most commonly, we utilize vision, audio, and language to perceive an event as they provide complementary information for robust reasoning. The closest approximation of this setup is video data as it provides both visual and audio information along with a text description as a caption. Recently, researchers have started to explore learning meaningful representations
+
+by leveraging multiple modalities to train efficient models at scale [5, 11, 42]. Such systems focus on representation learning that either improves features for each modality separately [5] or learns a joint multi-modal embedding [11, 42] space that enables various zero-shot tasks like retrieval or classification. However, given the inherent differences across the modalities, it is challenging to learn effective joint embeddings. Furthermore, the real-world data presents additional challenges like misalignment between modalities, leading to weak supervision.
+
+Current pre-training approaches in this area usually employ a contrastive objective [33] to learn the joint embeddings that pulls the cross-modal embeddings of a sample from the same temporal instance closer and pushes embeddings of other samples farther. Despite promising performances, these methods struggle with generalizability. This is particularly evident in previous approaches trained on HT100M [11, 42], which do well on the closely related downstream dataset YouCook2 but struggle to improve on the MSR-VTT dataset, which exhibits a relatively larger domain shift with respect to HT100M [42]. This is due to the contrastive objective's emphasis on strict alignment between modalities in the joint embedding space while ignoring the inherent weak alignment between different modalities [45], as well as the underlying semantic structure across samples [43, 50]. Recent works have tackled these issues, either by using joint multi-modal clustering [11] to preserve the semantic structure in the joint embedding space or by incorporating a reconstruction objective [11, 25] to retain modality-specific features in the joint embedding space, allowing for weak multi-modal alignment. However, the usual reconstruction objective trivially tries to retain most modality-specific features in the joint space, thus preventing the learning of optimal features for cross-modal tasks. And the multi-modal clustering approaches perform hard-clustering making it less flexible. Therefore, the limitations of the contrastive objective cannot be adequately addressed even after combining these independent objectives.
+
+To address this, we propose a semantic-structure
+
+preserving consistency loss (SSPC) to only retain information that is beneficial for both cross-modal embedding learning and retaining modality-specific semantic structure. In particular, for SSPC loss we consider each sample (e.g., a video clip) to be composed of multiple concepts: scene or objects involved in the downstream task. Therefore, the relationship between samples is multifaceted, representing both shared and unique concepts across samples. To capture this multifaceted relationship in a flexible manner, we propose to learn anchors (latent codes) and model the relationship between samples with respect to their relationships with these anchors. Therefore, these anchors act as a proxy to represent the modality-specific relationships between samples (semantic structure) which can be preserved using the proposed SSPC loss. Since we have no supervision to learn these anchors, we formulate this anchor learning problem as a many-to-many assignment problem, as modeling this multifaceted relationship simultaneously involves assigning multiple anchors to one sample and multiple samples to one anchor. Although there is a vast literature on solving the many-to-one assignment problem [9, 5, 38, 48], there is no efficient way to solve this many-to-many assignment problem.
+
+To this end, we propose a novel Multi-Assignment Sinkhorn-Knopp (Multi-SK) algorithm that iteratively optimizes the many-to-many anchor assignments for both the modality-specific embeddings (in input space) and modality-agnostic multi-modal embeddings (in joint embedding space). To allow for weak alignment between modalities, we select the dominant anchors for each sample to represent the relationship between different samples. Our proposed SSPC loss enforces consistency between the dominant anchor assignments at the input and joint embedding spaces to preserve the modality-specific semantic structure. To demonstrate the effectiveness of our proposed solution, we train our model on HT100M dataset and test on 6 zero-shot tasks on multiple downstream datasets and observe that our approach leads to state-of-the-art results in all settings.
+
+In summary, we make the following contributions: (i) We propose a flexible modality-specific semantic-structure-preserving approach to improve the generalizability of cross-modal features. (ii) We introduce Multi-Assignment Sinkhorn-Knopp, a novel algorithm to enable multiple assignments for flexible sample relationship modeling. (iii) Our proposed method outperforms the current state-of-the-art for multi-modal self-supervised representation learning on both in- and out-of domain datasets.
+
+# 2. Related Work
+
+# 2.1. Multi-Modal Learning
+
+With the availability of large-scale multi-modal datasets [31, 41, 7], multi-modal learning research has received a lot of attention. It comprises of vision-language
+
+learning [35, 51], vision-audio learning [2, 4, 6, 12, 46], video-audio-language learning [39, 11, 42], zero-shot learning [21, 28], cross-modal generation [37, 53, 27] and multi-modal multi-task learning [22]. Miech et al. [31] proposed a large-scale multi-modal dataset consisting of video, audio and text by collecting instructional videos from YouTube without requiring any human annotations. Note that the text is generated from audio using Automatic Speech Recognition (ASR) and has noisy alignment between the text and video. They also proposed a multimodal system to demonstrate the potential for learning video-text embedding via contrastive loss. To handle the noise in the dataset, Amrani et al. [3] proposed a noise estimation for multi-modal data via multi-modal density estimation. A noise-contrastive estimation approach in a multi-instance learning framework was proposed by Miech et al. [29]. XDC [2] performs clustering on audio-video for learning better features for each modality separately. These works utilize only two modalities for multi-modal learning, while others have explored utilizing audio, video, and text together for multi-modal learning. Multi-Modal versatile networks [1] was proposed to learn different embedding spaces for each combination of modalities. AVLNet [39] proposed to learn a shared embedding that maps all modalities to a single joint embedding space. Following this, MCN [11] proposed to perform joint clustering and reconstruction to learn joint embedding space. Note that, [11] performs multi-modal K-means clustering to learn hard semantic clusters. Unlike strict assignment in [11], we propose flexible learning with multiple assignments and separately for each modality. More recently, EAO [42] utilizes transformers and combinatorial fusion of modalities to learn the joint embedding with contrastive loss.
+
+Most of these works, utilize contrastive or clustering loss over fused multi-modal representation to learn the joint embedding space. By doing so, these models do not retain the modality-specific semantic structuring between samples encoded by the pre-trained modality specific backbones, hurting the generalization ability of the model. Additionally, recent works have reported that large-scale contrastive multimodal models (e.g., CLIP [36]) are somewhat robust to distributional shifts mainly due to diverse large-scale training data and prompt-engineering [15]. Therefore, our work focuses on making the pre-training objective robust to distributional shifts. In this context, we propose a novel approach to preserve the modality-specific semantic relationships in the joint embedding space by modeling the relationship between samples w.r.t learnable anchors. To enable flexible relationship modeling between samples, we learn multiple anchor assignments per sample, where anchors shared across samples model the commonality between them, and the distinct anchors between samples highlight the uniqueness of the samples.
+
+![](images/36b1353f0d482ab61ec574a70a19b3d49c9e6d43f912d0b1a2da048bb98354b6.jpg)  
+Figure 1: Overview of the proposed model. Given weakly aligned text, video, and audio, we first extract features using frozen modality-specific backbones. These features are then passed through a token projection layer to obtain input features  $(\mathbf{v},\mathbf{t},\mathbf{a})$  specific to each modality. Next, the modality-specific transformer models project the input features into a joint multi-modal representation space  $(\hat{\mathbf{v}},\hat{\mathbf{t}},\hat{\mathbf{a}})$ . The similarity between the input features  $(\mathbf{v},\mathbf{t},\mathbf{a})$  and input space anchors  $(\mathbf{Z}_{\mathbf{v}},\mathbf{Z}_{\mathbf{t}},\mathbf{Z}_{\mathbf{a}})$  (projected features  $(\hat{\mathbf{v}},\hat{\mathbf{t}},\hat{\mathbf{a}})$  and projected space anchors  $(\hat{\mathbf{Z}}_{\mathbf{v}},\hat{\mathbf{Z}}_{\mathbf{t}},\hat{\mathbf{Z}}_{\mathbf{a}}))$  is computed, and our Multi-SK algorithm  $(M(.))$  is used to optimize multiple anchor assignments per sample, as shown in the Input Space Assignments (Projected Space Assignments). We do this for each modality and enforce the respective consistency losses, but in this figure we only show modality-specific consistency loss for text anchors and cross-modal consistency between text and video modalities for brevity. LN, MHSA represents LayerNorm and Multi-Head Self-Attention.
+
+# 2.2. Sinkhorn-Knopp
+
+Recently, Sinkhorn-Knopp algorithm [44] has drawn huge attention because of its effectiveness in solving optimal-transport problems [23, 8]. Specifically, [13] proposed an entropic relaxation of the optimal transport problem which can be efficiently solved using Sinkhorn's matrix scaling algorithm. Many following works have since successfully utilized the Sinkhorn-Knopp algorithm to solve different label assignment problems framed as an optimal transport problems. For instance, SeLa [48] cast the unsupervised clustering problem as a pseudo-label assignment problem and used the Sinkhorn-Knopp algorithm to solve it. SeLaVi [5] extended this idea to self-supervised representation learning for multi-modal data where the cluster assignments between different modalities are swapped to encourage modality invariant representation learning. Similarly, SwAV [9] used the Sinkhorn-Knopp algorithm for self-supervised representation learning and proposed to swap the pseudo-labels for differently augmented versions of a sample and use soft assignments instead of hard pseudolabels. In contrast to these works, SuperGlue [40] used the Sinkhorn-Knopp algorithm to solve the correspondence
+
+problem between two sets of local features. Moreover, Sinkhorn-Knopp has been used in detection problems [17], where it was used to match the anchors with ground truths. Recently, UNO [16], and TRSSL [38] have successfully used the Sinkhorn-Knopp algorithm in solving novel class discovery and open-world semi-supervised learning problems, respectively. One key limitation of the traditional Sinkhorn-Knopp algorithm is that it cannot be directly utilized to compute multiple assignments necessary to perform multi-anchor based learning i.e. many-to-many assignments.
+
+Some prior works [18, 26] have attempted to solve many-to-many assignments in indirect ways. While authors of [18] use an intermediate graph to match groups of vertices from source to target graph, which can only perform group-to-group assignments and is inadequate for our problem that requires true many-to-many matching. [26] modifies the Sinkhorn-Knopp row and column constraints to obtain many-to-many assignment to model dense correspondences, however, we have found this approach to be inferior in solving the multiple anchor assignment problem. The modified Sinkhorn-Knopp constraints approach yields suboptimal results, as discussed in Sec. 4.4. To address these
+
+limitations, we propose a novel algorithm Multi-SK, which outperforms the modified Sinkhorn-Knopp constraints approach to get true many-to-many assignments.
+
+# 3. Method
+
+Given a set of multi-modal inputs  $\{\mathbf{t}^{(i)},\mathbf{v}^{(i)},\mathbf{a}^{(i)}\}_{i = 1}^{N}$  from N video clips, we learn modality-specific projection functions  $f_{t},f_{v},f_{a}$ , that transform t,v,a into a d-dimensional joint embedding space,  $\mathbb{R}^d$  to obtain  $\hat{\mathbf{t}},\hat{\mathbf{v}},\hat{\mathbf{a}}$  respectively. Our goal is to optimize the parameters of  $f_{t},f_{v},f_{a}$  in such a way that they maintain the semantic structure amongst samples from a particular modality in the joint embedding space, as discussed in Sec 3.1, and simultaneously brings the semantically related cross-modal inputs closer. In the following, first, we formulate our approach to modeling the relationship between samples using anchors in Sec 3.1, then in Sec 3.2, we discuss our novel Multi-SK algorithm to learn these anchors for representing sample relationships, and finally, we present the overall training objective to train the model in Sec 3.3.
+
+# 3.1. Modeling Sample Relationships with Anchors
+
+Our work aims to preserve the relationship between samples i.e.  $\mathbf{x}^{(i)},\mathbf{x}^{(j)}$  from a particular modality, to better generalize on the unseen data. To this end, we propose to model the relationship between samples using anchors, where the similarity of each sample w.r.t. anchors encodes the semantic structure of the feature space. Unlike clustering approaches, which involve hard-assignment to a specific cluster, our approach offers flexibility i.e., there can be shared and unique anchors across samples that define the relationship between them. To be particular, for each sample, we learn  $K$  anchors, and the similarity of assignments over these anchors represents the relationship (which we want to preserve) between samples from a particular modality.
+
+As the pre-trained features extracted from modality-specific backbones encode the semantic structure between samples within that particular modality, preserving such modality-specific relationships in the joint embedding space would boost the generalizability of the model. To achieve this, we define two sets of  $K$  learnable anchors  $\mathbf{z} = \{\mathbf{z}^{(i)}\}_{i=1}^{K}$ , and  $\hat{\mathbf{z}} = \{\hat{\mathbf{z}}^{(i)}\}_{i=1}^{K}$  to model the sample relationships before and after projecting them to the joint embedding space respectively. We repeat this for all the modalities,  $t, v, a$ . We propose to preserve the modality-specific semantic structure at the joint embedding space by enforcing consistency in anchor assignments before and after performing feature projections (Eq. 3).
+
+We still have one remaining challenge, i.e. how to learn/discover these anchors in an unsupervised manner. To this end, we cast the anchor discovery as a label assignment task with a uniform prior, i.e. each anchor will have an equal number of sample assignments. Additionally, to
+
+encourage flexible modeling, we enforce multiple anchor assignments per sample. However, it is difficult to estimate the exact number of anchors for each sample without additional prior information, thus we select the top  $K'$  anchors for each sample to effectively model the sample relationships. Even though this optimization task may seem like a difficult combinatorial problem, we model this combinatorial optimization task as an optimal transport problem. Following recent works [5, 9], one might assume the Sinkhorn-Knopp algorithm is a natural choice to solve this problem, however, the vanilla Sinkhorn-Knopp algorithm cannot handle multiple assignments per sample i.e. many-to-many assignment. To address this limitation, we propose Multi-Assignment Sinkhorn-Knopp algorithm, presented in the following.
+
+# 3.2. Multi-Assignment Sinkhorn-Knopp
+
+Given a sample matrix  $\mathbf{B}$  s.t.  $\mathbf{B} \in \mathbb{R}^{N \times d}$ , representing  $N$  samples and an anchor matrix  $\mathbf{Z}$  s.t.  $\mathbf{Z} \in \mathbb{R}^{K \times d}$ , of  $K$  anchor vectors, we obtain the similarity matrix  $\mathbf{S}$  s.t.  $\mathbf{S} = \mathbf{B}\mathbf{Z}^{\intercal}$  and  $\mathbf{S} \in \mathbb{R}^{N \times K}$ , where,  $\mathbf{S}_{ij}$  is the probability of assigning  $j$ th anchor to the  $i$ th sample vector. The goal is to find an anchor assignment matrix,  $\mathbf{Q}$ , such that it satisfies the following constraints: (i) a sample should be assigned to exactly  $K'$  anchors, to learn top  $K'$  anchor assignments per sample. (ii) anchor assignments must be equally partitioned i.e each anchor must be selected exactly  $N \times K' / K$  times, for uniform anchor assignment.
+
+To obtain such multiple anchor assignments per sample, we propose to create a substitute 3D assignment matrix  $\mathbf{Q}'$  s.t.  $\mathbf{Q}' \in \mathbb{R}^{K \times N \times K}$ . We also generate a 3D similarity matrix  $\mathbf{S}'$  from the 2D similarity matrix  $\mathbf{S}$  by introducing  $K$  channels (depth dimension) to get  $K \times N \times K$  matrix such that each channel is a scaled matrix of  $S$  with a predefined ranking between the channels enabling top  $K'$  anchor selection. Since we are only interested in selecting the top  $K'$  anchors, we set the first  $K'$  channels of  $\mathbf{S}'$  to be the same as  $\mathbf{S}$ . The remaining  $K - K'$  channels are set to  $\mu \mathbf{S}$ . Here,  $\mu$  is a damping factor s.t.  $0 < \mu < 1$ , to help select the top  $K'$  anchors for each sample. We discuss alternate designs for  $\mathbf{S}'$  generation in the Supplementary Sec. 2.
+
+The optimization objective of Multi-Assignment Sinkhorn-Knopp is to find an assignment matrix  $\mathbf{Q}'$  such that it satisfies our multi-anchor assignment constraints while maximizing similarity with the initial assignment/similarity matrix,  $\mathbf{S}'$ . This optimization problem is defined as:
+
+$$
+\begin{array}{l} \mathbf {Q} ^ {*}: \max  _ {\mathbf {Q} ^ {\prime}} <   \mathbf {Q} ^ {\prime}, \mathbf {S} ^ {\prime} > + \epsilon H (\mathbf {Q} ^ {\prime}) \\ H \left(\mathbf {Q} ^ {*}\right) = - \sum \mathbf {Q} _ {i j k} ^ {\prime} \log \mathbf {Q} _ {i j k} ^ {\prime} \tag {1} \\ \end{array}
+$$
+
+$\mathbf{Q}^*$  needs to satisfy the following constraints to be a valid solution for our multi-anchor assignment problem.
+
+-  $\mathbf{Q}^*$  Row Constraint: Within a channel, the sum of all el
+
+ements in a particular row must be equal to one. This is because we only want one anchor assignment for a sample in a particular channel.  $\forall i,k\sum_{j}\mathbf{Q}^{*}_{ijk} = 1$
+
+-  $\mathbf{Q}^*$  Column Constraint: In a channel, the sum of all elements in a column should be equal to  $N / K$ . This constraint enforces equal partitioning of anchor assignments.  $\forall j, k \sum_{i} \mathbf{Q}^{*}_{ijk} = N / K$ .  
+-  $\mathbf{Q}^*$  Depth constraint: Depth-wise sum should be equal to one for every sample and anchor combination. This constraint prevents selecting the same anchor across different channels.  $\forall i, j \sum_{k} \mathbf{Q}^{*}_{ijk} = 1$ .
+
+The traditional Sinkhorn-Knopp method uses an iterative matrix scaling algorithm that scales the rows and columns alternatively (with desired constraints) till the desired assignment matrix is obtained. We employ a similar scheme and extend the iterative scaling to the depth dimension for estimating  $\mathbf{Q}^*$ . We iteratively scale the rows, columns, and channels (depth dimension) till convergence. The final 2D assignment matrix  $\mathbf{Q}$  is computed by performing a depth-wise sum on the top  $K^{\prime}$  channels. We provide Pytorch-style pseudo-code for our Multi-Assignment Sinkhorn-Knop algorithm in the Supplementary Sec 3.
+
+# 3.3. Training Objective
+
+Semantic Structure Preserving Consistency Loss. To preserve the semantic structure of each modality,  $t, v, a$ , we apply consistency loss to enforce similar anchor assignments between input and joint embedding space. As the cross-modal contrastive loss in the joint embedding space tries to bring different modalities together, features in the joint embedding space from a particular modality should preserve the common anchors that exist in corresponding features from the other modalities. Therefore, we also apply cross-modal anchor consistency across all modalities as shown in Eq. 3. Since we are dealing with 3 input modalities, this cross-modal consistency results in 9 consistency constraints.
+
+Let's denote  $\mathcal{L}(\mathbf{t},\hat{\mathbf{v}},\mathbf{z}_t,\hat{\mathbf{z}}_t)$  as the consistency loss between text anchor assignment in the input space and the textual anchor assignment of the corresponding video features at the joint embedding space as shown below in Eq. 2:
+
+$$
+\begin{array}{l} \mathcal {L} (\mathbf {t}, \hat {\mathbf {v}}, \mathbf {z} _ {t}, \hat {\mathbf {z}} _ {t}) = \alpha_ {\mathbf {t}, \hat {\mathbf {v}}} g \left(\operatorname {s i m} \left(\mathbf {t}, \mathbf {z} _ {t}\right), M \left(\operatorname {s i m} \left(\hat {\mathbf {v}}, \hat {\mathbf {z}} _ {t}\right)\right)\right) (2) \\ + \beta_ {\mathbf {t}, \hat {\mathbf {v}}} g (\operatorname {s i m} (\hat {\mathbf {v}}, \hat {\mathbf {z}} _ {t})), M (\operatorname {s i m} (\mathbf {t}, \mathbf {z} _ {t}))), (2) \\ \end{array}
+$$
+
+Here,  $\mathbf{z}_t, \hat{\mathbf{z}}_t$  respectively represent input and output learnable anchor vectors for the text modality.  $g(.)$ , and  $M(.)$  respectively represent binary cross-entropy-with-logits loss and Multi-Assignment Sinkhorn-Knopp (discussed in Sec. 3.2),  $\alpha$  and  $\beta$  represents loss coefficients, and  $\mathrm{sim}(\mathbf{a},\mathbf{b}) = \exp (\mathbf{a.b} / \tau \| \mathbf{a}\| \| \mathbf{b}\|)$ ,  $\tau$  is the temperature hyperparameter of the similarity metric.
+
+Overall semantic structure preserving consistency loss for all the modalities is defined as:
+
+$$
+\mathcal {L} _ {s s p c} = \sum_ {m \in \{t, v, a \}} \sum_ {n \in \{t, v, a \}} \mathcal {L} (\mathbf {m}, \hat {\mathbf {n}}, \mathbf {z} _ {m}, \hat {\mathbf {z}} _ {m}). \tag {3}
+$$
+
+Contrastive Loss. Following [11, 42], we also use contrastive loss to bring cross-modal embeddings of the same sample closer while pushing away embeddings from other sample. For this, we use 3 pairwise single-modality constrastive losses,  $\mathcal{L}_{nce\_ tv}$ ,  $\mathcal{L}_{nce\_ ta}$ ,  $\mathcal{L}_{nce\_ va}$  between  $(t,v)$ ,  $(t,a)$ ,  $(v,a)$  respectively. Specifically, we use Noise Contrastive Estimation [33] with temperature  $\kappa$  as shown in Eq. 4.
+
+$$
+\mathcal {L} _ {n c e - x y} = - \log \frac {\exp \left(\mathbf {x} ^ {\intercal} \mathbf {y} / \kappa\right)}{\sum_ {i = 1} ^ {N} \exp \left(\mathbf {x} ^ {(i) \intercal} \mathbf {y} ^ {(i)} / \kappa\right)} \tag {4}
+$$
+
+The overall contrastive loss for all modalities is defined as  $\mathcal{L}_{nce} = \lambda_{tv}\mathcal{L}_{nce\_tv} + \lambda_{ta}\mathcal{L}_{nce\_ta} + \lambda_{va}\mathcal{L}_{nce\_va}$
+
+# Overall Loss
+
+The overall training objective is a combination of SSPC loss (Eq. 3) and contrastive loss (4):  $\mathcal{L}_f = \lambda_{sspc}\mathcal{L}_{sspc} + \lambda_{nce}\mathcal{L}_{nce}$ , where,  $\lambda_{sspc}$  and  $\lambda_{nce}$  are loss coefficients. By combining both losses, the model learns a more generic joint embedding space which preserves the modality-specific semantic structure by enforcing the anchor assignment similarity before and after feature projection and also brings representations of different modalities together by utilizing contrastive loss.
+
+# 4. Experiments
+
+# 4.1. Experimental Setup
+
+Backbones. For comparability, we follow the same setup of previous works [31, 39, 11, 42]. As visual backbone, we use a combination of 2D features from ResNet-152 [20] pretrained onImagenet [14], and 3D features from ResNeXt101 [19] pretrained on Kinetics [10]. The text backbone is GoogleNews pretrained Word2vec model [32]. These backbones are fixed and not finetuned during training. Following [11, 42], we use a trainable CNN with residual layers as an audio backbone. We provide additional details in the Supplementary Sec. 5.
+
+Data Sampling. We use a batch of 216 videos and randomly sample ten 8-second clips per video. If the sampled clip contains narration (95% clips), we use ASR time stamps to select clip borders. To disentangle high text-audio correlation in HT100M, we shift the audio clip randomly by 4 seconds with respect to the video and text boundaries.
+
+Projections. Following [42, 11, 39], we use a gated linear projection [30] to project features into common token space, as well as to project resulting tokens into shared embedding space. We set the dimension of the common token space to 4096 and of the shared embedding space to 6144.
+
+<table><tr><td rowspan="2">Method</td><td>Retrieval</td><td>Train</td><td>Visual</td><td colspan="3">Trainable BB</td><td colspan="4">MSR-VTT</td><td colspan="3">YouCook2</td><td></td></tr><tr><td></td><td>Dataset</td><td>BB</td><td>t</td><td>v</td><td>a</td><td>R@5↑</td><td>R@10↑</td><td>MedR↓</td><td>MeanR↓</td><td>R@5↑</td><td>R@10↑</td><td>MedR↓</td><td>MeanR↓</td></tr><tr><td>ActBERT [54]</td><td>t → v</td><td>HT100M</td><td>Res3D+Faster R-CNN</td><td></td><td></td><td></td><td>23.4</td><td>33.1</td><td>36</td><td>-</td><td>26.7</td><td>38.0</td><td>19</td><td>-</td></tr><tr><td>SupportSet [34]</td><td>t → v</td><td>HT100M</td><td>R152 + R(2+1)D-34</td><td>✓</td><td></td><td></td><td>23.0</td><td>31.1</td><td>31</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>HT100M [31]</td><td>t → v</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td></td><td>21.2</td><td>29.6</td><td>38</td><td>-</td><td>17.3</td><td>24.8</td><td>46</td><td>-</td></tr><tr><td>AVLNet [39]</td><td>t → v</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>24.7</td><td>34.2</td><td>-</td><td>-</td><td>21.1</td><td>29.6</td><td>-</td><td>-</td></tr><tr><td>EAO [42]</td><td>t → v</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>24.6</td><td>35.3</td><td>25</td><td>90.4</td><td>27.9</td><td>38.9</td><td>19</td><td>119.6</td></tr><tr><td>Ours</td><td>t → v</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>26.4</td><td>35.1</td><td>23</td><td>92.2</td><td>29.4</td><td>40.7</td><td>18</td><td>111.8</td></tr><tr><td>AVLNet [39]</td><td>v → t</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>27.2</td><td>35.7</td><td>25</td><td>86.5</td><td>22.8</td><td>32.9</td><td>30</td><td>142.2</td></tr><tr><td>EAO [42]</td><td>v → t</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>27.6</td><td>36.6</td><td>25</td><td>85</td><td>31.8</td><td>70.5</td><td>15</td><td>91.9</td></tr><tr><td>Ours</td><td>v → t</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>27.2</td><td>37.1</td><td>23</td><td>84.5</td><td>32</td><td>72</td><td>15</td><td>85.2</td></tr><tr><td>AVLNet [39]</td><td>t → v + a</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>19.2</td><td>27.4</td><td>47</td><td>-</td><td>36.1</td><td>44.3</td><td>16</td><td>-</td></tr><tr><td>MCN [11]</td><td>t → v + a</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>25.2</td><td>33.8</td><td>-</td><td>-</td><td>35.5</td><td>45.2</td><td>-</td><td>-</td></tr><tr><td>EAO [42]</td><td>t → v + a</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>23.3</td><td>33.2</td><td>29</td><td>94.8</td><td>38.5</td><td>49.2</td><td>11</td><td>82.7</td></tr><tr><td>Ours</td><td>t → v + a</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>25.1</td><td>34.5</td><td>26</td><td>91.8</td><td>39.4</td><td>50.1</td><td>10</td><td>83.3</td></tr><tr><td>AVLNet [39]</td><td>v + a → t</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>19</td><td>26.3</td><td>44</td><td>128.1</td><td>48.8</td><td>58.4</td><td>6</td><td>67.1</td></tr><tr><td>EAO [42]</td><td>v + a → t</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>21.8</td><td>31.4</td><td>28.5</td><td>98.9</td><td>49</td><td>60.9</td><td>6</td><td>43.8</td></tr><tr><td>Ours</td><td>v + a → t</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>24</td><td>32</td><td>27</td><td>95.9</td><td>48.8</td><td>61.3</td><td>6</td><td>43.5</td></tr></table>
+
+Table 1: Zero-shot Retrieval results on MSR-VTT/YouCook2. For fair comparison, we compare with models trained on text, video and audio. Retrieval column represents the evaluation task. BB=Backbone. Bold, underline represent highest and second-highest scores.
+
+We use a single transformer block with hidden size of 4096 with 64 heads and an MLP size of 4096. We set the number of anchors  $K$  as 64 and  $K'$  as 32 with damping factor  $\mu$  as 0.25. We train all models for 15 epochs using an Adam optimizer [24] with a learning rate of 5e-5, exponential decay of 0.9 and the temperature of cosine similarity  $(\tau)$  as 0.1. We maintain a memory-bank of size 5500 while performing Multi-SK. Following [1, 42], we set higher weight for loss terms that involve text-video in Eq. 3, i.e., for all text-video terms the weight is set to 1.0 and the rest of the weights are set to 0.1.
+
+# 4.2. Datasets, Tasks, Metrics
+
+Pretraining Dataset. We train our model on the HT100M dataset [31], which contains over 1 million instructional videos with automatically generated text narrations. The text narrations can be assumed to be noisy and to not always describe the video scene.
+
+Zero-shot Retrieval. We use MSR-VTT [47] and YouCook2 [52] datasets to evaluate the zero-shot retrieval capability of our model. We report performance on 4 retrieval tasks: (i) Text-to-Video retrieval, (ii) Video-to-Text retrieval, (iii) Text-to-Video-Audio retrieval, (iv) Video-Audio-to-Text retrieval. The YouCook2 dataset contains cooking videos from YouTube with human-annotated clips ( $\sim$  2 - 200 secs). For evaluation we use at maximum first 48 seconds of clip, since most clips are shorter than that. The MSR-VTT dataset contains human-annotated clips ( $\sim$  10 - 30 secs) on various topics and provides captions with natural language sentences. Following [31, 39, 11, 42], to evaluate our model on MSR-VTT, we use the 1k set of test clips [49], and for YouCook2, we use 3,350 validation clips [31]. To perform  $(t\rightarrow v + a)$  retrieval, we compute similarities by dot product between a text query  $t$  and all videos in the dataset using a averaged  $v + a$  representa
+
+tion for each video. We report standard recall metrics  $\mathbf{R}@\mathbf{5}$ ,  $\mathbf{R}@\mathbf{10}$ , median rank (MedR) and the mean rank(MeanR). Further, we also evaluate our model using CLIP backbone and report results in Sec. 1 of the Supplementary.
+
+Zero-Shot Full-Video Retrieval. Following [11], we evaluate Zero-Shot Full Video Retrieval from a set of captions on YouCook2 dataset. We report recall metrics following Caption averaging method [11] that finds maximal prediction over all the clips of video for each caption and averaging over set of captions in query leading to a single prediction for full video.
+
+Text-to-video Retrieval after Fine-tuning. We additionally evaluate the retrieval performance of the models fine-tuned on downstream tasks. We use 6783 clips from MSR-VTT (which contain audio) and 9586 clips from YouCook2 train datasets to fine-tune the model as proposed by [39].
+
+We also evaluate our model for Zero-shot Classification and report additional results in Sec. 1 of the Supplementary.
+
+# 4.3. Comparison with State-of-the-art
+
+Zero-shot Retrieval Tasks. In Tab. 1, we report the performance of the learned multi-modal representations on four zero-shot retrieval tasks, (i) Text-to-Video, (ii) Video-to-Text, (iii) Text-to-Video-Audio, (iv) Video-Audio-to-Text., on MSR-VTT and YouCook2 datasets. For a fair comparison, we only compare with models trained on all three modalities i.e. text, video and audio. In summary, our proposed method outperforms the current state-of-the-art methods by a noticeable margin on both the datasets. The results on the MSR-VTT dataset are particularly interesting since it demonstrates the generalizability of our model. In particular, HT100M consists of instructional videos and the textual descriptions are generated from audio using ASR. The text narration has noisy alignment with video and typically describes the steps in the video. YouCook2 shares
+
+<table><tr><td rowspan="2">Method</td><td>Retrieval</td><td>Train</td><td>Visual</td><td colspan="2">Trainable BB</td><td colspan="4">MSR-VTT</td><td colspan="4">YouCook2</td><td></td></tr><tr><td></td><td>Dataset</td><td>BB</td><td>t</td><td>v</td><td>a</td><td>R@5↑</td><td>R@10↑</td><td>MedR↓</td><td>MeanR↓</td><td>R@5↑</td><td>R@10↑</td><td>MedR↓</td><td>MeanR↓</td></tr><tr><td>AVLNet [39]</td><td>t → v</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>42.2</td><td>56.2</td><td>7</td><td>35.1</td><td>23.7</td><td>32.7</td><td>28</td><td>122.6</td></tr><tr><td>EAO [42]</td><td>t → v</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>47.7</td><td>59.3</td><td>6</td><td>35.6</td><td>33.9</td><td>45.8</td><td>13</td><td>70.7</td></tr><tr><td>Ours</td><td>t → v</td><td>HT100M</td><td>R152 + RX101</td><td></td><td></td><td>✓</td><td>48.7</td><td>60.6</td><td>5</td><td>33.1</td><td>35.6</td><td>48.1</td><td>12</td><td>64.9</td></tr></table>
+
+Table 2: Text-to-video retrieval results on MSR-VTT/YouCook2 in the fine-tune setting. For fair comparison, we compare with models trained on all 3 modalities i.e. text, video and audio. Bold, underline represent best and second-best scores.  
+
+<table><tr><td>Method</td><td>Aggregation</td><td>R@1↑</td><td>R@5↑</td><td>R@10↑</td></tr><tr><td>Random</td><td>-</td><td>0.23</td><td>1.15</td><td>2.32</td></tr><tr><td>HT100M [31]</td><td>Caption Avg.</td><td>43.1</td><td>68.6</td><td>79.1</td></tr><tr><td>MIL-NCE [29]</td><td>Caption Avg.</td><td>46.6</td><td>74.3</td><td>83.7</td></tr><tr><td>MCN [11]</td><td>Caption Avg.</td><td>53.4</td><td>75.0</td><td>81.4</td></tr><tr><td>EAO [42]</td><td>Caption Avg.</td><td>62.9</td><td>80.5</td><td>86.7</td></tr><tr><td>Ours</td><td>Caption Avg.</td><td>65.1</td><td>83.2</td><td>87.6</td></tr></table>
+
+domain similarity with HT100M as its videos are instructional format, and the text descriptions correspond to specific steps in the recipe. In contrast, MSR-VTT is not restricted to instructional videos and the text is typically a single sentence caption describing the whole video. As a result, there is a distributional shift between HT100M and MSR-VTT. Therefore, zero-shot retrieval on MSR-VTT has to overcome distributional shift, which is a crucial requirement for practical deployment. Our proposed method shows relatively higher improvement on this task validating the effectiveness of anchor-based learning.
+
+In case of text-to-video-audio retrieval, our method improves the Median and Mean rank of the baseline [42] by  $3\%$  on MSR-VTT along with gains on recall metrics  $R@5$  and  $R@10$ . For video-audio-to-text retrieval, our methods perform very well on MSR-VTT with  $3\%$  improvement on MeanR,  $2.2\%$  improvement on  $R@5$ . We also observe similar improvements on the YouCook2 dataset. Similarly, our proposed method outperforms the current state-of-the-art on most of the metrics for text-video-retrieval video-text-retrieval. Additionally, to compare our approach with text-video only model, we train our model on video and text to compare the performance with state-of-the-art in Sec 1 of the Supplementary.
+
+Zero-shot Full Video Retrieval. In Tab. 3, we report results for text-to-full-video retrieval task. Our approach outperforms prior works by  $2.2\%$ ,  $2.7\%$  on R@1 and R@5 respectively.
+
+Retrieval after Fine-tuning. We also evaluate the retrieval performance of our model after fine-tuning on the downstream datasets as shown in Tab. 2. For a fair comparison, we only report the baselines that use the same training split. We outperform state-of-the-art consistently on all the metrics and on both MSR-VTT & YouCook2 datasets
+
+as shown in Tab. 2.
+
+# 4.4. Ablation Studies
+
+First we analyze the impact of the proposed components and report results in Tab. 4, followed by the effect of number of anchors  $(K)$  &  $(K^{\prime})$  in Tab. 5.
+
+Table 3: Zero-Shot Text-to-Full Video retrieval on YouCook2. Averaging across captions is used to obtain video-level predictions.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">MSR-VTT</td><td colspan="2">YouCook2</td></tr><tr><td>R@5↑</td><td>R@10↑</td><td>R@5↑</td><td>R@10↑</td></tr><tr><td>Recon. + CL</td><td>23.1</td><td>32.4</td><td>37.8</td><td>48.7</td></tr><tr><td>w/o CM SSPCL</td><td>22.7</td><td>31.8</td><td>36.9</td><td>48.3</td></tr><tr><td>w/o SSPCL</td><td>23.8</td><td>33.3</td><td>37.9</td><td>48.8</td></tr><tr><td>Modified SK</td><td>23.4</td><td>31.3</td><td>37.9</td><td>48.3</td></tr><tr><td>Ours</td><td>25.1</td><td>34.5</td><td>39.4</td><td>50.1</td></tr></table>
+
+Table 4: Ablation studies showing the impact of various components for zero-shot retrieval task. Recon.=Reconstruction Loss, CM SSPCL=Cross-Modal SSPC Loss, SK=Sinkhorn-Knopp, CL=Contrastive Loss.
+
+Effect of Proposed Components. We report the results for this ablation in Tab. 4. In first row, we report results using reconstruction loss and contrastive loss, we notice that using reconstruction loss reduces the performance by  $2\%$  on all metrics indicating the effectiveness of the proposed SSPC loss. In the second row, we report results without our proposed cross-modal SSPC loss ('w/o CM SSPC'). We employ this loss to bring the cross-modal representations closer in the joint embedding space to obtain better performance in zero-shot cross-modal tasks. We notice that removing the cross-modal SSPC loss drastically decreases the zero-shot retrieval performance on both MSR-VTT and YouCook2 datasets,  $2.4\%$  and  $2.5\%$  drop in R@5 performance respectively. This empirically validates the effectiveness of the cross-modal SSPC loss in obtaining better cross-modal representations.
+
+In the third row, we analyze the effect of the proposed SSPC loss ('w/o SSPCL'). We apply the SSPC loss to retain modality-specific semantic structure from the pretrained models in the joint embedding space. To investigate its impact, we remove the anchor consistency between the modality-specific and joint embedding spaces. Instead, we enforce the anchor assignments before Multi-SK to be consistent with the Multi-SK optimized anchor as
+
+![](images/b2cc957fe4de615732ae3dfbc2a05de5dbf12c811084fd877125e2262a3ada0f.jpg)  
+(a) UMeazzCYlps (Holidays and Traditions)
+
+![](images/8263a7ab1f21a9d5ba54fa0df65dff181159df1cd90f2150283e72300b5f1d53.jpg)
+
+![](images/61fe488eb9ef28b0a46f636b11a59fdb546ec4494b06ce9f624e88b67cc4cfe5.jpg)
+
+![](images/2deac40873fcf07f5e27e410e2c90f9c34d5c2d674c8dbd2d806a878162499c0.jpg)  
+(b) fdUgKk-G5Tw (Food and Entertaining)
+
+![](images/96296052ffce8bfcf13cced27a5f7ebc8c2ffa957ebf7b27539e93f161741a84.jpg)
+
+![](images/6a69f6250264733e7faaad8f4cccfab0da2a185e5af60020327526c5625df11b.jpg)
+
+![](images/f9c1f037b7e0082a484beb440766b9e3808933f28d379e5e1d01b1bd205ba1be.jpg)
+
+![](images/57f91f9de5423015637a8942a535d6095aeb60cc0d73e585210e8e50ea094708.jpg)  
+(d) A0hURqF7h-c (Personal Care and Style)
+
+![](images/24c730676e7520321e775bf58e7956c85dc832f12e26efc1912c06d46d44371c.jpg)
+
+![](images/b4b48feb9d8852948609348aa4acd1ffb68a3512ae000e6de3bd7ed5e96af616.jpg)  
+(c) Video Anchor Assignments
+
+![](images/5eb9eb018f1a5cdd6437b2ab14e1e16f6076521c5f5a8a8b1778559db2862949.jpg)  
+Figure 2: Anchor assignments illustrated in this figure demonstrate the visual similarity between video samples from two related categories i.e. (a) Holidays and Tradition and (b) Food and Entertaining; and two different categories i.e. (d) Personal Care ans Style and (e) Finance and Business within the HowTo100M dataset. The visual similarity is also reflected in video anchor assignments (c) as most assigned anchors are similar with minor differences in assignments, thereby showcasing the flexibility and effectiveness of our approach. (d), (e) samples look very different and therefore the anchor assignments are also very different as shown in (f). Green cell  $\rightarrow$  Anchor assigned, Yellow  $\rightarrow$  Anchor not assigned. Difference in anchor assignments indicated in red.
+
+![](images/51d8ddbe79ea5b9f265920a633c0e65b28d235cdf409ee39c8ee4e284455718d.jpg)  
+(e) _Y51yX3KJoE (Finance and Business)
+
+![](images/90effec5408bbe895c6bb2857e4cf0a5e5319eef736e15b533f5ccddcbf34d57.jpg)
+
+![](images/6199cb8143c1a217676b9c629e752f1e33284306eea87dafd469e07455f56cd4.jpg)  
+(f) Video anchors Assignments
+
+<table><tr><td rowspan="2">K</td><td rowspan="2">K&#x27;</td><td colspan="2">MSR-VTT</td><td colspan="2">YouCook2</td></tr><tr><td>R@5↑</td><td>R@10↑</td><td>R@5↑</td><td>R@10↑</td></tr><tr><td>16</td><td>8</td><td>23.1</td><td>32</td><td>37.1</td><td>47.5</td></tr><tr><td>32</td><td>16</td><td>23.2</td><td>32.1</td><td>36.1</td><td>47.6</td></tr><tr><td>64</td><td>16</td><td>23.3</td><td>31.8</td><td>36.7</td><td>47.2</td></tr><tr><td>64</td><td>48</td><td>23.7</td><td>32.1</td><td>36.7</td><td>47.7</td></tr><tr><td>64(Ours)</td><td>32</td><td>25.1</td><td>34.5</td><td>39.4</td><td>50.1</td></tr></table>
+
+Table 5: Effect of different # of anchors on zero-shot retrieval.  $K \rightarrow$  # of anchors and  $K' \rightarrow$  # of selected anchors, respectively.
+
+signments from the same embedding space. We also apply the cross-modal SSPC loss to only isolate the impact of the proposed SSPC loss across feature projections. We observe that removing the anchor consistency between the modality-specific and joint embedding spaces reduces the zero-shot retrieval performance by a significant margin ( $\sim 1.5\%$  drop in R@5 performance), indicating the importance of SSPC loss in maintaining modality-specific semantic structure in the joint embedding space for better performance.
+
+Finally, in the fourth row of Tab. 4, we analyse the effectiveness of the proposed Multi-Assignment Sinkhorn-Knoppalgorithm. To this end, we modify the vanilla Sinkhorn-Knopp [13] to obtain multiple anchor assignments per sample (refer to Sec. 5 of the Supplementary
+
+for details). We notice that the modified Sinkhorn-Knopp does not perform well with a drop in performance ( $\sim 1.6\%$  drop in R@5) on both MSR-VTT and YouCook2 datasets on all the metrics indicating the significance of our proposed MULTI-SK algorithm.
+
+Effect of Number of Anchors. To analyze the effect of the number of anchors, we conduct experiments with different number of anchors,  $K$ , and different number of selected anchors,  $K'$ . We report the results in Tab. 5. We conduct experiments on both MSR-VTT and YouCook2 datasets. From the top half of Tab. 5, we observe that the performance of our proposed method improves with the increasing number of anchors. This is to be expected since a higher number of anchors have a higher representation learning capacity. We also observe that the method performs reasonably well even with a very small number of anchors showing the general effectiveness of our proposed solution. We notice that the performance of the proposed method improves as we select more anchors as shown in bottom half of Tab. 5. However, selecting a very large number of anchors (48 out of 64 for this experiment) adds more constraints leading to poor performance.
+
+# 4.5. Qualitative Analysis
+
+Here, first we present a fine-grained visual analysis of the learned anchors, followed by qualitative retrieval results.
+
+"Animated comic scene of guy cutting up food for dinner"
+
+"a woman holding a ribbon"
+
+"mix ingredients refrigerate"
+
+"add mutton pan"
+
+![](images/f7e7e48075b15f9cebdea991dca2cd685f88583ac000e5ef483866cdd5b9ec21.jpg)
+
+![](images/6fef8ff10153ebdf7ac5c77cc855c27afb983fdf4313f22ded1e74fa85fa38ab.jpg)
+
+![](images/512d1d94ddde5cf43663e4a2e81964b05b5a7cdddeecacf992d07d761e849b06.jpg)
+
+![](images/ab8c55517bc30b4c7669d599c469af248a5b5e51ff5f07ba41c1f57877828df4.jpg)
+
+![](images/d0e9f4499376fb190e25ebf377003bdc013496e0898fcf285a4aee1788216100.jpg)
+
+![](images/2606ee2d6bb550f4e50b4634782c64cb43f4270c6676bdfc28678fc1ae9cab97.jpg)
+
+![](images/338f294da97650e8d905b2ca23441cd769228c343f180bf1a2f58a7b7a4e2240.jpg)
+
+![](images/29b37b8697b4b205debcf43945958609e3896f11874ae2e6902d41eb45c53fd2.jpg)  
+Ours  
+Figure 3: Examples of Zero-Shot Text-to-Video Retrieval on MSR-VTT and YouCook2 datasets. Each row consists of Textual Query (left), and top-3 retrieved videos for our method (center) and the state-of-the-art method EAO [42](right). Match indicates correct video for the query.
+
+![](images/0cb0d4848fc921cc9b6e19727e77b9028c0e698a01a9a16bd45e24f76aeee2b9.jpg)
+
+![](images/ed979fe8cc4337ba93053f95e71750f1bc7a26534d33199d5f8127cca553dc1b.jpg)
+
+![](images/a658fa31a34597a89d52f8422fe597aad42e5d996fd7433d7224243c46167342.jpg)
+
+![](images/599738612238d178becabd032ef3bf12f1be5ee833e3f691923c2f63c7c539b4.jpg)
+
+![](images/e1af4cc6a71efce0bb02d25715f16f879b7a83012168aedf932c43e3be5d1b70.jpg)
+
+![](images/87c0cfcde940184f1dd6aeb55faae211d18f326ea0d359136eb4099f0ebfeeea.jpg)
+
+![](images/c867ea629bf90de482f897293f68428df213fce06ca51f8f1c8aeb0a598b9dc4.jpg)
+
+![](images/951cc2b16efba37ab8c6d6d9d3dc1b77bea082db667bd05ce5e81007f9f65907.jpg)  
+EAO [42]
+
+![](images/87ae398794fc54d31a8d14024f3f4bfe97471dba08d74bf625b58a5b64e3c8c9.jpg)
+
+![](images/314daf3ecc2444e258048b4629fd33d4ecdd430322a99058b0db9be5655dede1.jpg)
+
+![](images/5adc6043e40230bd6ac85dbe81b708b0840445aa16714abd37e90bff2a8abdf3.jpg)
+
+We show fine-grained analysis of the learned anchors in Fig. 2. For the purpose of this analysis, we visualize the anchor assignments as binary assignments. However, during training we use soft anchor assignments. In Fig. 2(c), we compare the anchor assignments for samples from similar categories. It can be seen that the videos are visually similar even though they belong to different categories and the anchor assignments for these two examples are able to capture the sample similarity. In Fig. 2(f), we compare the anchor assignments for videos from different categories and it can be seen that the anchor assignments are very different as expected. This further validates our claim that our proposed method can assign semantically meaningful anchors without any explicit supervision. Further, we show qualitative retrieval comparison with EAO [42] in Fig. 3. We present more qualitative analysis in Supplementary Sec. 4.
+
+# 5. Conclusion
+
+We proposed a novel approach that preserves the modality-specific semantic relationship between samples in the joint multi-modal embedding space. To this end, we propose a flexible sample relationship modeling approach by assigning multiple anchors to each sample, which captures both shared and unique aspects of samples. To obtain these assignments, we develop a novel Multi-Assignment Sinkhorn-Knopp (Multi-SK) algorithm, and also utilize the proposed anchor consistency loss to learn these anchors. Our qualitative results demonstrate that our learnt anchors correspond to meaningful semantic concepts. Our extensive experimentation demonstrates that the proposed approach improves generalizability by outperforming state-of-the-art methods on both in- and out-of-domain datasets. We also
+
+show that our method achieves state-of-the-art performance on multiple zero-shot tasks, and also outperforms when fine-tuned on downstream datasets.
+
+# Acknowledgements
+
+Research was sponsored by the Army Research Office and was accomplished under Grant Number W911NF-19-1-0356. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Office or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. Nina Shvetsova is supported by German Federal Ministry of Education and Research (BMBF) project STCL - 01IS22067.
+
+# References
+
+[1] Jean-Baptiste Alayrac, Adria Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. Advances in Neural Information Processing Systems, 33:25-37, 2020. 2, 6  
+[2] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. Advances in Neural Information Processing Systems, 33:9758-9770, 2020. 2  
+[3] Elad Amrani, Rami Ben-Ari, Daniel Rotman, and Alex Bronstein. Noise estimation using density estimation for self-supervised multimodal learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 6644-6652, 2021. 2
+
+[4] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In Proceedings of the IEEE International Conference on Computer Vision, pages 609-617, 2017. 2  
+[5] Yuki Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. Advances in Neural Information Processing Systems, 33:4660-4671, 2020. 1, 2, 3, 4  
+[6] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Sound-net: Learning sound representations from unlabeled video. Advances in neural information processing systems, 29, 2016. 2  
+[7] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 2  
+[8] Yann Brenier. D'écomposition polaire et r'arrangement monotone des champs de vecteurs. CR Acad. Sci. Paris Sér. I Math., 305:805-808, 1987. 3  
+[9] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 2, 3, 4  
+[10] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 5  
+[11] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie Boggust, Rameswar Panda, Brian Kingsbury, Rogerio Feris, David Harwath, et al. Multimodal clustering networks for self-supervised learning from unlabeled videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8012-8021, 2021. 1, 2, 5, 6, 7  
+[12] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 2  
+[13] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 3, 8  
+[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5  
+[15] Alex Fang, Gabriel Ilharco, Mitchell Wortsman, Yuhao Wan, Vaishaal Shankar, Achal Dave, and Ludwig Schmidt. Data determines distributional robustness in contrastive language image pre-training (clip). In International Conference on Machine Learning. PMLR, 2022. 2  
+[16] Enrico Fini, Enver Sangineto, Stéphane Lathuilière, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3
+
+[17] Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 303-312, 2021. 3  
+[18] Anca-Ioana Grapa, Laure Blanc-Feraud, Ellen van Obberghen-Schilling, and Xavier Descombes. Optimal transport vs many-to-many assignment for graph matching. In GRETSI 2019-XXVIIème Colloque francophone de traitement du signal et des images, 2019. 3  
+[19] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imageNet? In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6546-6555, 2018. 5  
+[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. cvpr. 2016. arXiv preprint arXiv:1512.03385, 2016. 5  
+[21] Dat Huynh and Ehsan Elhamifar. A shared multi-attention framework for multi-label zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8776–8786, 2020. 2  
+[22] Lukasz Kaiser, Aidan N Gomez, Noam Shazeer, Ashish Vaswani, Niki Parmar, Llion Jones, and Jakob Uszkoreit. One model to learn them all. arXiv preprint arXiv:1706.05137, 2017. 2  
+[23] Leonid Kantorovich. On translation of mass. In Dokl. AN SSSR, volume 37, page 20, 1942. 3  
+[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6  
+[25] Mingchao Li, Xiaoming Shi, Haitao Leng, Wei Zhoi, Haitao Zheng, and Kuncai Zhang. Learning semantic alignment with global modality reconstruction for video-language pretraining towards retrieval. Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 1  
+[26] Yanbin Liu, Linchao Zhu, Makoto Yamada, and Yi Yang. Semantic correspondence as an optimal transport problem. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4463-4472, 2020. 3  
+[27] Shuang Ma, Daniel McDuff, and Yale Song. Unpaired image-to-speech synthesis with multimodal information bottleneck. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 2  
+[28] Massimiliano Mancini, Muhammad Ferjad Naeem, Yongqin Xian, and Zeynep Akata. Open world compositional zero-shot learning. 2021. 2  
+[29] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 2, 7  
+[30] Antoine Miech, Ivan Laptev, and Josef Sivic. Learnable pooling with context gating for video classification. arXiv preprint arXiv:1706.06905, 2017. 5  
+[31] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching
+
+hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 2, 5, 6, 7  
+[32] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013. 5  
+[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 5  
+[34] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, Joao Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. arXiv preprint arXiv:2010.02824, 2020. 6  
+[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2  
+[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning. PMLR, 2021. 2  
+[37] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning. PMLR, 2016. 2  
+[38] Mamshad Nayeem Rizve, Navig Kardan, and Mubarak Shah. Towards realistic semi-supervised learning. In European Conference on Computer Vision, pages 437-455. Springer, 2022. 2, 3  
+[39] Andrew Rouditchenko, Angie Boggust, David Harwath, Brian Chen, Dhiraj Joshi, Samuel Thomas, Kartik Audhkhasi, Hilde Kuehne, Rameswar Panda, Rogerio Feris, et al. Avlnet: Learning audio-visual language representations from instructional videos. arXiv preprint arXiv:2006.09199, 2020. 2, 5, 6, 7  
+[40] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 3  
+[41] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 2  
+[42] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. 1, 2, 5, 6, 7, 9
+
+[43] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15638-15650, June 2022. 1  
+[44] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967. 3  
+[45] Zineng Tang, Jie Lei, and Mohit Bansal. DeCEMBERT: Learning from noisy instructional videos via dense captions and entropy minimization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Association for Computational Linguistics, 2021. 1  
+[46] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020. 2  
+[47] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 6  
+[48] Asano YM., Rupprecht C., and Vedaldi A. Self-labelling via simultaneous clustering and representation learning. In International Conference on Learning Representations, 2020. 2, 3  
+[49] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 6  
+[50] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634-23651, 2021. 1  
+[51] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 2  
+[52] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. 6  
+[53] Yipin Zhou, Zhaowen Wang, Chen Fang, Trung Bui, and Tamara L Berg. Visual to sound: Generating natural sound for videos in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 2  
+[54] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 6

2023/Preserving Modality Structure Improves Multi-Modal Learning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:49357fa12318fc9e37a0c72a0aec16d00666fd4751145870c7b36c2a43fe1c9f
+size 599605

2023/Preserving Tumor Volumes for Unsupervised Medical Image Registration/dbd41759-9a48-41b9-9206-f4ea354bc2a2_content_list.json

@@ -0,0 +1,1714 @@
+[
+    {
+        "type": "text",
+        "text": "Preserving Tumor Volumes for Unsupervised Medical Image Registration",
+        "text_level": 1,
+        "bbox": [
+            114,
+            130,
+            854,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Qihua Dong $^{1}$ , Hao Du $^{1}$ , Ying Song $^{2}$ , Yan Xu $^{3*}$ , Jing Liao $^{1*}$",
+        "bbox": [
+            248,
+            179,
+            718,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Department of Computer Science, City University of Hong Kong",
+        "bbox": [
+            223,
+            198,
+            745,
+            217
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College  $^{3}$ School of Biological Science and Medical Engineering, Beihang University",
+        "bbox": [
+            138,
+            215,
+            831,
+            270
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "dongqh078@gmail.com, haodu8-c@my.cityu.edu.hk, jingliao@cityu.edu.hk;",
+        "bbox": [
+            179,
+            272,
+            787,
+            287
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "songying1770@hotmail.com",
+        "bbox": [
+            377,
+            290,
+            591,
+            303
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "xuyan04@gmail.com",
+        "bbox": [
+            406,
+            308,
+            560,
+            321
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            356,
+            312,
+            372
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Medical image registration is a critical task that estimates the spatial correspondence between pairs of images. However, current traditional and deep-learning-based methods rely on similarity measures to generate a deforming field, which often results in disproportionate volume changes in dissimilar regions, especially in tumor regions. These changes can significantly alter the tumor size and underlying anatomy, which limits the practical use of image registration in clinical diagnosis. To address this issue, we have formulated image registration with tumors as a constraint problem that preserves tumor volumes while maximizing image similarity in other normal regions. Our proposed strategy involves a two-stage process. In the first stage, we use similarity-based registration to identify potential tumor regions by their volume change, generating a soft tumor mask accordingly. In the second stage, we propose a volume-preserving registration with a novel adaptive volume-preserving loss that penalizes the change in size adaptively based on the masks calculated from the previous stage. Our approach balances image similarity and volume preservation in different regions, i.e., normal and tumor regions, by using soft tumor masks to adjust the imposition of volume-preserving loss on each one. This ensures that the tumor volume is preserved during the registration process. We have evaluated our strategy on various datasets and network architectures, demonstrating that our method successfully preserves the tumor volume while achieving comparable registration results with state-of-the-art methods. Our codes is available at: https://dddraxxx.github.io/Volume-Preserving-Registration/.",
+        "bbox": [
+            75,
+            393,
+            473,
+            845
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            501,
+            356,
+            630,
+            372
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/3c773565a73bf98b9747bd76d71e915cf6c3c99ab9b4b4ae399abdb539649b8e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            526,
+            391,
+            686,
+            513
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/657ff349d7c2a03bbddd4c7a8472927f39480b62c2a453cd8914b61720acb5f5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            707,
+            391,
+            866,
+            513
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/5501f7c771a1a7977cd66eb7f46958891ea67ce1d7b7ca64e1a6c7819722c45a.jpg",
+        "image_caption": [
+            "Fixed Image",
+            "Moving Image",
+            "Figure 1. Red boxes represent the location of tumors in the moving image before warped. In image registration with tumors, similarity-based registration typically leads to significant alterations in tumor size while our volume-preserving registration is capable of preserving tumor anatomy."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            526,
+            530,
+            684,
+            651
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/cf51c67221ba2f991df800d0691c3d451ed0df7e8b666712eb207107aaa1d580.jpg",
+        "image_caption": [
+            "Similarity-Based Registration",
+            "Volume-Preserving Registration"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            707,
+            529,
+            866,
+            651
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Deformable image registration is a fundamental task that estimates non-linear spatial correspondences between two images. It is useful for medical image studies that involve 3D images of organs or tissues, such as MR brain scans and CT liver scans. Recently, a wide range of deep-learning-based methods are proposed in the field, with better performance and obvious speed-ups in inference time compared to the traditional registration methods [5, 38, 32, 14, 15].",
+        "bbox": [
+            496,
+            750,
+            892,
+            869
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Currently, most of the learning-based methods [5, 38] train a registration model to achieve high similarity in either",
+        "bbox": [
+            500,
+            869,
+            890,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding authors",
+        "bbox": [
+            94,
+            886,
+            223,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "21208",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "image intensity [38, 5] or anatomy label maps [14]. While this is useful in various medical applications, such as atlas-based segmentation and image fusion [8], problem emerges, however, when studying images with tumors [4, 12, 24]. Tracking tumor growth is a core task in cancer treatment, which can be used to evaluate the outcomes of radiotherapy and chemotherapy, and plan optimal postoperative treatment [12]. The process requires registration to align the anatomy of images from different periods of the patient, while preserving tumor properties e.g., the size. Traditional models using regular registration have struggled to accomplish this task [12]. We have observed that this remains an issue for current learning-based methods, and a viable solution has yet to be proposed.",
+        "bbox": [
+            75,
+            90,
+            472,
+            303
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 1, the volume of tumors significantly reduces when using popular unsupervised registration networks to align tumor images with atlas images. The issue is widespread since mainstream registration networks focus on optimizing the similarity between image pairs while ignoring tumor regions. Tumor regions often lack corresponding parts, and even for images of the same patient, the size and location of tumors may vary greatly. This can happen because tumors can change shape, size, location, or disappear over time. Therefore, in deformable settings, the size of tumors will change disproportionately to maximize the image similarity. This problem may be more severe when the registration network has better warping ability, leading to better performance in evaluations from previous works, as confirmed by our experiments in Section 4. The disproportionate change of tumor volume in deformable registration is lethal for clinical evaluation of tumor growth and greatly hinders the application of registration in clinics.",
+        "bbox": [
+            75,
+            304,
+            472,
+            578
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we re-formulate the deformable registration problem as a constraint problem that preserves the properties of tumors in the original images, i.e., the shape and the size of the tumor, while maximizing image similarity in non-tumor regions. For instance, in tumor image registration with atlas, besides the alignment of anatomical structures, the tumor should have similar morphing behaviours as its surrounding organs, i.e., the size of it should change proportionally to the size of the organ. However, this poses two major challenges for our strategy. Firstly, due to the limited availability of annotated data, most existing registration methods resort to unsupervised learning. Therefore, a key question is how to identify tumor regions in the unsupervised setting. Second, after the tumor regions have been identified, the challenge still remains of preserving tumor volumes while simultaneously maximizing similarity in other regions. Finding a way to balance these two competing objectives is crucial for achieving effective tumor image registration.",
+        "bbox": [
+            75,
+            580,
+            472,
+            867
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To tackle the challenges, we propose a novel two-stage strategy that operates in an unsupervised manner. In the first",
+        "bbox": [
+            76,
+            869,
+            470,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "stage, we leverage a similarity-based registration network to identify regions that undergo excessive volume change during registration. Such regions are indicative of the presence of tumors, as tumors tend to exhibit larger changes in volume due to a lack of correspondence between image pairs. The outcome of the first stage is a soft tumor mask that distinguishes normal and possible tumor regions. In the second stage, we introduce an adaptive volume-preserving loss to train a volume-preserving registration network. Based on the soft tumor mask obtained from the first stage, this loss is designed to adaptively penalize volume changes in different regions of the image. Possible tumor regions are assigned larger loss weights, and similarity loss weights are adjusted accordingly in the opposite direction. The application of the adaptive volume-preserving loss effectively balances similarity and volume preservation in different parts of the image.",
+        "bbox": [
+            496,
+            90,
+            893,
+            348
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "It should be noted that the soft masks from stage one do not need to be highly accurate to achieve desirable results, as our stage two is robust, which is demonstrated by our ablation experiments in section 4. In addition to evaluating the robustness of our proposed strategy on various datasets and frameworks, including CNN and transformer, as well as different types of imaging data such as brain and abdominal, MRI and CT images, we have also introduced the Square Tumor Size Ratio (STSR) metric to measure the preservation of tumor volumes. The results demonstrate that our methods achieve comparable warping performance metrics such as Dice coefficient and average landmark distances while effectively preserving tumor properties as measured by the STSR.",
+        "bbox": [
+            496,
+            349,
+            893,
+            559
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To conclude, our contribution includes: (1) We have reformulated image registration with tumors as a constraint problem that preserves tumor volumes while maximizes image similarity in other normal regions. (2) We have designed a two-stage process that the first stage performs regular registration to estimate tumor regions, and the second stage performs volume-preserving registration on tumor regions based on the adaptive volume-preserving loss. (3) We have proposed the STSR metric to measure the preservation of tumor volumes in registration, and have evaluated the effectiveness of our strategy on different network architectures and various datasets.",
+        "bbox": [
+            496,
+            561,
+            893,
+            742
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            758,
+            640,
+            776
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Learning-Based Deformable Image Registration",
+        "text_level": 1,
+        "bbox": [
+            500,
+            785,
+            890,
+            815
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Compared to conventional registration methods, deep networks have achieved remarkable registration accuracy and speed using various network structures, i.e., CNN and Transformer, in applications like atlas-image registration [29, 7, 16, 6, 32, 38, 5, 14, 15, 8]. They generally",
+        "bbox": [
+            496,
+            825,
+            893,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "21209",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "adopts unsupervised learning, utilizing image similarity on either image intensities [29, 7, 6, 38] or synthesized label maps [14, 15] as well as smoothness regularization to supervise the registration network [7, 6, 14, 15, 38]. Advancements are often made through the introduction of novel regularization such as cycle consistency [16] and geometry preservation [1], or through innovative network structures like multi-scale architectures [28, 23]. However, existing methods do not adequately consider cases where tumors exist. Tumors often have no corresponding part between image pairs, and therefore, optimizing for similarity can excessively change their volumes, which limits the application of registration networks in critical clinical applications like cancer treatment. To the best of our knowledge, we are the first to study this problem with learning-based methods and propose a metric to measure volume preservation specifically for tumors.",
+        "bbox": [
+            75,
+            90,
+            472,
+            349
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2. Deformable Image Registration with Tumors",
+        "text_level": 1,
+        "bbox": [
+            76,
+            358,
+            459,
+            376
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Conventional registration models often resolve registering images with tumor by (1) excluding the similarity measure on tumors [11, 18], which requires either massive manual segmentation [11] or initial seed to explicitly model tumor-growth [12, 18]. (2) jointly segmenting and registering images [9], which is time-consuming since it need to iteratively detect dissimilarity in image intensities and perform resection and retraction. (3) reconstructing quasi-normal images from pathological ones [13, 19], which is also slow since the reconstruction need to be iteratively improved.",
+        "bbox": [
+            75,
+            382,
+            472,
+            547
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Recently, learning-based methods for registration with tumors [24, 21, 33] are proposed for the release of BraTS-Reg dataset [4]. They aim to improve the spatial correspondence for non-tumor regions, thus either directly annotates landmarks to supervise registration [21] or introducing invertibility of the registration to mask non-inverted regions and ensure high correspondence in others [24, 33]. However, these methods assume the absence of tumors in one image of the registered pair, which may not be applicable in practical applications where we need to compare two tumor images from one patient to track the tumor growth. In this scenario, invertibility in tumors is still possible and thus, it is essential to consider the preservation of tumor properties during registration.",
+        "bbox": [
+            75,
+            549,
+            472,
+            760
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3. Unsupervised Tumor Segmentation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            771,
+            385,
+            787
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To ensure the preservation of tumor volumes during registration, it is crucial to first identify the locations of tumors. Current deep learning methods have made remarkable progress in unsupervised tumor segmentation, as demonstrated by recent studies [34, 35, 20]. These methods typically use GAN models or simulate tumors on healthy images to generate synthetic data, which is then used to",
+        "bbox": [
+            75,
+            795,
+            472,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "train the segmentation network. Although the well-trained network can achieve high segmentation accuracy, it often requires complex training strategies and additional models or testing steps to achieve the desired performance, as noted in recent research [35, 20]. However, in our approach, we do not rely on sophisticated segmentation methods to accurately segment tumors. Instead, our experiments, as presented in Section 4, demonstrate that our tumor mask estimation is sufficient for our strategy to learn a volume-preserving registration for tumors.",
+        "bbox": [
+            496,
+            90,
+            893,
+            243
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.4. Volume-Preserving Deformable Image Registration",
+        "text_level": 1,
+        "bbox": [
+            498,
+            250,
+            893,
+            280
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Volume-preserving registration has been studied extensively to improve the accuracy of anatomical registration. In pulmonary CT image registration, for example, tissue intensity changes between inspiration and expiration phases can be utilized as prior knowledge, and adding a volume-preserving constraint has been shown to be effective [31, 37]. In other cases, such as when soft tissues are incompressible or when images come from different modalities, a volume-preserving constraint can also serve as a helpful regularization [25, 30]. Our work, instead, focuses on using volume-preservation to maintain the properties of tumors during image registration. The proposed strategy imposes this constraint adaptively on different parts of the images in order to preserve tumor volumes and ensure anatomical accuracy, while previous works have either preserved volumes for the entire images [25, 30] or can use simple methods to determine the regions to be preserved [31, 37].",
+        "bbox": [
+            496,
+            287,
+            893,
+            546
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Methodology",
+        "text_level": 1,
+        "bbox": [
+            498,
+            556,
+            635,
+            574
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            500,
+            580,
+            640,
+            595
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Problem Setting. Given a moving image  $I_{m}$  and fixed image  $I_{f}$  defined over  $d$ -dimensional (d=3 in this paper) space  $\\Omega$ , deformable image registration aims to find a dense deformation filed  $\\phi : \\Omega \\to \\Omega$ . This field warps  $I_{m}$ , meaning it projects every pixel position  $x$  in warped image  $I_{w}$  to  $I_{m}$  to achieve alignment with  $I_{f}$ . In this paper, the task is to construct a registration network  $F_{\\theta}$  that takes input  $I_{m}, I_{f}$  to predict the deformation field  $\\phi$ .",
+        "bbox": [
+            496,
+            604,
+            893,
+            726
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\phi = F _ {\\theta} \\left(I _ {m}, I _ {f}\\right) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            647,
+            733,
+            890,
+            750
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nI _ {w} (x) = I _ {m} (\\phi (x) + x) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            616,
+            752,
+            890,
+            768
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nI _ {w} \\sim I _ {f} \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            638,
+            771,
+            890,
+            787
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Criteria. In general, a successful deformable image registration involves accurately aligning the anatomical structures while also ensuring a smooth deformation field. This is often achieved through the use of similarity-based objectives and regularization methods that promote smoothness. However, in the context of image registration with tumors, simply aligning the anatomy is not",
+        "bbox": [
+            496,
+            795,
+            893,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "21210",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/5cb5a1782ab687053e0200bb173bbb0a88140e957c50a252f9b6032916461fd0.jpg",
+        "image_caption": [
+            "Figure 2. Main Framework. Our strategy consists of two stages. In the first stage, a soft mask indicating tumor regions is estimated by analyzing the Jacobian matrix of the deformation field obtained by a registration network pre-trained on similarity loss. In the second stage, the soft mask is used to guide the calculation of both volume-preserving and similarity losses for training the volume-preserving registration network. This deformation field estimated by the stage two ensures that both the image similarity and the preservation of tumor sizes."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            95,
+            831,
+            397
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "sufficient. Preserving the volume of the tumors is crucial for accurately tracking tumor growth, as previously discussed. To be specific, given the moving image  $I_{m}$  with the tumor  $T_{m}$  and its surrounding organ  $O_{m}$  and the warped moving image  $I_{w}$  with  $T_{w}$  and  $O_{w}$ , the morphing of tumor  $T_{m}$  should be similar to that of  $O_{m}$ . Thus in respect of the volume preservation, the change of tumor size  $|T_{m}|$  should be proportionate to the change of the containing organ  $|O_{m}|$ , i.e.:",
+        "bbox": [
+            75,
+            489,
+            470,
+            625
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nT S R \\left(I _ {m}\\right) = \\frac {\\left| T _ {m} \\right|}{\\left| O _ {m} \\right|} \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            191,
+            633,
+            468,
+            666
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nT S R \\left(I _ {m}\\right) \\approx T S R \\left(I _ {w}\\right) \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            192,
+            667,
+            468,
+            686
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We therefore define the tumor size ratio (TSR) as  $\\frac{|T|}{|O|}$ , which should be preserved after registration, and the re-formulated registration objective is now maximizing image similarity while preserving the volumes of tumors, i.e., the TSR.",
+        "bbox": [
+            76,
+            698,
+            468,
+            763
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Overall Strategy",
+        "text_level": 1,
+        "bbox": [
+            76,
+            771,
+            240,
+            787
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The proposed strategy consists of two stages, as shown in Figure 2. The first stage estimates potential tumor regions and generates a soft mask for them. This is done by analyzing volume changes, specifically the Jacobian matrix, in the deformation field obtained by an existing similarity-based registration network. As tumor regions tend to undergo excessive volume change during registration, they can",
+        "bbox": [
+            75,
+            794,
+            468,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "be identified through this analysis. In the second stage, we perform volume-preserving registration for tumors using an adaptive volume-preserving loss. This loss is designed to guide the volume-preserving process effectively and is associated with the tumor masks estimated in the first stage. Through the assignment of distinct weights to diverse regions, including both normal and tumor regions, it is possible to preserve tumor volumes while also ensuring a high degree of similarity in other regions.",
+        "bbox": [
+            496,
+            489,
+            890,
+            625
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Importantly, our strategy can be applied to nearly any learning-based registration network without introducing any additional modules.",
+        "bbox": [
+            496,
+            626,
+            890,
+            671
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.1 Stage 1: Unsupervised Tumor Mask Estimation",
+        "text_level": 1,
+        "bbox": [
+            500,
+            693,
+            883,
+            709
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Following previous works [5, 6, 38], we first train a similarity-based registration network  $F_{\\theta_1}$  with parameters  $\\theta_1$  and fix its parameters. As shown in Figure 1, tumors in medical images can be detected by analyzing the change in volume through similarity-based registration. Our goal is to utilize this characteristic to facilitate the unsupervised estimation of the tumor mask for volume preservation during registration.",
+        "bbox": [
+            496,
+            718,
+            890,
+            838
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To determine the volume change at voxel  $x$ , we calculate the determinant of the Jacobian matrix of the deformation field  $(J_{\\theta})$  with model parameter  $\\theta$ . Our objective is to ensure that the volume change in each point of the tumor region is",
+        "bbox": [
+            496,
+            839,
+            893,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "21211",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "similar to the volume change of the organ, which is an indicator of the organ's size change (as shown in Equation 6). To measure the distance between the two, we calculate the number of times one ratio exceeds another, denoted as  $D_{\\theta}$ , using Equations 8 and 7 where  $D_{\\theta}'(x)$  represents the relative volume change of voxel  $x$  with respect to organ areas.",
+        "bbox": [
+            76,
+            90,
+            472,
+            183
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\frac {\\left| T _ {w} \\right|}{\\left| O _ {w} \\right|} / \\frac {\\left| T _ {m} \\right|}{\\left| O _ {m} \\right|} \\approx 1 \\Rightarrow \\frac {\\left| T _ {w} \\right|}{\\left| T _ {m} \\right|} / \\frac {\\left| O _ {w} \\right|}{\\left| O _ {m} \\right|} \\approx 1, \\text {f r o m E q .} 5, \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            96,
+            200,
+            468,
+            234
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nD _ {\\theta} ^ {\\prime} (x) = \\left| J _ {\\theta (x)} \\right| / \\frac {\\left| O _ {w} \\right|}{\\left| O _ {m} \\right|}, x \\in \\Omega , \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            166,
+            234,
+            468,
+            268
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nD _ {\\theta} (x) = \\max  \\left(D _ {\\theta} ^ {\\prime} (x), 1 / D _ {\\theta} ^ {\\prime} (x)\\right), \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            166,
+            270,
+            468,
+            287
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In practice, we estimate the organ masks  $O_{m}$  and  $O_{w}$  by warping a reference image with the ground truth (GT) organ segmentation. We randomly select the reference image from the training dataset, and the choice of reference image does not significantly impact the results because registration for organs achieves high accuracy, as evidenced by the dice value in table 1. By warping the reference image, we obtain the organ segmentation for the moving image  $O_{m}$ . We then calculate the organ segmentation for the warped image  $O_{w}$ .",
+        "bbox": [
+            76,
+            295,
+            468,
+            430
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To obtain a soft tumor mask  $STM$  from the distance  $D_{\\theta_1}$  defined in 8, we use a transformation function  $Tr$ . This function can be formulated as the following equation:",
+        "bbox": [
+            76,
+            431,
+            468,
+            476
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS T M (x) = T r \\left(D _ {\\theta_ {1}} (x)\\right), \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            184,
+            484,
+            468,
+            500
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Here,  $D_{\\theta_1}(x)$  represents the distance value at pixel  $x$ , and  $Tr$  transforms values from the range of  $[0,\\infty]$  to the range of  $[0,1]$ . A value of 1 indicates the presence of a tumor, while a value of 0 indicates a normal region. In practical implementation, we use  $Tr(x) = \\text{Sign}(5 \\cdot (x - 1.5))$  since the sigmoid function is widely adopted for converting values to the range of  $[0,1]$  and to our experience, the size of tumor regions averagely changes around 1.5 times. However, the choice of transformation functions is not exclusive, as demonstrated in our experiments in Section 4. We also set the regions outside the organ mask to 0.",
+        "bbox": [
+            76,
+            508,
+            468,
+            672
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We observed that some normal regions, particularly those close to the organ boundary, may also exhibit some volume changes after registration. As a result, these regions may be incorrectly included in the tumor mask, as shown in Figure 3 (a). To address this issue, we implemented a preregistration step to first align the edges and exclude volume changes caused by edge alignment. Specifically, in the preregistration step, we applied a bilateral filter to the moving image, which preserves the edges while smoothing the interior regions. We then used the registration network to register the filtered moving image with the fixed image in order to align their edges. After this, an unfiltered warped image was obtained by spatially transforming the unfiltered moving image with the deformation field. The second registration process was then performed on the unfiltered warped",
+        "bbox": [
+            76,
+            674,
+            470,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/ae5b9c79bfc3f87b3cbe338bdb6f8cec305e153515e9eefa29531f91634fa745.jpg",
+        "image_caption": [
+            "(a)Tumor Mask"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            526,
+            88,
+            686,
+            210
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/ee6ad0dfa4d53dd393bd7326b429ac9e80ca121224703f18d3a870c56fd5e7a6.jpg",
+        "image_caption": [
+            "(b)Tumor Mask after pre-registration",
+            "Figure 3. Visualization of the soft tumor mask. (a) is the estimated soft tumor mask without pre-registration and (b) is the one with pre-registration. The gray colors with intensity variations denote the soft mask, while the green and red colors represent the ground truth segmentation for the liver and tumors, respectively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            705,
+            88,
+            867,
+            210
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "image, which focused on the registration of interior regions and was able to accurately estimate a tumor mask, as shown in Figure 3 (b). Therefore, this pre-registration approach effectively selects the tumor regions while filtering out the non-tumor regions.",
+        "bbox": [
+            498,
+            324,
+            890,
+            400
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2.2 Stage 2: Volume-Preserving Registration for Tumors",
+        "text_level": 1,
+        "bbox": [
+            500,
+            424,
+            890,
+            453
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After obtaining the soft tumor mask  $STM$ , we can adaptively select tumor regions to maintain the size ratio. Here, we define the volume-preserving loss directly as the distance  $D_{\\theta_2}$ , as shown in Equation 8. By minimizing the distance, we can achieve smaller volume changes, indicating better preservation of tumor volume. The volume-preserving loss is multiplied by the soft tumor mask. This ensures that the volume-preserving constraints are adaptively applied to the tumor regions and not to the normal regions, which can be formulated as:",
+        "bbox": [
+            498,
+            465,
+            890,
+            616
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {\\mathrm {V P}} = \\frac {1}{\\left| I _ {m} \\right|} \\sum_ {x \\in \\Omega} D _ {\\theta_ {2}} (x) \\cdot S T M (x), \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            573,
+            628,
+            890,
+            665
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Here, in the second stage, the registration network takes the soft tumor mask as input, which serves as an indicator for the regions that need to be preserved in terms of volume. We follow the similarity and registration loss formulation in [38], but we use a weighted similarity loss to adapt to the volume-preserving loss, where  $Cov$  is the covariance between two images  $I_1, I_2$ ,  $\\hat{I}$  is the mean of  $I$ , and  $Cov'$  is the adapted version utilizing  $STM$  weight:",
+        "bbox": [
+            498,
+            678,
+            890,
+            799
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {C o v} ^ {\\prime} \\left[ I _ {1}, I _ {2} \\right] = \\frac {\\sum \\left(I _ {1} - \\hat {I _ {1}}\\right) \\left(I _ {2} - \\hat {I _ {2}}\\right) (1 - S T M)}{\\sum (1 - S T M)}, \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            513,
+            825,
+            890,
+            861
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {\\mathrm {w} - \\mathrm {s i m}} (x) = \\frac {\\operatorname {C o v} ^ {\\prime} \\left[ I _ {w} , I _ {f} \\right]}{\\sqrt {\\operatorname {C o v} \\left[ I _ {w} , I _ {w} \\right] \\operatorname {C o v} \\left[ I _ {f} , I _ {f} \\right]}}, \\tag {12}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            531,
+            862,
+            890,
+            898
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "21212",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Therefore, the training loss of the second stage is:",
+        "bbox": [
+            76,
+            90,
+            406,
+            106
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL = L _ {\\mathrm {w} \\text {- s i m}} + \\alpha_ {1} \\cdot L _ {\\mathrm {V P}} + \\alpha_ {2} \\cdot L _ {\\mathrm {r e g}}, \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            155,
+            119,
+            468,
+            135
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where we use the smoothness regularization from Voxel-Morph [5] as  $L_{reg}$ , and  $\\alpha_{1}$  and  $\\alpha_{2}$  as hyperparameters to control the relative importance of the volume-preserving loss and the regularization loss.",
+        "bbox": [
+            76,
+            147,
+            468,
+            208
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            222,
+            209,
+            239
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            76,
+            247,
+            282,
+            263
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation. Following the implementation of Recursive Cascaded Network (RCN) [38], we employed PyTorch and utilized the same similarity loss and regularization losses (for affine and deformable, respectively) for baseline methods. The weight of the volume-preserving losses is set to 0.1 while the weights of similarity loss is 1 and regularization loss is 0.1. Models were trained on one NVIDIA GeForce RTX 3090 GPU with a batch size of 4. During the training stage, we used the Adam optimizer [36] and ran for a total of 5 epochs, with 20000 iterations in each epoch. The learning rate was set to  $10^{-4}$ .",
+        "bbox": [
+            75,
+            271,
+            468,
+            436
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Baselines. We re-implemented our strategy on four SOTA learning-based models that have distinct network architectures: 1) recursive cascaded network with VTN base network (VTN) [38], which stacks CNN networks to achieve higher registration accuracy; 2) VoxelMorph [5], which is a widely used CNN network that utilize stationary vector field to register images; 3) TransMorph [6], which is a well-performed hybrid Transformer-ConvNet network; 4) LapIRN [23], which is a Laplacian Pyramid network designed to operate in a coarse-to-fine manner. We fit these learning-based networks into our strategy with minor efforts and achieve good performance. Additionally, we compare our proposed method with two state-of-the-art traditional methods for deformable image registration, SyN [2] (integrated in ANTs [3] with the affine stage) and B-spline [26] (integrated in Elastix [17] with the affine stage) following [38].",
+        "bbox": [
+            75,
+            467,
+            468,
+            724
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metrics. The evaluation of the proposed strategy is composed of five metrics: Dice score, landmark distance to measure the anatomical alignment, and Square Tumor Size Ratio metric to measure the volume preservation of tumors. More specifically, the Dice score is calculated using the following formula:",
+        "bbox": [
+            75,
+            724,
+            468,
+            815
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nD i c e \\left(I _ {1}, I _ {2}\\right) = \\frac {2 \\left| I _ {1} \\cap I _ {2} \\right|}{\\left| I _ {1} \\right| + \\left| I _ {2} \\right|}, \\tag {14}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            179,
+            825,
+            468,
+            859
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Additionally, we also use landmark annotations to measure the anatomical alignment between the warped image and",
+        "bbox": [
+            76,
+            869,
+            468,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "the fixed image. We calculate the average distance between the fixed image's landmarks and the warped landmarks of the moving image, as also introduced in RCN [38]. To quantify the diffeomorphism and smoothness of the deformation fields, we followed the methodology of previous works such as [7, 6, 39]. Here, two metrics were employed: the average percentage of voxels with non-positive Jacobian determinant  $(|J_{\\theta}| < 0(\\%))$  in the deformation fields, and the standard deviation of the Jacobian determinant  $(Std.|J_{\\theta}|)$ .",
+        "bbox": [
+            496,
+            90,
+            890,
+            242
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To measure the volume preservation of tumors, we propose Square Tumor Size Ratio metric (STSR), which is calculated from the change of tumor size ratio (TSR) defined in Equation 4:",
+        "bbox": [
+            496,
+            243,
+            890,
+            303
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS T S R = \\max  \\left(\\left(\\frac {T S R \\left(I _ {m}\\right)}{T S R \\left(I _ {w}\\right)}\\right), \\left(\\frac {T S R \\left(I _ {w}\\right)}{T S R \\left(I _ {m}\\right)}\\right)\\right) ^ {2}, \\tag {15}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            529,
+            316,
+            890,
+            349
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Here,  $I_{m}, I_{w}$  refers to the moving image and the warped moving image. The square term is used o account for larger changes in tumor volumes, which can have a more significant impact on the accuracy of tumor growth estimation in clinical applications. By utilizing the STSR metric, we are able to accurately evaluate the performance of our proposed strategy.",
+        "bbox": [
+            496,
+            362,
+            890,
+            467
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Preprocessing. In our study, we follow standard preprocessing steps as outlined in VTN [39] and VoxelMorph [5]. The raw scans are resampled into  $128 \\times 128 \\times 128$  voxels after cropping any unnecessary areas around the target object.",
+        "bbox": [
+            496,
+            469,
+            890,
+            544
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Datasets",
+        "text_level": 1,
+        "bbox": [
+            500,
+            558,
+            601,
+            571
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Training. To train our model, we utilized two publicly available datasets: (1) LiTS17 [10]: The LiTS dataset comprises 201 CT scans that were acquired using various CT scanners and devices. The resolution of the images in this dataset ranges from  $0.56\\mathrm{mm}$  to  $1.0\\mathrm{mm}$  in the axial direction and from  $0.45\\mathrm{mm}$  to  $6.0\\mathrm{mm}$  in the  $z$  direction. The number of slices in the  $z$  direction ranges from 42 to 1026. (2) BraTS20 [22]: It consists of 369 MRI scans of brain tumors, each with ground truth segmentation of four different tumor components. For training our neural network, we utilized the T1ce modality and treated all four components as tumors, while considering the entire brain as the surrounding organ.",
+        "bbox": [
+            496,
+            582,
+            890,
+            777
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Testing. For testing, we utilized a set of 10 cases of CT liver scans with metastatic tumors, and 10 MRI brain scans with metastatic tumors. The modality of MRI brain scans is T1ce. Each case included both pre- and post-treatment images of the same patients, resulting in a total of 20 scans for each set. Following [27], the segmentation of the scans, which contains the organ (liver or brain) and the tumor inside, were carefully annotated by three experts in the field.",
+        "bbox": [
+            496,
+            779,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "21213",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/2e42f463d4be0f91f0d1d06d5ddaca8fecaa2047459b42efd043552603df60a6.jpg",
+        "image_caption": [
+            "Fixed"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            99,
+            166,
+            169,
+            220
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/b1a7e7fddb152d1de4077cb4ad0753ac84c031c0007beba59b782801589a1a04.jpg",
+        "image_caption": [
+            "Moving"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            176,
+            166,
+            246,
+            220
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/51ab0beed0cf28f8fbcd4e4d1010e4d440278149705dab1eb4955ca24c0c5c1d.jpg",
+        "image_caption": [
+            "Fixed GT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            99,
+            232,
+            169,
+            286
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/a86685584960360e104e78114dcfe893f15d653fcdadd8bd6eb4597615b23e44.jpg",
+        "image_caption": [
+            "Moving GT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            176,
+            232,
+            246,
+            286
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/64e7e18e954134102460c267fd341a992df289e493aad2bed3fcc0c13f3c5175.jpg",
+        "image_caption": [
+            "eannnnnne nnnnne eannnnnne",
+            "VTN",
+            "Regular",
+            "Figure 4. Qualitative comparison of different volume-preserving (VP) methods trained on the Liver Tumor Segmentation (LiTS) dataset. The left side of the figure shows two sets of images: Fixed and Ground Truth (GT), and Moving and GT. The first row of the figure displays the warped moving image, while the second row illustrates the organ outlines in green and red for the moving and fixed images, respectively. The yellow overlay highlights the tumors. Our proposed volume-preserving (VP) method ensures the preservation of tumor volume while aligning the images, as demonstrated by reduced number of visible changes in tumor size. In the third row, the Jacobian matrix of the deformation field is visualized. The green and red lines represent the organ and tumor outlines, respectively. The white areas indicate a large Jacobian, which corresponds to a more significant change in volume. The method without volume-preserving loss demonstrates a larger white area in the tumor, indicating a greater volume change of tumor volume. The last row of the figure displays the deformation field. Due to space limitations, qualitative results for BraTS20 dataset are provided in the supplemental material."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            303,
+            99,
+            372,
+            325
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/26233b76521c2fd6aea02034266fb08e635933235c2f5a87bc53c8351c632a23.jpg",
+        "image_caption": [
+            "VTN",
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            374,
+            99,
+            442,
+            325
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f8e1a32aa61351d4c5d1c93e39ef98466c1ee3c707d4c8a1e3c41a95c57adb31.jpg",
+        "image_caption": [
+            "VXM",
+            "Regular"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            446,
+            99,
+            514,
+            325
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/258b2ba85af0dd80532b0a7eeba60664abc8102b7adde7e0677b65dc3c3b4db7.jpg",
+        "image_caption": [
+            "VXM",
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            516,
+            99,
+            584,
+            325
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/75787e95360e42407cf70ede43f41ee18f9ebe0b5a60e2dbcd184dd36a6858ad.jpg",
+        "image_caption": [
+            "Trans.",
+            "Regular"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            584,
+            99,
+            653,
+            325
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/13bbbab58f2c237f5207e5501ee9c2bdf3621ed2347b5de9ace53472bf452c20.jpg",
+        "image_caption": [
+            "Trans.",
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            655,
+            99,
+            723,
+            325
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/57311016b929b767ece3f36a86addd2f31f6c392c98128a9fc3a9ec2345870f9.jpg",
+        "image_caption": [
+            "LapIRN",
+            "Regular"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            727,
+            99,
+            795,
+            325
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/45b405b4edf9e1156fa4cb0bd953473ee8d7311edda6a4f32fc201939e30ea29.jpg",
+        "image_caption": [
+            "LapIRN",
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            800,
+            99,
+            870,
+            325
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/b07572ddd504240b53b2bcfaf87deccec765aafe2247e6352494f7cd6c1d3954.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 1. Quantitative comparison with state-of-the-art registration methods. Our strategy was applied to two datasets, namely LiTS17 and BraTS20, and three network architectures In addition, five metrics were employed for the evaluation process. The results of the comparison indicate that our proposed strategy can be well generalized to other state-of-the-art methods to preserve tumor volumes while achieving comparable or better performance on other metrics."
+        ],
+        "table_body": "<table><tr><td rowspan=\"2\">Network</td><td rowspan=\"2\">Methods</td><td colspan=\"5\">LiTs17</td><td colspan=\"5\">BraTS20</td></tr><tr><td>Dice ↑</td><td>Lm. Dist ↓</td><td>|Jθ| &lt; 0(%) ↓</td><td>Std. |Jθ| ↓</td><td>STSR ↓</td><td>Dice ↑</td><td>Lm. Dist ↓</td><td>|Jθ| &lt; 0(%) ↓</td><td>Std. |Jθ| ↓</td><td>STSR ↓</td></tr><tr><td>SyN</td><td>-</td><td>0.850</td><td>12.36</td><td>0.192</td><td>0.221</td><td>1.564</td><td>0.9560</td><td>4.01</td><td>0.276</td><td>0.226</td><td>1.512</td></tr><tr><td>B-Spline</td><td>-</td><td>0.853</td><td>14.22</td><td>0.003</td><td>0.288</td><td>1.330</td><td>0.9601</td><td>4.51</td><td>0.001</td><td>0.171</td><td>1.622</td></tr><tr><td rowspan=\"2\">VTN</td><td>Regular</td><td>0.912</td><td>10.46</td><td>3.545</td><td>2.048</td><td>2.330</td><td>0.9784</td><td>2.98</td><td>0.396</td><td>0.725</td><td>2.263</td></tr><tr><td>Ours</td><td>0.908</td><td>10.77</td><td>2.956</td><td>1.124</td><td>1.260</td><td>0.9738</td><td>3.10</td><td>0.596</td><td>0.285</td><td>1.416</td></tr><tr><td rowspan=\"2\">VXM</td><td>Regular</td><td>0.860</td><td>12.62</td><td>4.191</td><td>1.699</td><td>1.738</td><td>0.9701</td><td>3.56</td><td>0.024</td><td>0.448</td><td>1.929</td></tr><tr><td>Ours</td><td>0.857</td><td>12.43</td><td>1.681</td><td>0.691</td><td>1.241</td><td>0.9741</td><td>3.60</td><td>0.030</td><td>0.226</td><td>1.463</td></tr><tr><td rowspan=\"2\">TransMorph</td><td>Regular</td><td>0.859</td><td>13.35</td><td>6.377</td><td>1.906</td><td>1.785</td><td>0.9710</td><td>4.01</td><td>0.164</td><td>0.468</td><td>1.868</td></tr><tr><td>Ours</td><td>0.856</td><td>13.19</td><td>4.084</td><td>0.827</td><td>1.302</td><td>0.9758</td><td>3.98</td><td>0.019</td><td>0.194</td><td>1.319</td></tr><tr><td rowspan=\"2\">LapIRN</td><td>Regular</td><td>0.8933</td><td>10.23</td><td>5.301</td><td>1.788</td><td>2.360</td><td>0.9750</td><td>3.40</td><td>3.020</td><td>0.301</td><td>1.889</td></tr><tr><td>Ours</td><td>0.8893</td><td>10.55</td><td>4.120</td><td>0.988</td><td>1.645</td><td>0.9743</td><td>3.36</td><td>0.040</td><td>0.105</td><td>1.224</td></tr></table>",
+        "bbox": [
+            78,
+            500,
+            890,
+            655
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Landmarks were also annotated by three experts in the field to evaluate the registration accuracy of our strategy.",
+        "bbox": [
+            75,
+            736,
+            468,
+            767
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3. Main Results",
+        "text_level": 1,
+        "bbox": [
+            76,
+            782,
+            215,
+            797
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Quantitative Comparison. In Table 1, we present a comprehensive comparison of our proposed strategy with two traditional methods and three learning-based methods with different network architectures (CNN, cascaded CNN and hybrid transformer-CNN). We compare the methods using a similarity-based manner (regular) following the settings",
+        "bbox": [
+            75,
+            809,
+            468,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "in [38], as well as our volume-preserving manners. Our results demonstrate that our proposed strategy outperforms all other methods in terms of the STSR metrics, while achieving comparable results in Dice and Landmark Dist, and often better results in folding detection  $(|J_{\\theta}| < 0)$  and smoothness quality  $(Std. |J_{\\theta}|)$ . Furthermore, we observe that registration networks that achieve better performance in terms of previous metrics such as Dice tend to have worse performance in preserving tumor volumes (STSR). This observation highlights the need to balance these two objectives",
+        "bbox": [
+            496,
+            736,
+            893,
+            888
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "21214",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/5219a19dba0415b543f916c89125ddb38cd37e3049ede55a640d9c82564896d2.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 2. Effectiveness of adaptive volume-preserving loss in unsupervised registration frameworks on LiTs17 dataset. Three networks, namely VTN in [38], VXM in [5], and TransMorph in [6], were applied. The methods used in the study were regular unsupervised (first row), weighted similarity loss only based on the ground truth (GT) mask (second row), volume-preserving loss based on GT(third row), volume-preserving loss on the whole organ based on estimation (fourth row), and on the tumor based on estimation (ours)."
+        ],
+        "table_body": "<table><tr><td rowspan=\"2\">#</td><td rowspan=\"2\">Network</td><td rowspan=\"2\">Unsupervised</td><td rowspan=\"2\">Adaptive</td><td rowspan=\"2\">Region</td><td colspan=\"5\">LiTs17</td></tr><tr><td>Dice ↑</td><td>Lm. Dist ↓</td><td>|Jθ| &lt; 0(%) ↓</td><td>Std. |Jθ| ↓</td><td>STSR ↓</td></tr><tr><td>1</td><td rowspan=\"5\">VTN</td><td>✓</td><td>-</td><td>-</td><td>0.912</td><td>10.46</td><td>3.545</td><td>2.048</td><td>2.330</td></tr><tr><td>2</td><td>×</td><td>-</td><td>-</td><td>0.914</td><td>10.22</td><td>3.241</td><td>1.165</td><td>1.648</td></tr><tr><td>3</td><td>×</td><td>×</td><td>Tumor</td><td>0.911</td><td>10.74</td><td>2.883</td><td>1.017</td><td>1.205</td></tr><tr><td>4</td><td>✓</td><td>×</td><td>Organ</td><td>0.904</td><td>11.07</td><td>3.194</td><td>1.015</td><td>1.640</td></tr><tr><td>5</td><td>✓</td><td>✓</td><td>Organ</td><td>0.908</td><td>10.77</td><td>2.956</td><td>1.124</td><td>1.260</td></tr><tr><td>6</td><td rowspan=\"5\">VXM</td><td>✓</td><td>-</td><td>-</td><td>0.860</td><td>12.62</td><td>4.191</td><td>1.699</td><td>1.738</td></tr><tr><td>7</td><td>×</td><td>-</td><td>-</td><td>0.863</td><td>12.22</td><td>1.603</td><td>0.691</td><td>1.417</td></tr><tr><td>8</td><td>×</td><td>×</td><td>Tumor</td><td>0.859</td><td>12.38</td><td>1.567</td><td>0.698</td><td>1.291</td></tr><tr><td>9</td><td>✓</td><td>×</td><td>Organ</td><td>0.856</td><td>12.78</td><td>1.635</td><td>0.675</td><td>1.407</td></tr><tr><td>10</td><td>✓</td><td>✓</td><td>Organ</td><td>0.857</td><td>12.43</td><td>1.681</td><td>0.691</td><td>1.241</td></tr><tr><td>11</td><td rowspan=\"5\">TransMorph</td><td>✓</td><td>-</td><td>-</td><td>0.859</td><td>13.35</td><td>6.337</td><td>1.906</td><td>1.785</td></tr><tr><td>12</td><td>×</td><td>-</td><td>-</td><td>0.863</td><td>13.10</td><td>4.923</td><td>1.540</td><td>1.540</td></tr><tr><td>13</td><td>×</td><td>×</td><td>Tumor</td><td>0.858</td><td>13.21</td><td>4.401</td><td>0.877</td><td>1.227</td></tr><tr><td>14</td><td>✓</td><td>×</td><td>Organ</td><td>0.849</td><td>13.02</td><td>4.651</td><td>0.845</td><td>1.454</td></tr><tr><td>15</td><td>✓</td><td>✓</td><td>Organ</td><td>0.856</td><td>13.19</td><td>4.084</td><td>0.827</td><td>1.302</td></tr></table>",
+        "bbox": [
+            102,
+            88,
+            867,
+            325
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "in order to achieve competitive results.",
+        "bbox": [
+            78,
+            393,
+            331,
+            407
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Qualitative Comparison. Figure 5 presents a qualitative comparison of our proposed strategy with different state-of-the-art methods. Comparing the \"VTN/VXM/Trans./LapIRN regular\" column and the \"VTN/VXM/Trans./LapIRN ours\" column, it is evident that our method outperforms other methods in terms of volume preservation in the warped image  $(1_{\\mathrm{st}}$  row) and the corresponding deformation field  $(4_{\\mathrm{th}}$  row). Furthermore, in terms of visualization of the registration accuracy and the overlay of the tumor  $(2_{\\mathrm{nd}}$  row), our proposed method achieves better volume preservation with comparable accuracy compared to other methods. Lastly, the Jacobian determinant of the deformation field  $(3_{\\mathrm{rd}}$  row) reveals that regular methods tend to have a larger Jacobian in tumor regions, indicating greater volume change, compared to our method. This highlights the effectiveness of our volume-preserving approach in preserving tumor volumes while aligning the images.",
+        "bbox": [
+            78,
+            409,
+            467,
+            679
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.4. Ablation study",
+        "text_level": 1,
+        "bbox": [
+            78,
+            694,
+            223,
+            709
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Volume-Preserving Losses. In Table 2, we present the results of our evaluation of the effectiveness of various versions of volume-preserving losses. For each sub-table, the first row (#1, #6, and #11) represents using only the similarity loss, while the second row (#2, #7, and #12) utilizes the ground truth as a weighted similarity loss. The third row (#3, #8, and #13) uses the ground truth binary mask as a tumor mask, while the fourth row (#4, #9, and #14) uses the mask of the entire organ as a tumor mask. The fifth row (#5, #10, and #15) represents our proposed strategy.",
+        "bbox": [
+            78,
+            718,
+            467,
+            868
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "By comparing the first and second rows of each sub-table with the fifth row, we observe that our proposed volume-",
+        "bbox": [
+            78,
+            869,
+            467,
+            898
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "preserving loss can effectively preserve details and achieve superior performance across various metrics. The improvement from the third and fourth rows to the fifth row indicates the importance of adaptiveness in preserving the volume of the organ and tumor, respectively. It should be noted that we do not require pixel-wise ground truth for tumor volume preservation.",
+        "bbox": [
+            501,
+            393,
+            890,
+            497
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Robustness to Tumor Mask Estimation. In Figure 5, we showcase the robustness of our proposed strategy in the presence of noisy ground truth segmentation masks. To evaluate this, we randomly select points within organs to ensure that tumors have a certain Dice score with respect to the ground truth segmentation of tumors. As depicted in the figure, despite the variations in the noisy ground truth segmentation masks, our strategy's performance remains consistent. Specifically, when the tumor mask has a Dice score above a certain threshold, such as 0.15, the estimated tumor mask consistently achieves a Dice score of approximately 0.17. In contrast, using the whole organ as mask in volume-preserving loss causes degradation in performance as it only has a dice score of roughly 0.05. This highlights the necessity and efficiency of our tumor mask estimation in stage one.",
+        "bbox": [
+            503,
+            500,
+            890,
+            738
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusions",
+        "text_level": 1,
+        "bbox": [
+            503,
+            755,
+            624,
+            768
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In summary, this work addressed the problem of preserving tumor volumes while maximizing similarity during learning-based registration, which is crucial for tracking tumor growth. We proposed a two-stage unsupervised strategy that effectively preserves the size ratio of tumors, without requiring extra networks. Our approach involves generating a soft tumor mask in the first stage using a similarity-based registration network and incorporating",
+        "bbox": [
+            503,
+            780,
+            890,
+            898
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "21215",
+        "bbox": [
+            478,
+            946,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/3fa7600e12f5561e2b1165278e52e24243ef91cdb3f409a4ecb8e16d3efe6156.jpg",
+        "image_caption": [
+            "Figure 5. Robustness of our tumor mask estimation. We evaluated the robustness of our tumor mask estimation by running our strategy using RCN on the LiTS datasets with different noisy ground truth (GT) masks as tumor masks, each with specific Dice scores with tumors. We plotted the results on a graph with the Dice score on the x-axis and the STSR in the test set on the y-axis (where smaller values are better)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            93,
+            423,
+            243
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "an adaptive volume-preserving loss and a weighted similarity loss in the second stage to improve registration performance. To evaluate our strategy's performance, we introduced a new metric, the Square Tumor Size Ratio (STSR), which measures the preservation of tumor volume. Our proposed strategy was evaluated on various networks and demonstrated superior performance in tumor volume preservation, while achieving comparable results in other metrics. This highlights the effectiveness of our approach in tracking tumor growth and its ability to generalize well to different registration networks. Additionally, we noted that the volume preservation approach could be extended to other lesions besides tumors.",
+        "bbox": [
+            75,
+            375,
+            468,
+            569
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement: This work was supported by the HKSAR Innovation and Technology Commission (ITC) under ITF Project MHP/109/19, the National Natural Science Foundation in China under Grant 62022010, the Beijing Natural Science Foundation Haidian District Joint Fund in China under Grant L222032 and the Fundamental Research Funds for the Central Universities of China from the State Key Laboratory of Software Development Environment in Beihang University in China.",
+        "bbox": [
+            75,
+            585,
+            468,
+            723
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "21216",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            173,
+            104
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Moab Arar, Yiftach Ginger, Dov Danon, Amit H Bermano, and Daniel Cohen-Or. Unsupervised multi-modal image registration via geometry preserving image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13410–13419, 2020. 3",
+            "[2] Brian B Avants, Charles L Epstein, Murray Grossman, and James C Gee. Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain. Medical image analysis, 12(1):26-41, 2008. 6",
+            "[3] Brian B Avants, Nick Tustison, Gang Song, et al. Advanced normalization tools (ants). Insight j, 2(365):1-35, 2009. 6",
+            "[4] Bhakti Baheti, Diana Waldmannstetter, Satrajit Chakrabarty, Hamed Akbari, Michel Bilello, Benedikt Wiestler, Julian Schwarting, Evan Calabrese, Jeffrey Rudie, Syed Abidi, et al. The brain tumor sequence registration challenge: establishing correspondence between pre-operative and follow-up mri scans of diffuse glioma patients. arXiv preprint arXiv:2112.06979, 2021. 2, 3",
+            "[5] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. An unsupervised learning model for deformable medical image registration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9252-9260, 2018. 1, 2, 4, 6, 8",
+            "[6] Junyu Chen, Eric C Frey, Yufan He, William P Segars, Ye Li, and Yong Du. Transmorph: Transformer for unsupervised medical image registration. Medical image analysis, 82:102615, 2022. 2, 3, 4, 6, 8",
+            "[7] Jiashun Chen, Donghuan Lu, Yu Zhang, Dong Wei, Munan Ning, Xinyu Shi, Zhe Xu, and Yefeng Zheng. Deformer: Towards displacement field learning for unsupervised medical image registration. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2022: 25th International Conference, Singapore, September 18-22, 2022, Proceedings, Part VI, pages 141-151. Springer, 2022. 2, 3, 6",
+            "[8] Xiang Chen, Yan Xia, Nishant Ravikumar, and Alejandro F Frangi. Joint segmentation and discontinuity-preserving deformable registration: Application to cardiac cine-mr images. arXiv preprint arXiv:2211.13828, 2022. 2",
+            "[9] Nicha Chitphakdithai and James S. Duncan. Non-rigid registration with missing correspondences in preoperative and postresection brain images. 3",
+            "[10] Patrick Christ. Lits – liver tumor segmentation challenge (lits17). 2017. 6",
+            "[11] Olivier Clatz, Hervé Delingette, Ion-Florin Talos, Alexandra J. Golby, Ron Kikinis, Ferenc A. Jolesz, Nicholas Ayache, and Simon K. Warfield. Robust nonrigid registration to capture brain shift from intraoperative mri. IEEE Transactions on Medical Imaging, 2005. 3",
+            "[12] Ahmed Elazab, Yousry M Abdulazeem, Ahmed M Anter, Qingmao Hu, Tianfu Wang, and Baiying Lei. Macroscopic cerebral tumor growth modeling from medical images: A review. IEEE Access, 6:30663-30679, 2018. 2, 3"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[13] Xu Han, Xiao Yang, Stephen R. Aylward, Roland Kwitt, and Marc Niethammer. Efficient registration of pathological images: A joint pca/image-reconstruction approach. arXiv: Computer Vision and Pattern Recognition, 2017. 3",
+            "[14] Malte Hoffmann, Benjamin Billot, Douglas N Greve, Juan Eugenio Iglesias, Bruce Fischl, and Adrian V Dalca. Synth morph: learning contrast-invariant registration without acquired images. IEEE transactions on medical imaging, 41(3):543-558, 2021. 1, 2, 3",
+            "[15] Malte Hoffmann, Andrew Hoopes, Douglas N Greve, Bruce Fischl, and Adrian V Dalca. Anatomy-aware and acquisition-agnostic joint registration with synthmorph. arXiv preprint arXiv:2301.11329, 2023. 1, 2, 3",
+            "[16] Boah Kim, Dong Hwan Kim, Seong Ho Park, Jieun Kim, June-Goo Lee, and Jong Chul Ye. Cyclesmorph: cycle consistent unsupervised deformable image registration. Medical image analysis, 71:102036, 2021. 2, 3",
+            "[17] Stefan Klein, Marius Staring, Keelin Murphy, Max A Viergever, and Josien PW Pluim. Elastix: a toolbox for intensity-based medical image registration. IEEE transactions on medical imaging, 29(1):196-205, 2009. 6",
+            "[18] Dongjin Kwon, Marc Niethammer, Hamed Akbari, Michel Bilello, Christos Davatzikos, and Kilian M. Pohl. Portr: Preoperative and post-recurrence brain tumor registration. IEEE Transactions on Medical Imaging, 2014. 3",
+            "[19] Dongjin Kwon, Ke Zeng, Michel Bilello, and Christos Davatzikos. Estimating patient specific templates for preoperative and follow-up brain tumor registration. Lecture Notes in Computer Science, 2015. 3",
+            "[20] Fei Lyu, Mang Ye, Andy J Ma, Terry Cheuk-Fung Yip, Grace Lai-Hung Wong, and Pong C Yuen. Learning from synthetic ct images via test-time training for liver tumor segmentation. IEEE transactions on medical imaging, 41(9):2510-2520, 2022. 3",
+            "[21] Mingyuan Meng, Lei Bi, Dagan Feng, and Jinman Kim. Brain tumor sequence registration with non-iterative coarse-to-fine networks and dual deep supervision. arXiv preprint arXiv:2211.07876, 2022. 3",
+            "[22] Bjoern H Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al. The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging, 34(10):1993-2024, 2014. 6",
+            "[23] Tony CW Mok and Albert CS Chung. Large deformation diffeomorphic image registration with laplacian pyramid networks. 3, 6",
+            "[24] Tony CW Mok and Albert CS Chung. Unsupervised deformable image registration with absent correspondences in pre-operative and post-recurrence brain tumor mri scans. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2022: 25th International Conference, Singapore, September 18-22, 2022, Proceedings, Part VI, pages 25-35. Springer, 2022. 2, 3",
+            "[25] Torsten Rohlfing, Calvin R Maurer, David A Bluemke, and Michael A Jacobs. An alternating-constraints algorithm for volume-preserving non-rigid registration of contrast-"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "21217",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "enhanced mr breast images. In Biomedical Image Registration: Second International Workshop, WBIR 2003, Philadelphia, PA, USA, June 23-24, 2003. Revised Papers 2, pages 291-300. Springer, 2003. 3",
+            "[26] Daniel Rueckert, Luke I Sonoda, Carmel Hayes, Derek LG Hill, Martin O Leach, and David J Hawkes. Nonrigid registration using free-form deformations: application to breast mr images. IEEE transactions on medical imaging, 18(8):712-721, 1999. 6",
+            "[27] Siyuan Shan, Wen Yan, Xiaqing Guo, Eric I Chang, Yubo Fan, Yan Xu, et al. Unsupervised end-to-end learning for deformable medical image registration. arXiv preprint arXiv:1711.08608, 2017. 6",
+            "[28] Shuwei Shao, Zhongcai Pei, Weihai Chen, Wentao Zhu, Xingming Wu, and Baochang Zhang. A multi-scale unsupervised learning for deformable image registration. International Journal of Computer Assisted Radiology and Surgery, pages 1-10, 2022. 3",
+            "[29] Jiacheng Shi, Yuting He, Youyong Kong, Jean-Louis Coa-trieux, Huazhong Shu, Guanyu Yang, and Shuo Li. Xmorpher: Full transformer for deformable medical image registration via cross attention. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2022: 25th International Conference, Singapore, September 18-22, 2022, Proceedings, Part VI, pages 217-226. Springer, 2022. 2, 3",
+            "[30] Christine Tanner, Julia A. Schnabel, Andreas Degenhard, Andy D. Castellano-Smith, Carmel Hayes, Martin O. Leach, D. R. Hose, Derek L. G. Hill, and David J. Hawkes. Validation of volume-preserving non-rigid registration: Application to contrast-enhanced mr-mammography. Lecture Notes in Computer Science, 2002. 3",
+            "[31] Di Wang, Yue Pan, Oguz C. Durmeric, Joseph M. Reinhardt, Eric A. Hoffman, Joyce D. Schroeder, and Gary E. Christensen. Plosl: Population learning followed by one shot learning pulmonary image registration using tissue volume preserving and vesselness constraints. Medical Image Analysis, 2022. 3",
+            "[32] Yibo Wang, Wen Qian, Mengqi Li, and Xuming Zhang. A transformer-based network for deformable medical image registration. In Artificial Intelligence: Second CAAI International Conference, CICAI 2022, Beijing, China, August 27-28, 2022, Revised Selected Papers, Part I, pages 502-513. Springer, 2022. 1, 2",
+            "[33] Marek Wodzinski, Artur Jurgas, Niccolo Marini, Manfredo Atzori, and Henning Muller. Unsupervised method for intra-patient registration of brain magnetic resonance images based on objective function weighting by inverse consistency: Contribution to the brats-reg challenge. arXiv preprint arXiv:2211.07386, 2022. 3",
+            "[34] Qingsong Yao, Li Xiao, Peihang Liu, and S Kevin Zhou. Label-free segmentation of Covid-19 lesions in lung ct. IEEE transactions on medical imaging, 40(10):2808-2819, 2021. 3",
+            "[35] Xiaoman Zhang, Weidi Xie, Chaoqin Huang, Ya Zhang, Xin Chen, Qi Tian, and Yanfeng Wang. Self-supervised tumor segmentation with sim2real adaptation. IEEE Journal of Biomedical and Health Informatics, 2023. 3"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            900
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[36] Zijun Zhang. Improved adam optimizer for deep neural networks. In 2018 IEEE/ACM 26th international symposium on quality of service (IWQoS), pages 1-2. IEEE, 2018. 6",
+            "[37] Bowen Zhao, G. Christensen, Joo Hyun Song, Yue Pan, Sarah E. Gerard, Joseph M. Reinhardt, Kaifang Du, Taylor J. Patton, John E. Bayouth, and Geoffrey D. Hugo. Tissuevolume preserving deformable image registration for 4dct pulmonary images. 2016 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 481-489, 2016. 3",
+            "[38] Shengyu Zhao, Yue Dong, Eric I Chang, Yan Xu, et al. Recursive cascaded networks for unsupervised medical image registration. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10600-10610, 2019. 1, 2, 3, 4, 5, 6, 7, 8",
+            "[39] Shengyu Zhao, Tingfung Lau, Ji Luo, I Eric, Chao Chang, and Yan Xu. Unsupervised 3d end-to-end medical image registration with volume tweening network. IEEE journal of biomedical and health informatics, 24(5):1394-1404, 2019. 6"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            371
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "21218",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2023/Preserving Tumor Volumes for Unsupervised Medical Image Registration/dbd41759-9a48-41b9-9206-f4ea354bc2a2_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab59fc64231b15ecf9f05da79caf857b48309dfd1d77cda550a77fb0fe627558
+size 1086763

2023/Preserving Tumor Volumes for Unsupervised Medical Image Registration/full.md

@@ -0,0 +1,351 @@
+# Preserving Tumor Volumes for Unsupervised Medical Image Registration
+
+Qihua Dong $^{1}$ , Hao Du $^{1}$ , Ying Song $^{2}$ , Yan Xu $^{3*}$ , Jing Liao $^{1*}$
+
+$^{1}$ Department of Computer Science, City University of Hong Kong
+
+$^{2}$ National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College  $^{3}$ School of Biological Science and Medical Engineering, Beihang University
+
+dongqh078@gmail.com, haodu8-c@my.cityu.edu.hk, jingliao@cityu.edu.hk;
+
+songying1770@hotmail.com
+
+xuyan04@gmail.com
+
+# Abstract
+
+Medical image registration is a critical task that estimates the spatial correspondence between pairs of images. However, current traditional and deep-learning-based methods rely on similarity measures to generate a deforming field, which often results in disproportionate volume changes in dissimilar regions, especially in tumor regions. These changes can significantly alter the tumor size and underlying anatomy, which limits the practical use of image registration in clinical diagnosis. To address this issue, we have formulated image registration with tumors as a constraint problem that preserves tumor volumes while maximizing image similarity in other normal regions. Our proposed strategy involves a two-stage process. In the first stage, we use similarity-based registration to identify potential tumor regions by their volume change, generating a soft tumor mask accordingly. In the second stage, we propose a volume-preserving registration with a novel adaptive volume-preserving loss that penalizes the change in size adaptively based on the masks calculated from the previous stage. Our approach balances image similarity and volume preservation in different regions, i.e., normal and tumor regions, by using soft tumor masks to adjust the imposition of volume-preserving loss on each one. This ensures that the tumor volume is preserved during the registration process. We have evaluated our strategy on various datasets and network architectures, demonstrating that our method successfully preserves the tumor volume while achieving comparable registration results with state-of-the-art methods. Our codes is available at: https://dddraxxx.github.io/Volume-Preserving-Registration/.
+
+# 1. Introduction
+
+![](images/3c773565a73bf98b9747bd76d71e915cf6c3c99ab9b4b4ae399abdb539649b8e.jpg)
+
+![](images/657ff349d7c2a03bbddd4c7a8472927f39480b62c2a453cd8914b61720acb5f5.jpg)
+
+![](images/5501f7c771a1a7977cd66eb7f46958891ea67ce1d7b7ca64e1a6c7819722c45a.jpg)  
+Fixed Image  
+Moving Image  
+Figure 1. Red boxes represent the location of tumors in the moving image before warped. In image registration with tumors, similarity-based registration typically leads to significant alterations in tumor size while our volume-preserving registration is capable of preserving tumor anatomy.
+
+![](images/cf51c67221ba2f991df800d0691c3d451ed0df7e8b666712eb207107aaa1d580.jpg)  
+Similarity-Based Registration  
+Volume-Preserving Registration
+
+Deformable image registration is a fundamental task that estimates non-linear spatial correspondences between two images. It is useful for medical image studies that involve 3D images of organs or tissues, such as MR brain scans and CT liver scans. Recently, a wide range of deep-learning-based methods are proposed in the field, with better performance and obvious speed-ups in inference time compared to the traditional registration methods [5, 38, 32, 14, 15].
+
+Currently, most of the learning-based methods [5, 38] train a registration model to achieve high similarity in either
+
+image intensity [38, 5] or anatomy label maps [14]. While this is useful in various medical applications, such as atlas-based segmentation and image fusion [8], problem emerges, however, when studying images with tumors [4, 12, 24]. Tracking tumor growth is a core task in cancer treatment, which can be used to evaluate the outcomes of radiotherapy and chemotherapy, and plan optimal postoperative treatment [12]. The process requires registration to align the anatomy of images from different periods of the patient, while preserving tumor properties e.g., the size. Traditional models using regular registration have struggled to accomplish this task [12]. We have observed that this remains an issue for current learning-based methods, and a viable solution has yet to be proposed.
+
+As shown in Figure 1, the volume of tumors significantly reduces when using popular unsupervised registration networks to align tumor images with atlas images. The issue is widespread since mainstream registration networks focus on optimizing the similarity between image pairs while ignoring tumor regions. Tumor regions often lack corresponding parts, and even for images of the same patient, the size and location of tumors may vary greatly. This can happen because tumors can change shape, size, location, or disappear over time. Therefore, in deformable settings, the size of tumors will change disproportionately to maximize the image similarity. This problem may be more severe when the registration network has better warping ability, leading to better performance in evaluations from previous works, as confirmed by our experiments in Section 4. The disproportionate change of tumor volume in deformable registration is lethal for clinical evaluation of tumor growth and greatly hinders the application of registration in clinics.
+
+In this paper, we re-formulate the deformable registration problem as a constraint problem that preserves the properties of tumors in the original images, i.e., the shape and the size of the tumor, while maximizing image similarity in non-tumor regions. For instance, in tumor image registration with atlas, besides the alignment of anatomical structures, the tumor should have similar morphing behaviours as its surrounding organs, i.e., the size of it should change proportionally to the size of the organ. However, this poses two major challenges for our strategy. Firstly, due to the limited availability of annotated data, most existing registration methods resort to unsupervised learning. Therefore, a key question is how to identify tumor regions in the unsupervised setting. Second, after the tumor regions have been identified, the challenge still remains of preserving tumor volumes while simultaneously maximizing similarity in other regions. Finding a way to balance these two competing objectives is crucial for achieving effective tumor image registration.
+
+To tackle the challenges, we propose a novel two-stage strategy that operates in an unsupervised manner. In the first
+
+stage, we leverage a similarity-based registration network to identify regions that undergo excessive volume change during registration. Such regions are indicative of the presence of tumors, as tumors tend to exhibit larger changes in volume due to a lack of correspondence between image pairs. The outcome of the first stage is a soft tumor mask that distinguishes normal and possible tumor regions. In the second stage, we introduce an adaptive volume-preserving loss to train a volume-preserving registration network. Based on the soft tumor mask obtained from the first stage, this loss is designed to adaptively penalize volume changes in different regions of the image. Possible tumor regions are assigned larger loss weights, and similarity loss weights are adjusted accordingly in the opposite direction. The application of the adaptive volume-preserving loss effectively balances similarity and volume preservation in different parts of the image.
+
+It should be noted that the soft masks from stage one do not need to be highly accurate to achieve desirable results, as our stage two is robust, which is demonstrated by our ablation experiments in section 4. In addition to evaluating the robustness of our proposed strategy on various datasets and frameworks, including CNN and transformer, as well as different types of imaging data such as brain and abdominal, MRI and CT images, we have also introduced the Square Tumor Size Ratio (STSR) metric to measure the preservation of tumor volumes. The results demonstrate that our methods achieve comparable warping performance metrics such as Dice coefficient and average landmark distances while effectively preserving tumor properties as measured by the STSR.
+
+To conclude, our contribution includes: (1) We have reformulated image registration with tumors as a constraint problem that preserves tumor volumes while maximizes image similarity in other normal regions. (2) We have designed a two-stage process that the first stage performs regular registration to estimate tumor regions, and the second stage performs volume-preserving registration on tumor regions based on the adaptive volume-preserving loss. (3) We have proposed the STSR metric to measure the preservation of tumor volumes in registration, and have evaluated the effectiveness of our strategy on different network architectures and various datasets.
+
+# 2. Related Work
+
+# 2.1. Learning-Based Deformable Image Registration
+
+Compared to conventional registration methods, deep networks have achieved remarkable registration accuracy and speed using various network structures, i.e., CNN and Transformer, in applications like atlas-image registration [29, 7, 16, 6, 32, 38, 5, 14, 15, 8]. They generally
+
+adopts unsupervised learning, utilizing image similarity on either image intensities [29, 7, 6, 38] or synthesized label maps [14, 15] as well as smoothness regularization to supervise the registration network [7, 6, 14, 15, 38]. Advancements are often made through the introduction of novel regularization such as cycle consistency [16] and geometry preservation [1], or through innovative network structures like multi-scale architectures [28, 23]. However, existing methods do not adequately consider cases where tumors exist. Tumors often have no corresponding part between image pairs, and therefore, optimizing for similarity can excessively change their volumes, which limits the application of registration networks in critical clinical applications like cancer treatment. To the best of our knowledge, we are the first to study this problem with learning-based methods and propose a metric to measure volume preservation specifically for tumors.
+
+# 2.2. Deformable Image Registration with Tumors
+
+Conventional registration models often resolve registering images with tumor by (1) excluding the similarity measure on tumors [11, 18], which requires either massive manual segmentation [11] or initial seed to explicitly model tumor-growth [12, 18]. (2) jointly segmenting and registering images [9], which is time-consuming since it need to iteratively detect dissimilarity in image intensities and perform resection and retraction. (3) reconstructing quasi-normal images from pathological ones [13, 19], which is also slow since the reconstruction need to be iteratively improved.
+
+Recently, learning-based methods for registration with tumors [24, 21, 33] are proposed for the release of BraTS-Reg dataset [4]. They aim to improve the spatial correspondence for non-tumor regions, thus either directly annotates landmarks to supervise registration [21] or introducing invertibility of the registration to mask non-inverted regions and ensure high correspondence in others [24, 33]. However, these methods assume the absence of tumors in one image of the registered pair, which may not be applicable in practical applications where we need to compare two tumor images from one patient to track the tumor growth. In this scenario, invertibility in tumors is still possible and thus, it is essential to consider the preservation of tumor properties during registration.
+
+# 2.3. Unsupervised Tumor Segmentation
+
+To ensure the preservation of tumor volumes during registration, it is crucial to first identify the locations of tumors. Current deep learning methods have made remarkable progress in unsupervised tumor segmentation, as demonstrated by recent studies [34, 35, 20]. These methods typically use GAN models or simulate tumors on healthy images to generate synthetic data, which is then used to
+
+train the segmentation network. Although the well-trained network can achieve high segmentation accuracy, it often requires complex training strategies and additional models or testing steps to achieve the desired performance, as noted in recent research [35, 20]. However, in our approach, we do not rely on sophisticated segmentation methods to accurately segment tumors. Instead, our experiments, as presented in Section 4, demonstrate that our tumor mask estimation is sufficient for our strategy to learn a volume-preserving registration for tumors.
+
+# 2.4. Volume-Preserving Deformable Image Registration
+
+Volume-preserving registration has been studied extensively to improve the accuracy of anatomical registration. In pulmonary CT image registration, for example, tissue intensity changes between inspiration and expiration phases can be utilized as prior knowledge, and adding a volume-preserving constraint has been shown to be effective [31, 37]. In other cases, such as when soft tissues are incompressible or when images come from different modalities, a volume-preserving constraint can also serve as a helpful regularization [25, 30]. Our work, instead, focuses on using volume-preservation to maintain the properties of tumors during image registration. The proposed strategy imposes this constraint adaptively on different parts of the images in order to preserve tumor volumes and ensure anatomical accuracy, while previous works have either preserved volumes for the entire images [25, 30] or can use simple methods to determine the regions to be preserved [31, 37].
+
+# 3. Methodology
+
+# 3.1. Preliminaries
+
+Problem Setting. Given a moving image  $I_{m}$  and fixed image  $I_{f}$  defined over  $d$ -dimensional (d=3 in this paper) space  $\Omega$ , deformable image registration aims to find a dense deformation filed  $\phi : \Omega \to \Omega$ . This field warps  $I_{m}$ , meaning it projects every pixel position  $x$  in warped image  $I_{w}$  to  $I_{m}$  to achieve alignment with  $I_{f}$ . In this paper, the task is to construct a registration network  $F_{\theta}$  that takes input  $I_{m}, I_{f}$  to predict the deformation field  $\phi$ .
+
+$$
+\phi = F _ {\theta} \left(I _ {m}, I _ {f}\right) \tag {1}
+$$
+
+$$
+I _ {w} (x) = I _ {m} (\phi (x) + x) \tag {2}
+$$
+
+$$
+I _ {w} \sim I _ {f} \tag {3}
+$$
+
+Evaluation Criteria. In general, a successful deformable image registration involves accurately aligning the anatomical structures while also ensuring a smooth deformation field. This is often achieved through the use of similarity-based objectives and regularization methods that promote smoothness. However, in the context of image registration with tumors, simply aligning the anatomy is not
+
+![](images/5cb5a1782ab687053e0200bb173bbb0a88140e957c50a252f9b6032916461fd0.jpg)  
+Figure 2. Main Framework. Our strategy consists of two stages. In the first stage, a soft mask indicating tumor regions is estimated by analyzing the Jacobian matrix of the deformation field obtained by a registration network pre-trained on similarity loss. In the second stage, the soft mask is used to guide the calculation of both volume-preserving and similarity losses for training the volume-preserving registration network. This deformation field estimated by the stage two ensures that both the image similarity and the preservation of tumor sizes.
+
+sufficient. Preserving the volume of the tumors is crucial for accurately tracking tumor growth, as previously discussed. To be specific, given the moving image  $I_{m}$  with the tumor  $T_{m}$  and its surrounding organ  $O_{m}$  and the warped moving image  $I_{w}$  with  $T_{w}$  and  $O_{w}$ , the morphing of tumor  $T_{m}$  should be similar to that of  $O_{m}$ . Thus in respect of the volume preservation, the change of tumor size  $|T_{m}|$  should be proportionate to the change of the containing organ  $|O_{m}|$ , i.e.:
+
+$$
+T S R \left(I _ {m}\right) = \frac {\left| T _ {m} \right|}{\left| O _ {m} \right|} \tag {4}
+$$
+
+$$
+T S R \left(I _ {m}\right) \approx T S R \left(I _ {w}\right) \tag {5}
+$$
+
+We therefore define the tumor size ratio (TSR) as  $\frac{|T|}{|O|}$ , which should be preserved after registration, and the re-formulated registration objective is now maximizing image similarity while preserving the volumes of tumors, i.e., the TSR.
+
+# 3.2. Overall Strategy
+
+The proposed strategy consists of two stages, as shown in Figure 2. The first stage estimates potential tumor regions and generates a soft mask for them. This is done by analyzing volume changes, specifically the Jacobian matrix, in the deformation field obtained by an existing similarity-based registration network. As tumor regions tend to undergo excessive volume change during registration, they can
+
+be identified through this analysis. In the second stage, we perform volume-preserving registration for tumors using an adaptive volume-preserving loss. This loss is designed to guide the volume-preserving process effectively and is associated with the tumor masks estimated in the first stage. Through the assignment of distinct weights to diverse regions, including both normal and tumor regions, it is possible to preserve tumor volumes while also ensuring a high degree of similarity in other regions.
+
+Importantly, our strategy can be applied to nearly any learning-based registration network without introducing any additional modules.
+
+# 3.2.1 Stage 1: Unsupervised Tumor Mask Estimation
+
+Following previous works [5, 6, 38], we first train a similarity-based registration network  $F_{\theta_1}$  with parameters  $\theta_1$  and fix its parameters. As shown in Figure 1, tumors in medical images can be detected by analyzing the change in volume through similarity-based registration. Our goal is to utilize this characteristic to facilitate the unsupervised estimation of the tumor mask for volume preservation during registration.
+
+To determine the volume change at voxel  $x$ , we calculate the determinant of the Jacobian matrix of the deformation field  $(J_{\theta})$  with model parameter  $\theta$ . Our objective is to ensure that the volume change in each point of the tumor region is
+
+similar to the volume change of the organ, which is an indicator of the organ's size change (as shown in Equation 6). To measure the distance between the two, we calculate the number of times one ratio exceeds another, denoted as  $D_{\theta}$ , using Equations 8 and 7 where  $D_{\theta}'(x)$  represents the relative volume change of voxel  $x$  with respect to organ areas.
+
+$$
+\frac {\left| T _ {w} \right|}{\left| O _ {w} \right|} / \frac {\left| T _ {m} \right|}{\left| O _ {m} \right|} \approx 1 \Rightarrow \frac {\left| T _ {w} \right|}{\left| T _ {m} \right|} / \frac {\left| O _ {w} \right|}{\left| O _ {m} \right|} \approx 1, \text {f r o m E q .} 5, \tag {6}
+$$
+
+$$
+D _ {\theta} ^ {\prime} (x) = \left| J _ {\theta (x)} \right| / \frac {\left| O _ {w} \right|}{\left| O _ {m} \right|}, x \in \Omega , \tag {7}
+$$
+
+$$
+D _ {\theta} (x) = \max  \left(D _ {\theta} ^ {\prime} (x), 1 / D _ {\theta} ^ {\prime} (x)\right), \tag {8}
+$$
+
+In practice, we estimate the organ masks  $O_{m}$  and  $O_{w}$  by warping a reference image with the ground truth (GT) organ segmentation. We randomly select the reference image from the training dataset, and the choice of reference image does not significantly impact the results because registration for organs achieves high accuracy, as evidenced by the dice value in table 1. By warping the reference image, we obtain the organ segmentation for the moving image  $O_{m}$ . We then calculate the organ segmentation for the warped image  $O_{w}$ .
+
+To obtain a soft tumor mask  $STM$  from the distance  $D_{\theta_1}$  defined in 8, we use a transformation function  $Tr$ . This function can be formulated as the following equation:
+
+$$
+S T M (x) = T r \left(D _ {\theta_ {1}} (x)\right), \tag {9}
+$$
+
+Here,  $D_{\theta_1}(x)$  represents the distance value at pixel  $x$ , and  $Tr$  transforms values from the range of  $[0,\infty]$  to the range of  $[0,1]$ . A value of 1 indicates the presence of a tumor, while a value of 0 indicates a normal region. In practical implementation, we use  $Tr(x) = \text{Sign}(5 \cdot (x - 1.5))$  since the sigmoid function is widely adopted for converting values to the range of  $[0,1]$  and to our experience, the size of tumor regions averagely changes around 1.5 times. However, the choice of transformation functions is not exclusive, as demonstrated in our experiments in Section 4. We also set the regions outside the organ mask to 0.
+
+We observed that some normal regions, particularly those close to the organ boundary, may also exhibit some volume changes after registration. As a result, these regions may be incorrectly included in the tumor mask, as shown in Figure 3 (a). To address this issue, we implemented a preregistration step to first align the edges and exclude volume changes caused by edge alignment. Specifically, in the preregistration step, we applied a bilateral filter to the moving image, which preserves the edges while smoothing the interior regions. We then used the registration network to register the filtered moving image with the fixed image in order to align their edges. After this, an unfiltered warped image was obtained by spatially transforming the unfiltered moving image with the deformation field. The second registration process was then performed on the unfiltered warped
+
+![](images/ae5b9c79bfc3f87b3cbe338bdb6f8cec305e153515e9eefa29531f91634fa745.jpg)  
+(a)Tumor Mask
+
+![](images/ee6ad0dfa4d53dd393bd7326b429ac9e80ca121224703f18d3a870c56fd5e7a6.jpg)  
+(b)Tumor Mask after pre-registration  
+Figure 3. Visualization of the soft tumor mask. (a) is the estimated soft tumor mask without pre-registration and (b) is the one with pre-registration. The gray colors with intensity variations denote the soft mask, while the green and red colors represent the ground truth segmentation for the liver and tumors, respectively.
+
+image, which focused on the registration of interior regions and was able to accurately estimate a tumor mask, as shown in Figure 3 (b). Therefore, this pre-registration approach effectively selects the tumor regions while filtering out the non-tumor regions.
+
+# 3.2.2 Stage 2: Volume-Preserving Registration for Tumors
+
+After obtaining the soft tumor mask  $STM$ , we can adaptively select tumor regions to maintain the size ratio. Here, we define the volume-preserving loss directly as the distance  $D_{\theta_2}$ , as shown in Equation 8. By minimizing the distance, we can achieve smaller volume changes, indicating better preservation of tumor volume. The volume-preserving loss is multiplied by the soft tumor mask. This ensures that the volume-preserving constraints are adaptively applied to the tumor regions and not to the normal regions, which can be formulated as:
+
+$$
+L _ {\mathrm {V P}} = \frac {1}{\left| I _ {m} \right|} \sum_ {x \in \Omega} D _ {\theta_ {2}} (x) \cdot S T M (x), \tag {10}
+$$
+
+Here, in the second stage, the registration network takes the soft tumor mask as input, which serves as an indicator for the regions that need to be preserved in terms of volume. We follow the similarity and registration loss formulation in [38], but we use a weighted similarity loss to adapt to the volume-preserving loss, where  $Cov$  is the covariance between two images  $I_1, I_2$ ,  $\hat{I}$  is the mean of  $I$ , and  $Cov'$  is the adapted version utilizing  $STM$  weight:
+
+$$
+\operatorname {C o v} ^ {\prime} \left[ I _ {1}, I _ {2} \right] = \frac {\sum \left(I _ {1} - \hat {I _ {1}}\right) \left(I _ {2} - \hat {I _ {2}}\right) (1 - S T M)}{\sum (1 - S T M)}, \tag {11}
+$$
+
+$$
+L _ {\mathrm {w} - \mathrm {s i m}} (x) = \frac {\operatorname {C o v} ^ {\prime} \left[ I _ {w} , I _ {f} \right]}{\sqrt {\operatorname {C o v} \left[ I _ {w} , I _ {w} \right] \operatorname {C o v} \left[ I _ {f} , I _ {f} \right]}}, \tag {12}
+$$
+
+Therefore, the training loss of the second stage is:
+
+$$
+L = L _ {\mathrm {w} \text {- s i m}} + \alpha_ {1} \cdot L _ {\mathrm {V P}} + \alpha_ {2} \cdot L _ {\mathrm {r e g}}, \tag {13}
+$$
+
+where we use the smoothness regularization from Voxel-Morph [5] as  $L_{reg}$ , and  $\alpha_{1}$  and  $\alpha_{2}$  as hyperparameters to control the relative importance of the volume-preserving loss and the regularization loss.
+
+# 4. Experiments
+
+# 4.1. Experimental Settings
+
+Implementation. Following the implementation of Recursive Cascaded Network (RCN) [38], we employed PyTorch and utilized the same similarity loss and regularization losses (for affine and deformable, respectively) for baseline methods. The weight of the volume-preserving losses is set to 0.1 while the weights of similarity loss is 1 and regularization loss is 0.1. Models were trained on one NVIDIA GeForce RTX 3090 GPU with a batch size of 4. During the training stage, we used the Adam optimizer [36] and ran for a total of 5 epochs, with 20000 iterations in each epoch. The learning rate was set to  $10^{-4}$ .
+
+Baselines. We re-implemented our strategy on four SOTA learning-based models that have distinct network architectures: 1) recursive cascaded network with VTN base network (VTN) [38], which stacks CNN networks to achieve higher registration accuracy; 2) VoxelMorph [5], which is a widely used CNN network that utilize stationary vector field to register images; 3) TransMorph [6], which is a well-performed hybrid Transformer-ConvNet network; 4) LapIRN [23], which is a Laplacian Pyramid network designed to operate in a coarse-to-fine manner. We fit these learning-based networks into our strategy with minor efforts and achieve good performance. Additionally, we compare our proposed method with two state-of-the-art traditional methods for deformable image registration, SyN [2] (integrated in ANTs [3] with the affine stage) and B-spline [26] (integrated in Elastix [17] with the affine stage) following [38].
+
+Evaluation Metrics. The evaluation of the proposed strategy is composed of five metrics: Dice score, landmark distance to measure the anatomical alignment, and Square Tumor Size Ratio metric to measure the volume preservation of tumors. More specifically, the Dice score is calculated using the following formula:
+
+$$
+D i c e \left(I _ {1}, I _ {2}\right) = \frac {2 \left| I _ {1} \cap I _ {2} \right|}{\left| I _ {1} \right| + \left| I _ {2} \right|}, \tag {14}
+$$
+
+Additionally, we also use landmark annotations to measure the anatomical alignment between the warped image and
+
+the fixed image. We calculate the average distance between the fixed image's landmarks and the warped landmarks of the moving image, as also introduced in RCN [38]. To quantify the diffeomorphism and smoothness of the deformation fields, we followed the methodology of previous works such as [7, 6, 39]. Here, two metrics were employed: the average percentage of voxels with non-positive Jacobian determinant  $(|J_{\theta}| < 0(\%))$  in the deformation fields, and the standard deviation of the Jacobian determinant  $(Std.|J_{\theta}|)$ .
+
+To measure the volume preservation of tumors, we propose Square Tumor Size Ratio metric (STSR), which is calculated from the change of tumor size ratio (TSR) defined in Equation 4:
+
+$$
+S T S R = \max  \left(\left(\frac {T S R \left(I _ {m}\right)}{T S R \left(I _ {w}\right)}\right), \left(\frac {T S R \left(I _ {w}\right)}{T S R \left(I _ {m}\right)}\right)\right) ^ {2}, \tag {15}
+$$
+
+Here,  $I_{m}, I_{w}$  refers to the moving image and the warped moving image. The square term is used o account for larger changes in tumor volumes, which can have a more significant impact on the accuracy of tumor growth estimation in clinical applications. By utilizing the STSR metric, we are able to accurately evaluate the performance of our proposed strategy.
+
+Preprocessing. In our study, we follow standard preprocessing steps as outlined in VTN [39] and VoxelMorph [5]. The raw scans are resampled into  $128 \times 128 \times 128$  voxels after cropping any unnecessary areas around the target object.
+
+# 4.2. Datasets
+
+Training. To train our model, we utilized two publicly available datasets: (1) LiTS17 [10]: The LiTS dataset comprises 201 CT scans that were acquired using various CT scanners and devices. The resolution of the images in this dataset ranges from  $0.56\mathrm{mm}$  to  $1.0\mathrm{mm}$  in the axial direction and from  $0.45\mathrm{mm}$  to  $6.0\mathrm{mm}$  in the  $z$  direction. The number of slices in the  $z$  direction ranges from 42 to 1026. (2) BraTS20 [22]: It consists of 369 MRI scans of brain tumors, each with ground truth segmentation of four different tumor components. For training our neural network, we utilized the T1ce modality and treated all four components as tumors, while considering the entire brain as the surrounding organ.
+
+Testing. For testing, we utilized a set of 10 cases of CT liver scans with metastatic tumors, and 10 MRI brain scans with metastatic tumors. The modality of MRI brain scans is T1ce. Each case included both pre- and post-treatment images of the same patients, resulting in a total of 20 scans for each set. Following [27], the segmentation of the scans, which contains the organ (liver or brain) and the tumor inside, were carefully annotated by three experts in the field.
+
+![](images/2e42f463d4be0f91f0d1d06d5ddaca8fecaa2047459b42efd043552603df60a6.jpg)  
+Fixed
+
+![](images/b1a7e7fddb152d1de4077cb4ad0753ac84c031c0007beba59b782801589a1a04.jpg)  
+Moving
+
+![](images/51ab0beed0cf28f8fbcd4e4d1010e4d440278149705dab1eb4955ca24c0c5c1d.jpg)  
+Fixed GT
+
+![](images/a86685584960360e104e78114dcfe893f15d653fcdadd8bd6eb4597615b23e44.jpg)  
+Moving GT
+
+![](images/64e7e18e954134102460c267fd341a992df289e493aad2bed3fcc0c13f3c5175.jpg)  
+eannnnnne nnnnne eannnnnne  
+VTN  
+Regular  
+Figure 4. Qualitative comparison of different volume-preserving (VP) methods trained on the Liver Tumor Segmentation (LiTS) dataset. The left side of the figure shows two sets of images: Fixed and Ground Truth (GT), and Moving and GT. The first row of the figure displays the warped moving image, while the second row illustrates the organ outlines in green and red for the moving and fixed images, respectively. The yellow overlay highlights the tumors. Our proposed volume-preserving (VP) method ensures the preservation of tumor volume while aligning the images, as demonstrated by reduced number of visible changes in tumor size. In the third row, the Jacobian matrix of the deformation field is visualized. The green and red lines represent the organ and tumor outlines, respectively. The white areas indicate a large Jacobian, which corresponds to a more significant change in volume. The method without volume-preserving loss demonstrates a larger white area in the tumor, indicating a greater volume change of tumor volume. The last row of the figure displays the deformation field. Due to space limitations, qualitative results for BraTS20 dataset are provided in the supplemental material.
+
+![](images/26233b76521c2fd6aea02034266fb08e635933235c2f5a87bc53c8351c632a23.jpg)  
+VTN  
+Ours
+
+![](images/f8e1a32aa61351d4c5d1c93e39ef98466c1ee3c707d4c8a1e3c41a95c57adb31.jpg)  
+VXM  
+Regular
+
+![](images/258b2ba85af0dd80532b0a7eeba60664abc8102b7adde7e0677b65dc3c3b4db7.jpg)  
+VXM  
+Ours
+
+![](images/75787e95360e42407cf70ede43f41ee18f9ebe0b5a60e2dbcd184dd36a6858ad.jpg)  
+Trans.  
+Regular
+
+![](images/13bbbab58f2c237f5207e5501ee9c2bdf3621ed2347b5de9ace53472bf452c20.jpg)  
+Trans.  
+Ours
+
+![](images/57311016b929b767ece3f36a86addd2f31f6c392c98128a9fc3a9ec2345870f9.jpg)  
+LapIRN  
+Regular
+
+![](images/45b405b4edf9e1156fa4cb0bd953473ee8d7311edda6a4f32fc201939e30ea29.jpg)  
+LapIRN  
+Ours
+
+<table><tr><td rowspan="2">Network</td><td rowspan="2">Methods</td><td colspan="5">LiTs17</td><td colspan="5">BraTS20</td></tr><tr><td>Dice ↑</td><td>Lm. Dist ↓</td><td>|Jθ| &lt; 0(%) ↓</td><td>Std. |Jθ| ↓</td><td>STSR ↓</td><td>Dice ↑</td><td>Lm. Dist ↓</td><td>|Jθ| &lt; 0(%) ↓</td><td>Std. |Jθ| ↓</td><td>STSR ↓</td></tr><tr><td>SyN</td><td>-</td><td>0.850</td><td>12.36</td><td>0.192</td><td>0.221</td><td>1.564</td><td>0.9560</td><td>4.01</td><td>0.276</td><td>0.226</td><td>1.512</td></tr><tr><td>B-Spline</td><td>-</td><td>0.853</td><td>14.22</td><td>0.003</td><td>0.288</td><td>1.330</td><td>0.9601</td><td>4.51</td><td>0.001</td><td>0.171</td><td>1.622</td></tr><tr><td rowspan="2">VTN</td><td>Regular</td><td>0.912</td><td>10.46</td><td>3.545</td><td>2.048</td><td>2.330</td><td>0.9784</td><td>2.98</td><td>0.396</td><td>0.725</td><td>2.263</td></tr><tr><td>Ours</td><td>0.908</td><td>10.77</td><td>2.956</td><td>1.124</td><td>1.260</td><td>0.9738</td><td>3.10</td><td>0.596</td><td>0.285</td><td>1.416</td></tr><tr><td rowspan="2">VXM</td><td>Regular</td><td>0.860</td><td>12.62</td><td>4.191</td><td>1.699</td><td>1.738</td><td>0.9701</td><td>3.56</td><td>0.024</td><td>0.448</td><td>1.929</td></tr><tr><td>Ours</td><td>0.857</td><td>12.43</td><td>1.681</td><td>0.691</td><td>1.241</td><td>0.9741</td><td>3.60</td><td>0.030</td><td>0.226</td><td>1.463</td></tr><tr><td rowspan="2">TransMorph</td><td>Regular</td><td>0.859</td><td>13.35</td><td>6.377</td><td>1.906</td><td>1.785</td><td>0.9710</td><td>4.01</td><td>0.164</td><td>0.468</td><td>1.868</td></tr><tr><td>Ours</td><td>0.856</td><td>13.19</td><td>4.084</td><td>0.827</td><td>1.302</td><td>0.9758</td><td>3.98</td><td>0.019</td><td>0.194</td><td>1.319</td></tr><tr><td rowspan="2">LapIRN</td><td>Regular</td><td>0.8933</td><td>10.23</td><td>5.301</td><td>1.788</td><td>2.360</td><td>0.9750</td><td>3.40</td><td>3.020</td><td>0.301</td><td>1.889</td></tr><tr><td>Ours</td><td>0.8893</td><td>10.55</td><td>4.120</td><td>0.988</td><td>1.645</td><td>0.9743</td><td>3.36</td><td>0.040</td><td>0.105</td><td>1.224</td></tr></table>
+
+Table 1. Quantitative comparison with state-of-the-art registration methods. Our strategy was applied to two datasets, namely LiTS17 and BraTS20, and three network architectures In addition, five metrics were employed for the evaluation process. The results of the comparison indicate that our proposed strategy can be well generalized to other state-of-the-art methods to preserve tumor volumes while achieving comparable or better performance on other metrics.
+
+Landmarks were also annotated by three experts in the field to evaluate the registration accuracy of our strategy.
+
+# 4.3. Main Results
+
+Quantitative Comparison. In Table 1, we present a comprehensive comparison of our proposed strategy with two traditional methods and three learning-based methods with different network architectures (CNN, cascaded CNN and hybrid transformer-CNN). We compare the methods using a similarity-based manner (regular) following the settings
+
+in [38], as well as our volume-preserving manners. Our results demonstrate that our proposed strategy outperforms all other methods in terms of the STSR metrics, while achieving comparable results in Dice and Landmark Dist, and often better results in folding detection  $(|J_{\theta}| < 0)$  and smoothness quality  $(Std. |J_{\theta}|)$ . Furthermore, we observe that registration networks that achieve better performance in terms of previous metrics such as Dice tend to have worse performance in preserving tumor volumes (STSR). This observation highlights the need to balance these two objectives
+
+<table><tr><td rowspan="2">#</td><td rowspan="2">Network</td><td rowspan="2">Unsupervised</td><td rowspan="2">Adaptive</td><td rowspan="2">Region</td><td colspan="5">LiTs17</td></tr><tr><td>Dice ↑</td><td>Lm. Dist ↓</td><td>|Jθ| &lt; 0(%) ↓</td><td>Std. |Jθ| ↓</td><td>STSR ↓</td></tr><tr><td>1</td><td rowspan="5">VTN</td><td>✓</td><td>-</td><td>-</td><td>0.912</td><td>10.46</td><td>3.545</td><td>2.048</td><td>2.330</td></tr><tr><td>2</td><td>×</td><td>-</td><td>-</td><td>0.914</td><td>10.22</td><td>3.241</td><td>1.165</td><td>1.648</td></tr><tr><td>3</td><td>×</td><td>×</td><td>Tumor</td><td>0.911</td><td>10.74</td><td>2.883</td><td>1.017</td><td>1.205</td></tr><tr><td>4</td><td>✓</td><td>×</td><td>Organ</td><td>0.904</td><td>11.07</td><td>3.194</td><td>1.015</td><td>1.640</td></tr><tr><td>5</td><td>✓</td><td>✓</td><td>Organ</td><td>0.908</td><td>10.77</td><td>2.956</td><td>1.124</td><td>1.260</td></tr><tr><td>6</td><td rowspan="5">VXM</td><td>✓</td><td>-</td><td>-</td><td>0.860</td><td>12.62</td><td>4.191</td><td>1.699</td><td>1.738</td></tr><tr><td>7</td><td>×</td><td>-</td><td>-</td><td>0.863</td><td>12.22</td><td>1.603</td><td>0.691</td><td>1.417</td></tr><tr><td>8</td><td>×</td><td>×</td><td>Tumor</td><td>0.859</td><td>12.38</td><td>1.567</td><td>0.698</td><td>1.291</td></tr><tr><td>9</td><td>✓</td><td>×</td><td>Organ</td><td>0.856</td><td>12.78</td><td>1.635</td><td>0.675</td><td>1.407</td></tr><tr><td>10</td><td>✓</td><td>✓</td><td>Organ</td><td>0.857</td><td>12.43</td><td>1.681</td><td>0.691</td><td>1.241</td></tr><tr><td>11</td><td rowspan="5">TransMorph</td><td>✓</td><td>-</td><td>-</td><td>0.859</td><td>13.35</td><td>6.337</td><td>1.906</td><td>1.785</td></tr><tr><td>12</td><td>×</td><td>-</td><td>-</td><td>0.863</td><td>13.10</td><td>4.923</td><td>1.540</td><td>1.540</td></tr><tr><td>13</td><td>×</td><td>×</td><td>Tumor</td><td>0.858</td><td>13.21</td><td>4.401</td><td>0.877</td><td>1.227</td></tr><tr><td>14</td><td>✓</td><td>×</td><td>Organ</td><td>0.849</td><td>13.02</td><td>4.651</td><td>0.845</td><td>1.454</td></tr><tr><td>15</td><td>✓</td><td>✓</td><td>Organ</td><td>0.856</td><td>13.19</td><td>4.084</td><td>0.827</td><td>1.302</td></tr></table>
+
+Table 2. Effectiveness of adaptive volume-preserving loss in unsupervised registration frameworks on LiTs17 dataset. Three networks, namely VTN in [38], VXM in [5], and TransMorph in [6], were applied. The methods used in the study were regular unsupervised (first row), weighted similarity loss only based on the ground truth (GT) mask (second row), volume-preserving loss based on GT(third row), volume-preserving loss on the whole organ based on estimation (fourth row), and on the tumor based on estimation (ours).
+
+in order to achieve competitive results.
+
+Qualitative Comparison. Figure 5 presents a qualitative comparison of our proposed strategy with different state-of-the-art methods. Comparing the "VTN/VXM/Trans./LapIRN regular" column and the "VTN/VXM/Trans./LapIRN ours" column, it is evident that our method outperforms other methods in terms of volume preservation in the warped image  $(1_{\mathrm{st}}$  row) and the corresponding deformation field  $(4_{\mathrm{th}}$  row). Furthermore, in terms of visualization of the registration accuracy and the overlay of the tumor  $(2_{\mathrm{nd}}$  row), our proposed method achieves better volume preservation with comparable accuracy compared to other methods. Lastly, the Jacobian determinant of the deformation field  $(3_{\mathrm{rd}}$  row) reveals that regular methods tend to have a larger Jacobian in tumor regions, indicating greater volume change, compared to our method. This highlights the effectiveness of our volume-preserving approach in preserving tumor volumes while aligning the images.
+
+# 4.4. Ablation study
+
+Volume-Preserving Losses. In Table 2, we present the results of our evaluation of the effectiveness of various versions of volume-preserving losses. For each sub-table, the first row (#1, #6, and #11) represents using only the similarity loss, while the second row (#2, #7, and #12) utilizes the ground truth as a weighted similarity loss. The third row (#3, #8, and #13) uses the ground truth binary mask as a tumor mask, while the fourth row (#4, #9, and #14) uses the mask of the entire organ as a tumor mask. The fifth row (#5, #10, and #15) represents our proposed strategy.
+
+By comparing the first and second rows of each sub-table with the fifth row, we observe that our proposed volume-
+
+preserving loss can effectively preserve details and achieve superior performance across various metrics. The improvement from the third and fourth rows to the fifth row indicates the importance of adaptiveness in preserving the volume of the organ and tumor, respectively. It should be noted that we do not require pixel-wise ground truth for tumor volume preservation.
+
+Robustness to Tumor Mask Estimation. In Figure 5, we showcase the robustness of our proposed strategy in the presence of noisy ground truth segmentation masks. To evaluate this, we randomly select points within organs to ensure that tumors have a certain Dice score with respect to the ground truth segmentation of tumors. As depicted in the figure, despite the variations in the noisy ground truth segmentation masks, our strategy's performance remains consistent. Specifically, when the tumor mask has a Dice score above a certain threshold, such as 0.15, the estimated tumor mask consistently achieves a Dice score of approximately 0.17. In contrast, using the whole organ as mask in volume-preserving loss causes degradation in performance as it only has a dice score of roughly 0.05. This highlights the necessity and efficiency of our tumor mask estimation in stage one.
+
+# 5. Conclusions
+
+In summary, this work addressed the problem of preserving tumor volumes while maximizing similarity during learning-based registration, which is crucial for tracking tumor growth. We proposed a two-stage unsupervised strategy that effectively preserves the size ratio of tumors, without requiring extra networks. Our approach involves generating a soft tumor mask in the first stage using a similarity-based registration network and incorporating
+
+![](images/3fa7600e12f5561e2b1165278e52e24243ef91cdb3f409a4ecb8e16d3efe6156.jpg)  
+Figure 5. Robustness of our tumor mask estimation. We evaluated the robustness of our tumor mask estimation by running our strategy using RCN on the LiTS datasets with different noisy ground truth (GT) masks as tumor masks, each with specific Dice scores with tumors. We plotted the results on a graph with the Dice score on the x-axis and the STSR in the test set on the y-axis (where smaller values are better).
+
+an adaptive volume-preserving loss and a weighted similarity loss in the second stage to improve registration performance. To evaluate our strategy's performance, we introduced a new metric, the Square Tumor Size Ratio (STSR), which measures the preservation of tumor volume. Our proposed strategy was evaluated on various networks and demonstrated superior performance in tumor volume preservation, while achieving comparable results in other metrics. This highlights the effectiveness of our approach in tracking tumor growth and its ability to generalize well to different registration networks. Additionally, we noted that the volume preservation approach could be extended to other lesions besides tumors.
+
+Acknowledgement: This work was supported by the HKSAR Innovation and Technology Commission (ITC) under ITF Project MHP/109/19, the National Natural Science Foundation in China under Grant 62022010, the Beijing Natural Science Foundation Haidian District Joint Fund in China under Grant L222032 and the Fundamental Research Funds for the Central Universities of China from the State Key Laboratory of Software Development Environment in Beihang University in China.
+
+# References
+
+[1] Moab Arar, Yiftach Ginger, Dov Danon, Amit H Bermano, and Daniel Cohen-Or. Unsupervised multi-modal image registration via geometry preserving image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13410–13419, 2020. 3  
+[2] Brian B Avants, Charles L Epstein, Murray Grossman, and James C Gee. Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain. Medical image analysis, 12(1):26-41, 2008. 6  
+[3] Brian B Avants, Nick Tustison, Gang Song, et al. Advanced normalization tools (ants). Insight j, 2(365):1-35, 2009. 6  
+[4] Bhakti Baheti, Diana Waldmannstetter, Satrajit Chakrabarty, Hamed Akbari, Michel Bilello, Benedikt Wiestler, Julian Schwarting, Evan Calabrese, Jeffrey Rudie, Syed Abidi, et al. The brain tumor sequence registration challenge: establishing correspondence between pre-operative and follow-up mri scans of diffuse glioma patients. arXiv preprint arXiv:2112.06979, 2021. 2, 3  
+[5] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. An unsupervised learning model for deformable medical image registration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9252-9260, 2018. 1, 2, 4, 6, 8  
+[6] Junyu Chen, Eric C Frey, Yufan He, William P Segars, Ye Li, and Yong Du. Transmorph: Transformer for unsupervised medical image registration. Medical image analysis, 82:102615, 2022. 2, 3, 4, 6, 8  
+[7] Jiashun Chen, Donghuan Lu, Yu Zhang, Dong Wei, Munan Ning, Xinyu Shi, Zhe Xu, and Yefeng Zheng. Deformer: Towards displacement field learning for unsupervised medical image registration. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2022: 25th International Conference, Singapore, September 18-22, 2022, Proceedings, Part VI, pages 141-151. Springer, 2022. 2, 3, 6  
+[8] Xiang Chen, Yan Xia, Nishant Ravikumar, and Alejandro F Frangi. Joint segmentation and discontinuity-preserving deformable registration: Application to cardiac cine-mr images. arXiv preprint arXiv:2211.13828, 2022. 2  
+[9] Nicha Chitphakdithai and James S. Duncan. Non-rigid registration with missing correspondences in preoperative and postresection brain images. 3  
+[10] Patrick Christ. Lits – liver tumor segmentation challenge (lits17). 2017. 6  
+[11] Olivier Clatz, Hervé Delingette, Ion-Florin Talos, Alexandra J. Golby, Ron Kikinis, Ferenc A. Jolesz, Nicholas Ayache, and Simon K. Warfield. Robust nonrigid registration to capture brain shift from intraoperative mri. IEEE Transactions on Medical Imaging, 2005. 3  
+[12] Ahmed Elazab, Yousry M Abdulazeem, Ahmed M Anter, Qingmao Hu, Tianfu Wang, and Baiying Lei. Macroscopic cerebral tumor growth modeling from medical images: A review. IEEE Access, 6:30663-30679, 2018. 2, 3
+
+[13] Xu Han, Xiao Yang, Stephen R. Aylward, Roland Kwitt, and Marc Niethammer. Efficient registration of pathological images: A joint pca/image-reconstruction approach. arXiv: Computer Vision and Pattern Recognition, 2017. 3  
+[14] Malte Hoffmann, Benjamin Billot, Douglas N Greve, Juan Eugenio Iglesias, Bruce Fischl, and Adrian V Dalca. Synth morph: learning contrast-invariant registration without acquired images. IEEE transactions on medical imaging, 41(3):543-558, 2021. 1, 2, 3  
+[15] Malte Hoffmann, Andrew Hoopes, Douglas N Greve, Bruce Fischl, and Adrian V Dalca. Anatomy-aware and acquisition-agnostic joint registration with synthmorph. arXiv preprint arXiv:2301.11329, 2023. 1, 2, 3  
+[16] Boah Kim, Dong Hwan Kim, Seong Ho Park, Jieun Kim, June-Goo Lee, and Jong Chul Ye. Cyclesmorph: cycle consistent unsupervised deformable image registration. Medical image analysis, 71:102036, 2021. 2, 3  
+[17] Stefan Klein, Marius Staring, Keelin Murphy, Max A Viergever, and Josien PW Pluim. Elastix: a toolbox for intensity-based medical image registration. IEEE transactions on medical imaging, 29(1):196-205, 2009. 6  
+[18] Dongjin Kwon, Marc Niethammer, Hamed Akbari, Michel Bilello, Christos Davatzikos, and Kilian M. Pohl. Portr: Preoperative and post-recurrence brain tumor registration. IEEE Transactions on Medical Imaging, 2014. 3  
+[19] Dongjin Kwon, Ke Zeng, Michel Bilello, and Christos Davatzikos. Estimating patient specific templates for preoperative and follow-up brain tumor registration. Lecture Notes in Computer Science, 2015. 3  
+[20] Fei Lyu, Mang Ye, Andy J Ma, Terry Cheuk-Fung Yip, Grace Lai-Hung Wong, and Pong C Yuen. Learning from synthetic ct images via test-time training for liver tumor segmentation. IEEE transactions on medical imaging, 41(9):2510-2520, 2022. 3  
+[21] Mingyuan Meng, Lei Bi, Dagan Feng, and Jinman Kim. Brain tumor sequence registration with non-iterative coarse-to-fine networks and dual deep supervision. arXiv preprint arXiv:2211.07876, 2022. 3  
+[22] Bjoern H Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al. The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging, 34(10):1993-2024, 2014. 6  
+[23] Tony CW Mok and Albert CS Chung. Large deformation diffeomorphic image registration with laplacian pyramid networks. 3, 6  
+[24] Tony CW Mok and Albert CS Chung. Unsupervised deformable image registration with absent correspondences in pre-operative and post-recurrence brain tumor mri scans. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2022: 25th International Conference, Singapore, September 18-22, 2022, Proceedings, Part VI, pages 25-35. Springer, 2022. 2, 3  
+[25] Torsten Rohlfing, Calvin R Maurer, David A Bluemke, and Michael A Jacobs. An alternating-constraints algorithm for volume-preserving non-rigid registration of contrast-
+
+enhanced mr breast images. In Biomedical Image Registration: Second International Workshop, WBIR 2003, Philadelphia, PA, USA, June 23-24, 2003. Revised Papers 2, pages 291-300. Springer, 2003. 3  
+[26] Daniel Rueckert, Luke I Sonoda, Carmel Hayes, Derek LG Hill, Martin O Leach, and David J Hawkes. Nonrigid registration using free-form deformations: application to breast mr images. IEEE transactions on medical imaging, 18(8):712-721, 1999. 6  
+[27] Siyuan Shan, Wen Yan, Xiaqing Guo, Eric I Chang, Yubo Fan, Yan Xu, et al. Unsupervised end-to-end learning for deformable medical image registration. arXiv preprint arXiv:1711.08608, 2017. 6  
+[28] Shuwei Shao, Zhongcai Pei, Weihai Chen, Wentao Zhu, Xingming Wu, and Baochang Zhang. A multi-scale unsupervised learning for deformable image registration. International Journal of Computer Assisted Radiology and Surgery, pages 1-10, 2022. 3  
+[29] Jiacheng Shi, Yuting He, Youyong Kong, Jean-Louis Coa-trieux, Huazhong Shu, Guanyu Yang, and Shuo Li. Xmorpher: Full transformer for deformable medical image registration via cross attention. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2022: 25th International Conference, Singapore, September 18-22, 2022, Proceedings, Part VI, pages 217-226. Springer, 2022. 2, 3  
+[30] Christine Tanner, Julia A. Schnabel, Andreas Degenhard, Andy D. Castellano-Smith, Carmel Hayes, Martin O. Leach, D. R. Hose, Derek L. G. Hill, and David J. Hawkes. Validation of volume-preserving non-rigid registration: Application to contrast-enhanced mr-mammography. Lecture Notes in Computer Science, 2002. 3  
+[31] Di Wang, Yue Pan, Oguz C. Durmeric, Joseph M. Reinhardt, Eric A. Hoffman, Joyce D. Schroeder, and Gary E. Christensen. Plosl: Population learning followed by one shot learning pulmonary image registration using tissue volume preserving and vesselness constraints. Medical Image Analysis, 2022. 3  
+[32] Yibo Wang, Wen Qian, Mengqi Li, and Xuming Zhang. A transformer-based network for deformable medical image registration. In Artificial Intelligence: Second CAAI International Conference, CICAI 2022, Beijing, China, August 27-28, 2022, Revised Selected Papers, Part I, pages 502-513. Springer, 2022. 1, 2  
+[33] Marek Wodzinski, Artur Jurgas, Niccolo Marini, Manfredo Atzori, and Henning Muller. Unsupervised method for intra-patient registration of brain magnetic resonance images based on objective function weighting by inverse consistency: Contribution to the brats-reg challenge. arXiv preprint arXiv:2211.07386, 2022. 3  
+[34] Qingsong Yao, Li Xiao, Peihang Liu, and S Kevin Zhou. Label-free segmentation of Covid-19 lesions in lung ct. IEEE transactions on medical imaging, 40(10):2808-2819, 2021. 3  
+[35] Xiaoman Zhang, Weidi Xie, Chaoqin Huang, Ya Zhang, Xin Chen, Qi Tian, and Yanfeng Wang. Self-supervised tumor segmentation with sim2real adaptation. IEEE Journal of Biomedical and Health Informatics, 2023. 3
+
+[36] Zijun Zhang. Improved adam optimizer for deep neural networks. In 2018 IEEE/ACM 26th international symposium on quality of service (IWQoS), pages 1-2. IEEE, 2018. 6  
+[37] Bowen Zhao, G. Christensen, Joo Hyun Song, Yue Pan, Sarah E. Gerard, Joseph M. Reinhardt, Kaifang Du, Taylor J. Patton, John E. Bayouth, and Geoffrey D. Hugo. Tissuevolume preserving deformable image registration for 4dct pulmonary images. 2016 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 481-489, 2016. 3  
+[38] Shengyu Zhao, Yue Dong, Eric I Chang, Yan Xu, et al. Recursive cascaded networks for unsupervised medical image registration. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10600-10610, 2019. 1, 2, 3, 4, 5, 6, 7, 8  
+[39] Shengyu Zhao, Tingfung Lau, Ji Luo, I Eric, Chao Chang, and Yan Xu. Unsupervised 3d end-to-end medical image registration with volume tweening network. IEEE journal of biomedical and health informatics, 24(5):1394-1404, 2019. 6

2023/Preserving Tumor Volumes for Unsupervised Medical Image Registration/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:017f5b78605291cd68bb9109cd800abd1b4cf2e7dc0be39cedcf5977e7b0a32b
+size 507721

2023/Pretrained Language Models as Visual Planners for Human Assistance/057af23b-bb0a-449d-acad-c434ed40a938_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:21dafe893b03856348540e4cf0c60fe77cba31daf64cd743695cff282d7f8da5
+size 1222911

2023/Pretrained Language Models as Visual Planners for Human Assistance/full.md

@@ -0,0 +1,377 @@
+# Pretrained Language Models as Visual Planners for Human Assistance
+
+Dhruvesh Patel $^{1,2}$  Hamid Eghbalzadeh $^{1}$  Nitin Kamra $^{1}$  Michael Louis Iuzzolino $^{1}$  Unnat Jain $^{1*}$  Ruta Desai $^{1*†}$ $^{1}$ Meta  $^{2}$ UMass Amherst https://github.com/facebookresearch/vlamp
+
+# Abstract
+
+In our pursuit of advancing multi-modal AI assistants capable of guiding users to achieve complex multi-step goals, we propose the task of 'Visual Planning for Assistance (VPA)'. Given a succinct natural language goal, e.g., "make a shelf", and a video of the user's progress so far, the aim of VPA is to devise a plan, i.e. a sequence of actions such as "sand shelf", "paint shelf", etc. to realize the specified goal. This requires assessing the user's progress from the (untrimmed) video, and relating it to the requirements of natural language goal, i.e. which actions to select and in what order? Consequently, this requires handling long video history and arbitrarily complex action dependencies. To address these challenges, we decompose VPA into video action segmentation and forecasting. Importantly, we experiment by formulating the forecasting step as a multimodal sequence modeling problem, allowing us to leverage the strength of pre-trained LMs (as the sequence model). This novel approach, which we call Visual Language Model based Planner (VLaMP), outperforms baselines across a suite of metrics that gauge the quality of the generated plans. Furthermore, through comprehensive ablations, we also isolate the value of each component - language pretraining, visual observations, and goal information. We have open-sourced all the data, model checkpoints, and training code.
+
+# 1. Introduction
+
+Imagine assembling a new piece of furniture or following a new recipe for a dinner party. To achieve such a goal, you might follow a manual or a video tutorial, going back and forth as you perform the steps. Instead of fumbling through a manual, imagine an assistive agent capable of being invoked through natural language, having the ability to understand human actions, and providing actionable multi-step guidance for achieving your desired goal. Such multi-
+
+![](images/1ec0e91231c15a34a02907bea4c7c18d3395949578a3dd3ca32d00ace36a79d4.jpg)  
+Figure 1: Visual Planning for Assistance overview (top) and general methodology (bottom). Given a user-specified, natural language goal ("build a shelf") and corresponding visual history of the user's progress till now, VPA involves predicting a sequence of actions, to assist the user towards the goal ("cut", "sand", "paint"). Our approach is based on multi-modal sequence modeling where we reuse pre-trained video segmentation and language models.
+
+modal assistive agents should be able to reason human activities from visual observations, contextualize them to the goal at hand, and plan future actions for providing guidance.
+
+To quantify the progress and aid the development of such multi-modal neural models, we need a intuitive task, as illustrated in the example in Fig. 1 (top). We call this Visual Planning for Assistance (VPA) that we detail next. Given a user-specified goal in natural language ("build shelf") and corresponding video observations of the user's progress towards this goal, the task objective should be to generate the ordered sequence of next actions towards achieving the goal ("cut  $\rightarrow$  sand  $\rightarrow$  paint"). We base VPA off instructional YouTube videos of such procedural activities from large, open-sourced datasets - CrossTask [89] and COIN [73]. This is a natural choice as procedural human activities (cooking, assembly, repair, etc.) are a perfect source of multi-step and complex sequence of actions, where humans routinely seek guidance. Realistic nature
+
+of VPA makes it particularly challenging. Operating on untrimmed videos requires dealing with potentially many irrelevant background frames, chunking actions in the videos will also be imperative. Another challenge is the validity of the plan of actions i.e. they must respect the constraints of the activity – the shelf shouldn't be painted before sanding.
+
+The natural next question is - what's a good way to tackle VPA? Marrying research from video forecasting & anticipation with embodied AI [19, 17], we cast VPA as goal-conditioned task planning. We approach VPA by utilizing video action segmentation and transformer-based neural sequence modeling - the former allows us to deal with long video history and the latter can handle arbitrary sequential constraints [86, 39]. The formulation (particularly, transformer-based neural sequence modeling) allows us to tap into the prowess of pre-trained language models (PTLMs). These PTLMs contain useful priors about action-action similarity, action-goal association, and action ordering [7, 30, 58], that our formulation can piggy-back off. Particularly, our model - Visual Language Model Planner (VLaMP) - conditions the generated plan onto the visual history by using a transformer-based mapper network that projects embeddings corresponding to visual history into the input space of the LM. Using VPA as the testbed, we show that VLaMP outperforms a range of standardized baselines. We undertake head-on ablations to quantify effect of each component of VLaMP - language pre-training of the LM, the visual history, and goal description. We believe the modularity of our approach can allow researchers to swap components with their own, to make rapid progress towards VPA.
+
+In summary, our main contributions are: (1) a new task VPA on human interaction videos capturing unique aspects of real-world vision-powered assistive agents; (2) a general-purpose methodology for VPA, which allows leveraging pre-trained multi-modal encoders and sequence models; (3) an instantiation of this methodology (we call VLaMP), where we reuse sequence priors from PTLM and investigate its efficacy across two, well-studied datasets of procedural human activities.
+
+# 2. Related Work
+
+Forecasting in Videos. VPA entails future action sequence prediction for a video and is closely related to anticipation in video understanding, including future localization [22, 40, 78], future frame prediction [77, 46, 53], next active object estimation [21, 5, 25, 67, 4, 24, 44], as well as short- and long-term action anticipation [66, 15, 20, 23, 25, 65, 64, 57, 45, 1, 52, 50, 11, 79]. While short-term forecasting approaches, such as [66], are limited to predicting the single next action occurring only a few seconds into the future, our work focuses on predicting sequences of actions over longer time horizons (on the order of minutes)
+
+<table><tr><td></td><td>Action Prediction</td><td>Goal Modality</td><td>Visual Reasoning</td></tr><tr><td>Action anticipation [14]</td><td>Single</td><td>None</td><td>Video-based</td></tr><tr><td>Action forecasting [25]</td><td>Multiple, ordered</td><td>None</td><td>Video-based</td></tr><tr><td>Procedural planning [9]</td><td>Multiple, ordered</td><td>Vision</td><td>Image-based</td></tr><tr><td>VPA (ours)</td><td>Multiple, ordered</td><td>Natural Language</td><td>Video-based</td></tr></table>
+
+Table 1: VPA vs. prior tasks related to forecasting of real-world human activities. We predict an ordered sequence of actions given video history with a focus on natural language goal-conditioning for human-assistive applications.
+
+into the future. The long-term action (LTA) forecasting benchmarking was recently established on a small subset of Ego4D [25], and although VPA is similarly devised to predict a temporally ordered sequence of actions conditioned on video-based visual observations, our approach differs in that the LTA task does no goal-conditioning. Our task incorporates goal-conditioning on long term sequence prediction as we argue this will drive the development of critical system aspects, such as virtual assistants that afford human interaction via the natural language goals. Additionally, in contrast to LTA, we focus on a wide range of goal-oriented activities available within the CrossTask and COIN datasets. Tab. 1 highlights the differences between VPA and other forecasting tasks.
+
+Procedural Planning Approaches. VPA is similar to the task of procedure planning [9], wherein given a starting and terminating visual observation, the aim is to predict the actions that would transform the state from starting to terminating. However, we argue that due to the unavailability of the terminating visual state, the procedure planning task is not useful in a real-world assistance setting. While several recent works introduce novel models for procedure planning, these models cannot be deployed for VPA either because they rely heavily on the availability of visual goal [70, 87, 6], or assume access to true action history [49]. Transformers for Decision Making. VLaMP autoregressively predicts future states and/or observations and actions and is similar in spirit to sequence models for decision making such as GATO [61], Decision Transformer [12], and Trajectory Transformer [35]. VLaMP extends such models to work with egocentric observations that humans observe in the day-to-day activities.
+
+Planning in Embodied AI using LMs. Past research has leveraged PTLMs for task planning in real world for embodied agents [30, 68, 41, 7, 42, 31, 18, 16, 36]. Many of these works focus on converting a high-level task or instruction into sequence of low-level steps and then ground them in the environment using either affordance functions [7], visual feedback [31], or multimodal prompts [36, 18]. Owing to their focus on robotic agents, these works focus on predominately pick-place tasks. Instead VPA is focused on
+
+complex tasks, in which humans might require assistance in the form of recommended future actions. Consequently, VLaMP requires grounding and reasoning of much more complex states and actions, which it accomplishes by finetuning a pretrained LM on multimodal sequences of (visual) observations and (text-based) actions.
+
+Multi-modal LMs. Recent works have successfully used the transformer architecture for modeling multi-modal sequences that have visual tokens. For instance, [47, 59, 2, 48, 29, 18] train large transformer based multi-modal sequence prediction models. While [3, 74, 56] focus on adapting pre-trained LMs to work with visual tokens by aligning the representation spaces of the two modalities. VLaMP's approach of modeling sequence of visual and textual representations using PTLMs is similar in spirit to these, i.e. VLaMP also uses a mapper network to align video representations to LM's token space and jointly learns the mapper with LM finetuning for VPA. However, in contrast to previous works, VLaMP predicts the visual tokens autoregressively at inference time to enable forecasting of the state for planning. Consequently, VLaMP's token prediction loss is multi-modal, unlike most multi-modal LMs that only use token prediction loss for text tokens.
+
+# 3. Visual Planning for Assistance
+
+Here, we introduce the task of Visual Planning for Assistance (VPA), towards enabling multi-step guidance to humans in their real-world activities. We instantiate VPA for procedural activities, where humans routinely seek assistance. In this section, we include the definition of VPA, and describe the evaluation protocol.
+
+# 3.1. Task Definition
+
+The following two intuitive inputs are given to any model performing VPA.
+
+Goal Prompt  $(G)$ . The natural language description (in short phrase) of the user's goal, emulating a typical user's request for assistance for a day-to-day task. Examples include, "build a shelf" and "change a tire".
+
+Visual History  $(V_{t})$ . An untrimmed video that provides context about the user's progress towards a goal from the start till time, say  $t$ . We assume that  $V_{t}$  contains  $k$  actions or steps  $\{a_1,\ldots ,a_k\}$  pertaining to the goal. However, VPA doesn't have access to  $\{a_1,\dots ,a_k\}$  or  $k$  and must work with  $V_{t}$ .
+
+Given these two inputs,  $V_{t}$  and  $G$ , the objective of VPA is to generate a plan  $\mathcal{T}$ . The plan is a sequence of actions that should be executed (in the next steps) to assist the user in achieving the goal  $G$ . Concretely, the prediction is denoted by  $\mathcal{T} = (a_{k + 1},\dots ,a_{k + l})$ , where  $a_{i}$  are represented in natural language but come from a closed set  $\mathcal{A}$ . Here,  $l\leq K$  denotes the number of future actions that should be predicted,
+
+out of the  $K$  number of remaining actions required to accomplish the goal. For our shelf-building example, the correct  $\mathcal{T} =$  (sand shelf, paint shelf, attach shelf), capturing the remaining 3 future actions (here,  $l = N = 3$ ).
+
+In day-to-day activities, we request assistance for goals in natural language. However, prior works procedural planning [9, 6] assume access to visual goal state. This is not a realistic assumption for an AI agent assisting humans. Hence, we purposefully relax this assumption, making our formulation of the task significantly more practical. Moreover, an agent with access to the visual modality, should be able to improve its plan by filtering out relevant information from the raw video stream of the progress. These are the two central assumptions around which we formulate the task of VPA.
+
+# 3.2. Evaluation
+
+Open-Sourced Video Data. We leverage existing datasets CrossTask [89] and COIN [73], originally developed to enable video action understanding for VPA, based on the following three requirements:
+
+- Rich diversity of activities from multiple domains: The data from different domains such as cooking, assembly etc., enables testing of VPA models in a more generalized manner.  
+- Goal-oriented activities consisting of long sequences of actions: Since the objective is to generate  $l$  future actions, activities in these datasets that require diverse sequences of actions e.g., "making a pancake", instead of "running" are more suitable for VPA.  
+- Action annotations from a fixed closed set of actions: Action labels described using verb-noun from a finite set [14], which are temporally aligned with the videos. This makes evaluating the accuracy of  $\mathcal{T}$  prediction straightforward. Specifically, free-form, narration-style descriptions of actions that are available in recent video datasets aid in efficient multi-modal representation learning for video understanding [25, 54]. However, evaluating the quality of action sequences towards goal achievement, where each action is described in free-form natural language, is non-trivial. We leave the instantiation of VPA with free-form natural language actions as future work.  
+Table 2 summarizes the features from CrossTask and COIN, aligned with the above requirements.
+
+Metrics. The planning performance of a VPA model is measured by comparing the generated plan  $\hat{\mathcal{T}} = (\hat{a}_{k + 1},\dots ,\hat{a}_{k + l})$  to the ground truth plan  $\mathcal{T}$  for  $l$  actions in the future, given  $V_{t}$  and  $G$ . Here  $\hat{a}_{k + i}$  denotes the prediction for the  $k + i$ -th step given history till  $k$ -th step. Consistent with community practices [9, 6], we use the following metrics, listed in decreasing order of strictness: success rate (SR), mean accuracy (mAcc), mean intersection over union
+
+<table><tr><td>Dataset</td><td># train videos</td><td># test videos</td><td># test samples</td><td>actions per video</td><td># goals</td><td># domains</td></tr><tr><td>CrossTask [89]</td><td>1756</td><td>752</td><td>4123</td><td>7.6 ± 4.3</td><td>18</td><td>3</td></tr><tr><td>COIN [73]</td><td>9428</td><td>1047</td><td>2011</td><td>3.9 ± 2.4</td><td>180</td><td>12</td></tr></table>
+
+Table 2: VPA datasets. We evaluate VPA on two existing video datasets containing multiple goal-oriented procedural activities from varied domains. Such activities contain sequences of multiple actions making them ideal for VPA.
+
+(mIOU). Success rate requires an exact match between all actions and their sequence between  $\hat{\mathcal{T}}$  and  $\mathcal{T}$ . Mean accuracy is the accuracy of the actions at each step. Unlike success rate, mean accuracy does not require a  $100\%$  match to ground truth. Instead it considers matching at each individual step. Lastly, mean intersection over union captures the cases where the model predicts the steps correctly, but fails to identify the correct order. Concretely,
+
+$$
+\mathrm {m I O U} _ {l} = \frac {\left| \hat {a} _ {\{k + 1 : k + l \}} \bigcap a _ {\{k + 1 : k + l \}} \right|}{\left| \hat {a} _ {\{k + 1 : k + l \}} \bigcup a _ {\{k + 1 : k + l \}} \right|}, \tag {1}
+$$
+
+$$
+\mathrm {m A c c} _ {l} = \frac {1}{l} \sum_ {i = 1} ^ {l} \mathbb {1} [ \hat {a} _ {k + i,} = a _ {k + i} ], \tag {2}
+$$
+
+$$
+\mathrm {S R} _ {l} = \prod_ {i = 1} ^ {l} \mathbb {1} [ \hat {a} _ {k + i} = a _ {k + i} ], \tag {3}
+$$
+
+where  $\mathbb{1}[\cdot ]$  is the identity function, which is 1 when the condition in its input is true, and 0, and  $\hat{a}_{\{k + 1:k + l\}}$  denotes the set of  $l$  future actions in  $\tilde{\mathcal{T}}$ , i.e., a sequence but disregarding the order. To complement the above metrics, we also measure the accuracy of predicting the next action i.e. nAcc, as defined in (4), where 'n' stands for next. Note that all metrics are averaged over the test set details of which are included in Sec. 5.
+
+$$
+\mathrm {n A c c} = \mathbb {1} [ \hat {a} _ {k + 1} = a _ {k + 1} ] \tag {4}
+$$
+
+# 4. Visual LM Planner
+
+VPA can be viewed as a sequential decision making problem, where the model (say,  $\pi$ ) predicting the sequence of next actions is a policy conditioned on the visual history  $V_{t}$ , serving as (partially-observed) state, and goal prompt  $G$ . Inspired by the offline learning formulation closest to 'learning from offline demonstrations' [27, 76, 37, 81], in VLaMP, we formulate VPA as a goal-conditioned, multimodal sequence prediction problem. This formulation allows us to leverage high-capacity sequence models like
+
+Transformers [75], which have been successfully applied to sequential decision making [35, 12]. Furthermore, inspired by the recent success of pre-trained transformer language models (PTLMs) on such tasks, [62, 30, 12], we propose to leverage PTLMs as the sequence model in our formulation. In the remainder of this section we describe our approach for VPA, called Visual Language Model based Planner (VLaMP), consisting of a segmentation module and PTLM based sequence prediction module.
+
+# 4.1. Planning with Segmentation and Forecasting
+
+We define  $\pi$  as a goal-conditioned, multi-modal sequence prediction problem:
+
+$$
+\pi = P \left(a _ {k + 1}, a _ {k + 2}, \dots \mid V _ {t}, G\right). \tag {5}
+$$
+
+where  $\pi$  models the probability of goal-relevant and valid future actions sequences conditioned on  $V_{t}$  and  $G$ .
+
+Modeling the multi-modal sequence in Eq. 5 is computationally expensive and difficult to scale because of the high-dimensional state space of raw untrimmed video  $V_{t}$ . Also, there is limited data to learn the distribution over valid action sequences for goals in real-world applications. Tackling these challenges, we argue that the latent space of factors influencing the future plan can indeed be expressed in lesser dimensions. Particularly, to ensure scalability for handling raw videos and sample efficiency in learning valid action sequence distributions, we decompose our policy  $\pi$  into two modules. The first module is video segmentation, which converts the untrimmed video history  $V_{t}$  into a sequence of video segments i.e. a segment history  $S_{k} = (s_{1},\dots ,s_{k})$ , where each segment corresponds to an action  $a_{i}$  that occurred in the video. The second module enables forecasting i.e. it transforms the output of the segmentation module and generates the plan (see Fig. 1 (bottom)). While a probabilistic formulation of this decomposition can be expressed as:
+
+$$
+\pi = \sum_ {S _ {k}} \underbrace {P \left(a _ {k + 1} , a _ {k + 2} , \cdots \mid S _ {k} , G\right)} _ {\text {F o r e c a s t i n g}} \underbrace {P \left(S _ {k} \mid V _ {t}\right)} _ {\text {S e g m e n t a t i o n}}, \tag {6}
+$$
+
+where the segment history  $S_{k}$  is a latent variable, the summation over all possible segment histories is intractable. We, however, use Eq. 6 as the guiding expression to formulate the input and output of both modules. Next, we include the technical details of both of these modules.
+
+# 4.2. Segmentation Module
+
+This module splits the untrimmed video history  $V_{t}$  into segment history  $S_{k}$  of multiple segments, each segment
+
+![](images/087d67a1de51d94c1b34f3dfb37dc87fc6774ac21ebb63bfdfaf6e3967206c84.jpg)  
+Figure 2: VLaMP - Segmentation Module. The untrimmed visual history  $V_{t}$  is converted into segments, each consisting of observation  $o_{i}$  and the action  $a_{i}$ . The observation  $o_{i}$  is the collection of video frames of  $\delta$  seconds around the start time stamp  $t_{i}$  of the corresponding action  $a_{i}$ . Two such segments are shown here.
+
+corresponding to an action. The segmentation is done using a video-action segmentation model in three steps: preprocessing, classification, and consolidation. In the preprocessing step, the raw video frames from  $V_{t}$  are bundled into fixed-length window clips  $c_{i}$ , each of length 1 second, to obtain  $V_{t} = (c_{1},\dots ,c_{t})$ . In the classification step, a video-action segmentation model is used to output the most probable action for each clip  $c_{i}$ , which can be denoted as  $\tilde{A}_t = (\tilde{a}_1,\tilde{a}_2,\ldots ,\tilde{a}_t)$ . Finally, in the consolidation steps, we convert  $\tilde{A}_t$  into a form that can be used by the forecasting module; this form consists of two sequences: action history  $A_{k}$  and observation history  $O_{k}$ . To this end, same actions in consecutive seconds in  $\tilde{A}_t$  are consolidated to form the action history  $A_{k} = (a_{1},\dots ,a_{k})$ . As illustrated in Fig. 2, assuming  $t_i$  denotes the starting timestamp for  $a_{i}$ , we also extract video frames from  $t_i - \delta /2$  to  $t_i + \delta /2$  to obtain an observation window  $o_i$  corresponding to  $a_{i}$ , and consequently the full observation history  $O_{k} = (o_{1},\dots,o_{k})$ . The resultant segment history is termed  $S_{k} = ((o_{1},a_{1}),\dots ,(o_{k},a_{k}))$ , summarized in Fig. 2.
+
+# 4.3. Forecasting Module
+
+The usefulness of  $\pi$ 's decomposition expressed in Eq. (6) becomes apparent now. Modeling the segmentation module's output as segment history  $S_{k}$ , where each segment consisting of action and observations, allows writing the output of the forecasting module in an autoregressive manner:
+
+$$
+\begin{array}{l} P \left(a _ {k + 1}, a _ {k + 2}, \dots \mid o _ {1}, a _ {1}, \dots , o _ {k}, a _ {k}, G\right) \\ = \prod_ {i > 0} \sum_ {o _ {k + i}} P \left(o _ {k + i}, a _ {k + i} \mid o _ {1}, a _ {1}, \dots , o _ {k}, a _ {k}, G\right). \tag {7} \\ \end{array}
+$$
+
+Illustrated in Fig. 3, this autoregressive expression allows the possibility of using any sequence-to-sequence neural
+
+network as the forecasting module in combination with pretrained text and video encoders for representing action and observation history  $A_{k}$  and  $O_{k}$ . This general-purpose framework allows the use of any neural sequence model – LSTM [28], GRU [13], Transformers [75], etc. We choose to instantiate the forecasting module for  $\pi$  using a pretrained transformer-based LM. Next we present the details of the encoders for the two modalities and the LM based sequence model.
+
+Action encoder  $(f_{\mathrm{act}})$  Each action  $a_i$  in  $A_k$  is encoded by  $f_{\mathrm{act}}$  and the output is denoted by  $\alpha_i$ . Concretely, token embeddings are expressed as:
+
+$$
+\begin{array}{l} \left(\alpha_ {1}, \dots , \alpha_ {k}\right) = \left(f _ {\text {a c t}} \left(a _ {1}\right), \dots , f _ {\text {a c t}} \left(a _ {k}\right)\right), \text {w h e r e} \\ \alpha_ {i} = \left(\alpha_ {i} ^ {1}, \dots , \alpha_ {i} ^ {r _ {i}}\right) \in \mathbb {R} ^ {r _ {i} \times d} \tag {8} \\ \end{array}
+$$
+
+As we illustrate in Fig. 3 (left), each action  $a_i$  is tokenized into  $r_i$  tokens using appropriate tokenizer for the LM, the tokens are indexed using the vocabulary of the LM, and are represented using an embedding lookup from the token embeddings of the LM to produce  $\alpha_i$ . Here,  $r_i$  is the number of tokens and  $d$  is the dimensionality of token embeddings.
+
+Observation encoder  $(f_{\mathrm{obs}})$ . Visual cues play an integral role in knowing what actions lie ahead, towards achieving the goal prompt  $G$ . To this end, the visual observations  $O_k$  are encoded and play a critical role in the planner. Recall, the visual observation history  $O_k$  comprises of  $o_i$  corresponding to action  $a_i$ , each of  $\delta$  frames. As illustrated in Fig. 3 (middle), we transform each  $o_i$  employing the widely-adopted S3D backbone [84]  $f_{\mathrm{S3D^*}}$  (* denotes backbone is frozen). We must project visual encodings to a shared latent space of action (language) embeddings described before  $(\alpha_i)$ . To this end, we map S3D features via a trainable transformer mapper  $f_{\mathrm{map}}$ . Concretely,
+
+$$
+\begin{array}{l} \left(\beta_ {1}, \dots , \beta_ {k}\right) = \left(f _ {\mathrm {o b s}} \left(o _ {1}\right), \dots , f _ {\mathrm {o b s}} \left(o _ {k}\right)\right), \text {w h e r e} \\ \beta_ {i} = \left(\beta_ {i} ^ {1}, \dots , \beta_ {i} ^ {\delta}\right) \in \mathbb {R} ^ {\delta \times d} \text {a n d} f _ {\mathrm {o b s}} = f _ {\mathrm {S 3 D} ^ {*}} \circ f _ {\mathrm {m a p}} \tag {9} \\ \end{array}
+$$
+
+Overall, as we show in Fig. 3 (right), the resultant encoded sequence of representation for  $S_{k}$  is thus
+
+$$
+f _ {\text {e n c}} (S _ {k}) = \left(\beta_ {1}, \alpha_ {1}, \dots , \beta_ {k}, \alpha_ {k}\right) = H _ {k},
+$$
+
+Sequence model  $(f_{\mathrm{seq}})$ . Given the above encoding for the segment history  $S_{k}$ , the role of the sequence model is to predict a representation of the next token, that would in return enable VLaMP plan generation capabilities for VPA. Importantly, in the process of generating sequence of future actions autoregressively, we would also need to generate the representations of 'future observations'. Therefore, as we shown in Fig. 3, our sequence model, which consists of the transformer layers of a PTLM, also produces representations for vision (in addition to the necessary action tokens). Before proceeding further, we pause and introduce additional notation for the sequence model, which
+
+![](images/7b9f0c8b9772067c8dede95c30a238593029dd56f40d5f96be0718231733c1c2.jpg)  
+Figure 3: VLaMP - Forecasting Module. As shown in left and middle, actions and observations obtained from the segmentation module are encoded using appropriate modality encoders. The observation encoder leverages pretrained video encoder for observations, while also learning a mapper that aligns the representations from observations with actions. As shown on the right, VLaMP uses a joint sequence model on top of interleaved action (blue) and observation (gray) representations to forecast autoregressively the next representation.
+
+![](images/0ada3fca083539da731052742542fd4ea73f4bbc817e7d19b6f807cdaddccf2d.jpg)
+
+![](images/d581d41982481ea46b81e5bcf8d3975617081858294f2818bc57ff561951c912.jpg)  
+Figure 4: Tokenized Sequence with Masks. The encoded sequence of representations for  $k = 2$  segments, denoted alternatively using modality agnostic notation of  $H_{2}$  and mask  $M$  for next token prediction training.
+
+will make the subsequent explanation for training and inference easier to follow. As shown in Figure 4, we alternatively denote the sequence of representations  $(\beta_{1},\alpha_{1},\ldots ,\beta_{k},\alpha_{k})$  by  $H_{k} = (h_{1},\dots,h_{n})$ , with  $n = k\delta +\sum_{i = 1}^{k}r_{i}$ . A binary mask  $M = (m_{1},\dots,m_{n})$ , where  $m_{i}$  is 1 if the corresponding representation is for an action and 0 otherwise, can help obtain necessary action or visual observations. With this notation, given first  $j$  representations denoted as  $h_{1:j}$ , one step of the sequence model produces the representation for  $j + 1$ , i.e.,  $f_{\mathrm{seq}}(h_{1:j}) = \hat{h}_{j + 1}$ .
+
+# 4.4. Training
+
+The joint training of the segmentation and forecasting modules following Eq. (6) is intractable. But, by exploiting the availability of unpaired training data, we approximate Eq. (6) by feeding in the output of the segmentation module to the forecasting module and training them separately, each on their respective labeled data. The video-action segmentation model is trained utilizing the VideoCLIP setup [85], where in the segmentation model performs classification to predict the action for each second of the video. The forecasting model is trained by adopting the
+
+next representation prediction objective. Unlike vanilla LM pretraining, however, we also need to train for predicting visual representations in addition to text (action). Therefore, we use two different losses  $L_{\mathrm{act}}$  and  $L_{\mathrm{obs}}$  for text and visual representations respectively. Specifically,  $L_{\mathrm{act}}$  is the conventional cross-entropy loss over the LM's vocabulary  $V_{\mathrm{LM}}$  for the action representations while  $L_{\mathrm{obs}}$  is the mean-squared error between the predicted and the ground truth observation representations. The total loss is the sum of both the loss terms as shown in Eq. (11).
+
+$$
+L = - \sum_ {j = 1} ^ {n} m _ {j} L _ {\mathrm {a c t}} (\hat {h} _ {j}) + (1 - m _ {j}) L _ {\mathrm {o b s}} (\hat {h} _ {j}), \text {w h e r e}
+$$
+
+$$
+L _ {\mathrm {a c t}} = h _ {j} \cdot \hat {h} _ {j} - \log \sum_ {p = 1} ^ {| V _ {\mathrm {L M}} |} \exp \left(h _ {p} \cdot \hat {h} _ {j}\right); \tag {10}
+$$
+
+$$
+L _ {\mathrm {o b s}} = \frac {\| h _ {j} - \hat {h} _ {j} \| _ {2} ^ {2}}{d} \tag {11}
+$$
+
+In order to have a stable training, we use ground truth action history to construct  $S_{k}$  (and subsequently  $H_{k}$ ) instead of the output of the segmentation module. Appendix B provides further details on loss and optimizers for training.
+
+Inference. We next detail the inference procedure for VLaMP. Recall that we use  $1:n$  to denote a sequence of  $n$  representations (i.e,  $h_{1:n} = (h_1,\ldots ,h_n)$ ). Additionally, we denote the concatenation operator over two representation sequences by " $\diamond$ ". With this notation at hand, we define the score of an action  $a\in \mathcal{A}$  for following history  $h_{1:n}$  as
+
+$$
+\phi \left(h _ {1: n} \diamond f _ {\mathrm {a c t}} (a)\right) = \sum_ {j = 1} ^ {r _ {a}} a ^ {j} \cdot f _ {\mathrm {s e q}} \left(h _ {1: n} \diamond a ^ {1: j}\right), \tag {12}
+$$
+
+where  $\cdot$  is the vector dot product, and  $f_{\mathrm{act}}(a) = \alpha^{1:r_a} = (\alpha^1,\ldots ,\alpha^{r_a})$  is the sequence of encoded representations
+
+for action  $a$ . In other words, this score is the sum of unnormalized log-probability under the sequence model using the standard softmax distribution. We use this scoring function with to perform beam search (detailed inference algorithm: Algo. 1 in App. B).
+
+# 5. Experiments
+
+We instantiate VLaMP's segmentation module utilizing VideoCLIP [85] that has been fine-tuned on COIN and CrossTask. Similarly, we instantiate VLaMP's sequence model  $f_{\mathrm{seq}}$  (in the forecasting module) by GPT2 [60]. We utilize the open-sourced model weights of GPT2 (from HuggingFace [83]). For inference, we use Algo. 1 presented in Appendix B, with beam size  $B = 10$  for CrossTask and a  $B = 3$  for COIN.
+
+# 5.1. Data and Baselines
+
+Data. For a video  $V$  with goal  $G$ , both CrossTask [89] and COIN [73] provide annotations of the form  $\{a_k, (t_k, t'_k)\}_{k=1}^K$ , where  $a_k$  are the actions in the video, and  $t_k$  (resp.  $t'_k$ ) are the start (resp. end) timestamps for  $a_k$ . Given an annotated video consisting of  $K$  steps, we generate  $K - l$  examples, each with input  $x_k = (G, V_{t_k})$  and output  $y_k = (a_{k+1}, \ldots, a_{k+l})$ , for  $k = 1, \ldots, K - l$  (leaving at least  $l$  steps to predict in each example). Therefore from  $M$  videos, we generated  $N = \sum_{m=1}^{M} (K_m - l)$  examples, where  $K_m$  is the number of steps in the  $m$ -th video, forming a dataset  $\mathcal{D} = \{x^{(j)}, y^{(j)}\}_{j=1}^N$  suitable for VPA (total number of samples in shown in Table 2).
+
+Baselines. As a first step towards benchmarking, we utilize two heuristic baselines - a random baseline and most probable action baseline. Additionally, we also adopt a variety of strong goal-conditioned models. The procedure planning task is most relevant task from the literature to VPA, therefore, we adapt (details in App. A.1) the widely used DDN model introduced by Chang et al. [9], which is an established model in many procedural planning benchmarks in prior works [6, 87, 70]. Building bridges to the research community working on Ego4D's Long Term action Anticipation benchmark (LTA) [25], we also evaluate their best performing baseline. Finally, we include a prompt-based GPT-3 planner. Zooming out, these baselines span several paradigms: prompt-based, and random baselines are learning-free, the most-probable and random baselines are non-neural, while DDN and LTA baselines are learned and
+
+neural. Succinct descriptions of baselines are included next (details are deferred to Appendix A):
+
+- Random: Predicts the plan by picking all  $l$  actions uniformly randomly from the set of all actions  $\mathcal{A}$ .  
+- Random  $w/$  goal: A stronger baseline; for each goal  $G$ , we allow privilege access to a set of applicable actions to that goal  $\mathcal{A}_G \subseteq \mathcal{A}$ , and predicts the plan by randomly picking actions from the restricted set.  
+- Most probable action: Given the previous action  $a_j$ , picks the most probable next action  $a_j$  according conditional probability  $Pr(a_i | a_j)$  obtained using frequency count from the training.  
+- Most probable action  $w/$  goal: Akin to random  $w/$  goal baseline, we also evaluate a goal-conditioned most probable action baseline, that uses a goal-specific set of actions  $\mathcal{A}_G \subset \mathcal{A}$  during sampling. Since the most probable baselines, provide a probability distribution over the actions, we also employ beam search (for fairness, with the same beam size same as VLaMP), and pick the highest scoring plan.  
+- Dual Dynamics Network (DDN) [9]: Accounting for the difference in task definition, i.e. lack of visual goal, a direct application of DDN inference algorithm was not possible. So we keep DDN's network structure but use Algo.1 for inference on VPA.  
+- Ego4D Long Term Action Anticipation (LTA) [25]: The best performing baseline for LTA uses a SlowFast visual encoder followed by a transformer aggregator. We adapt this baseline using the S3D encoder for a fair apples-to-apples comparison with VLaMP.  
+- Prompt-based GPT-3 Planner: Huang et al. [30] utilize a LLM as zero-shot planners, we also experiment with prompting a frozen pretrained large language model i.e. GPT-3 for VPA (additional details in App. A.2).
+
+# 5.2. Quantitative Results
+
+In the following, we include quantitative findings of benchmarking methods on two video data sources (Tab. 3) and head-on ablations (Tab. 4).
+
+Improved performance across video datasets. As we show in Tab. 3, VLaMP significantly outperforms the baselines for both CrossTask and COIN. DDN that is customized for procedural tasks performs significantly better than heuristic baselines leading to a mAcc boost from  $12.7 \rightarrow 24.1\%$  (row 2 and 5,  $l = 4$ , Tab. 3). With our novel decomposition and pretraining objective, VLaMP outperforms DDN with a further bump up from  $24.1 \rightarrow 31.7\%$  (row 5 and 7).
+
+Steady gains in short & long horizon predictions. As length of prediction  $l$  increases ( $l = 1$  to  $l = 4$ ), the performance naturally decreases, across all baselines and tasks. With this in mind and zooming in on COIN re
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">l = 1</td><td colspan="2">l = 3</td><td colspan="2">l = 4</td><td></td></tr><tr><td>{n/m}Acc</td><td>SR</td><td>mAcc</td><td>mIOU</td><td>SR</td><td>mAcc</td><td>mIOU</td></tr><tr><td colspan="8">CrossTask Video Dataset [89]</td></tr><tr><td>Random</td><td>0.9</td><td>0.0</td><td>0.9</td><td>1.5</td><td>0.0</td><td>0.9</td><td>1.9</td></tr><tr><td>Random w/ goal</td><td>13.2</td><td>0.3</td><td>13.4</td><td>23.6</td><td>0.0</td><td>12.7</td><td>27.8</td></tr><tr><td>Most probable</td><td>10.4</td><td>1.7</td><td>6.1</td><td>9.9</td><td>1.3</td><td>5.5</td><td>13.9</td></tr><tr><td>Most probable w/ goal</td><td>12.4</td><td>2.4</td><td>8.9</td><td>15.5</td><td>1.5</td><td>7.9</td><td>20.5</td></tr><tr><td>DDN [9]</td><td>33.4</td><td>6.8</td><td>25.8</td><td>35.2</td><td>3.6</td><td>24.1</td><td>37.0</td></tr><tr><td>LTA [25]</td><td>26.9</td><td>2.4</td><td>24.0</td><td>35.2</td><td>1.2</td><td>21.7</td><td>36.8</td></tr><tr><td>VLaMP (ours)</td><td>50.6</td><td>10.3</td><td>35.3</td><td>44.0</td><td>4.4</td><td>31.7</td><td>43.4</td></tr><tr><td colspan="8">COIN Video Dataset [73]</td></tr><tr><td>Random</td><td>0.1</td><td>0.0</td><td>0.1</td><td>0.2</td><td>0.0</td><td>0.1</td><td>0.2</td></tr><tr><td>Random w/ goal</td><td>24.5</td><td>1.7</td><td>21.4</td><td>42.7</td><td>0.3</td><td>20.1</td><td>47.7</td></tr><tr><td>Most probable</td><td>0.7</td><td>1.6</td><td>4.3</td><td>6.8</td><td>1.6</td><td>8.2</td><td>15.3</td></tr><tr><td>Most probable w/ goal</td><td>23.9</td><td>10.9</td><td>18.0</td><td>24.9</td><td>9.1</td><td>16.3</td><td>32.2</td></tr><tr><td>DDN [9]</td><td>29.3</td><td>10.1</td><td>22.3</td><td>32.2</td><td>7.0</td><td>21.0</td><td>37.3</td></tr><tr><td>GPT-3 prompt-based [30]</td><td>19.3</td><td>1.7</td><td>11.1</td><td>19.5</td><td>0.0</td><td>10.5</td><td>21.0</td></tr><tr><td>VLaMP (ours)</td><td>45.2</td><td>18.3</td><td>39.2</td><td>56.6</td><td>9.0</td><td>35.2</td><td>54.2</td></tr></table>
+
+sults, we observe large gains of VLaMP over DDN (next best method). Particularly, a relative improvement of  $54\%$ $(29.3 \rightarrow 45.2)$  in mAcc for  $l = 1$  (row 7 and 8, Tab. 3) and  $68\%$ $(21.0 \rightarrow 35.2)$  for  $l = 4$  is demonstrated by VLaMP over DDN.
+
+Goal-conditioning is crucial. A key difference between the task formulations of LTA and VPA is goal conditioning (see Sec. 2). Comparing rows for LTA model and VLaMP-there is almost  $2\mathrm{x}$  gap  $(27\%$  vs.  $51\%$ $l = 1$  Tab.3). This underscores the importance of goal-conditioning. To second this inference, in Tab. 4 (rows 1 and 2), we precisely ablate the effect of providing the goal as a textual description for the basic (last-observation-only) model. The only difference is goal prompt  $G$  , which increases mAcc performance from  $44.5\rightarrow 53.1\%$  for  $l = 1$  and  $28.3\to 34.7\%$  for  $l = 3$
+
+Goal-conditioned random and most probable baseline come close to DDN on easy metrics. For fair comparisons, we include heuristic baselines. We observe the performance of random  $w/$  goal and most probable  $w/$  goal comes close to DDN, when evaluated using lenient metrics like mAcc and mIOU (see COIN results in Tab. 3). Note, that these two baselines enjoy the privileged access to (a much smaller) 'relevant actions set' for a given goal. On an average, the relevant or feasible action set is smaller for COIN than CrossTask videos. This apriori access to feasible actions makes random  $w/$  goal and most probable  $w/$  goal competitive (albeit, slightly unfair) baselines.
+
+Prompt-based LLM planner is not competitive. We evaluate the prompt-based GPT-3 planner on COIN and find
+
+Table 3: Performance on different datasets and horizons. The mean of various metrics (Sec. 3.1) obtained using 5 runs with different random seeds (std. errors are provide in Appendix E). Note that the action and observation history are the output of the separately finetuned video-action segmentation model and hence are noisy compared to the ground truth history.  
+
+<table><tr><td></td><td>G</td><td>Ak</td><td>Ok</td><td colspan="2">l = 3</td><td colspan="2">l = 1</td></tr><tr><td></td><td></td><td></td><td></td><td>SR</td><td>mAcc</td><td>mIOU</td><td>nAcc</td></tr><tr><td>1</td><td>X</td><td>X</td><td>ok</td><td>6.8 ± 0.3</td><td>28.3 ± 1.9</td><td>34.8 ± 2.0</td><td>44.5 ± 3.8</td></tr><tr><td>2</td><td>✓</td><td>X</td><td>ok</td><td>8.9 ± 0.2</td><td>34.7 ± 0.7</td><td>41.6 ± 0.8</td><td>53.1 ± 2.0</td></tr><tr><td>3</td><td>✓</td><td>✓</td><td>X</td><td>14.9 ± 0.3</td><td>37.8 ± 0.4</td><td>50.8 ± 0.6</td><td>48.0 ± 0.2</td></tr><tr><td>4</td><td>✓</td><td>✓</td><td>Ok</td><td>15.2 ± 0.3</td><td>43.5 ± 0.8</td><td>51.4 ± 0.9</td><td>64.8 ± 0.9</td></tr><tr><td>R</td><td>✓</td><td>✓</td><td>Ok</td><td>10.7 ± 0.2</td><td>36.5 ± 0.7</td><td>41.7 ± 0.6</td><td>61.4 ± 1.8</td></tr></table>
+
+Table 4: Role of different inputs and LM pre-training in VLaMP.  $G$ ,  $A_{k}$ ,  $O_{k}$  denote the three inputs to VLaMP: goal, action history and observation history, respectively. Shorthand of  $o_{k}$  means only the most recent observation of the full history ( $O_{k}$ ) is used. Mean  $\pm$  std. error over 5 random seeds on CrossTask are reported. The row R corresponds to the forecasting module of VLaMP trained with random initialization (same model architecture) as opposed to initialization using the weights of a pre-trained LM.
+
+that the prompt based model performs significantly worse than VLaMP. This highlights that VPA is not easy to solve using just a PTLM and prompting.
+
+# 5.3. Ablations and Error Analysis
+
+We undertake head-on ablations to quantify effect of each component of VLaMP - action history, the visual history, and language pre-training of the LM. To remove confounding factors, in Tab. 4 we use the ground truth output for the segmentation module. Following this, we include detailed error analysis.
+
+Action and observation history improve complementary planning metrics. As seen in all the rows with action history Tab. 4 (i.e.  $A_{k}$ ), the provision of action history increases difficult metrics such as SR. In comparing rows 2 and 3, we see that SR increases relatively by  $67\%$  ( $8.9 \rightarrow 14.9$ ), simply by using action history even without access to any past observation. However, the lack of observation history affects nAcc, which drops relatively by  $10\%$  ( $53.1 \rightarrow 48.0$ ) when observation history is swapped by action history between rows 2 and 3. This implies that observation history is important to predict the person's state in the task and what they might do next.
+
+Priors from the pre-trained LM improve performance. In Tab. 4 row  $\mathbf{R}$ , we report the performance of VLaMP when its forecasting module is trained with random weight initialization, i.e., the transformer of the same architecture trained from scratch instead of LM pre-training. We find the performance in row  $\mathbf{R}$  is thus much lower than that of VLaMP with pre-trained LM shown in row 4. This underscores the crucial importance of LM pre-training.
+
+Segmentation errors are detrimental. While the accuracy of video-action segmentation module (particularly, VideoCLIP model) is quite good -  $80.2\%$  for CrossTask  $(68.7\%)$
+
+for COIN), we believe this should be a major focus for future research. Why? The effect of this segmentation error in VLaMP's performance is quantified by comparing VLaMP row in Tab. 3 (CrossTask) with corresponding performance assuming perfect segmentation (Tab. 4, row 4). A relative  $50\%$  gap in success and big gaps in all other metrics.
+
+![](images/1a3965aa5361235b0aeb78132271c2c58450cde177458ff8458ef9f46b8fe7fb.jpg)  
+Figure 5: Effect of segmentation errors. Performance of VLaMP on CrossTask with classification errors  $(\%)$  in the video-action segmentation. As the errors increase, the performance gap between the full model and the one without access to observations increases.
+
+Visual observation history mitigates effect of video-action segmentation errors. Errors by segmentation module lead to mis-classification of actions, in turn leading to erroneous action history. To precisely study the effect of erroneous action history on VLaMP's performance, we perform controlled experiments wherein we add noise in the ground truth segmentation. This helps us tune desired segmentation error, irrespective of VideoCLIP's accuracy (in segmentation module). We achieve this by replacing a calculated  $\%$  of ground-truth actions by random actions. Decreasing performance, with increasing segmentation classification error, is expected and observed in Fig. 5. Further, we compare two models variations, VLaMP $(G, A_k, O_k)$  i.e. the full model (uses both action and observation history) and VLaMP $(A, A_k)$  that does not have visual observation history. As we increase segmentation classification error, the gap in the performance of these model variations increases. We infer that visual observations add robustness in VPA against video-action segmentation errors.
+
+Errors do not drop as more history is provided. In Fig. 6, we show mAcc w.r.t. the number of steps  $k$  in the history i.e. length of  $(O_k, A_k)$ . As more information is available (due to more steps in the history), one may expect the accuracy to improve. However, this isn't the trend for many tasks. For e.g., note the accuracy drops for the task "add oil to your
+
+![](images/983469d45deed55ff647f7b9a6fea5e7bdadcd974307ea4747d1428276bdfc34.jpg)  
+Figure 6: Accuracy in the tail of long activities is challenging. Longer history  $(O_{k},A_{k})$  do not necessary lead to better accuracy. Further longer histories occur with increasing rarity, leading to learning challenges. (Plot shows mAcc for  $l = 3$  CrossTask vs. the #steps in history  $k$ )
+
+car” w.r.t  $k$ . Digging deeper, we find two reasons for this pattern. First, we find that tasks with long history like “add oil to your car” tend to have repetitive steps towards the end of their action trajectories, making it difficult to predict the precise number and the pattern of such repetitions in VPA. Second is simply a data issue – long trajectories are also exponentially less frequent in the dataset. This forms the tail of the data distribution of action sequences in the dataset (see App. E for statistics). Consequently, prediction in this tail of activities, is a pertinent challenge for future research.
+
+# 6. Conclusion
+
+Visual Planning for Assistance is a new and intuitive formulation to support planning from natural visual observations for assisting humans in day-to-day activities. We benchmark VPA using standardized suite of prior baselines and a new multi-modal sequence modeling formation of VLaMP. The novel decomposition of a VLaMP policy into video action segmentation and forecasting leads to several efficiency and modeling benefits, that benefit the community, even beyond VPA. Particularly, this allows to leverage pre-trained LM that leads to significant performance gains. Alternative decompositions, self-supervised pre-training objectives for PTLMs, modeling multiple actors, and tapping into other modalities (audio, camera metadata, etc.) [26, 55, 43, 72, 32, 88] are interesting ways forward for the community, on the exciting task of VPA.
+
+Acknowledgements: The authors are indebted to Brian Chen for his efforts to open-source the code and LTA experiments, and to Gideon Stocek for supporting with the cluster management. We thank Vasu Sharma, Tushar Nagarajan, and Homanga Bharadhwaj for their writing feedback. We thank Weichao Mao for early discussions and Asli Celikyilmaz for her constant guidance.
+
+# References
+
+[1] Yazan Abu Farha, Alexander Richard, and Juergen Gall. When will you do what?-anticipating temporal occurrences of activities. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5343-5352, 2018. 2  
+[2] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, et al. Cm3: A causal masked multimodal model of the internet. arXiv preprint arXiv:2201.07520, 2022. 3  
+[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 3  
+[4] Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. Affordances from human videos as a versatile representation for robotics. 2023. 2  
+[5] Gedas Bertasius, Hyun Soo Park, X Yu Stella, and Jianbo Shi. First-person action-object detection with egonet. In RSS, 2017. 2  
+[6] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 2, 3, 7  
+[7] Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, et al. Do as i can, not as i say: Grounding language in robotic affordances. In 6th Annual Conference on Robot Learning, 2022. 2  
+[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. 14  
+[9] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In European Conference on Computer Vision, pages 334–350. Springer, 2020. 2, 3, 7, 8, 14, 16  
+[10] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In ECCV, 2020. 4  
+[11] Guo Chen, Sen Xing, Zhe Chen, Yi Wang, Kunchang Li, Yizhuo Li, Yi Liu, Jiahao Wang, Yin-Dong Zheng, Bingkun Huang, et al. Internvideo-ego4d: A pack of champion solutions to ego4d challenges. arXiv preprint arXiv:2211.09529, 2022. 2  
+[12] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. NeurIPS, 2021. 2, 4  
+[13] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine
+
+translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. 5  
+[14] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European Conference on Computer Vision (ECCV), pages 720-736, 2018. 2, 3  
+[15] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision. arXiv preprint arXiv:2006.13256, 2020. 2  
+[16] Ishita Dasgupta, Christine Kaeser-Chen, Kenneth Marino, Arun Ahuja, Sheila Babayan, Felix Hill, and Rob Fergus. Collaborating with language models for embodied reasoning. arXiv preprint arXiv:2302.00763, 2023. 2  
+[17] Matt Deitke, Dhruv Batra, Yonatan Bisk, Tommaso Campari, Angel X Chang, Devendra Singh Chaplot, Changan Chen, Claudia Pérez D'Arpino, Kiana Ehsani, Ali Farhadi, et al. Retrospectives on the embodied ai workshop. arXiv preprint arXiv:2210.06849, 2022. 2  
+[18] Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. Palm-e: An embodied multimodal language model. In arXiv preprint arXiv:2303.03378, 2023. 2, 3  
+[19] Jiafei Duan, Samson Yu, Hui Li Tan, Hongyuan Zhu, and Cheston Tan. A survey of embodied ai: From simulators to research tasks. IEEE Transactions on Emerging Topics in Computational Intelligence, 2022. 2  
+[20] Yazan Abu Farha, Qiuhong Ke, Bernt Schiele, and Juergen Gall. Long-term anticipation of activities with cycle consistency. arXiv preprint arXiv:2009.01142, 2020. 2  
+[21] Antonino Furnari, Sebastiano Battiato, Kristen Grauman, and Giovanni Maria Farinella. Next-active-object prediction from egocentric videos. Journal of Visual Communication and Image Representation, 2017. 2  
+[22] Antonino Furnari and Giovanni Maria Farinella. What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In ICCV, 2019. 2  
+[23] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13505-13515, 2021. 2  
+[24] Mohit Goyal, Sahil Modi, Rishabh Goyal, and Saurabh Gupta. Human hands as probes for interactive object understanding. In CVPR, 2022. 2  
+[25] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonio Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2, 3, 7, 8, 16, 17
+
+[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 9  
+[27] Todd Hester, Matej Vecerik, Olivier Pietquin, Marc Lanctot, Tom Schaul, Bilal Piot, Dan Horgan, John Quan, Andrew Sendonaris, Ian Osband, et al. Deep q-learning from demonstrations. In AAAI, 2018. 4  
+[28] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 5, 14, 15  
+[29] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, et al. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 3  
+[30] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. ICML, 2022. 2, 4, 7, 8, 14  
+[31] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608, 2022. 2  
+[32] Unnat Jain, Iou-Jen Liu, Svetlana Lazebnik, Aniruddha Kembhavi, Luca Weihs, and Alexander G Schwing. Gridtopix: Training embodied agents with minimal supervision. In ICCV, 2021. 9  
+[33] Unnat Jain, Luca Weihs, Eric Kolve, Ali Farhadi, Svetlana Lazebnik, Aniruddha Kembhavi, and Alexander G. Schwing. A cordial sync: Going beyond marginal policies for multiagent embodied tasks. In ECCV, 2020. 4  
+[34] Unnat Jain, Luca Weihs, Eric Kolve, Mohammad Rastegari, Svetlana Lazebnik, Ali Farhadi, Alexander G. Schwing, and Aniruddha Kembhavi. Two body problem: Collaborative visual task completion. In CVPR, 2019. 4  
+[35] Michael Janner, Qiyang Li, and Sergey Levine. Offline reinforcement learning as one big sequence modeling problem. NeurIPS, 2021. 2, 4  
+[36] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2022. 2  
+[37] Bingyi Kang, Zequn Jie, and Jiashi Feng. Policy optimization with demonstrations. In ICML, 2018. 4  
+[38] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017. 4  
+[39] Anastasis Kratsios, Behnoosh Zamanlooy, Tianlin Liu, and Ivan Dokmanic. Universal Approximation Under Constraints is Possible with Transformers, Feb. 2022. arXiv:2110.03303 [cs, math]. 2  
+[40] Tian Lan, Tsung-Chuan Chen, and Silvio Savarese. A hierarchical representation for future action prediction. In ECCV, 2014. 2
+
+[41] Shuang Li, Xavier Puig, Yilun Du, Clinton Wang, Ekin Akyurek, Antonio Torralba, Jacob Andreas, and Igor Mordatch. Pre-trained language models for interactive decision-making. arXiv preprint arXiv:2202.01771, 2022. 2  
+[42] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. arXiv preprint arXiv:2209.07753, 2022. 2  
+[43] Iou-Jen Liu, Unnat Jain, Raymond A Yeh, and Alexander Schwing. Cooperative exploration for multi-agent deep reinforcement learning. In ICML, 2021. 9  
+[44] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In CVPR, 2022. 2  
+[45] Siyuan Brandon Loh, Debaditya Roy, and Basura Fernando. Long-term action forecasting using multi-headed attention-based variational recurrent neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2419-2427, 2022. 2  
+[46] William Lotter, Gabriel Kreiman, and David Cox. Deep predictive coding networks for video prediction and unsupervised learning. arXiv preprint arXiv:1605.08104, 2016. 2  
+[47] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3  
+[48] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916, 2022. 3  
+[49] Weichao Mao, Ruta Desai, Michael Louis Iuzzolino, and Nitin Kamra. Action dynamics task graphs for learning plannable representations of procedural tasks. arXiv preprint arXiv:2302.05330, 2023. 2  
+[50] Esteve Valls Mascaró, Hyemin Ahn, and Dongheui Lee. Intention-conditioned long-term human egocentric action anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6048-6057, 2023. 2  
+[51] Esteve Valls Mascaro, Hyemin Ahn, and Dongheui Lee. Intention-conditioned long-term human egocentric action anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6048-6057, January 2023. 17  
+[52] Nazanin Mehrasa, Akash Abdu Jyothi, Thibaut Durand, Jiawei He, Leonid Sigal, and Greg Mori. A variational autoencoder model for stochastic point processes. in 2019 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3160-3169. 2  
+[53] Russell Mendonca, Shikhar Bahl, and Deepak Pathak. Structured world models from human videos. RSS, 2023. 2  
+[54] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 3
+
+[55] Himangi Mittal, Pedro Morgado, Unnat Jain, and Abhinav Gupta. Learning state-aware visual representations from audible interactions. In NeurIPS, 2022. 9  
+[56] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 3  
+[57] Tushar Nagarajan, Yanghao Li, Christoph Feichtenhofer, and Kristen Grauman. Ego-topo: Environment affordances from egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 163-172, 2020. 2  
+[58] Fabian Paischer, Thomas Adler, Vihang Patil, Angela Bitton-Nemling, Markus Holzleitner, Sebastian Lehner, Hamid Eghbal-Zadeh, and Sepp Hochreiter. History compression via language models in reinforcement learning. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 17156-17185. PMLR, 17-23 Jul 2022. 2  
+[59] Alexander Pashevich, Cordelia Schmid, and Chen Sun. Episodic transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15942-15952, 2021. 3  
+[60] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. https://d4mucfpkseyw.cloudfront.net/better-language-models/language_models_are_unsupervisedMULTITASK_learners.pdf, 2019.7  
+[61] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022. 2  
+[62] Keisuke Sakaguchi, Chandra Bhagavatula, Ronan Le Bras, Niket Tandon, Peter Clark, and Yejin Choi. proscript: Partially ordered scripts generation via pre-trained language models. arXiv preprint arXiv:2104.08251, 2021. 4  
+[63] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. ICCV, 2019. 4  
+[64] Fadime Sener, Rishabh Saraf, and Angela Yao. Transferring knowledge from text to video: Zero-shot anticipation for procedural actions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2  
+[65] Fadime Sener, Dipika Singhania, and Angela Yao. Temporal aggregate representations for long-range video understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 154-171. Springer, 2020. 2  
+[66] Fadime Sener and Angela Yao. Zero-shot anticipation for instructional activities. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 862-871, 2019. 2
+
+[67] Dandan Shan, Jiaqi Geng, Michelle Shu, and David Fouhey. Understanding human hands in contact at internet scale. In CVPR, 2020. 2  
+[68] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. arXiv preprint arXiv:2209.11302, 2022. 2  
+[69] Sanjana Srivastava, Chengshu Li, Michael Lingelbach, Roberto Martin-Martin, Fei Xia, Kent Elliott Vainio, Zheng Lian, Cem Gokmen, Shyamal Buch, Karen Liu, et al. Behavior: Benchmark for everyday household activities in virtual, interactive, and ecological environments. In Conference on Robot Learning, pages 477-490. PMLR, 2022. 4  
+[70] Jiankai Sun, De-An Huang, Bo Lu, Yun-Hui Liu, Bolei Zhou, and Animesh Garg. Plate: Visually-grounded planning with transformers in procedural tasks. IEEE Robotics and Automation Letters, 7(2):4924-4930, 2022. 2, 7, 15  
+[71] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems, 34:251-266, 2021. 4  
+[72] Andrew Szot, Unnat Jain, Dhruv Batra, Zsolt Kira, Ruta Desai, and Akshara Rai. Adaptive coordination for social embodied rearrangement. In ICML, 2023. 9  
+[73] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 1, 3, 4, 7, 8  
+[74] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 3  
+[75] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 4, 5, 15  
+[76] Mel Vecerik, Todd Hester, Jonathan Scholz, Fumin Wang, Olivier Pietquin, Bilal Piot, Nicolas Heess, Thomas Rothörl, Thomas Lampe, and Martin Riedmiller. Leveraging demonstrations for deep reinforcement learning on robotics problems with sparse rewards. arXiv preprint arXiv:1707.08817, 2017.4  
+[77] Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. arXiv preprint arXiv:1706.08033, 2017. 2  
+[78] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Anticipating visual representations from unlabeled video. In CVPR, 2016. 2  
+[79] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection without exocentric pretraining. arXiv preprint arXiv:2301.01380, 2023. 2
+
+[80] Saim Wani, Shivansh Patel, Unnat Jain, Angel Chang, and Manolis Savva. Multion: Benchmarking semantic map memory using multi-object navigation. NeurIPS, 2020. 4  
+[81] Luca Weihs, Unnat Jain, Iou-Jen Liu, Jordi Salvador, Svetlana Lazebnik, Aniruddha Kembhavi, and Alexander Schwing. Bridging the imitation gap by adaptive insubordination. In NeurIPS, 2021. 4  
+[82] Luca Weihs, Jordi Salvador, Klemen Kotar, Unnat Jain, Kuo-Hao Zeng, Roozbeh Mottaghi, and Aniruddha Kembhavi. Allenact: A framework for embodied ai research. arXiv preprint arXiv:2008.12760, 2020.4  
+[83] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38-45, 2020. 7  
+[84] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 318–335, Cham, 2018. Springer International Publishing. 5, 14  
+[85] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 6, 7  
+[86] Chulhee Yun, Srinadh Bhojanapalli, Ankit Singh Rawat, Sashank Reddi, and Sanjiv Kumar. Are Transformers universal approximators of sequence-to-sequence functions? Feb. 2022. 2  
+[87] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2938-2948, 2022. 2, 7, 15  
+[88] Chenhao Zheng, Ayush Shrivastava, and Andrew Owens. Exif as language: Learning cross-modal associations between images and camera metadata. In CVPR, 2023. 9  
+[89] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 1, 3, 4, 7, 8

2023/Pretrained Language Models as Visual Planners for Human Assistance/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ba0f3d03f4688f27606bab3d469fdc6da02578366f1b32f1ba3624fd6998475
+size 395643

2023/Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models/438697ca-0d24-4663-a1e4-f11b5f684ada_content_list.json

@@ -0,0 +1,1763 @@
+[
+    {
+        "type": "text",
+        "text": "Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models",
+        "text_level": 1,
+        "bbox": [
+            135,
+            130,
+            836,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zangwei Zheng $^{1}$  Mingyuan Ma $^{2}$  Kai Wang $^{1}$  Ziheng Qin $^{1}$  Xiangyu Yue $^{3}$  Yang You $^{1}$ $^{1}$ National University of Singapore  $^{2}$ UC Berkeley  $^{3}$ The Chinese University of Hong Kong",
+        "bbox": [
+            114,
+            202,
+            846,
+            244
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$  {zangwei, kai.wang, zihengq, youy}  $@$  comp.nus.edu.sg  $^{2}$ mamingyuan2001@berkeley.edu  $^{3}$ xyyue@ie.cuhk.edu.hk",
+        "bbox": [
+            102,
+            250,
+            866,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            301,
+            313,
+            316
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Continual learning (CL) can help pre-trained vision-language models efficiently adapt to new or under-trained data distributions without re-training. Nevertheless, during the continual training of the Contrastive Language-Image Pre-training (CLIP) model, we observe that the model's zero-shot transfer ability significantly degrades due to catastrophic forgetting. Existing CL methods can mitigate forgetting by replaying previous data. However, since the CLIP dataset is private, replay methods cannot access the pre-training dataset. In addition, replaying data of previously learned downstream tasks can enhance their performance but comes at the cost of sacrificing zero-shot performance. To address this challenge, we propose a novel method ZSCL to prevent zero-shot transfer degradation in the continual learning of vision-language models in both feature and parameter space. In the feature space, a reference dataset is introduced for distillation between the current and initial models. The reference dataset should have semantic diversity but no need to be labeled, seen in pre-training, or matched image-text pairs. In parameter space, we prevent a large parameter shift by averaging weights during the training. We propose a more challenging Multi-domain Task Incremental Learning (MTIL) benchmark to evaluate different methods, where tasks are from various domains instead of class-separated in a single dataset. Our method outperforms other methods in the traditional class-incremental learning setting and the MTIL by  $9.7\\%$  average score. Our code locates at https://github.com/Thunderbee/ZSCL.",
+        "bbox": [
+            75,
+            333,
+            473,
+            771
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            800,
+            209,
+            815
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Most deep learning models can access all the data during training [27, 16, 12]. If we want to expand a model's knowledge, such as learning a newly found animal species [41], we can re-train the model by adding new classes to the training dataset. However, re-training a large model is",
+        "bbox": [
+            75,
+            824,
+            468,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/c75e1ac3f0a441aaf322335722a181e73f7abc48d7a8e895bf06511d5e2d2e42.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            301,
+            874,
+            362
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/92b6c94e2754d253747798a82044fa5a64cabb9b7c6aac94cce33f61f2e5f62c.jpg",
+        "image_caption": [
+            "(a) Comparison between traditional CL and CL with a pre-trained vision-language model"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            364,
+            874,
+            473
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/c10d291ca81dad6cd99c3f2ff141da5b1ae32d191427ee18e0d17d2af6230b2d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            493,
+            687,
+            616
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/8ba2320b7369d63df75ed8058fd16d5665eb588e431df7dda0453ed1636163e1.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            707,
+            493,
+            875,
+            616
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/ed563a5c3d2b32f87328dfb825d0c8cc14c019a03e0d56a799d5d00ddf5c9f04.jpg",
+        "image_caption": [
+            "(b) Performance of different methods on preventing forgetting phenomenon"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            617,
+            686,
+            719
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/10428b9b041ff7ba9f51c846f7aa5a3ff0160dd0426e8cfc094ca94774103d1d.jpg",
+        "image_caption": [
+            "Figure 1. a) Conventional CL learns distinct task-specific heads, while CL with vision-language models can predict both learned tasks and out-of-distribution tasks. b) Accuracy (\\%) changes during CL of four datasets on 11 datasets. Our method is superior to others in preventing the forgetting of both zero-shot transfer ability and new knowledge."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            709,
+            617,
+            875,
+            719
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "costly. In contrast, continual learning (CL) [33, 47] incrementally learns task one after another. It can reduce",
+        "bbox": [
+            496,
+            869,
+            892,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            807,
+            47
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "19125",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "this cost by only learning the new data, thereby presenting itself as an efficient alternative to conventional learning methods [43, 34, 5]. Nonetheless, a model tends to forget previous information catastrophically when learning new tasks [47, 33, 59]. The \"catastrophic forgetting\" phenomenon is a great challenge for CL.",
+        "bbox": [
+            75,
+            90,
+            470,
+            181
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recently, vision-language models have shown powerful zero-shot transfer ability [44, 23, 32]. They can give zero-shot predictions without any training examples of a task. However, the performance on some tasks is poor due to insufficient relevant image-text pairs in the pre-training datasets. For example, it is difficult for CLIP [44] to distinguish among digital numbers, with an accuracy on MNIST [8] below  $60\\%$  much lower than a naively trained CNN [29]. If we want to widen the knowledge in the vision-language model by re-training, the computational cost is too large (e.g., CLIP is pretrained on 400 million image-text pairs). Fine-tuning downstream tasks achieves high performance, but one model for a task takes much memory, and the model is not reusable. Prompt learning [64, 63] keeps the backbone parameters unchanged. However, it is only effective with limited training data due to a limited prompt length [64, 24]. In contrast, continual learning makes learning new knowledge a lifelong process for the vision-language model. The continually learned model can handle any image-text input and can be further used for downstream tasks [11, 53].",
+        "bbox": [
+            75,
+            185,
+            473,
+            501
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We find that existing CL methods hardly prevent the forgetting phenomenon for zero-shot transfer ability in continual learning of a pre-trained vision-language model. As shown in Fig. 1 (a), the CL with a pre-trained vision-language model differs from the traditional one. Besides forgetting previously learned task knowledge, the CLIP-based CL suffers from forgetting pre-training knowledge, namely a degradation of zero-shot transfer ability. For the replay-based CL methods [47, 50, 36, 21, 28, 42], the dataset during pre-training may be private and inaccessible during fine-tuning. For distillation-based CL methods [33, 10, 13, 11], they do not lay enough emphasis on the pre-trained model. On the one hand, a large model state change hinders tasks thereafter from using high-quality feature representations. On the other hand, it significantly degrades zero-shot performance on unseen datasets.",
+        "bbox": [
+            75,
+            505,
+            470,
+            747
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our method ZSCL protects the Zero-Shot transfer ability during Continual Learning. We view the knowledge stored in the pre-training model from two perspectives: a well-learned feature space and a good value in the parameter space. In feature space, we re-design previous distillation loss [18, 47] with different loss styles, teacher models, and data sources. We find the original CLIP model, as opposed to the newly acquired model, is the best option for the teacher model. Instead of using data collected from previous tasks [47] or current task [18], we find a reference",
+        "bbox": [
+            75,
+            750,
+            470,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "dataset with diverse semantics (e.g., images sampled from ImageNet) is a good option for distillation loss. The reference images need not be labeled or matched with the text. Preserving the relative similarity between reference images and texts makes the feature space deviate little from the original. In the parameter space, WiSE-FT [58] proposes interpolating the initial and fine-tuned model for better performance. Inspired by this, we ensemble the weights throughout continuous training to prevent a significant shift from the initial CLIP, which can be seen as interpolating models of different zero-shot transfer and downstream task performance tradeoffs. The weight ensemble method is more stable and not sensitive to hyper-parameters.",
+        "bbox": [
+            496,
+            90,
+            890,
+            287
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To better evaluate our method, we propose a new benchmark Multi-domain Task Incremental Learning (MTIL). Previous CL tasks are crafted by separating classes in one dataset [14, 60, 65], where the images and classes are within a single domain. In contrast, MTIL consists of data from different sources requiring different expert knowledge. It comprises 11 tasks ranging from animal species to aircraft series recognition. As displayed in Fig. 1 (b), when sequentially training CLIP on 11 datasets, the drop in the performance of task  $i$  after training task  $i$  is the traditional forgetting phenomenon. The degradation in the accuracy compared to the original zero-shot one before training task  $i$  represents the forgetting in zero-shot transfer ability. Our method better protects the zero-shot transfer ability and preserves the learned knowledge. We outperform previous methods in both conventional class-incremental learning and MTIL settings. In Fig. 1 (b),",
+        "bbox": [
+            496,
+            287,
+            892,
+            542
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To summarize, our contributions are as follows:",
+        "bbox": [
+            517,
+            544,
+            834,
+            558
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We investigate continual training with the vision-language model and demonstrate the importance of preserving zero-shot transfer ability. A more challenging benchmark MTIL is proposed to evaluate CL methods where the tasks come from distinct domains.",
+            "- We propose a novel method ZSCL to mitigate the catastrophic forgetting problem in continual learning of the vision-language model by distillation in the feature space and weight ensemble in the parameter space.",
+            "- The proposed ZSCL outperforms all state-of-the-art methods across multiple benchmark datasets. On 10 steps CL of CIFAR100 and TinyImageNet, our method outperforms the best of previous ones by  $7.7\\%$  and  $6.0\\%$  for the Last accuracy. On MTCL, ZSCL outperforms others by  $10.9\\%$  on Transfer and  $9.7\\%$  on Avg. scores."
+        ],
+        "bbox": [
+            517,
+            569,
+            890,
+            830
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            844,
+            640,
+            859
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Vision-Language Models. Inspired by the success of language foundation models such as GPT-3 [3] and T5 [45], a",
+        "bbox": [
+            500,
+            869,
+            890,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "19126",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "series of work pre-train vision-language models on large-scale image-text datasets [31, 44, 23]. Among them, Contrastive Language-Vision Pre-training [44] achieves remarkable performance on various downstream tasks. It concentrates on aligning images and texts to acquire a joint embedding space. The CLIP model contains an image encoder [16, 12] and a text encoder [9]. During pre-training, contrastive learning is performed in which a paired image-text is a positive pair while image and text from different image-text pairs form a negative pair. For inference, the closest text embedding for the image is chosen as the prediction. Vision-language models can give zero-shot predictions on unseen tasks with a robust zero-shot transfer ability on various downstream tasks.",
+        "bbox": [
+            75,
+            90,
+            472,
+            303
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Continual Learning Methods. Most existing continual learning methods can be categorized into four groups: parameter expansion, memory replay, distillation loss, and parameter regularization. Parameter expansion methods such as DyTox [14] and DEN [61] introduce new parameters for new tasks. As we want to achieve a more powerful CLIP model at the end of CL, we do not change the architecture of the CLIP model. Memory replay methods [50, 36, 42, 28, 40] including iCaRL [47] and SER [21] keep a memory for exemplars from previously learned tasks. However, pre-training datasets are too large for choosing exemplars or may not be available at downstream training, and downstream data are not good exemplars for preserving the pretraining knowledge. Distillation loss such as LwF [33], LwM [10], LwF-VR [11], and PodNet [13] aligns current output space with previous ones, whereas distillation based on current tasks are not strong enough to maintain foundational knowledge. For the CLIP model, we find that general images, even if never seen by the model, plus sentences with enough semantics, can be a good \"replay\" for the distillation loss. The parameter regularization loss restricts the flexibility of model parameters by training loss [25, 62, 1] or weight averaging [30, 58]. Although this type of strategy performs badly compared to other ways in previous research [55, 2], we found it helpful for CLIP continual learning. The limited parameter space prevents the model from diverging significantly from its original state.",
+        "bbox": [
+            75,
+            321,
+            472,
+            731
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Vision-Language Models for Downstream Tasks. Many works propose different training strategies of vision-language models for better performance on downstream tasks, such as CoOp [64], CLIP-Adapter [15] and WiSE-FT [58]. However, very few attempts at continual learning exist. While [52, 56] focus on CL in the pretrain of VL and [39] trains VL from scratch using a small dataset, our problem setting differs as we address CL in downstream tasks with a pretrained VL. Currently, most VL models are trained directly on an accessible large",
+        "bbox": [
+            75,
+            750,
+            472,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "dataset, yet integrating knowledge continuously into a pretrained VL is a practical necessity. None of these studies address the zero-shot transfer degradation phenomenon, instead focusing on the traditional forgetting of learned knowledge. Recently, Thengane et al. [53] shows CLIP zero-shot prediction achieves state-of-the-art performance in CL settings even without any training. LwF-VR [11] is a modified LwF method for the CLIP model where random-generated sentences are used for distillation loss. However, it only considers the feature space, and the distillation with random sentences cannot protect the vision backbone. Differently, we re-examine what should be used for distillation in the feature space and combine the parameter space weight ensemble to provide better performance for the vision-language model continual learning.",
+        "bbox": [
+            496,
+            90,
+            893,
+            320
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Approach",
+        "text_level": 1,
+        "bbox": [
+            500,
+            329,
+            609,
+            345
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            500,
+            353,
+            640,
+            369
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Continual Learning. Given  $n$  tasks  $[\\mathcal{T}^1, \\mathcal{T}^2, \\dots, \\mathcal{T}^n]$ , continual training is conducted in sequence on each task  $\\mathcal{T}^i = (\\mathcal{D}^i, C^i)$ ,  $i = 1, \\dots, n$ . Here,  $\\mathcal{D}^i$  represents the task dataset  $\\{(x_j^i, y_j^i)\\}_{j=1}^{N_i}$ , where  $x_j^i$  is an image,  $y_j^i$  is a one-hot vector indicating the ground truth, and  $N_i$  is the number of images in the dataset. Class names  $C^i = \\{c_j^i\\}_{j=1}^{m_i}$  maps the label of an image to an object name, where  $m_i$  is the number of classes for task  $\\mathcal{T}^i$ . The objective of continual learning is to achieve good performance on all tasks.",
+        "bbox": [
+            496,
+            376,
+            893,
+            513
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Two continual learning settings are focused on in this study [55]. In task-incremental learning, at inference, the image  $\\pmb{x}$  to be predicted is given with its task identity  $t$ , so the model only needs to distinguish between different classes in  $C^t$ . In class-incremental learning, the task identity  $t$  is not given. Thus we need to predict with the combined class set  $C = \\bigcup_{i=1}^{n} C^i$ .",
+        "bbox": [
+            496,
+            513,
+            893,
+            619
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "CLIP model. The CLIP model contains an image encoder  $f_{i}$  and a text encoder  $f_{t}$ . The inference process of the CLIP model for image classification is as follows. First, for task  $\\mathcal{T}^i$ , each class  $c$  is transformed into a sentence by a template like \"a photo of  $\\{c\\}$ \". Then  $f_{t}$  encodes the classes into text embeddings  $\\{\\pmb{t}_{j}^{i}\\}_{j = 1}^{m_{i}}$ . An image encoder encodes input image  $\\pmb{x}_k$ . The similarity score between the image embedding and text embeddings are calculated as  $s_{k,j}^{i} = \\langle f_{i}(\\pmb{x}_{k}),\\pmb{t}_{j}^{i}\\rangle$  where  $\\langle \\cdot ,\\cdot \\rangle$  denotes the cosine similarity. The class with the largest similarity score is the prediction for the image.",
+        "bbox": [
+            496,
+            636,
+            893,
+            787
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To fine-tune the CLIP model for downstream tasks, cross-entropy loss CE is applied to the similarity score with a temperature scaling:",
+        "bbox": [
+            496,
+            787,
+            893,
+            833
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathrm {C E} (\\tau \\cdot \\boldsymbol {s} _ {i}, \\boldsymbol {y} _ {i}), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            596,
+            838,
+            890,
+            880
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\tau$  is a parameter learned during the pre-training.",
+        "bbox": [
+            500,
+            885,
+            861,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "19127",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/4329bff0d636ac16c1e23263b5d689cb7dbc7b4b289e45b30094e3f0c1b8cdce.jpg",
+        "image_caption": [
+            "Figure 2. Illustration of ZSCL in feature space. Fig. 3(a) shows the pipeline of distillation. The original and the current model encode the reference images and texts, respectively. The probability distribution of images and texts with respect to each other is distilled. Fig. 3(b) displays how distillation loss preserves the feature space. Compared with the reference dataset, the features of fine-tuning tasks lie in a small subspace. The distillation loss preserves the structure of the feature space by maintaining relative distances."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            93,
+            478,
+            253
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/f03a900cd5ed94daddb29f1585000c813ccaa2207a4462c7b9f5c88529204ab8.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            496,
+            93,
+            893,
+            257
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Distillation in Feature Space",
+        "text_level": 1,
+        "bbox": [
+            76,
+            351,
+            331,
+            368
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Well-learned feature space for aligned images and texts enables vision-language models' strong zero-shot transfer ability. It also facilitates the learning of downstream tasks. Compared with the pre-training dataset, downstream datasets lie in a small scope in the feature space (shown in Fig. 2(b)). Direct fine-tuning of downstream tasks greatly distorts the feature distribution of out-of-distribution data, which leads to a significant drop in zero-shot prediction performance.",
+        "bbox": [
+            75,
+            375,
+            468,
+            508
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "While the cross-entropy improves the performance by altering fine-tuned feature subspace, we need a new regularization to preserve the potential out-of-distribution feature space. The relative similarity between one image and different texts is:",
+        "bbox": [
+            75,
+            510,
+            468,
+            585
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\boldsymbol {p} = \\operatorname {S o f t m a x} \\left(\\boldsymbol {s} _ {1}, \\dots , \\boldsymbol {s} _ {m}\\right). \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            179,
+            599,
+            468,
+            616
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We hope the above similarity distribution is stable during fine-tuning for all potential images and texts. Given a teacher model  $\\overline{f}$ , distillation loss can be applied to penalize changes from the original distribution:",
+        "bbox": [
+            75,
+            626,
+            468,
+            686
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t . i m g}} = \\operatorname {C E} (\\boldsymbol {p}, \\bar {\\boldsymbol {p}}) = - \\sum_ {j = 1} ^ {m} \\boldsymbol {p} _ {j} \\cdot \\log \\bar {\\boldsymbol {p}} _ {j}, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            135,
+            698,
+            468,
+            738
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\overline{\\pmb{p}}$  is the distribution given by the teacher model.",
+        "bbox": [
+            76,
+            750,
+            437,
+            763
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Although the distillation form has been widely used in previous methods [18, 33, 47], they are applied to continual learning from scratch. We investigate different components of the distillation loss for enhancing pre-trained vision-language models.",
+        "bbox": [
+            75,
+            763,
+            468,
+            839
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Three components are discussed in this paper in detail: the data source, the teacher model, and the loss design. Sec. 5.1 shows the performance for the different choices. First, for the data source, LwF [33] uses data from the cur",
+        "bbox": [
+            75,
+            839,
+            468,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "rent task, while iCarl [47] carefully selects data from previous tasks. However, data from downstream tasks span a small subspace and are not good enough to preserve the whole feature space. An ideal choice is the pre-training dataset. However, the CLIP pre-training dataset is private, and the size is too large to load. To solve this challenge, we introduce the reference dataset. A reference dataset is a publicly available image dataset with enough semantics. Enough semantics can be seen as random sampling in the whole feature space. The texts can be related class names, unrelated sentences, or even random tokens. This is because text semantics are easier to sample, and we need not ground truth to calculate the distance between one image to sufficient text samples spread among feature space.",
+        "bbox": [
+            496,
+            352,
+            890,
+            564
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For the teacher model, [33, 47] use the model after learning task i-1 and before learning task i as the teacher model. During the continual training of the vision-language model, the feature space deviates gradually from the initial model. Using fine-tuned models as teacher models enlarges this change. In contrast, we find using the pre-trained model as a teacher model not only preserves the zero-shot transfer ability but also takes advantage of well-learned feature space for better downstream performance.",
+        "bbox": [
+            496,
+            575,
+            890,
+            712
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Finally, previous distillation loss is applied on traditional backbones, where a classification head gives the probability for different labels. For the vision-language model, the probability is calculated based on the relative distance between images and texts. Thus, in addition to Eq. (3), we impose regularization  $\\mathcal{L}_{\\mathrm{dist\\_txt}}$  on the distances from a text to a batch of images. The whole framework is shown in Fig. 2 (a) with the following training loss:",
+        "bbox": [
+            496,
+            724,
+            890,
+            847
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c e}} + \\lambda \\cdot \\left(\\mathcal {L} _ {\\mathrm {l w f . i m g}} + \\mathcal {L} _ {\\mathrm {l w f . t x t}}\\right). \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            580,
+            885,
+            890,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "19128",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/eb3d24a46c63689d28133592e77841b07804ba7d89e7f0c1fa29b32d5c02b588.jpg",
+        "image_caption": [
+            "Figure 3. Models during training contain different tradeoffs between zero-shot and new task performance. Points for FT are sampled every 100 iterations, and the ones for WiSE-FT means different  $\\alpha$  choices. WE ensembles models during training and achieve better performance."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            87,
+            433,
+            263
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3. Weight Ensemble in Parameter Space",
+        "text_level": 1,
+        "bbox": [
+            76,
+            366,
+            403,
+            383
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Machine learning models integrate learned knowledge in their parameters. To mitigate the forgetting problem, a series of works [25, 62, 1] impose regularization losses on the changes of parameters. Weight consolidation (WC) [25] imposes the following loss:",
+        "bbox": [
+            75,
+            390,
+            470,
+            465
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {W C}} = \\sum_ {i} \\left(\\theta_ {i} - \\bar {\\theta} _ {i}\\right) ^ {2}. \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            196,
+            474,
+            468,
+            506
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\theta$  is the parameters of the current model, and  $\\overline{\\theta}$  is the reference ones. Although this regularization prevents forgetting, it hinders learning new tasks in a challenging CL setting.",
+        "bbox": [
+            75,
+            516,
+            468,
+            577
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Apart from regularization losses, another method in parameter space is to ensemble different model weights. Model soup [57] averages weights of multiple fine-tuned models to improve the model's robustness but introduces additional training costs. WiSE-FT [58] propose a weighted average between fine-tuned model and the original model to improve the out-of-distribution prediction robustness:",
+        "bbox": [
+            75,
+            578,
+            468,
+            683
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf (x; (1 - \\alpha) \\cdot \\theta_ {0} + \\alpha \\cdot \\theta_ {1}), \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            181,
+            693,
+            468,
+            709
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\theta_0$  is the original model and  $\\theta_{1}$  is the fine-tuned one. However, this method is hyper-parameter-sensitive where different  $\\alpha$  gives different tradeoffs between zero-shot transfer ability and downstream task performance (blue line in Fig. 3.",
+        "bbox": [
+            75,
+            719,
+            468,
+            795
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Inspired by this, we extend the weighted average to the CL setting. The motivation for the weighted average is to prevent fine-tuning from losing too much knowledge in the original model. As training goes by (green line in Fig. 3), the model performs better on new tasks while losing accuracy on previous ones. The models among training compose a sequence of different learning-forgetting tradeoffs.",
+        "bbox": [
+            75,
+            795,
+            470,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/c1c8038706070e6074aff8433598727beaf8d68f06790075ab76e09402bafbd0.jpg",
+        "image_caption": [
+            "(a)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            108,
+            578,
+            179
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/7612a0001597f0d99113646e1932e6b9a2dfdfa5b9fc31e1b66a56d4a369c262.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            593,
+            108,
+            666,
+            179
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/816e3786999b886a7661b676ba1153e0a0d8b60c6ea218d1e5871216e42d0f37.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            508,
+            188,
+            576,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/952c17a62f4e4c386253f86e5ba5a7e8676eb7209cde979b79afbd3ab84319b1.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            594,
+            188,
+            666,
+            260
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/81e23c68abd350a661469da9d431633c4fcb113cc4ec578215ac9647b344ca15.jpg",
+        "image_caption": [
+            "(b)",
+            "Figure 4. Fig.(a): examples of tasks from different domains in MTIL benchmark. Fig.(b): illustration of calculating metrics Transfer, Avg. and Last during continual learning."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            684,
+            104,
+            890,
+            287
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Instead of interpolating only between the initial and the final model, our method weight ensemble (WE) averages the weights in the sequence during the training time:",
+        "bbox": [
+            498,
+            364,
+            892,
+            411
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\theta} _ {t} = \\left\\{ \\begin{array}{l l} \\theta_ {0} & t = 0 \\\\ \\frac {1}{t + 1} \\theta_ {t} + \\frac {t}{t + 1} \\cdot \\hat {\\theta} _ {t - 1} & \\text {e v e r y I i t e r a t i o n s} \\end{array} . \\right. \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            537,
+            419,
+            890,
+            460
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where model weight sampling happens every  $I$  iteration. The method is related to Stochastic Weight Averaging (SWA) [22], but we do not use a modified learning rate schedule here because instead of getting better generalization ability, WE aim to give an improved learning-forgetting tradeoff. As shown in Fig. 3, WE achieves better performance on downstream tasks than WiSE-FT. In addition, while WiSE-FT is sensitive to different values of  $\\alpha$ , our method is much more robust under different hyperparameter  $(I)$  choices.",
+        "bbox": [
+            498,
+            469,
+            890,
+            619
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Multi-domain Task Incremental Learning",
+        "text_level": 1,
+        "bbox": [
+            498,
+            633,
+            874,
+            650
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Conventional Continual Learning Benchmark. A benchmark consisting of several tasks is needed to evaluate different methods for continual learning. Most previous benchmarks are built by separating classes in a single dataset, such as MNIST [55], CIFAR100 [14], TinyImageNet [60], and ImageNet [60, 65]. We also evaluate our method with traditional benchmarks. In CIFAR100 [26], classes are separated into groups to build tasks. Suppose the dataset has  $m$  classes, a  $k$ -step setting means we learn  $m / k$  classes in each new task. The CIFAR100 contains 100 classes, and the setting of 10, 20, and 50 steps are visited. For TinyImageNet with  $m = 200$ , the first step learns 100 classes, and the rest is learned with 5, 10, and 20 steps. As for ImageNet-100, we have two settings: ImageNet-100-B0, which includes the same amount of classes for each step, and ImageNet-100-B50, which has",
+        "bbox": [
+            496,
+            657,
+            890,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "19129",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/bd47b8cc97b89c696691abf6145c7fa7b7975a21209b7a49d0a2c12b27dbebbe.jpg",
+        "table_caption": [
+            "Table 1. Ablation experiments. Default settings are marked in gray, which uses image and text distillation loss with the initial CLIP model on 100k ImageNet images and texts generated from ImageNet classes with a simple template.",
+            "(a) Continual learning loss."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>loss</td><td>Transfer</td><td>Avg.</td><td>Last</td></tr><tr><td>CE only</td><td>44.6</td><td>55.9</td><td>77.3</td></tr><tr><td>Feat. Dist.</td><td>47.6</td><td>58.7</td><td>77.1</td></tr><tr><td>Image-only</td><td>56.5</td><td>68.9</td><td>82.1</td></tr><tr><td>Text-only</td><td>56.7</td><td>69.0</td><td>82.6</td></tr><tr><td>Both</td><td>56.8</td><td>69.2</td><td>83.0</td></tr></table>",
+        "bbox": [
+            86,
+            146,
+            336,
+            224
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/aa1d0b2e84d0d8395d5958f4119eb749f0b5b6e3f4a5e35a6c7cbb93ff77a63a.jpg",
+        "table_caption": [
+            "(d) Teacher model."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>source</td><td>Transfer</td><td>Avg.</td><td>Last</td></tr><tr><td>Initial</td><td>56.8</td><td>69.2</td><td>83.0</td></tr><tr><td>n-1</td><td>53.9</td><td>66.6</td><td>80.7</td></tr><tr><td>WiSE(0.5)</td><td>56.4</td><td>68.9</td><td>82.9</td></tr><tr><td>WiSE(0.8)</td><td>56.2</td><td>67.8</td><td>81.3</td></tr></table>",
+        "bbox": [
+            86,
+            241,
+            341,
+            308
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/e7a53235fc55eabdceac0aed247a87eb840a49f999bd06615b82055e6889be59.jpg",
+        "table_caption": [
+            "(b) Data sources for replay."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>source</td><td>Transfer</td><td>Avg.</td><td>Last</td></tr><tr><td>current</td><td>56.7</td><td>66.5</td><td>80.2</td></tr><tr><td>ImageNet</td><td>56.8</td><td>69.2</td><td>83.0</td></tr><tr><td>CC</td><td>57.2</td><td>68.5</td><td>80.9</td></tr><tr><td>CIFAR100</td><td>55.2</td><td>65.9</td><td>80.7</td></tr><tr><td>Flowers</td><td>54.7</td><td>66.0</td><td>80.8</td></tr></table>",
+        "bbox": [
+            367,
+            147,
+            620,
+            224
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/1bdffdf8de7307efbe544e64fd3f83a288a35b2bfb186dc9758440726f83378b.jpg",
+        "table_caption": [
+            "(e) # images for replay."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>#image</td><td>Transfer</td><td>Avg.</td><td>Last</td></tr><tr><td>1M</td><td>58.7</td><td>70.1</td><td>83.2</td></tr><tr><td>100k</td><td>56.8</td><td>69.2</td><td>83.0</td></tr><tr><td>10k</td><td>57.8</td><td>68.7</td><td>81.2</td></tr><tr><td>1k</td><td>56.3</td><td>67.6</td><td>80.8</td></tr></table>",
+        "bbox": [
+            367,
+            241,
+            620,
+            306
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/89a67456fdb955935a5376d6ef17aa5c518dd36a1e1af1bfedd4d3b100627246.jpg",
+        "table_caption": [
+            "(c) Text sources for replay."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>source</td><td>Transfer</td><td>Avg.</td><td>Last</td></tr><tr><td>current</td><td>51.8</td><td>64.9</td><td>82.0</td></tr><tr><td>prev. all</td><td>54.0</td><td>70.2</td><td>83.7</td></tr><tr><td>1k classes (IN)</td><td>56.8</td><td>69.2</td><td>83.0</td></tr><tr><td>13k Sent. (CC)</td><td>58.9</td><td>70.5</td><td>84.0</td></tr><tr><td>1k Rand. Sent.</td><td>58.7</td><td>70.2</td><td>83.8</td></tr></table>",
+        "bbox": [
+            645,
+            147,
+            898,
+            224
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/1fbef6a5fee68628c0b246ae13a2e24e77befe915ade7a66d22af3cc956111c2.jpg",
+        "table_caption": [
+            "(f) # image classes for replay."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>#class</td><td>Transfer</td><td>Avg.</td><td>Last</td></tr><tr><td>1000</td><td>56.8</td><td>67.6</td><td>83.0</td></tr><tr><td>100</td><td>56.7</td><td>67.3</td><td>82.3</td></tr><tr><td>10</td><td>53.8</td><td>66.4</td><td>81.0</td></tr><tr><td>1</td><td>53.1</td><td>65.5</td><td>80.5</td></tr></table>",
+        "bbox": [
+            645,
+            241,
+            898,
+            306
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/c391686ce697cfd0b0b4f8435ebbc81c674df95a2fd0850f219f8c907a0ccd04.jpg",
+        "table_caption": [
+            "Table 2. Ablation study of different components for ZSCL."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Transfer</td><td>Δ</td><td>Avg.</td><td>Δ</td><td>Last</td><td>Δ</td></tr><tr><td colspan=\"7\">CLIP ViT-B/16@224px</td></tr><tr><td>Zero-shot</td><td>69.4</td><td>0.0</td><td>65.3</td><td>0.0</td><td>65.3</td><td>0.0</td></tr><tr><td>Continual Learning</td><td>44.6</td><td>-24.8</td><td>55.9</td><td>-9.4</td><td>77.3</td><td>+12.0</td></tr><tr><td>+ Distillation</td><td>58.9</td><td>-10.5</td><td>70.5</td><td>+5.2</td><td>83.8</td><td>+18.5</td></tr><tr><td>+ WiSE-FT (best α)</td><td>61.7</td><td>-7.7</td><td>71.6</td><td>+6.3</td><td>83.3</td><td>+18.0</td></tr><tr><td>+ WE (ZSCL*)</td><td>62.2</td><td>-7.2</td><td>72.6</td><td>+7.3</td><td>84.5</td><td>+19.2</td></tr><tr><td>+ WC</td><td>67.6</td><td>-1.8</td><td>74.5</td><td>+9.2</td><td>83.2</td><td>+17.9</td></tr><tr><td>+ WiSE-FT</td><td>67.7</td><td>-1.7</td><td>74.2</td><td>+8.9</td><td>81.9</td><td>+16.6</td></tr><tr><td>+ WE (ZSCL)</td><td>68.1</td><td>-1.3</td><td>75.4</td><td>+10.1</td><td>83.6</td><td>+18.3</td></tr></table>",
+        "bbox": [
+            78,
+            357,
+            473,
+            483
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "50 classes for the first step, and the remaining 50 classes are observed progressively over the next 10 stages. For ImageNet [7], we investigate a 10-step setting, which learns 100 new classes per task.",
+        "bbox": [
+            75,
+            508,
+            470,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "MTIL Benchmark. Different classes from one dataset share the common image source and a similar style [48, 17]. Thus, we propose Multi-domain Task Incremental Learning (MTIL), a cross-domain version of task incremental learning. Different tasks are collected from different domains, requiring different domain knowledge for humans to achieve high accuracy. Our MTIL benchmark consists of 11 tasks (detailed in supplementary materials), as some of the tasks illustrated in Fig. 4 (a). The MTIL benchmark is very challenging with a total number of 1,201 classes. Two orders are used for the evaluation; the first one is alphabet order (Order-I): Aircraft, Caltech101, CIFAR100, DTD, EuroSAT, Flowers, Food, MNIST, OxfordPet, StanfordCars, SUN397. And the second one is a random order (Order-II): StanfordCars, Food, MNIST, OxfordPet, Flowers, SUN397, Aircraft, Caltech101, DTD, EuroSAT, CIFAR100. Experiments are done in Order I by default.",
+        "bbox": [
+            75,
+            590,
+            470,
+            849
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metrics. The metrics of MTIL are illustrated in Fig. 4(b), where rows represent training steps and a",
+        "bbox": [
+            75,
+            869,
+            470,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "column shows performance for one dataset. For conventional continual learning, only scores under the diagonal are meaningful, since they cannot give zero-shot predictions on unseen tasks. In contrast, the zeros-hot transfer ability enables a vision-language model to provide predictions for all datasets. The average accuracy on all datasets among all timestamps is the \"Avg\" metric. The \"Last\" metric is the average performance of all tasks after CL. The \"Transfer\" metric is the average task performance in the upper-right triangle of the matrix. Every task's performance is first averaged to equal the weight of each dataset. It measures to what extent the zero-shot transfer ability is preserved. Before learning task  $i$ , tasks not earlier than  $i$  are not finetuned. Thus, their performance is an indicator of zero-shot transfer ability.",
+        "bbox": [
+            496,
+            333,
+            893,
+            561
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            575,
+            633,
+            593
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation. We use CLIP [44] model with image encoder ViT-B/16 [12]. We conduct training with AdamW [37] optimizer and use a label smoothing [38] technique for a better baseline result. For multi-domain task continual learning, we train 1K iterations for each task, while for class-incremental learning, we follow the same evaluation setting in [14]. More implementation details can be found in the supplementary material.",
+        "bbox": [
+            496,
+            601,
+            893,
+            723
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. Main Properties",
+        "text_level": 1,
+        "bbox": [
+            500,
+            734,
+            661,
+            750
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We ablate our method in feature-space in Tab. 1, and different choices for parameter-space regularization in Tab. 2. Several interesting characteristics are noted.",
+        "bbox": [
+            496,
+            757,
+            890,
+            804
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Continual learning loss. In Tab. 1a, several types of loss for feature space are tested. The Feature Distance penalizing absolute distances achieves a low accuracy. Distillation loss on probability distribution regularizes relative distance in the feature space. With image-only or text-only distilla",
+        "bbox": [
+            496,
+            825,
+            893,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "19130",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/486e0281799167fdc39ad25fe08f6ed1648465a543e6415e44a1f33130cf66aa.jpg",
+        "image_caption": [
+            "Figure 5. t-SNE on five models' outputs together of Aircraft datasets after MTCL training: only our model maintains a similar feature distribution to the original CLIP ones with minor shift, while the rest significantly distort the feature space."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            90,
+            243,
+            205
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7784c328f4cdf2b9bb2ed7f173c38e0d41ae1044f2237fb671b5c4c8254a9dc8.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            248,
+            90,
+            405,
+            205
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/09b430fabc5fe55f1527a5618eccc3aa9a80fbc30f1e07b877ca20d9e3d1d5a5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            408,
+            90,
+            566,
+            205
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/eb90c5de160688afb191eadaa39b87a75136e54a5cc737262f85f987cb769e3e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            571,
+            90,
+            728,
+            205
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e8277e097f9a1ced82aa1b0687e58001beb7132aad61f8a706e47c69a5d84737.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            732,
+            90,
+            885,
+            205
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/93a36b72b943f7cd603380a954c6668578e979820d207fe6599536d19724b99c.jpg",
+        "table_caption": [
+            "Table 3. Comparison of different methods on MTIL in Order I."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Transfer</td><td>Δ</td><td>Avg.</td><td>Δ</td><td>Last</td><td>Δ</td></tr><tr><td colspan=\"7\">CLIP ViT-B/16@224px</td></tr><tr><td>Zero-shot</td><td>69.4</td><td>0.0</td><td>65.3</td><td>0.0</td><td>65.3</td><td>0.0</td></tr><tr><td>Continual Learning</td><td>44.6</td><td>-24.8</td><td>55.9</td><td>-9.4</td><td>77.3</td><td>+12.0</td></tr><tr><td>LwF [33]</td><td>56.9</td><td>-12.5</td><td>64.7</td><td>-0.6</td><td>74.6</td><td>+9.0</td></tr><tr><td>iCaRL [47]</td><td>50.4</td><td>-19.0</td><td>65.7</td><td>+0.4</td><td>80.1</td><td>+14.8</td></tr><tr><td>LwF-VR [11]</td><td>57.2</td><td>-12.2</td><td>65.1</td><td>-0.2</td><td>76.6</td><td>+11.3</td></tr><tr><td>WiSE-FT [58]</td><td>52.3</td><td>-17.1</td><td>60.7</td><td>-4.6</td><td>77.7</td><td>+12.4</td></tr><tr><td>ZSCL* (Ours)</td><td>62.2</td><td>-7.2</td><td>72.6</td><td>+7.3</td><td>84.5</td><td>+19.2</td></tr><tr><td>ZSCL (Ours)</td><td>68.1</td><td>-1.3</td><td>75.4</td><td>+10.1</td><td>83.6</td><td>+18.3</td></tr></table>",
+        "bbox": [
+            78,
+            292,
+            470,
+            426
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "tion, the Transfer, Avg., and Last accuracy all improve. A further boost in performance occurs with both of the distillation losses.",
+        "bbox": [
+            75,
+            450,
+            468,
+            496
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Data source for replay. Tabs. 1b, 1c, 1e and 1f seek the standard for a good reference dataset. As shown in Tab. 1b, distillation on images of current tasks achieves a good Transfer score. However, it hinders the learning on new tasks and results in a low Avg. and Last score. Images in a specific domain (e.g., Flowers) are also not good choices. General images in ImageNet and Conceptual Caption (CC) datasets are examples of good reference datasets. These images are easily available by a web crawler [54]. Text with more semantics can improve performance (shown in Tab. 1c). When using sentences from the Conceptual Caption dataset [49], or even sentences randomly generated from the CLIP vocabulary, there is no ground truth target among the texts for the image from the ImageNet dataset. However, they all achieve an improvement due to more semantics. The reference image dataset does not need to be labeled, matched with the text, or seen by the CLIP model.",
+        "bbox": [
+            75,
+            521,
+            468,
+            777
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Enough semantics in the reference image dataset boosts the distillation performance. In Tabs. 1e and 1f, a smaller number of images and classes all lead to a degradation in the performance. Fewer classes of images in the reference dataset have a worse impact on the performance compared with the image numbers. To keep a reasonable memory buffer, we randomly sample  $100\\mathrm{k}$  images from ImageNet for MTIL and use texts from CC dataset. For class-",
+        "bbox": [
+            75,
+            779,
+            470,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/a3c81c468e3153ab25efe2bc7736f185000675360c5f0818bbee81b4145c0144.jpg",
+        "table_caption": [
+            "Table 4. Comparison of different methods on MTIL in Order II."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Transfer</td><td>Δ</td><td>Avg.</td><td>Δ</td><td>Last</td><td>Δ</td></tr><tr><td colspan=\"7\">CLIP ViT-B/16@224px</td></tr><tr><td>Zero-shot</td><td>65.4</td><td>0.0</td><td>65.3</td><td>0.0</td><td>65.3</td><td>0.0</td></tr><tr><td>Continual Learning</td><td>46.6</td><td>-18.8</td><td>56.2</td><td>-9.1</td><td>67.4</td><td>+2.1</td></tr><tr><td>LwF [33]</td><td>53.2</td><td>-12.2</td><td>62.2</td><td>-5.2</td><td>71.9</td><td>+6.6</td></tr><tr><td>iCaRL [47]</td><td>50.9</td><td>-14.5</td><td>56.9</td><td>-8.4</td><td>71.6</td><td>+6.3</td></tr><tr><td>LwF-VR [11]</td><td>53.1</td><td>-12.3</td><td>60.6</td><td>-7.4</td><td>68.3</td><td>+0.9</td></tr><tr><td>WiSE-FT [58]</td><td>51.0</td><td>-14.4</td><td>61.5</td><td>-5.9</td><td>72.2</td><td>+6.9</td></tr><tr><td>ZSCL*</td><td>59.8</td><td>-5.6</td><td>71.8</td><td>+6.5</td><td>83.3</td><td>+18.0</td></tr><tr><td>ZSCL</td><td>64.2</td><td>-1.2</td><td>74.5</td><td>+9.2</td><td>83.4</td><td>+18.1</td></tr></table>",
+        "bbox": [
+            500,
+            292,
+            893,
+            426
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "incremental learning, conceptual caption dataset's validation set (28k images) is used to avoid information leakage.",
+        "bbox": [
+            498,
+            452,
+            890,
+            483
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Teacher model. Tab. 1d shows the performance of distillation loss with different teacher models. Unlike conventional continual learning, the teacher model should not be the one trained on the previous task; otherwise, the deviation from the initial CLIP in the previous task may be amplified. In contrast, with the initial CLIP as the teacher model, not only is the zero-shot performance improved but the Mean and Last scores are also boosted, indicating that preserving high-quality feature space facilitates continual learning.",
+        "bbox": [
+            496,
+            502,
+            892,
+            654
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Parameter-space regularization. In Tab. 2, we experiment with three different parameter-space regularizations. We experiment with two variants of WiSE-FT [58]: the weighted average between the current model with the initial CLIP or the one at the previous task. The result shows the latter one is a better choice because keeping weight averaging with initial CLIP loses the newly learned knowledge. We experiment with different  $\\alpha$  choices for WiSE, and the best result is reported. While distillation loss improves the whole performance, the parameter-space regularization further protects the zero-shot transfer ability with a higher Transfer. Among the three parameter-space regularizations, only WE achieve a better Last score. WC greatly improves the Transfer scores with a lower Last score. A combination of the weight consolidation loss with weight",
+        "bbox": [
+            496,
+            674,
+            893,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "19131",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/9c645337f982c4049b97a0b2cbd146fe3651998e8ffb2db64402112ddaccf5f9.jpg",
+        "table_caption": [
+            "Table 5. Transfer, Avg., and Last scores (%) of different continue training methods on MTIL benchmark."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Aircraft</td><td>Caltech101</td><td>CIFAR100</td><td>DTD</td><td>EuroSAT</td><td>Flowers</td><td>Food</td><td>MNIST</td><td>OxfordPet</td><td>Cars</td><td>SUN397</td></tr><tr><td colspan=\"12\">CLIP ViT-B/16@224px</td></tr><tr><td>Zero-shot</td><td>24.3</td><td>88.4</td><td>68.2</td><td>44.6</td><td>54.9</td><td>71.0</td><td>88.5</td><td>59.4</td><td>89.0</td><td>64.7</td><td>65.2</td></tr><tr><td>Fine-tune</td><td>62.0</td><td>95.1</td><td>89.6</td><td>79.5</td><td>98.9</td><td>97.5</td><td>92.7</td><td>99.6</td><td>94.7</td><td>89.6</td><td>81.8</td></tr><tr><td colspan=\"12\">Transfer</td></tr><tr><td>Continual-FT</td><td></td><td>67.1</td><td>46.0</td><td>32.1</td><td>35.6</td><td>35.0</td><td>57.7</td><td>44.1</td><td>60.8</td><td>20.5</td><td>46.6</td></tr><tr><td>LwF [33]</td><td></td><td>74.5</td><td>56.9</td><td>39.1</td><td>51.1</td><td>52.6</td><td>72.8</td><td>60.6</td><td>75.1</td><td>30.3</td><td>55.9</td></tr><tr><td>iCaRL [47]</td><td></td><td>56.6</td><td>44.6</td><td>32.7</td><td>39.3</td><td>46.6</td><td>68.0</td><td>46.0</td><td>77.4</td><td>31.9</td><td>60.5</td></tr><tr><td>LwF-VR [11]</td><td></td><td>77.1</td><td>61.0</td><td>40.5</td><td>45.3</td><td>54.4</td><td>74.6</td><td>47.9</td><td>76.7</td><td>36.3</td><td>58.6</td></tr><tr><td>WiSE-FT [58]</td><td></td><td>73.5</td><td>55.6</td><td>35.6</td><td>41.5</td><td>47.0</td><td>68.3</td><td>53.9</td><td>69.3</td><td>26.8</td><td>51.9</td></tr><tr><td>Dist. only</td><td></td><td>80.1</td><td>62.2</td><td>40.2</td><td>39.9</td><td>58.1</td><td>80.8</td><td>53.4</td><td>74.6</td><td>38.1</td><td>61.9</td></tr><tr><td>ZSCL*</td><td></td><td>78.3</td><td>64.0</td><td>42.9</td><td>45.2</td><td>63.5</td><td>84.2</td><td>56.1</td><td>78.9</td><td>44.1</td><td>64.3</td></tr><tr><td>ZSCL</td><td></td><td>86.0</td><td>67.4</td><td>45.4</td><td>50.4</td><td>69.1</td><td>87.6</td><td>61.8</td><td>86.8</td><td>60.1</td><td>66.8</td></tr><tr><td colspan=\"12\">Avg.</td></tr><tr><td>Continual-FT</td><td>25.5</td><td>81.5</td><td>59.1</td><td>53.2</td><td>64.7</td><td>51.8</td><td>63.2</td><td>64.3</td><td>69.7</td><td>31.8</td><td>49.7</td></tr><tr><td>LwF [33]</td><td>36.3</td><td>86.9</td><td>72.0</td><td>59.0</td><td>73.7</td><td>60.0</td><td>73.6</td><td>74.8</td><td>80.0</td><td>37.3</td><td>58.1</td></tr><tr><td>iCaRL [47]</td><td>35.5</td><td>89.2</td><td>72.2</td><td>60.6</td><td>68.8</td><td>70.0</td><td>78.2</td><td>62.3</td><td>81.8</td><td>41.2</td><td>62.5</td></tr><tr><td>LwF-VR [11]</td><td>29.6</td><td>87.7</td><td>74.4</td><td>59.5</td><td>72.4</td><td>63.6</td><td>77.0</td><td>66.7</td><td>81.2</td><td>43.7</td><td>60.7</td></tr><tr><td>WiSE-FT [58]</td><td>26.7</td><td>86.5</td><td>64.3</td><td>57.1</td><td>65.7</td><td>58.7</td><td>71.1</td><td>70.5</td><td>75.8</td><td>36.9</td><td>54.6</td></tr><tr><td>Dist. only</td><td>48.1</td><td>90.6</td><td>79.8</td><td>63.2</td><td>75.6</td><td>72.5</td><td>84.7</td><td>70.2</td><td>79.8</td><td>46.9</td><td>63.7</td></tr><tr><td>ZSCL*</td><td>50.7</td><td>90.9</td><td>79.8</td><td>63.8</td><td>76.6</td><td>77.3</td><td>87.0</td><td>71.9</td><td>83.0</td><td>52.0</td><td>65.9</td></tr><tr><td>ZSCL</td><td>45.1</td><td>92.0</td><td>80.1</td><td>64.3</td><td>79.5</td><td>81.6</td><td>89.6</td><td>75.2</td><td>88.9</td><td>64.7</td><td>68.0</td></tr><tr><td colspan=\"12\">Last</td></tr><tr><td>Continual-FT</td><td>31.0</td><td>89.3</td><td>65.8</td><td>67.3</td><td>88.9</td><td>71.1</td><td>85.6</td><td>99.6</td><td>92.9</td><td>77.3</td><td>81.1</td></tr><tr><td>LwF [33]</td><td>26.3</td><td>87.5</td><td>71.9</td><td>66.6</td><td>79.9</td><td>66.9</td><td>83.8</td><td>99.6</td><td>92.1</td><td>66.1</td><td>80.4</td></tr><tr><td>iCaRL [47]</td><td>35.8</td><td>93.0</td><td>77.0</td><td>70.2</td><td>83.3</td><td>88.5</td><td>90.4</td><td>86.7</td><td>93.2</td><td>81.2</td><td>81.9</td></tr><tr><td>LwF-VR [11]</td><td>20.5</td><td>89.8</td><td>72.3</td><td>67.6</td><td>85.5</td><td>73.8</td><td>85.7</td><td>99.6</td><td>93.1</td><td>73.3</td><td>80.9</td></tr><tr><td>WiSE-FT [58]</td><td>27.2</td><td>90.8</td><td>68.0</td><td>68.9</td><td>86.9</td><td>74.0</td><td>87.6</td><td>99.6</td><td>92.6</td><td>77.8</td><td>81.3</td></tr><tr><td>Dist. only</td><td>43.3</td><td>91.9</td><td>81.3</td><td>72.4</td><td>95.1</td><td>90.5</td><td>90.4</td><td>99.7</td><td>92.5</td><td>85.1</td><td>81.8</td></tr><tr><td>ZSCL*</td><td>46.0</td><td>92.3</td><td>81.2</td><td>72.4</td><td>93.0</td><td>92.1</td><td>90.8</td><td>99.6</td><td>93.3</td><td>86.6</td><td>81.7</td></tr><tr><td>ZSCL</td><td>40.6</td><td>92.2</td><td>81.3</td><td>70.5</td><td>94.8</td><td>90.5</td><td>91.9</td><td>98.7</td><td>93.9</td><td>85.3</td><td>80.2</td></tr></table>",
+        "bbox": [
+            107,
+            114,
+            864,
+            561
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "ensemble achieves a better tradeoff between Transfer and Last value. While  $\\mathrm{ZSCL}^*$ , a variant without the WC loss, obtains the highest Last score, the ZSCL with WC loss outperforms it with  $2.8\\%$  Avg. and  $5.9\\%$  Transfer scores.",
+        "bbox": [
+            75,
+            585,
+            470,
+            648
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2. Multi-domain Task Incremental Learning",
+        "text_level": 1,
+        "bbox": [
+            76,
+            662,
+            434,
+            680
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Tab. 3 displays the performance of different methods on the MTIL benchmark, and Tab. 5 presents the detailed Transfer, Avg, and Last metrics on each dataset. Zero-shot denotes the zero-shot prediction performance of the initial CLIP model, and Fine-tune means the direct fine-tuning accuracy on each dataset, which can be seen as an upperbound where no forgetting phenomenon happens. Continual learning uses cross-entropy loss to learn each dataset sequentially, where there is a significant forgetting issue on both zero-shot predictions (Transfer drops by  $24.8\\%$ ) and newly learned knowledge (Last drops by  $9.4\\%$ ). Previous methods improve the Last performance slightly and cannot maintain a high zero-shot prediction performance. Without WC,  $\\mathrm{ZSCL}^*$  achieves the best Last scores, outperforming",
+        "bbox": [
+            75,
+            688,
+            470,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "the previous best one by  $4.4\\%$ . Our method ZSCL improves  $9.1\\%$  on Transfer accuracy, with only  $1.3\\%$  drops compared with the initial CLIP model, and achieves a  $10.1\\%$  gain in the Avg. accuracy.",
+        "bbox": [
+            498,
+            585,
+            890,
+            648
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure Fig. 5 provides a visualization of the feature space on Aircraft of the original CLIP and four methods trained at the end of MTCL (t-SNE conducted only once on all features collected). This shows our method is capable of maintaining the pretrained feature distribution with a small averaged feature distance and thus preserving zero-shot performance.",
+        "bbox": [
+            496,
+            652,
+            893,
+            758
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The result of the MTIL method in Order-II is presented in Tab. 4. Our method surpasses previous methods in another order setting of the MTIL benchmark. A similar conclusion holds from the results of MTIL Order-II compared with MTIL Order-I. Our method ZSCL outperforms others by  $9.2\\%$  on the Avg. score and  $18.1\\%$  on the Last score with only a  $1.2\\%$  performance loss on the Transfer score. This shows our approach can work for different orders of the multi-domain task incremental learning. In addition,",
+        "bbox": [
+            496,
+            763,
+            893,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "19132",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/d650b00a134f12b88b67cba437bcb07263b2d4a1ef65ad84a9d666ca2cd3cbb3.jpg",
+        "table_caption": [
+            "Table 6. Comparison of state-of-the-art CL methods on CIFAR100 benchmark in class-incremental setting."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">10 steps</td><td colspan=\"2\">20 steps</td><td colspan=\"2\">50 steps</td></tr><tr><td>Avg</td><td>Last</td><td>Avg</td><td>Last</td><td>Avg</td><td>Last</td></tr><tr><td>UCIR [19]</td><td>58.66</td><td>43.39</td><td>58.17</td><td>40.63</td><td>56.86</td><td>37.09</td></tr><tr><td>BiC [59]</td><td>68.80</td><td>53.54</td><td>66.48</td><td>47.02</td><td>62.09</td><td>41.04</td></tr><tr><td>RPSNet [46]</td><td>68.60</td><td>57.05</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>PODNet [13]</td><td>58.03</td><td>41.05</td><td>53.97</td><td>35.02</td><td>51.19</td><td>32.99</td></tr><tr><td>DER [60]</td><td>74.64</td><td>64.35</td><td>73.98</td><td>62.55</td><td>72.05</td><td>59.76</td></tr><tr><td>DyTox+ [14]</td><td>74.10</td><td>62.34</td><td>71.62</td><td>57.43</td><td>68.90</td><td>51.09</td></tr><tr><td>CLIP [44]</td><td>74.47</td><td>65.92</td><td>75.20</td><td>65.74</td><td>75.67</td><td>65.94</td></tr><tr><td>FT</td><td>65.46</td><td>53.23</td><td>59.69</td><td>43.13</td><td>39.23</td><td>18.89</td></tr><tr><td>LwF [33]</td><td>65.86</td><td>48.04</td><td>60.64</td><td>40.56</td><td>47.69</td><td>32.90</td></tr><tr><td>iCaRL [47]</td><td>79.35</td><td>70.97</td><td>73.32</td><td>64.55</td><td>71.28</td><td>59.07</td></tr><tr><td>LwF-VR [11]</td><td>78.81</td><td>70.75</td><td>74.54</td><td>63.54</td><td>71.02</td><td>59.45</td></tr><tr><td>ZSCL (Ours)</td><td>82.15</td><td>73.65</td><td>80.39</td><td>69.58</td><td>79.92</td><td>67.36</td></tr><tr><td>Impr</td><td>+7.68</td><td>+7.73</td><td>+5.19</td><td>+3.84</td><td>+3.95</td><td>+1.42</td></tr></table>",
+        "bbox": [
+            76,
+            128,
+            470,
+            343
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "compared with Order-I, previous methods achieve a much lower Last score (e.g., for Continual-Learning, Order-I has  $77.3\\%$ , while Order-II has  $65.3\\%$ ). With ZSCL, the Last score is similar ( $83.6\\%$  compared with  $83.4\\%$ ). This shows our method is more robust towards different training orders.",
+        "bbox": [
+            75,
+            378,
+            468,
+            455
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.3. Class Incremental Learning",
+        "text_level": 1,
+        "bbox": [
+            76,
+            462,
+            328,
+            479
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We evaluate our methods on conventional class incremental learning. Tabs. 6 and 7 display results on CIFAR100 and TinyImageNet, respectively. We re-implement some previous methods with a CLIP backbone (after CLIP in the table), while others using a special network design cannot be easily adapted. Although zero-shot CLIP prediction achieves a good result on these benchmarks, continual learning with direct fine-tuning or LwF [33] degrades the performance greatly, especially under the setting of a large step number. This demonstrates a severe catastrophic forgetting phenomenon in fine-tuning the CLIP model. Our method consistently improves the performance of the CLIP model on both Avg. and Last scores with a large gap towards previous ones.",
+        "bbox": [
+            75,
+            484,
+            470,
+            696
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6. Limitation and Future Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            709,
+            341,
+            724
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our proposed method has a limitation that a reference dataset is needed. A promising direction of the work is to preserve the zero-shot transfer ability without the need for an outside dataset. For example, we may generate a synthetic image dataset as the reference dataset. Methods like [51] can synthesize datasets from a network.",
+        "bbox": [
+            75,
+            734,
+            468,
+            824
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The deep learning community tends to build large models with a huge dataset [3, 6], including vision-language models [44, 20]. As re-training cost upsurges, continual learning is an efficient approach for updating these models with custom usage.",
+        "bbox": [
+            75,
+            825,
+            468,
+            902
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/6479759bcc82c523d84cd2ef77a9de148b862d90a7cef21e31d139d08fbb221e.jpg",
+        "table_caption": [
+            "Table 7. Comparison of different methods on TinyImageNet splits in class-incremental settings with 100 base classes."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">5 steps</td><td colspan=\"2\">10 steps</td><td colspan=\"2\">20 steps</td></tr><tr><td>Avg</td><td>Last</td><td>Avg</td><td>Last</td><td>Avg</td><td>Last</td></tr><tr><td>EWC [25]</td><td>19.01</td><td>6.00</td><td>15.82</td><td>3.79</td><td>12.35</td><td>4.73</td></tr><tr><td>EEIL [4]</td><td>47.17</td><td>35.12</td><td>45.03</td><td>34.64</td><td>40.41</td><td>29.72</td></tr><tr><td>UCIR [19]</td><td>50.30</td><td>39.42</td><td>48.58</td><td>37.29</td><td>42.84</td><td>30.85</td></tr><tr><td>MUC [35]</td><td>32.23</td><td>19.20</td><td>26.67</td><td>15.33</td><td>21.89</td><td>10.32</td></tr><tr><td>PASS [65]</td><td>49.54</td><td>41.64</td><td>47.19</td><td>39.27</td><td>42.01</td><td>32.93</td></tr><tr><td>DyTox [14]</td><td>55.58</td><td>47.23</td><td>52.26</td><td>42.79</td><td>46.18</td><td>36.21</td></tr><tr><td>CLIP [44]</td><td>69.62</td><td>65.30</td><td>69.55</td><td>65.59</td><td>69.49</td><td>65.30</td></tr><tr><td>FT</td><td>61.54</td><td>46.66</td><td>57.05</td><td>41.54</td><td>54.62</td><td>44.55</td></tr><tr><td>LwF [33]</td><td>60.97</td><td>48.77</td><td>57.60</td><td>44.00</td><td>54.79</td><td>42.26</td></tr><tr><td>iCaRL [47]</td><td>77.02</td><td>70.39</td><td>73.48</td><td>65.97</td><td>69.65</td><td>64.68</td></tr><tr><td>LwF-VR [11]</td><td>77.56</td><td>70.89</td><td>74.12</td><td>67.05</td><td>69.94</td><td>63.89</td></tr><tr><td>ZSCL (Ours)</td><td>80.27</td><td>73.57</td><td>78.61</td><td>71.62</td><td>77.18</td><td>68.30</td></tr><tr><td>Impr</td><td>+10.65</td><td>+8.27</td><td>+9.06</td><td>+6.03</td><td>+7.69</td><td>+3.00</td></tr></table>",
+        "bbox": [
+            500,
+            128,
+            893,
+            343
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In some cases, we want to correct wrong information in the pre-training dataset or update old information with the latest one. How to conduct this task with a reference dataset is left as future work.",
+        "bbox": [
+            498,
+            383,
+            890,
+            443
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Lately, a noticeable trend involves the creation of multimodality models utilizing large language models, showcasing encouraging outcomes in tasks such as Visual Question Answering (VQA). Expanding our approach to encompass the next token prediction task remains an avenue for future exploration and research.",
+        "bbox": [
+            498,
+            445,
+            890,
+            536
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            555,
+            619,
+            571
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we investigate continual learning with the vision-language model. We propose a better continual learning algorithm to protect the zero-shot transfer ability in the vision-language model learned in the pre-training stage. Our algorithm mitigates the catastrophic forgetting in both feature space and parameter space. In feature space, distilling the initial model on a reference dataset significantly boosts the model's performance. In parameter space, weight ensemble among different training stages alleviates the forgetting issue. We propose a more challenging Multidomain Task Incremental Learning (MTIL) benchmark to evaluate the continual learning methods better. On both conventional and new benchmarks, our method achieves state-of-the-art performance.",
+        "bbox": [
+            496,
+            582,
+            890,
+            792
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            500,
+            813,
+            666,
+            830
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Yang You's research group is being sponsored by the NUS startup grant (Presidential Young Professorship), Singapore MOE Tier-1 grant, ByteDance grant, ARCTIC grant, SMI grant, and Alibaba grant.",
+        "bbox": [
+            498,
+            839,
+            890,
+            902
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "19133",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            174,
+            106
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV), pages 139-154, 2018. 3, 5",
+            "[2] Eden Belouadah, Adrian Popescu, and Ioannis Kanellos. A comprehensive study of class incremental learning algorithms for visual tasks. Neural Networks, 135:38-54, 2021. 3",
+            "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2, 9",
+            "[4] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 9",
+            "[5] Zhou Daquan, Kai Wang, Jianyang Gu, Xiangyu Peng, Dongze Lian, Yifan Zhang, Yang You, and Jiashi Feng. Dataset quantization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2",
+            "[6] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, et al. Scaling vision transformers to 22 billion parameters. arXiv preprint arXiv:2302.05442, 2023. 9",
+            "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6",
+            "[8] Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29(6):141-142, 2012. 2",
+            "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3",
+            "[10] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5138-5146, 2019. 2, 3",
+            "[11] Yuxuan Ding, Lingqiao Liu, Chunna Tian, Jingyuan Yang, and Haoxuan Ding. Don't stop learning: Towards continual learning for the clip model. arXiv preprint arXiv:2207.09248, 2022. 2, 3, 7, 8, 9",
+            "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 6",
+            "[13] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In European Con-"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            901
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "ference on Computer Vision, pages 86-102. Springer, 2020. 2,3,9",
+            "[14] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285-9295, 2022. 2, 3, 5, 6, 9",
+            "[15] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 3",
+            "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 3",
+            "[17] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 6",
+            "[18] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 2, 4",
+            "[19] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 831-839, 2019. 9",
+            "[20] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, et al. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 9",
+            "[21] David Isele and Akansel Cosgun. Selective experience replay for lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 2, 3",
+            "[22] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018. 5",
+            "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2, 3",
+            "[24] Woojeong Jin, Yu Cheng, Yelong Shen, Weizhu Chen, and Xiang Ren. A good prompt is worth millions of parameters? low-resource prompt-based learning for vision-language models. arXiv preprint arXiv:2110.08484, 2021. 2",
+            "[25] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 3, 5, 9"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            901
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "19134",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[26] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 5",
+            "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1",
+            "[28] Frantzeska Lavda, Jason Ramapuram, Magda Gregorova, and Alexandros Kalousis. Continual classification learning using generative models. arXiv preprint arXiv:1810.10612, 2018. 2, 3",
+            "[29] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 2",
+            "[30] Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming catastrophic forgetting by incremental moment matching. Advances in neural information processing systems, 30, 2017. 3",
+            "[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 3",
+            "[32] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 2022. 2",
+            "[33] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1, 2, 3, 4, 7, 8, 9",
+            "[34] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. ICCV-2023, 2023. 2",
+            "[35] Yu Liu, Sarah Parisot, Gregory Slabaugh, Xu Jia, Ales Leonardis, and Tinne Tuytelaars. More classifiers, less forgetting: A generic multi-classifier paradigm for incremental learning. In European Conference on Computer Vision, pages 699-716. Springer, 2020. 9",
+            "[36] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2, 3",
+            "[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6",
+            "[38] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 6",
+            "[39] Zixuan Ni, Longhui Wei, Siliang Tang, Yueting Zhuang, and Qi Tian. Continual vision-language representation learning with off-diagonal information. arXiv preprint arXiv:2305.07437, 2023. 3",
+            "[40] Oleksiy Ostapenko, Timothee Lesort, Pau Rodríguez, Md Rifat Arefin, Arthur Douillard, Irina Rish, and Laurent Charlin. Continual learning with foundation models: An empirical study of latent replay. In Conference on Lifelong Learning Agents, pages 60–91. PMLR, 2022. 3",
+            "[41] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3498-3505, 2012. 1"
+        ],
+        "bbox": [
+            78,
+            90,
+            467,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[42] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In European conference on computer vision, pages 524-540. Springer, 2020. 2, 3",
+            "[43] Ziheng Qin, Kai Wang, Zangwei Zheng, Jianyang Gu, Xi-angyu Peng, Daquan Zhou, and Yang You. Infobatch: Lossless training speed up by unbiased dynamic data pruning. arXiv preprint arXiv:2303.04947, 2023. 2",
+            "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 6, 9",
+            "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67, 2020. 2",
+            "[46] Jathushan Rajasegaran, Munawar Hayat, Salman Khan, Fahad Shahbaz Khan, Ling Shao, and Ming-Hsuan Yang. An adaptive random path selection approach for incremental learning. arXiv preprint arXiv:1906.01120, 2019. 9",
+            "[47] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1, 2, 3, 4, 7, 8, 9",
+            "[48] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 6",
+            "[49] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 7",
+            "[50] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 2, 3",
+            "[51] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9374-9384, 2021. 9",
+            "[52] Tejas Srinivasan, Ting-Yun Chang, Leticia Pinto Alva, Georgios Chochlakis, Mohammad Rostami, and Jesse Thomason. Climb: A continual learning benchmark for vision-and-language tasks. Advances in Neural Information Processing Systems, 35:29440-29453, 2022. 3",
+            "[53] Vishal Thengane, Salman Khan, Munawar Hayat, and Fahad Khan. Clip model is an efficient continual learner. arXiv preprint arXiv:2210.03114, 2022. 2, 3",
+            "[54] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 7"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "19135",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[55] Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. arXiv preprint arXiv:1904.07734, 2019. 3, 5",
+            "[56] Zhen Wang, Liu Liu, Yiqun Duan, Yajing Kong, and Dacheng Tao. Continual learning with lifelong vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 171-181, 2022. 3",
+            "[57] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 5",
+            "[58] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 2, 3, 5, 7, 8",
+            "[59] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 2, 9",
+            "[60] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3014-3023, 2021. 2, 5, 9",
+            "[61] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017. 3",
+            "[62] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International Conference on Machine Learning, pages 3987-3995. PMLR, 2017. 3, 5",
+            "[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2",
+            "[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3",
+            "[65] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and Cheng-Lin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871-5880, 2021. 2, 5, 9"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            837
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "19136",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 11
+    }
+]

2023/Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models/438697ca-0d24-4663-a1e4-f11b5f684ada_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1f48688e50341e819a3c505322d7926d42f811d9b37c429f9dbbb2f73a99d9b
+size 1351288

2023/Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models/full.md

@@ -0,0 +1,362 @@
+# Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models
+
+Zangwei Zheng $^{1}$  Mingyuan Ma $^{2}$  Kai Wang $^{1}$  Ziheng Qin $^{1}$  Xiangyu Yue $^{3}$  Yang You $^{1}$ $^{1}$ National University of Singapore  $^{2}$ UC Berkeley  $^{3}$ The Chinese University of Hong Kong
+
+$^{1}$  {zangwei, kai.wang, zihengq, youy}  $@$  comp.nus.edu.sg  $^{2}$ mamingyuan2001@berkeley.edu  $^{3}$ xyyue@ie.cuhk.edu.hk
+
+# Abstract
+
+Continual learning (CL) can help pre-trained vision-language models efficiently adapt to new or under-trained data distributions without re-training. Nevertheless, during the continual training of the Contrastive Language-Image Pre-training (CLIP) model, we observe that the model's zero-shot transfer ability significantly degrades due to catastrophic forgetting. Existing CL methods can mitigate forgetting by replaying previous data. However, since the CLIP dataset is private, replay methods cannot access the pre-training dataset. In addition, replaying data of previously learned downstream tasks can enhance their performance but comes at the cost of sacrificing zero-shot performance. To address this challenge, we propose a novel method ZSCL to prevent zero-shot transfer degradation in the continual learning of vision-language models in both feature and parameter space. In the feature space, a reference dataset is introduced for distillation between the current and initial models. The reference dataset should have semantic diversity but no need to be labeled, seen in pre-training, or matched image-text pairs. In parameter space, we prevent a large parameter shift by averaging weights during the training. We propose a more challenging Multi-domain Task Incremental Learning (MTIL) benchmark to evaluate different methods, where tasks are from various domains instead of class-separated in a single dataset. Our method outperforms other methods in the traditional class-incremental learning setting and the MTIL by  $9.7\%$  average score. Our code locates at https://github.com/Thunderbee/ZSCL.
+
+# 1. Introduction
+
+Most deep learning models can access all the data during training [27, 16, 12]. If we want to expand a model's knowledge, such as learning a newly found animal species [41], we can re-train the model by adding new classes to the training dataset. However, re-training a large model is
+
+![](images/c75e1ac3f0a441aaf322335722a181e73f7abc48d7a8e895bf06511d5e2d2e42.jpg)
+
+![](images/92b6c94e2754d253747798a82044fa5a64cabb9b7c6aac94cce33f61f2e5f62c.jpg)  
+(a) Comparison between traditional CL and CL with a pre-trained vision-language model
+
+![](images/c10d291ca81dad6cd99c3f2ff141da5b1ae32d191427ee18e0d17d2af6230b2d.jpg)
+
+![](images/8ba2320b7369d63df75ed8058fd16d5665eb588e431df7dda0453ed1636163e1.jpg)
+
+![](images/ed563a5c3d2b32f87328dfb825d0c8cc14c019a03e0d56a799d5d00ddf5c9f04.jpg)  
+(b) Performance of different methods on preventing forgetting phenomenon
+
+![](images/10428b9b041ff7ba9f51c846f7aa5a3ff0160dd0426e8cfc094ca94774103d1d.jpg)  
+Figure 1. a) Conventional CL learns distinct task-specific heads, while CL with vision-language models can predict both learned tasks and out-of-distribution tasks. b) Accuracy (\%) changes during CL of four datasets on 11 datasets. Our method is superior to others in preventing the forgetting of both zero-shot transfer ability and new knowledge.
+
+costly. In contrast, continual learning (CL) [33, 47] incrementally learns task one after another. It can reduce
+
+this cost by only learning the new data, thereby presenting itself as an efficient alternative to conventional learning methods [43, 34, 5]. Nonetheless, a model tends to forget previous information catastrophically when learning new tasks [47, 33, 59]. The "catastrophic forgetting" phenomenon is a great challenge for CL.
+
+Recently, vision-language models have shown powerful zero-shot transfer ability [44, 23, 32]. They can give zero-shot predictions without any training examples of a task. However, the performance on some tasks is poor due to insufficient relevant image-text pairs in the pre-training datasets. For example, it is difficult for CLIP [44] to distinguish among digital numbers, with an accuracy on MNIST [8] below  $60\%$  much lower than a naively trained CNN [29]. If we want to widen the knowledge in the vision-language model by re-training, the computational cost is too large (e.g., CLIP is pretrained on 400 million image-text pairs). Fine-tuning downstream tasks achieves high performance, but one model for a task takes much memory, and the model is not reusable. Prompt learning [64, 63] keeps the backbone parameters unchanged. However, it is only effective with limited training data due to a limited prompt length [64, 24]. In contrast, continual learning makes learning new knowledge a lifelong process for the vision-language model. The continually learned model can handle any image-text input and can be further used for downstream tasks [11, 53].
+
+We find that existing CL methods hardly prevent the forgetting phenomenon for zero-shot transfer ability in continual learning of a pre-trained vision-language model. As shown in Fig. 1 (a), the CL with a pre-trained vision-language model differs from the traditional one. Besides forgetting previously learned task knowledge, the CLIP-based CL suffers from forgetting pre-training knowledge, namely a degradation of zero-shot transfer ability. For the replay-based CL methods [47, 50, 36, 21, 28, 42], the dataset during pre-training may be private and inaccessible during fine-tuning. For distillation-based CL methods [33, 10, 13, 11], they do not lay enough emphasis on the pre-trained model. On the one hand, a large model state change hinders tasks thereafter from using high-quality feature representations. On the other hand, it significantly degrades zero-shot performance on unseen datasets.
+
+Our method ZSCL protects the Zero-Shot transfer ability during Continual Learning. We view the knowledge stored in the pre-training model from two perspectives: a well-learned feature space and a good value in the parameter space. In feature space, we re-design previous distillation loss [18, 47] with different loss styles, teacher models, and data sources. We find the original CLIP model, as opposed to the newly acquired model, is the best option for the teacher model. Instead of using data collected from previous tasks [47] or current task [18], we find a reference
+
+dataset with diverse semantics (e.g., images sampled from ImageNet) is a good option for distillation loss. The reference images need not be labeled or matched with the text. Preserving the relative similarity between reference images and texts makes the feature space deviate little from the original. In the parameter space, WiSE-FT [58] proposes interpolating the initial and fine-tuned model for better performance. Inspired by this, we ensemble the weights throughout continuous training to prevent a significant shift from the initial CLIP, which can be seen as interpolating models of different zero-shot transfer and downstream task performance tradeoffs. The weight ensemble method is more stable and not sensitive to hyper-parameters.
+
+To better evaluate our method, we propose a new benchmark Multi-domain Task Incremental Learning (MTIL). Previous CL tasks are crafted by separating classes in one dataset [14, 60, 65], where the images and classes are within a single domain. In contrast, MTIL consists of data from different sources requiring different expert knowledge. It comprises 11 tasks ranging from animal species to aircraft series recognition. As displayed in Fig. 1 (b), when sequentially training CLIP on 11 datasets, the drop in the performance of task  $i$  after training task  $i$  is the traditional forgetting phenomenon. The degradation in the accuracy compared to the original zero-shot one before training task  $i$  represents the forgetting in zero-shot transfer ability. Our method better protects the zero-shot transfer ability and preserves the learned knowledge. We outperform previous methods in both conventional class-incremental learning and MTIL settings. In Fig. 1 (b),
+
+To summarize, our contributions are as follows:
+
+- We investigate continual training with the vision-language model and demonstrate the importance of preserving zero-shot transfer ability. A more challenging benchmark MTIL is proposed to evaluate CL methods where the tasks come from distinct domains.  
+- We propose a novel method ZSCL to mitigate the catastrophic forgetting problem in continual learning of the vision-language model by distillation in the feature space and weight ensemble in the parameter space.  
+- The proposed ZSCL outperforms all state-of-the-art methods across multiple benchmark datasets. On 10 steps CL of CIFAR100 and TinyImageNet, our method outperforms the best of previous ones by  $7.7\%$  and  $6.0\%$  for the Last accuracy. On MTCL, ZSCL outperforms others by  $10.9\%$  on Transfer and  $9.7\%$  on Avg. scores.
+
+# 2. Related Work
+
+Vision-Language Models. Inspired by the success of language foundation models such as GPT-3 [3] and T5 [45], a
+
+series of work pre-train vision-language models on large-scale image-text datasets [31, 44, 23]. Among them, Contrastive Language-Vision Pre-training [44] achieves remarkable performance on various downstream tasks. It concentrates on aligning images and texts to acquire a joint embedding space. The CLIP model contains an image encoder [16, 12] and a text encoder [9]. During pre-training, contrastive learning is performed in which a paired image-text is a positive pair while image and text from different image-text pairs form a negative pair. For inference, the closest text embedding for the image is chosen as the prediction. Vision-language models can give zero-shot predictions on unseen tasks with a robust zero-shot transfer ability on various downstream tasks.
+
+Continual Learning Methods. Most existing continual learning methods can be categorized into four groups: parameter expansion, memory replay, distillation loss, and parameter regularization. Parameter expansion methods such as DyTox [14] and DEN [61] introduce new parameters for new tasks. As we want to achieve a more powerful CLIP model at the end of CL, we do not change the architecture of the CLIP model. Memory replay methods [50, 36, 42, 28, 40] including iCaRL [47] and SER [21] keep a memory for exemplars from previously learned tasks. However, pre-training datasets are too large for choosing exemplars or may not be available at downstream training, and downstream data are not good exemplars for preserving the pretraining knowledge. Distillation loss such as LwF [33], LwM [10], LwF-VR [11], and PodNet [13] aligns current output space with previous ones, whereas distillation based on current tasks are not strong enough to maintain foundational knowledge. For the CLIP model, we find that general images, even if never seen by the model, plus sentences with enough semantics, can be a good "replay" for the distillation loss. The parameter regularization loss restricts the flexibility of model parameters by training loss [25, 62, 1] or weight averaging [30, 58]. Although this type of strategy performs badly compared to other ways in previous research [55, 2], we found it helpful for CLIP continual learning. The limited parameter space prevents the model from diverging significantly from its original state.
+
+Vision-Language Models for Downstream Tasks. Many works propose different training strategies of vision-language models for better performance on downstream tasks, such as CoOp [64], CLIP-Adapter [15] and WiSE-FT [58]. However, very few attempts at continual learning exist. While [52, 56] focus on CL in the pretrain of VL and [39] trains VL from scratch using a small dataset, our problem setting differs as we address CL in downstream tasks with a pretrained VL. Currently, most VL models are trained directly on an accessible large
+
+dataset, yet integrating knowledge continuously into a pretrained VL is a practical necessity. None of these studies address the zero-shot transfer degradation phenomenon, instead focusing on the traditional forgetting of learned knowledge. Recently, Thengane et al. [53] shows CLIP zero-shot prediction achieves state-of-the-art performance in CL settings even without any training. LwF-VR [11] is a modified LwF method for the CLIP model where random-generated sentences are used for distillation loss. However, it only considers the feature space, and the distillation with random sentences cannot protect the vision backbone. Differently, we re-examine what should be used for distillation in the feature space and combine the parameter space weight ensemble to provide better performance for the vision-language model continual learning.
+
+# 3. Approach
+
+# 3.1. Preliminaries
+
+Continual Learning. Given  $n$  tasks  $[\mathcal{T}^1, \mathcal{T}^2, \dots, \mathcal{T}^n]$ , continual training is conducted in sequence on each task  $\mathcal{T}^i = (\mathcal{D}^i, C^i)$ ,  $i = 1, \dots, n$ . Here,  $\mathcal{D}^i$  represents the task dataset  $\{(x_j^i, y_j^i)\}_{j=1}^{N_i}$ , where  $x_j^i$  is an image,  $y_j^i$  is a one-hot vector indicating the ground truth, and  $N_i$  is the number of images in the dataset. Class names  $C^i = \{c_j^i\}_{j=1}^{m_i}$  maps the label of an image to an object name, where  $m_i$  is the number of classes for task  $\mathcal{T}^i$ . The objective of continual learning is to achieve good performance on all tasks.
+
+Two continual learning settings are focused on in this study [55]. In task-incremental learning, at inference, the image  $\pmb{x}$  to be predicted is given with its task identity  $t$ , so the model only needs to distinguish between different classes in  $C^t$ . In class-incremental learning, the task identity  $t$  is not given. Thus we need to predict with the combined class set  $C = \bigcup_{i=1}^{n} C^i$ .
+
+CLIP model. The CLIP model contains an image encoder  $f_{i}$  and a text encoder  $f_{t}$ . The inference process of the CLIP model for image classification is as follows. First, for task  $\mathcal{T}^i$ , each class  $c$  is transformed into a sentence by a template like "a photo of  $\{c\}$ ". Then  $f_{t}$  encodes the classes into text embeddings  $\{\pmb{t}_{j}^{i}\}_{j = 1}^{m_{i}}$ . An image encoder encodes input image  $\pmb{x}_k$ . The similarity score between the image embedding and text embeddings are calculated as  $s_{k,j}^{i} = \langle f_{i}(\pmb{x}_{k}),\pmb{t}_{j}^{i}\rangle$  where  $\langle \cdot ,\cdot \rangle$  denotes the cosine similarity. The class with the largest similarity score is the prediction for the image.
+
+To fine-tune the CLIP model for downstream tasks, cross-entropy loss CE is applied to the similarity score with a temperature scaling:
+
+$$
+\mathcal {L} _ {\mathrm {C E}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \mathrm {C E} (\tau \cdot \boldsymbol {s} _ {i}, \boldsymbol {y} _ {i}), \tag {1}
+$$
+
+where  $\tau$  is a parameter learned during the pre-training.
+
+![](images/4329bff0d636ac16c1e23263b5d689cb7dbc7b4b289e45b30094e3f0c1b8cdce.jpg)  
+Figure 2. Illustration of ZSCL in feature space. Fig. 3(a) shows the pipeline of distillation. The original and the current model encode the reference images and texts, respectively. The probability distribution of images and texts with respect to each other is distilled. Fig. 3(b) displays how distillation loss preserves the feature space. Compared with the reference dataset, the features of fine-tuning tasks lie in a small subspace. The distillation loss preserves the structure of the feature space by maintaining relative distances.
+
+![](images/f03a900cd5ed94daddb29f1585000c813ccaa2207a4462c7b9f5c88529204ab8.jpg)
+
+# 3.2. Distillation in Feature Space
+
+Well-learned feature space for aligned images and texts enables vision-language models' strong zero-shot transfer ability. It also facilitates the learning of downstream tasks. Compared with the pre-training dataset, downstream datasets lie in a small scope in the feature space (shown in Fig. 2(b)). Direct fine-tuning of downstream tasks greatly distorts the feature distribution of out-of-distribution data, which leads to a significant drop in zero-shot prediction performance.
+
+While the cross-entropy improves the performance by altering fine-tuned feature subspace, we need a new regularization to preserve the potential out-of-distribution feature space. The relative similarity between one image and different texts is:
+
+$$
+\boldsymbol {p} = \operatorname {S o f t m a x} \left(\boldsymbol {s} _ {1}, \dots , \boldsymbol {s} _ {m}\right). \tag {2}
+$$
+
+We hope the above similarity distribution is stable during fine-tuning for all potential images and texts. Given a teacher model  $\overline{f}$ , distillation loss can be applied to penalize changes from the original distribution:
+
+$$
+\mathcal {L} _ {\text {d i s t . i m g}} = \operatorname {C E} (\boldsymbol {p}, \bar {\boldsymbol {p}}) = - \sum_ {j = 1} ^ {m} \boldsymbol {p} _ {j} \cdot \log \bar {\boldsymbol {p}} _ {j}, \tag {3}
+$$
+
+where  $\overline{\pmb{p}}$  is the distribution given by the teacher model.
+
+Although the distillation form has been widely used in previous methods [18, 33, 47], they are applied to continual learning from scratch. We investigate different components of the distillation loss for enhancing pre-trained vision-language models.
+
+Three components are discussed in this paper in detail: the data source, the teacher model, and the loss design. Sec. 5.1 shows the performance for the different choices. First, for the data source, LwF [33] uses data from the cur
+
+rent task, while iCarl [47] carefully selects data from previous tasks. However, data from downstream tasks span a small subspace and are not good enough to preserve the whole feature space. An ideal choice is the pre-training dataset. However, the CLIP pre-training dataset is private, and the size is too large to load. To solve this challenge, we introduce the reference dataset. A reference dataset is a publicly available image dataset with enough semantics. Enough semantics can be seen as random sampling in the whole feature space. The texts can be related class names, unrelated sentences, or even random tokens. This is because text semantics are easier to sample, and we need not ground truth to calculate the distance between one image to sufficient text samples spread among feature space.
+
+For the teacher model, [33, 47] use the model after learning task i-1 and before learning task i as the teacher model. During the continual training of the vision-language model, the feature space deviates gradually from the initial model. Using fine-tuned models as teacher models enlarges this change. In contrast, we find using the pre-trained model as a teacher model not only preserves the zero-shot transfer ability but also takes advantage of well-learned feature space for better downstream performance.
+
+Finally, previous distillation loss is applied on traditional backbones, where a classification head gives the probability for different labels. For the vision-language model, the probability is calculated based on the relative distance between images and texts. Thus, in addition to Eq. (3), we impose regularization  $\mathcal{L}_{\mathrm{dist\_txt}}$  on the distances from a text to a batch of images. The whole framework is shown in Fig. 2 (a) with the following training loss:
+
+$$
+\mathcal {L} = \mathcal {L} _ {\mathrm {c e}} + \lambda \cdot \left(\mathcal {L} _ {\mathrm {l w f . i m g}} + \mathcal {L} _ {\mathrm {l w f . t x t}}\right). \tag {4}
+$$
+
+![](images/eb3d24a46c63689d28133592e77841b07804ba7d89e7f0c1fa29b32d5c02b588.jpg)  
+Figure 3. Models during training contain different tradeoffs between zero-shot and new task performance. Points for FT are sampled every 100 iterations, and the ones for WiSE-FT means different  $\alpha$  choices. WE ensembles models during training and achieve better performance.
+
+# 3.3. Weight Ensemble in Parameter Space
+
+Machine learning models integrate learned knowledge in their parameters. To mitigate the forgetting problem, a series of works [25, 62, 1] impose regularization losses on the changes of parameters. Weight consolidation (WC) [25] imposes the following loss:
+
+$$
+\mathcal {L} _ {\mathrm {W C}} = \sum_ {i} \left(\theta_ {i} - \bar {\theta} _ {i}\right) ^ {2}. \tag {5}
+$$
+
+where  $\theta$  is the parameters of the current model, and  $\overline{\theta}$  is the reference ones. Although this regularization prevents forgetting, it hinders learning new tasks in a challenging CL setting.
+
+Apart from regularization losses, another method in parameter space is to ensemble different model weights. Model soup [57] averages weights of multiple fine-tuned models to improve the model's robustness but introduces additional training costs. WiSE-FT [58] propose a weighted average between fine-tuned model and the original model to improve the out-of-distribution prediction robustness:
+
+$$
+f (x; (1 - \alpha) \cdot \theta_ {0} + \alpha \cdot \theta_ {1}), \tag {6}
+$$
+
+where  $\theta_0$  is the original model and  $\theta_{1}$  is the fine-tuned one. However, this method is hyper-parameter-sensitive where different  $\alpha$  gives different tradeoffs between zero-shot transfer ability and downstream task performance (blue line in Fig. 3.
+
+Inspired by this, we extend the weighted average to the CL setting. The motivation for the weighted average is to prevent fine-tuning from losing too much knowledge in the original model. As training goes by (green line in Fig. 3), the model performs better on new tasks while losing accuracy on previous ones. The models among training compose a sequence of different learning-forgetting tradeoffs.
+
+![](images/c1c8038706070e6074aff8433598727beaf8d68f06790075ab76e09402bafbd0.jpg)  
+(a)
+
+![](images/7612a0001597f0d99113646e1932e6b9a2dfdfa5b9fc31e1b66a56d4a369c262.jpg)
+
+![](images/816e3786999b886a7661b676ba1153e0a0d8b60c6ea218d1e5871216e42d0f37.jpg)
+
+![](images/952c17a62f4e4c386253f86e5ba5a7e8676eb7209cde979b79afbd3ab84319b1.jpg)
+
+![](images/81e23c68abd350a661469da9d431633c4fcb113cc4ec578215ac9647b344ca15.jpg)  
+(b)  
+Figure 4. Fig.(a): examples of tasks from different domains in MTIL benchmark. Fig.(b): illustration of calculating metrics Transfer, Avg. and Last during continual learning.
+
+Instead of interpolating only between the initial and the final model, our method weight ensemble (WE) averages the weights in the sequence during the training time:
+
+$$
+\hat {\theta} _ {t} = \left\{ \begin{array}{l l} \theta_ {0} & t = 0 \\ \frac {1}{t + 1} \theta_ {t} + \frac {t}{t + 1} \cdot \hat {\theta} _ {t - 1} & \text {e v e r y I i t e r a t i o n s} \end{array} . \right. \tag {7}
+$$
+
+where model weight sampling happens every  $I$  iteration. The method is related to Stochastic Weight Averaging (SWA) [22], but we do not use a modified learning rate schedule here because instead of getting better generalization ability, WE aim to give an improved learning-forgetting tradeoff. As shown in Fig. 3, WE achieves better performance on downstream tasks than WiSE-FT. In addition, while WiSE-FT is sensitive to different values of  $\alpha$ , our method is much more robust under different hyperparameter  $(I)$  choices.
+
+# 4. Multi-domain Task Incremental Learning
+
+Conventional Continual Learning Benchmark. A benchmark consisting of several tasks is needed to evaluate different methods for continual learning. Most previous benchmarks are built by separating classes in a single dataset, such as MNIST [55], CIFAR100 [14], TinyImageNet [60], and ImageNet [60, 65]. We also evaluate our method with traditional benchmarks. In CIFAR100 [26], classes are separated into groups to build tasks. Suppose the dataset has  $m$  classes, a  $k$ -step setting means we learn  $m / k$  classes in each new task. The CIFAR100 contains 100 classes, and the setting of 10, 20, and 50 steps are visited. For TinyImageNet with  $m = 200$ , the first step learns 100 classes, and the rest is learned with 5, 10, and 20 steps. As for ImageNet-100, we have two settings: ImageNet-100-B0, which includes the same amount of classes for each step, and ImageNet-100-B50, which has
+
+Table 1. Ablation experiments. Default settings are marked in gray, which uses image and text distillation loss with the initial CLIP model on 100k ImageNet images and texts generated from ImageNet classes with a simple template.  
+(a) Continual learning loss.  
+
+<table><tr><td>loss</td><td>Transfer</td><td>Avg.</td><td>Last</td></tr><tr><td>CE only</td><td>44.6</td><td>55.9</td><td>77.3</td></tr><tr><td>Feat. Dist.</td><td>47.6</td><td>58.7</td><td>77.1</td></tr><tr><td>Image-only</td><td>56.5</td><td>68.9</td><td>82.1</td></tr><tr><td>Text-only</td><td>56.7</td><td>69.0</td><td>82.6</td></tr><tr><td>Both</td><td>56.8</td><td>69.2</td><td>83.0</td></tr></table>
+
+(d) Teacher model.  
+
+<table><tr><td>source</td><td>Transfer</td><td>Avg.</td><td>Last</td></tr><tr><td>Initial</td><td>56.8</td><td>69.2</td><td>83.0</td></tr><tr><td>n-1</td><td>53.9</td><td>66.6</td><td>80.7</td></tr><tr><td>WiSE(0.5)</td><td>56.4</td><td>68.9</td><td>82.9</td></tr><tr><td>WiSE(0.8)</td><td>56.2</td><td>67.8</td><td>81.3</td></tr></table>
+
+(b) Data sources for replay.  
+
+<table><tr><td>source</td><td>Transfer</td><td>Avg.</td><td>Last</td></tr><tr><td>current</td><td>56.7</td><td>66.5</td><td>80.2</td></tr><tr><td>ImageNet</td><td>56.8</td><td>69.2</td><td>83.0</td></tr><tr><td>CC</td><td>57.2</td><td>68.5</td><td>80.9</td></tr><tr><td>CIFAR100</td><td>55.2</td><td>65.9</td><td>80.7</td></tr><tr><td>Flowers</td><td>54.7</td><td>66.0</td><td>80.8</td></tr></table>
+
+(e) # images for replay.  
+
+<table><tr><td>#image</td><td>Transfer</td><td>Avg.</td><td>Last</td></tr><tr><td>1M</td><td>58.7</td><td>70.1</td><td>83.2</td></tr><tr><td>100k</td><td>56.8</td><td>69.2</td><td>83.0</td></tr><tr><td>10k</td><td>57.8</td><td>68.7</td><td>81.2</td></tr><tr><td>1k</td><td>56.3</td><td>67.6</td><td>80.8</td></tr></table>
+
+(c) Text sources for replay.  
+
+<table><tr><td>source</td><td>Transfer</td><td>Avg.</td><td>Last</td></tr><tr><td>current</td><td>51.8</td><td>64.9</td><td>82.0</td></tr><tr><td>prev. all</td><td>54.0</td><td>70.2</td><td>83.7</td></tr><tr><td>1k classes (IN)</td><td>56.8</td><td>69.2</td><td>83.0</td></tr><tr><td>13k Sent. (CC)</td><td>58.9</td><td>70.5</td><td>84.0</td></tr><tr><td>1k Rand. Sent.</td><td>58.7</td><td>70.2</td><td>83.8</td></tr></table>
+
+(f) # image classes for replay.  
+
+<table><tr><td>#class</td><td>Transfer</td><td>Avg.</td><td>Last</td></tr><tr><td>1000</td><td>56.8</td><td>67.6</td><td>83.0</td></tr><tr><td>100</td><td>56.7</td><td>67.3</td><td>82.3</td></tr><tr><td>10</td><td>53.8</td><td>66.4</td><td>81.0</td></tr><tr><td>1</td><td>53.1</td><td>65.5</td><td>80.5</td></tr></table>
+
+Table 2. Ablation study of different components for ZSCL.  
+
+<table><tr><td>Method</td><td>Transfer</td><td>Δ</td><td>Avg.</td><td>Δ</td><td>Last</td><td>Δ</td></tr><tr><td colspan="7">CLIP ViT-B/16@224px</td></tr><tr><td>Zero-shot</td><td>69.4</td><td>0.0</td><td>65.3</td><td>0.0</td><td>65.3</td><td>0.0</td></tr><tr><td>Continual Learning</td><td>44.6</td><td>-24.8</td><td>55.9</td><td>-9.4</td><td>77.3</td><td>+12.0</td></tr><tr><td>+ Distillation</td><td>58.9</td><td>-10.5</td><td>70.5</td><td>+5.2</td><td>83.8</td><td>+18.5</td></tr><tr><td>+ WiSE-FT (best α)</td><td>61.7</td><td>-7.7</td><td>71.6</td><td>+6.3</td><td>83.3</td><td>+18.0</td></tr><tr><td>+ WE (ZSCL*)</td><td>62.2</td><td>-7.2</td><td>72.6</td><td>+7.3</td><td>84.5</td><td>+19.2</td></tr><tr><td>+ WC</td><td>67.6</td><td>-1.8</td><td>74.5</td><td>+9.2</td><td>83.2</td><td>+17.9</td></tr><tr><td>+ WiSE-FT</td><td>67.7</td><td>-1.7</td><td>74.2</td><td>+8.9</td><td>81.9</td><td>+16.6</td></tr><tr><td>+ WE (ZSCL)</td><td>68.1</td><td>-1.3</td><td>75.4</td><td>+10.1</td><td>83.6</td><td>+18.3</td></tr></table>
+
+50 classes for the first step, and the remaining 50 classes are observed progressively over the next 10 stages. For ImageNet [7], we investigate a 10-step setting, which learns 100 new classes per task.
+
+MTIL Benchmark. Different classes from one dataset share the common image source and a similar style [48, 17]. Thus, we propose Multi-domain Task Incremental Learning (MTIL), a cross-domain version of task incremental learning. Different tasks are collected from different domains, requiring different domain knowledge for humans to achieve high accuracy. Our MTIL benchmark consists of 11 tasks (detailed in supplementary materials), as some of the tasks illustrated in Fig. 4 (a). The MTIL benchmark is very challenging with a total number of 1,201 classes. Two orders are used for the evaluation; the first one is alphabet order (Order-I): Aircraft, Caltech101, CIFAR100, DTD, EuroSAT, Flowers, Food, MNIST, OxfordPet, StanfordCars, SUN397. And the second one is a random order (Order-II): StanfordCars, Food, MNIST, OxfordPet, Flowers, SUN397, Aircraft, Caltech101, DTD, EuroSAT, CIFAR100. Experiments are done in Order I by default.
+
+Evaluation Metrics. The metrics of MTIL are illustrated in Fig. 4(b), where rows represent training steps and a
+
+column shows performance for one dataset. For conventional continual learning, only scores under the diagonal are meaningful, since they cannot give zero-shot predictions on unseen tasks. In contrast, the zeros-hot transfer ability enables a vision-language model to provide predictions for all datasets. The average accuracy on all datasets among all timestamps is the "Avg" metric. The "Last" metric is the average performance of all tasks after CL. The "Transfer" metric is the average task performance in the upper-right triangle of the matrix. Every task's performance is first averaged to equal the weight of each dataset. It measures to what extent the zero-shot transfer ability is preserved. Before learning task  $i$ , tasks not earlier than  $i$  are not finetuned. Thus, their performance is an indicator of zero-shot transfer ability.
+
+# 5. Experiments
+
+Implementation. We use CLIP [44] model with image encoder ViT-B/16 [12]. We conduct training with AdamW [37] optimizer and use a label smoothing [38] technique for a better baseline result. For multi-domain task continual learning, we train 1K iterations for each task, while for class-incremental learning, we follow the same evaluation setting in [14]. More implementation details can be found in the supplementary material.
+
+# 5.1. Main Properties
+
+We ablate our method in feature-space in Tab. 1, and different choices for parameter-space regularization in Tab. 2. Several interesting characteristics are noted.
+
+Continual learning loss. In Tab. 1a, several types of loss for feature space are tested. The Feature Distance penalizing absolute distances achieves a low accuracy. Distillation loss on probability distribution regularizes relative distance in the feature space. With image-only or text-only distilla
+
+![](images/486e0281799167fdc39ad25fe08f6ed1648465a543e6415e44a1f33130cf66aa.jpg)  
+Figure 5. t-SNE on five models' outputs together of Aircraft datasets after MTCL training: only our model maintains a similar feature distribution to the original CLIP ones with minor shift, while the rest significantly distort the feature space.
+
+![](images/7784c328f4cdf2b9bb2ed7f173c38e0d41ae1044f2237fb671b5c4c8254a9dc8.jpg)
+
+![](images/09b430fabc5fe55f1527a5618eccc3aa9a80fbc30f1e07b877ca20d9e3d1d5a5.jpg)
+
+![](images/eb90c5de160688afb191eadaa39b87a75136e54a5cc737262f85f987cb769e3e.jpg)
+
+![](images/e8277e097f9a1ced82aa1b0687e58001beb7132aad61f8a706e47c69a5d84737.jpg)
+
+Table 3. Comparison of different methods on MTIL in Order I.  
+
+<table><tr><td>Method</td><td>Transfer</td><td>Δ</td><td>Avg.</td><td>Δ</td><td>Last</td><td>Δ</td></tr><tr><td colspan="7">CLIP ViT-B/16@224px</td></tr><tr><td>Zero-shot</td><td>69.4</td><td>0.0</td><td>65.3</td><td>0.0</td><td>65.3</td><td>0.0</td></tr><tr><td>Continual Learning</td><td>44.6</td><td>-24.8</td><td>55.9</td><td>-9.4</td><td>77.3</td><td>+12.0</td></tr><tr><td>LwF [33]</td><td>56.9</td><td>-12.5</td><td>64.7</td><td>-0.6</td><td>74.6</td><td>+9.0</td></tr><tr><td>iCaRL [47]</td><td>50.4</td><td>-19.0</td><td>65.7</td><td>+0.4</td><td>80.1</td><td>+14.8</td></tr><tr><td>LwF-VR [11]</td><td>57.2</td><td>-12.2</td><td>65.1</td><td>-0.2</td><td>76.6</td><td>+11.3</td></tr><tr><td>WiSE-FT [58]</td><td>52.3</td><td>-17.1</td><td>60.7</td><td>-4.6</td><td>77.7</td><td>+12.4</td></tr><tr><td>ZSCL* (Ours)</td><td>62.2</td><td>-7.2</td><td>72.6</td><td>+7.3</td><td>84.5</td><td>+19.2</td></tr><tr><td>ZSCL (Ours)</td><td>68.1</td><td>-1.3</td><td>75.4</td><td>+10.1</td><td>83.6</td><td>+18.3</td></tr></table>
+
+tion, the Transfer, Avg., and Last accuracy all improve. A further boost in performance occurs with both of the distillation losses.
+
+Data source for replay. Tabs. 1b, 1c, 1e and 1f seek the standard for a good reference dataset. As shown in Tab. 1b, distillation on images of current tasks achieves a good Transfer score. However, it hinders the learning on new tasks and results in a low Avg. and Last score. Images in a specific domain (e.g., Flowers) are also not good choices. General images in ImageNet and Conceptual Caption (CC) datasets are examples of good reference datasets. These images are easily available by a web crawler [54]. Text with more semantics can improve performance (shown in Tab. 1c). When using sentences from the Conceptual Caption dataset [49], or even sentences randomly generated from the CLIP vocabulary, there is no ground truth target among the texts for the image from the ImageNet dataset. However, they all achieve an improvement due to more semantics. The reference image dataset does not need to be labeled, matched with the text, or seen by the CLIP model.
+
+Enough semantics in the reference image dataset boosts the distillation performance. In Tabs. 1e and 1f, a smaller number of images and classes all lead to a degradation in the performance. Fewer classes of images in the reference dataset have a worse impact on the performance compared with the image numbers. To keep a reasonable memory buffer, we randomly sample  $100\mathrm{k}$  images from ImageNet for MTIL and use texts from CC dataset. For class-
+
+Table 4. Comparison of different methods on MTIL in Order II.  
+
+<table><tr><td>Method</td><td>Transfer</td><td>Δ</td><td>Avg.</td><td>Δ</td><td>Last</td><td>Δ</td></tr><tr><td colspan="7">CLIP ViT-B/16@224px</td></tr><tr><td>Zero-shot</td><td>65.4</td><td>0.0</td><td>65.3</td><td>0.0</td><td>65.3</td><td>0.0</td></tr><tr><td>Continual Learning</td><td>46.6</td><td>-18.8</td><td>56.2</td><td>-9.1</td><td>67.4</td><td>+2.1</td></tr><tr><td>LwF [33]</td><td>53.2</td><td>-12.2</td><td>62.2</td><td>-5.2</td><td>71.9</td><td>+6.6</td></tr><tr><td>iCaRL [47]</td><td>50.9</td><td>-14.5</td><td>56.9</td><td>-8.4</td><td>71.6</td><td>+6.3</td></tr><tr><td>LwF-VR [11]</td><td>53.1</td><td>-12.3</td><td>60.6</td><td>-7.4</td><td>68.3</td><td>+0.9</td></tr><tr><td>WiSE-FT [58]</td><td>51.0</td><td>-14.4</td><td>61.5</td><td>-5.9</td><td>72.2</td><td>+6.9</td></tr><tr><td>ZSCL*</td><td>59.8</td><td>-5.6</td><td>71.8</td><td>+6.5</td><td>83.3</td><td>+18.0</td></tr><tr><td>ZSCL</td><td>64.2</td><td>-1.2</td><td>74.5</td><td>+9.2</td><td>83.4</td><td>+18.1</td></tr></table>
+
+incremental learning, conceptual caption dataset's validation set (28k images) is used to avoid information leakage.
+
+Teacher model. Tab. 1d shows the performance of distillation loss with different teacher models. Unlike conventional continual learning, the teacher model should not be the one trained on the previous task; otherwise, the deviation from the initial CLIP in the previous task may be amplified. In contrast, with the initial CLIP as the teacher model, not only is the zero-shot performance improved but the Mean and Last scores are also boosted, indicating that preserving high-quality feature space facilitates continual learning.
+
+Parameter-space regularization. In Tab. 2, we experiment with three different parameter-space regularizations. We experiment with two variants of WiSE-FT [58]: the weighted average between the current model with the initial CLIP or the one at the previous task. The result shows the latter one is a better choice because keeping weight averaging with initial CLIP loses the newly learned knowledge. We experiment with different  $\alpha$  choices for WiSE, and the best result is reported. While distillation loss improves the whole performance, the parameter-space regularization further protects the zero-shot transfer ability with a higher Transfer. Among the three parameter-space regularizations, only WE achieve a better Last score. WC greatly improves the Transfer scores with a lower Last score. A combination of the weight consolidation loss with weight
+
+Table 5. Transfer, Avg., and Last scores (%) of different continue training methods on MTIL benchmark.  
+
+<table><tr><td>Method</td><td>Aircraft</td><td>Caltech101</td><td>CIFAR100</td><td>DTD</td><td>EuroSAT</td><td>Flowers</td><td>Food</td><td>MNIST</td><td>OxfordPet</td><td>Cars</td><td>SUN397</td></tr><tr><td colspan="12">CLIP ViT-B/16@224px</td></tr><tr><td>Zero-shot</td><td>24.3</td><td>88.4</td><td>68.2</td><td>44.6</td><td>54.9</td><td>71.0</td><td>88.5</td><td>59.4</td><td>89.0</td><td>64.7</td><td>65.2</td></tr><tr><td>Fine-tune</td><td>62.0</td><td>95.1</td><td>89.6</td><td>79.5</td><td>98.9</td><td>97.5</td><td>92.7</td><td>99.6</td><td>94.7</td><td>89.6</td><td>81.8</td></tr><tr><td colspan="12">Transfer</td></tr><tr><td>Continual-FT</td><td></td><td>67.1</td><td>46.0</td><td>32.1</td><td>35.6</td><td>35.0</td><td>57.7</td><td>44.1</td><td>60.8</td><td>20.5</td><td>46.6</td></tr><tr><td>LwF [33]</td><td></td><td>74.5</td><td>56.9</td><td>39.1</td><td>51.1</td><td>52.6</td><td>72.8</td><td>60.6</td><td>75.1</td><td>30.3</td><td>55.9</td></tr><tr><td>iCaRL [47]</td><td></td><td>56.6</td><td>44.6</td><td>32.7</td><td>39.3</td><td>46.6</td><td>68.0</td><td>46.0</td><td>77.4</td><td>31.9</td><td>60.5</td></tr><tr><td>LwF-VR [11]</td><td></td><td>77.1</td><td>61.0</td><td>40.5</td><td>45.3</td><td>54.4</td><td>74.6</td><td>47.9</td><td>76.7</td><td>36.3</td><td>58.6</td></tr><tr><td>WiSE-FT [58]</td><td></td><td>73.5</td><td>55.6</td><td>35.6</td><td>41.5</td><td>47.0</td><td>68.3</td><td>53.9</td><td>69.3</td><td>26.8</td><td>51.9</td></tr><tr><td>Dist. only</td><td></td><td>80.1</td><td>62.2</td><td>40.2</td><td>39.9</td><td>58.1</td><td>80.8</td><td>53.4</td><td>74.6</td><td>38.1</td><td>61.9</td></tr><tr><td>ZSCL*</td><td></td><td>78.3</td><td>64.0</td><td>42.9</td><td>45.2</td><td>63.5</td><td>84.2</td><td>56.1</td><td>78.9</td><td>44.1</td><td>64.3</td></tr><tr><td>ZSCL</td><td></td><td>86.0</td><td>67.4</td><td>45.4</td><td>50.4</td><td>69.1</td><td>87.6</td><td>61.8</td><td>86.8</td><td>60.1</td><td>66.8</td></tr><tr><td colspan="12">Avg.</td></tr><tr><td>Continual-FT</td><td>25.5</td><td>81.5</td><td>59.1</td><td>53.2</td><td>64.7</td><td>51.8</td><td>63.2</td><td>64.3</td><td>69.7</td><td>31.8</td><td>49.7</td></tr><tr><td>LwF [33]</td><td>36.3</td><td>86.9</td><td>72.0</td><td>59.0</td><td>73.7</td><td>60.0</td><td>73.6</td><td>74.8</td><td>80.0</td><td>37.3</td><td>58.1</td></tr><tr><td>iCaRL [47]</td><td>35.5</td><td>89.2</td><td>72.2</td><td>60.6</td><td>68.8</td><td>70.0</td><td>78.2</td><td>62.3</td><td>81.8</td><td>41.2</td><td>62.5</td></tr><tr><td>LwF-VR [11]</td><td>29.6</td><td>87.7</td><td>74.4</td><td>59.5</td><td>72.4</td><td>63.6</td><td>77.0</td><td>66.7</td><td>81.2</td><td>43.7</td><td>60.7</td></tr><tr><td>WiSE-FT [58]</td><td>26.7</td><td>86.5</td><td>64.3</td><td>57.1</td><td>65.7</td><td>58.7</td><td>71.1</td><td>70.5</td><td>75.8</td><td>36.9</td><td>54.6</td></tr><tr><td>Dist. only</td><td>48.1</td><td>90.6</td><td>79.8</td><td>63.2</td><td>75.6</td><td>72.5</td><td>84.7</td><td>70.2</td><td>79.8</td><td>46.9</td><td>63.7</td></tr><tr><td>ZSCL*</td><td>50.7</td><td>90.9</td><td>79.8</td><td>63.8</td><td>76.6</td><td>77.3</td><td>87.0</td><td>71.9</td><td>83.0</td><td>52.0</td><td>65.9</td></tr><tr><td>ZSCL</td><td>45.1</td><td>92.0</td><td>80.1</td><td>64.3</td><td>79.5</td><td>81.6</td><td>89.6</td><td>75.2</td><td>88.9</td><td>64.7</td><td>68.0</td></tr><tr><td colspan="12">Last</td></tr><tr><td>Continual-FT</td><td>31.0</td><td>89.3</td><td>65.8</td><td>67.3</td><td>88.9</td><td>71.1</td><td>85.6</td><td>99.6</td><td>92.9</td><td>77.3</td><td>81.1</td></tr><tr><td>LwF [33]</td><td>26.3</td><td>87.5</td><td>71.9</td><td>66.6</td><td>79.9</td><td>66.9</td><td>83.8</td><td>99.6</td><td>92.1</td><td>66.1</td><td>80.4</td></tr><tr><td>iCaRL [47]</td><td>35.8</td><td>93.0</td><td>77.0</td><td>70.2</td><td>83.3</td><td>88.5</td><td>90.4</td><td>86.7</td><td>93.2</td><td>81.2</td><td>81.9</td></tr><tr><td>LwF-VR [11]</td><td>20.5</td><td>89.8</td><td>72.3</td><td>67.6</td><td>85.5</td><td>73.8</td><td>85.7</td><td>99.6</td><td>93.1</td><td>73.3</td><td>80.9</td></tr><tr><td>WiSE-FT [58]</td><td>27.2</td><td>90.8</td><td>68.0</td><td>68.9</td><td>86.9</td><td>74.0</td><td>87.6</td><td>99.6</td><td>92.6</td><td>77.8</td><td>81.3</td></tr><tr><td>Dist. only</td><td>43.3</td><td>91.9</td><td>81.3</td><td>72.4</td><td>95.1</td><td>90.5</td><td>90.4</td><td>99.7</td><td>92.5</td><td>85.1</td><td>81.8</td></tr><tr><td>ZSCL*</td><td>46.0</td><td>92.3</td><td>81.2</td><td>72.4</td><td>93.0</td><td>92.1</td><td>90.8</td><td>99.6</td><td>93.3</td><td>86.6</td><td>81.7</td></tr><tr><td>ZSCL</td><td>40.6</td><td>92.2</td><td>81.3</td><td>70.5</td><td>94.8</td><td>90.5</td><td>91.9</td><td>98.7</td><td>93.9</td><td>85.3</td><td>80.2</td></tr></table>
+
+ensemble achieves a better tradeoff between Transfer and Last value. While  $\mathrm{ZSCL}^*$ , a variant without the WC loss, obtains the highest Last score, the ZSCL with WC loss outperforms it with  $2.8\%$  Avg. and  $5.9\%$  Transfer scores.
+
+# 5.2. Multi-domain Task Incremental Learning
+
+Tab. 3 displays the performance of different methods on the MTIL benchmark, and Tab. 5 presents the detailed Transfer, Avg, and Last metrics on each dataset. Zero-shot denotes the zero-shot prediction performance of the initial CLIP model, and Fine-tune means the direct fine-tuning accuracy on each dataset, which can be seen as an upperbound where no forgetting phenomenon happens. Continual learning uses cross-entropy loss to learn each dataset sequentially, where there is a significant forgetting issue on both zero-shot predictions (Transfer drops by  $24.8\%$ ) and newly learned knowledge (Last drops by  $9.4\%$ ). Previous methods improve the Last performance slightly and cannot maintain a high zero-shot prediction performance. Without WC,  $\mathrm{ZSCL}^*$  achieves the best Last scores, outperforming
+
+the previous best one by  $4.4\%$ . Our method ZSCL improves  $9.1\%$  on Transfer accuracy, with only  $1.3\%$  drops compared with the initial CLIP model, and achieves a  $10.1\%$  gain in the Avg. accuracy.
+
+Figure Fig. 5 provides a visualization of the feature space on Aircraft of the original CLIP and four methods trained at the end of MTCL (t-SNE conducted only once on all features collected). This shows our method is capable of maintaining the pretrained feature distribution with a small averaged feature distance and thus preserving zero-shot performance.
+
+The result of the MTIL method in Order-II is presented in Tab. 4. Our method surpasses previous methods in another order setting of the MTIL benchmark. A similar conclusion holds from the results of MTIL Order-II compared with MTIL Order-I. Our method ZSCL outperforms others by  $9.2\%$  on the Avg. score and  $18.1\%$  on the Last score with only a  $1.2\%$  performance loss on the Transfer score. This shows our approach can work for different orders of the multi-domain task incremental learning. In addition,
+
+Table 6. Comparison of state-of-the-art CL methods on CIFAR100 benchmark in class-incremental setting.  
+
+<table><tr><td rowspan="2">Methods</td><td colspan="2">10 steps</td><td colspan="2">20 steps</td><td colspan="2">50 steps</td></tr><tr><td>Avg</td><td>Last</td><td>Avg</td><td>Last</td><td>Avg</td><td>Last</td></tr><tr><td>UCIR [19]</td><td>58.66</td><td>43.39</td><td>58.17</td><td>40.63</td><td>56.86</td><td>37.09</td></tr><tr><td>BiC [59]</td><td>68.80</td><td>53.54</td><td>66.48</td><td>47.02</td><td>62.09</td><td>41.04</td></tr><tr><td>RPSNet [46]</td><td>68.60</td><td>57.05</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>PODNet [13]</td><td>58.03</td><td>41.05</td><td>53.97</td><td>35.02</td><td>51.19</td><td>32.99</td></tr><tr><td>DER [60]</td><td>74.64</td><td>64.35</td><td>73.98</td><td>62.55</td><td>72.05</td><td>59.76</td></tr><tr><td>DyTox+ [14]</td><td>74.10</td><td>62.34</td><td>71.62</td><td>57.43</td><td>68.90</td><td>51.09</td></tr><tr><td>CLIP [44]</td><td>74.47</td><td>65.92</td><td>75.20</td><td>65.74</td><td>75.67</td><td>65.94</td></tr><tr><td>FT</td><td>65.46</td><td>53.23</td><td>59.69</td><td>43.13</td><td>39.23</td><td>18.89</td></tr><tr><td>LwF [33]</td><td>65.86</td><td>48.04</td><td>60.64</td><td>40.56</td><td>47.69</td><td>32.90</td></tr><tr><td>iCaRL [47]</td><td>79.35</td><td>70.97</td><td>73.32</td><td>64.55</td><td>71.28</td><td>59.07</td></tr><tr><td>LwF-VR [11]</td><td>78.81</td><td>70.75</td><td>74.54</td><td>63.54</td><td>71.02</td><td>59.45</td></tr><tr><td>ZSCL (Ours)</td><td>82.15</td><td>73.65</td><td>80.39</td><td>69.58</td><td>79.92</td><td>67.36</td></tr><tr><td>Impr</td><td>+7.68</td><td>+7.73</td><td>+5.19</td><td>+3.84</td><td>+3.95</td><td>+1.42</td></tr></table>
+
+compared with Order-I, previous methods achieve a much lower Last score (e.g., for Continual-Learning, Order-I has  $77.3\%$ , while Order-II has  $65.3\%$ ). With ZSCL, the Last score is similar ( $83.6\%$  compared with  $83.4\%$ ). This shows our method is more robust towards different training orders.
+
+# 5.3. Class Incremental Learning
+
+We evaluate our methods on conventional class incremental learning. Tabs. 6 and 7 display results on CIFAR100 and TinyImageNet, respectively. We re-implement some previous methods with a CLIP backbone (after CLIP in the table), while others using a special network design cannot be easily adapted. Although zero-shot CLIP prediction achieves a good result on these benchmarks, continual learning with direct fine-tuning or LwF [33] degrades the performance greatly, especially under the setting of a large step number. This demonstrates a severe catastrophic forgetting phenomenon in fine-tuning the CLIP model. Our method consistently improves the performance of the CLIP model on both Avg. and Last scores with a large gap towards previous ones.
+
+# 6. Limitation and Future Work
+
+Our proposed method has a limitation that a reference dataset is needed. A promising direction of the work is to preserve the zero-shot transfer ability without the need for an outside dataset. For example, we may generate a synthetic image dataset as the reference dataset. Methods like [51] can synthesize datasets from a network.
+
+The deep learning community tends to build large models with a huge dataset [3, 6], including vision-language models [44, 20]. As re-training cost upsurges, continual learning is an efficient approach for updating these models with custom usage.
+
+Table 7. Comparison of different methods on TinyImageNet splits in class-incremental settings with 100 base classes.  
+
+<table><tr><td rowspan="2">Methods</td><td colspan="2">5 steps</td><td colspan="2">10 steps</td><td colspan="2">20 steps</td></tr><tr><td>Avg</td><td>Last</td><td>Avg</td><td>Last</td><td>Avg</td><td>Last</td></tr><tr><td>EWC [25]</td><td>19.01</td><td>6.00</td><td>15.82</td><td>3.79</td><td>12.35</td><td>4.73</td></tr><tr><td>EEIL [4]</td><td>47.17</td><td>35.12</td><td>45.03</td><td>34.64</td><td>40.41</td><td>29.72</td></tr><tr><td>UCIR [19]</td><td>50.30</td><td>39.42</td><td>48.58</td><td>37.29</td><td>42.84</td><td>30.85</td></tr><tr><td>MUC [35]</td><td>32.23</td><td>19.20</td><td>26.67</td><td>15.33</td><td>21.89</td><td>10.32</td></tr><tr><td>PASS [65]</td><td>49.54</td><td>41.64</td><td>47.19</td><td>39.27</td><td>42.01</td><td>32.93</td></tr><tr><td>DyTox [14]</td><td>55.58</td><td>47.23</td><td>52.26</td><td>42.79</td><td>46.18</td><td>36.21</td></tr><tr><td>CLIP [44]</td><td>69.62</td><td>65.30</td><td>69.55</td><td>65.59</td><td>69.49</td><td>65.30</td></tr><tr><td>FT</td><td>61.54</td><td>46.66</td><td>57.05</td><td>41.54</td><td>54.62</td><td>44.55</td></tr><tr><td>LwF [33]</td><td>60.97</td><td>48.77</td><td>57.60</td><td>44.00</td><td>54.79</td><td>42.26</td></tr><tr><td>iCaRL [47]</td><td>77.02</td><td>70.39</td><td>73.48</td><td>65.97</td><td>69.65</td><td>64.68</td></tr><tr><td>LwF-VR [11]</td><td>77.56</td><td>70.89</td><td>74.12</td><td>67.05</td><td>69.94</td><td>63.89</td></tr><tr><td>ZSCL (Ours)</td><td>80.27</td><td>73.57</td><td>78.61</td><td>71.62</td><td>77.18</td><td>68.30</td></tr><tr><td>Impr</td><td>+10.65</td><td>+8.27</td><td>+9.06</td><td>+6.03</td><td>+7.69</td><td>+3.00</td></tr></table>
+
+In some cases, we want to correct wrong information in the pre-training dataset or update old information with the latest one. How to conduct this task with a reference dataset is left as future work.
+
+Lately, a noticeable trend involves the creation of multimodality models utilizing large language models, showcasing encouraging outcomes in tasks such as Visual Question Answering (VQA). Expanding our approach to encompass the next token prediction task remains an avenue for future exploration and research.
+
+# 7. Conclusion
+
+In this paper, we investigate continual learning with the vision-language model. We propose a better continual learning algorithm to protect the zero-shot transfer ability in the vision-language model learned in the pre-training stage. Our algorithm mitigates the catastrophic forgetting in both feature space and parameter space. In feature space, distilling the initial model on a reference dataset significantly boosts the model's performance. In parameter space, weight ensemble among different training stages alleviates the forgetting issue. We propose a more challenging Multidomain Task Incremental Learning (MTIL) benchmark to evaluate the continual learning methods better. On both conventional and new benchmarks, our method achieves state-of-the-art performance.
+
+# Acknowledgements
+
+Yang You's research group is being sponsored by the NUS startup grant (Presidential Young Professorship), Singapore MOE Tier-1 grant, ByteDance grant, ARCTIC grant, SMI grant, and Alibaba grant.
+
+# References
+
+[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV), pages 139-154, 2018. 3, 5  
+[2] Eden Belouadah, Adrian Popescu, and Ioannis Kanellos. A comprehensive study of class incremental learning algorithms for visual tasks. Neural Networks, 135:38-54, 2021. 3  
+[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2, 9  
+[4] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 9  
+[5] Zhou Daquan, Kai Wang, Jianyang Gu, Xiangyu Peng, Dongze Lian, Yifan Zhang, Yang You, and Jiashi Feng. Dataset quantization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2  
+[6] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, et al. Scaling vision transformers to 22 billion parameters. arXiv preprint arXiv:2302.05442, 2023. 9  
+[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6  
+[8] Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29(6):141-142, 2012. 2  
+[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3  
+[10] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5138-5146, 2019. 2, 3  
+[11] Yuxuan Ding, Lingqiao Liu, Chunna Tian, Jingyuan Yang, and Haoxuan Ding. Don't stop learning: Towards continual learning for the clip model. arXiv preprint arXiv:2207.09248, 2022. 2, 3, 7, 8, 9  
+[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 6  
+[13] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In European Con-
+
+ference on Computer Vision, pages 86-102. Springer, 2020. 2,3,9  
+[14] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285-9295, 2022. 2, 3, 5, 6, 9  
+[15] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 3  
+[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 3  
+[17] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 6  
+[18] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 2, 4  
+[19] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 831-839, 2019. 9  
+[20] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, et al. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 9  
+[21] David Isele and Akansel Cosgun. Selective experience replay for lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 2, 3  
+[22] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018. 5  
+[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2, 3  
+[24] Woojeong Jin, Yu Cheng, Yelong Shen, Weizhu Chen, and Xiang Ren. A good prompt is worth millions of parameters? low-resource prompt-based learning for vision-language models. arXiv preprint arXiv:2110.08484, 2021. 2  
+[25] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 3, 5, 9
+
+[26] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 5  
+[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1  
+[28] Frantzeska Lavda, Jason Ramapuram, Magda Gregorova, and Alexandros Kalousis. Continual classification learning using generative models. arXiv preprint arXiv:1810.10612, 2018. 2, 3  
+[29] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 2  
+[30] Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming catastrophic forgetting by incremental moment matching. Advances in neural information processing systems, 30, 2017. 3  
+[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 3  
+[32] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 2022. 2  
+[33] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1, 2, 3, 4, 7, 8, 9  
+[34] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. ICCV-2023, 2023. 2  
+[35] Yu Liu, Sarah Parisot, Gregory Slabaugh, Xu Jia, Ales Leonardis, and Tinne Tuytelaars. More classifiers, less forgetting: A generic multi-classifier paradigm for incremental learning. In European Conference on Computer Vision, pages 699-716. Springer, 2020. 9  
+[36] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2, 3  
+[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6  
+[38] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 6  
+[39] Zixuan Ni, Longhui Wei, Siliang Tang, Yueting Zhuang, and Qi Tian. Continual vision-language representation learning with off-diagonal information. arXiv preprint arXiv:2305.07437, 2023. 3  
+[40] Oleksiy Ostapenko, Timothee Lesort, Pau Rodríguez, Md Rifat Arefin, Arthur Douillard, Irina Rish, and Laurent Charlin. Continual learning with foundation models: An empirical study of latent replay. In Conference on Lifelong Learning Agents, pages 60–91. PMLR, 2022. 3  
+[41] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3498-3505, 2012. 1
+
+[42] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In European conference on computer vision, pages 524-540. Springer, 2020. 2, 3  
+[43] Ziheng Qin, Kai Wang, Zangwei Zheng, Jianyang Gu, Xi-angyu Peng, Daquan Zhou, and Yang You. Infobatch: Lossless training speed up by unbiased dynamic data pruning. arXiv preprint arXiv:2303.04947, 2023. 2  
+[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 6, 9  
+[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67, 2020. 2  
+[46] Jathushan Rajasegaran, Munawar Hayat, Salman Khan, Fahad Shahbaz Khan, Ling Shao, and Ming-Hsuan Yang. An adaptive random path selection approach for incremental learning. arXiv preprint arXiv:1906.01120, 2019. 9  
+[47] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1, 2, 3, 4, 7, 8, 9  
+[48] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 6  
+[49] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 7  
+[50] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 2, 3  
+[51] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9374-9384, 2021. 9  
+[52] Tejas Srinivasan, Ting-Yun Chang, Leticia Pinto Alva, Georgios Chochlakis, Mohammad Rostami, and Jesse Thomason. Climb: A continual learning benchmark for vision-and-language tasks. Advances in Neural Information Processing Systems, 35:29440-29453, 2022. 3  
+[53] Vishal Thengane, Salman Khan, Munawar Hayat, and Fahad Khan. Clip model is an efficient continual learner. arXiv preprint arXiv:2210.03114, 2022. 2, 3  
+[54] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 7
+
+[55] Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. arXiv preprint arXiv:1904.07734, 2019. 3, 5  
+[56] Zhen Wang, Liu Liu, Yiqun Duan, Yajing Kong, and Dacheng Tao. Continual learning with lifelong vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 171-181, 2022. 3  
+[57] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 5  
+[58] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 2, 3, 5, 7, 8  
+[59] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 2, 9  
+[60] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3014-3023, 2021. 2, 5, 9  
+[61] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017. 3  
+[62] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International Conference on Machine Learning, pages 3987-3995. PMLR, 2017. 3, 5  
+[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2  
+[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3  
+[65] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and Cheng-Lin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871-5880, 2021. 2, 5, 9

2023/Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:465b863ef1b8357943637e91cb430f7f659bc85d9012ef20687d532a19b3becd
+size 764821

2023/Prior-guided Source-free Domain Adaptation for Human Pose Estimation/67d12eb1-f1e4-4b90-aba2-4771e3bf7c39_content_list.json

@@ -0,0 +1,2011 @@
+[
+    {
+        "type": "text",
+        "text": "Prior-guided Source-free Domain Adaptation for Human Pose Estimation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            130,
+            854,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Dripta S. Raychaudhuri $^{1,2*}$  Calvin-Khang Ta $^{1}$  Arindam Dutta $^{1}$  Rohit Lal $^{1}$  Amit K. Roy-Chowdhury $^{1}$ $^{1}$ University of California, Riverside  $^{2}$ AWS AI Labs",
+        "bbox": [
+            86,
+            186,
+            898,
+            228
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{drayc001@, cta003@, adutt020@, rla1011@, amertc@ece.}ucr.edu",
+        "bbox": [
+            223,
+            233,
+            763,
+            250
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            300,
+            313,
+            316
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Domain adaptation methods for 2D human pose estimation typically require continuous access to the source data during adaptation, which can be challenging due to privacy, memory, or computational constraints. To address this limitation, we focus on the task of source-free domain adaptation for pose estimation, where a source model must adapt to a new target domain using only unlabeled target data. Although recent advances have introduced source-free methods for classification tasks, extending them to the regression task of pose estimation is non-trivial. In this paper, we present Prior-guided Self-training (POST), a pseudolabeling approach that builds on the popular Mean Teacher framework to compensate for the distribution shift. POST leverages prediction-level and feature-level consistency between a student and teacher model against certain image transformations. In the absence of source data, POST utilizes a human pose prior that regularizes the adaptation process by directing the model to generate more accurate and anatomically plausible pose pseudo-labels. Despite being simple and intuitive, our framework can deliver significant performance gains compared to applying the source model directly to the target data, as demonstrated in our extensive experiments and ablation studies. In fact, our approach achieves comparable performance to recent state-of-the-art methods that use source data for adaptation.",
+        "bbox": [
+            75,
+            333,
+            473,
+            710
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            742,
+            209,
+            757
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Human pose estimation is a fundamental task in computer vision that involves determining the precise locations of keypoints, such as joints, on a human body in an image or video [1]. The growing need for pose estimation in various applications such as action recognition [2], human-computer interaction [3], and video surveillance [4] has driven the rapid development of highly accurate deep learning tech",
+        "bbox": [
+            75,
+            768,
+            470,
+            875
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/f269fd9a6d1df8c4a7d32d89b1b306e6b1a67e8a5f63ca42de4197fc02660e7b.jpg",
+        "image_caption": [
+            "Figure 1. Problem setup. Existing UDA methods for pose estimation rely on a labeled source dataset while adapting to an unlabeled target dataset. However, privacy concerns surrounding the use of personally identifiable information in these labeled datasets, as well as the significant storage and computational requirements, can limit access to such data. Hence, our work focuses on source-free UDA of pose estimation models."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            285,
+            890,
+            512
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "niques. However, the challenge of obtaining large annotated datasets for training, compounded with the susceptibility to a performance decline in the face of data distribution shifts still poses limitations for current pose estimation models.",
+        "bbox": [
+            496,
+            642,
+            893,
+            703
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To overcome these limitations, recent studies have focused on unsupervised domain adaptation (UDA) of pose estimators [5, 6]. UDA allows for transferring a pose estimation model trained on a source domain, where labeled data is available, to a target domain where labeled data is unavailable. Despite improved and robust pose estimation, the requirement of simultaneous access to both source and target domains during adaptation hinders real-world implementation. For instance, the labeled source data may not be accessible post-deployment due to privacy or proprietary issues. This is particularly relevant for human pose datasets, which contain personally identifiable information (PII) [7]. Furthermore, adaptation using the entire source data might",
+        "bbox": [
+            496,
+            704,
+            893,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            807,
+            47
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Currently at AWS AI Labs. Work done while the author was at UCR.",
+        "bbox": [
+            98,
+            886,
+            467,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "14996",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "be infeasible due to both memory and computational constraints. In light of these issues, we focus on source-free UDA of human pose estimation models.",
+        "bbox": [
+            76,
+            90,
+            470,
+            136
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Concretely, our objective is to adapt a 2D human pose estimation model to a new target domain utilizing only its trained parameters and unlabeled target data. This presents a major challenge as the absence of source data for regularization can cause catastrophic forgetting. While recent advances have introduced methods to tackle this issue in classification tasks [8, 9, 10], extending them to the regression task of pose estimation is non-trivial. To address this challenge, we introduce Prior-guided Self-training (POST), a self-training regression framework that employs a human pose prior to effectively guide the adaptation process in the absence of source data. An overview of our problem setup is shown in Figure 1.",
+        "bbox": [
+            75,
+            138,
+            470,
+            335
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our approach builds on the Mean Teacher [11] framework, which uses consistency in the prediction space of a student and teacher model to produce trustworthy pseudo-labels and learn from the unlabeled target domain. To achieve this, we create two augmented views of each target image, varying in scale, spatial context, and color statistics. Aligned pose predictions from both models in both views are then obtained, and consistency between the predictions across the different views is encouraged to facilitate prediction space adaptation. However, our empirical results show that relying solely on consistency in the output space is insufficient when supervision from the source data is lacking. To address this, we also introduce feature space adaptation, which aims to encourage consistency across features extracted from the two separate views. We adopt the Barlow Twins [12] approach to accomplish this. Specifically, we seek to make the cross-correlation matrix calculated from a pair of feature embeddings as close to the identity matrix as possible.",
+        "bbox": [
+            75,
+            338,
+            473,
+            611
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In addition to the adaptation across both outputs and features, we employ a human pose prior that models the full manifold of plausible poses in some high-dimensional pose space to refine possible noisy pseudo-labels that may arise during self-training. The plausible poses are represented as points on the manifold surface, with zero distance from it, while non-plausible poses are located outside the surface, with a non-zero distance from it. This manifold is learned using a high-dimensional neural field, similar to Pose-NDF [13]. The pose prior acts as a strong regularizer, directing the model to generate more accurate pose pseudolabels on the target data and leading to improved adaptation. The learning of this prior requires an auxiliary dataset of plausible human poses, but this does not compromise the privacy aspect of our framework as the prior does not make use of RGB images. In addition, it is worth noting that the prior can be trained offline, separately from the adaptation process. This not only saves computational resources but also reduces the amount of storage required. Compared to storing entire",
+        "bbox": [
+            75,
+            613,
+            470,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "images, it is much more efficient to store pose coordinates, which requires approximately  $3000 \\times$  less memory.",
+        "bbox": [
+            498,
+            90,
+            893,
+            121
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Main contributions. To summarize, our primary contributions are as follows:",
+        "bbox": [
+            498,
+            122,
+            893,
+            151
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We address the problem of adapting a human pose estimation model to a target domain consisting of unlabeled data, without access to the original source dataset. This ameliorates the privacy concern associated with the current domain adaptive pose estimation methods.",
+            "- We introduce Prior-guided Self-training (POST), a simple source-free unsupervised adaptation algorithm. POST leverages both prediction-level and feature-level consistency, in addition to a human pose prior, to drive self-training for improved adaptation to the target domain.",
+            "- We evaluate our method qualitatively and quantitatively on three challenging domain adaptive scenarios, demonstrating comparable performance to existing UDA methods that have access to the source data."
+        ],
+        "bbox": [
+            500,
+            152,
+            893,
+            367
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Works",
+        "text_level": 1,
+        "bbox": [
+            500,
+            383,
+            650,
+            398
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Pose Estimation. 2D human pose estimation aims to locate human anatomical keypoints, such as the elbow and knee. Prior works can be categorized into two primary frameworks: the top-down framework and the bottom-up framework. Top-down methods [14, 15, 16, 17, 18, 19, 1] first detect each person from the image and then perform single-person pose estimation on each bounding box independently. On the other hand, bottom-up methods [20, 21, 22, 23, 24, 25, 26] predict keypoints of each person directly in an end-to-end manner. Typical bottom-up methods consist of two steps: predicting keypoint heatmaps and grouping the detected keypoints into separate poses. In this work, we focus on the bottom-up framework for efficiency purposes and adopt the Simple Baseline [18] architecture following [6] to ensure fair comparisons with prior domain adaptation algorithms.",
+        "bbox": [
+            498,
+            409,
+            893,
+            636
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Unsupervised Domain Adaptation. UDA methods have been extensively applied to a broad range of computer vision tasks, including image classification [27], semantic segmentation [28], object detection [29], and reinforcement learning [30] to tackle the issue of data distribution shift. Most approaches aim to align the source and target data distributions through techniques such as maximum mean discrepancy [31] and adversarial learning [32, 27]. Another line of research utilizes image translation methods to perform adaptation by transforming the source images into the target domain [33, 34]. More recently, there has been a surge of interest in adaptation using only a pre-trained source model due to privacy and memory storage concerns related to the source data. These include techniques such as information maximization [8, 9], pseudo-labeling [35, 36] and self-supervision [37]. Compared to other tasks, domain adaptation for regression tasks, such as pose estimation, remains",
+        "bbox": [
+            496,
+            643,
+            895,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "14997",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "relatively unexplored.",
+        "bbox": [
+            76,
+            90,
+            223,
+            104
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Domain Adaptive Pose Estimation. UDA methods for pose estimation have explored various techniques for overcoming the domain gap, including adversarial feature alignment and pseudo-labeling. RegDA [5] estimates the domain discrepancy by evaluating false predictions on the target domain and minimizes it. Mu et al. [38] proposed consistency regularization with respect to transformations and temporal consistency learning within a video. Li et al. [39] proposed a refinement module and a self-feedback loop to obtain reliable pseudo-labels. Recently, Kim et al. [6] introduced a unified framework for both human and animal keypoint detection, which aligns representations using input-level and output-level cues. Typically, these methods require access to the source data, which may raise data privacy, memory and computation concerns. In contrast, our method addresses the domain adaptation problem in a source-free setting.",
+        "bbox": [
+            75,
+            112,
+            473,
+            354
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Prior-guided Self-training",
+        "text_level": 1,
+        "bbox": [
+            76,
+            368,
+            320,
+            386
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our work investigates source-free domain adaptation for 2D human pose estimation. In the pose estimation task, given an input image  $\\pmb{x} \\in \\mathbb{R}^{H \\times W \\times 3}$ , the goal is to predict the corresponding  $y \\in \\mathbb{R}^{K \\times 2}$  representing the 2D coordinates of  $K$  keypoints using a pose regression model  $f$ . In this paper, we assume access to a pre-trained model, denoted by  $f_{S}$ , as well as  $N$  unlabeled images  $\\mathcal{D} = \\{\\pmb{x}_i\\}_{i=1}^N$  from a target domain  $\\mathcal{T}$ . Our goal is to adapt the source model to the target such that it performs better on images drawn from the target distribution than when directly using the source model on the target images.",
+        "bbox": [
+            75,
+            393,
+            470,
+            559
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Overview. In the absence of source data, we propose to use self-training to adapt the source pose estimation model to the target domain. However, self-training methods are prone to error accumulation, particularly when labeled data is absent to act as regularization. Hence, we introduce POST, an enhanced self-training strategy that employs three essential ideas to prevent such errors:",
+        "bbox": [
+            75,
+            564,
+            468,
+            669
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. A weight-averaged teacher model is used to generate the pseudo-labels for self-training. This ensures better retention of the source knowledge within the teacher model by reducing the effect of updating the weights via noisy pseudo-labels (Section 3.1.1).",
+            "2. In addition to adaptation over the output space via pseudolabels, the model is adapted in the feature space as well. For each target image, two aligned predictive views are generated via data augmentation, and consistency across features extracted from the two separate views is encouraged via a contrastive learning strategy (Section 3.1.2).",
+            "3. A human pose prior is used to regularize the adaptation by directing the model to generate more accurate and anatomically plausible pose pseudo-labels (Section 3.2). An overview of our framework is presented in Figure 2."
+        ],
+        "bbox": [
+            76,
+            670,
+            470,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Self-training via Mean Teacher",
+        "text_level": 1,
+        "bbox": [
+            500,
+            90,
+            774,
+            107
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Motivated by research suggesting that weight-averaged models over training steps tend to perform better than the final model [6, 39], we utilize the Mean Teacher framework [11] to generate pseudo-labels for self-training. The framework involves creating two identical models, a teacher model  $f_{\\mathrm{tea}}$  and a student model  $f_{\\mathrm{stu}}$ , both of which are initialized with the pre-trained network  $f_{S}$  at time step  $t = 0$ . At each subsequent time step  $t$ , the student model parameters  $\\theta$  are updated by backpropagating the supervisory signals provided by the teacher model. The parameters of the teacher model  $\\tilde{\\theta}$  are updated via an exponential moving average (EMA) of the student model parameters:",
+        "bbox": [
+            496,
+            113,
+            893,
+            295
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\tilde {\\theta} _ {t} = \\alpha \\tilde {\\theta} _ {t - 1} + (1 - \\alpha) \\theta_ {t}, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            609,
+            305,
+            893,
+            325
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\alpha$  denotes the smoothing coefficient which is set to 0.999 by default. The EMA update prevents the teacher model from overfitting to noisy pseudo-labels during the initial rounds of self-training, thereby, preserving the source knowledge. This is especially advantageous in our scenario where source data is unavailable to regularize the adaptation.",
+        "bbox": [
+            496,
+            335,
+            893,
+            426
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In the following sections, we demonstrate how to adapt  $f_{\\mathrm{stu}}$  using supervisory signals from  $f_{\\mathrm{tea}}$  on both the feature space and the pose prediction space.",
+        "bbox": [
+            496,
+            426,
+            893,
+            472
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1.1 Prediction Space Adaptation",
+        "text_level": 1,
+        "bbox": [
+            498,
+            492,
+            754,
+            508
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "At each time step  $t$ , we apply two different data augmentations,  $\\mathcal{A}_1$  and  $\\mathcal{A}_2$ , to a target image  $\\pmb{x}$  to generate two views. We then obtain the keypoint heatmap corresponding to the first transformed image  $\\tilde{h}^t = f_{\\mathrm{tea}}^t (\\mathcal{A}_1(\\pmb {x}))$ . Here  $\\tilde{h}^t\\in \\mathbb{R}^{K\\times H'\\times W'}$  denotes the spatial likelihood of the  $K$  different keypoints on each channel. The pseudolabel from the teacher model is generated by obtaining the coordinates which produce the maximum activations  $\\hat{y}^t = \\operatorname{argmax}_u\\tilde{h}_{[:,u]}^t$ .",
+        "bbox": [
+            496,
+            516,
+            893,
+            652
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To reduce the influence of erroneous pseudo-labels in our training process, we utilize a confidence threshold to discard potentially unreliable labels. Specifically, we only retain the keypoint activations with the top  $p\\%$  maximum values among all activations and discard the rest. We set the threshold  $\\tau$  accordingly to reflect this, thereby ensuring that only the most confident labels are used for training.",
+        "bbox": [
+            496,
+            652,
+            893,
+            757
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Following prior work on supervised pose estimation, we first convert the pseudo-labels to normalized Gaussian heatmaps [40] and then use the mean squared error (MSE) loss to update the student model over batches of target images sampled from  $\\mathcal{D}$ :",
+        "bbox": [
+            496,
+            758,
+            893,
+            834
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {o u t}} = \\frac {1}{| \\mathcal {B} |} \\sum_ {\\mathbf {x} \\in \\mathcal {B}} \\sum_ {k = 1} ^ {K} \\mathbb {1} \\left(\\hat {h} _ {k} ^ {t} \\geq \\tau\\right) \\| \\tilde {\\mathcal {A}} _ {1} \\left(\\hat {h} _ {k} ^ {t}\\right) - \\tilde {\\mathcal {A}} _ {2} \\left(h _ {k} ^ {t}\\right) \\| _ {2}. \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            511,
+            845,
+            893,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "14998",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/c556debaba324cdf504fd1ec865cb5e134057372a5fa29e9517f0c4813dae355.jpg",
+        "image_caption": [
+            "Figure 2. Framework overview. Our approach builds on the Mean Teacher framework and performs adaptation both in the pose prediction space using  $\\mathcal{L}_{\\mathrm{out}}$ , and the feature space using  $\\mathcal{L}_{\\mathrm{feat}}$ . This is supplemented by a human pose prior  $g$  that scores the predicted pseudo-labels in terms of plausibility. These scores are used to regularize the adaptation process in the absence of labeled source data via  $\\mathcal{L}_{\\mathrm{prior}}$ . The student model  $f_{\\mathrm{stu}}$  is trained by the combination of the three losses, while the teacher model  $f_{\\mathrm{tea}}$  is updated with the exponential moving average (EMA) of the weights of the student model."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            138,
+            88,
+            831,
+            367
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We denote the inverse functions of the chosen augmentations  $\\mathcal{A}_1$  and  $\\mathcal{A}_2$  as  $\\tilde{\\mathcal{A}}_1$  and  $\\tilde{\\mathcal{A}}_2$ , respectively. The heatmap predicted by the student model for the  $k$ -th keypoint is represented by  $h_k^t = f_{\\mathrm{stu}}^t (\\mathcal{A}_2(\\pmb {x}))^k$ , while the heatmap generated from the pseudo-labels predicted by the teacher model is represented by  $\\hat{h}_k^t = L(\\hat{y}^t)^k$ . Here,  $L(\\cdot)$  represents the heatmap generating function and  $\\mathcal{B}$  denotes a batch of target images.",
+        "bbox": [
+            75,
+            465,
+            473,
+            573
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1.2 Feature Apace Adaptation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            589,
+            316,
+            606
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Pose estimation models primarily rely on a high-to-low-resolution feature encoder to generate low-resolution representations, which are then used to recover high-resolution pose heatmaps [18]. Since the upsampling process is inherently noisy, providing intermediate supervision to explicitly adapt the features, in addition to adapting the output pose keypoints as shown in the previous section, can be beneficial [17]. While previous work has demonstrated the benefits of joint adaptation across the output and feature space via adversarial learning in tasks such as semantic segmentation [41], this has been exclusively focused on scenarios where source data is available. Here, we propose an alternative way to accomplish feature space adaptation in the absence of source data via contrastive learning.",
+        "bbox": [
+            75,
+            613,
+            470,
+            824
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We begin by creating two different views of each target image  $\\mathbf{x}$  using a pair of sampled augmentations  $\\mathcal{A}_1$  and  $\\mathcal{A}_2$ , as previously shown. Next, considering the pose estimation model as a composition of a feature encoder and an output regressor, i.e.,  $f = \\mathrm{Dec} \\circ \\mathrm{Enc}$ , we extract",
+        "bbox": [
+            75,
+            825,
+            472,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "the augmentation reversed feature maps from the teacher model  $\\tilde{z} = \\tilde{\\mathcal{A}}_1(\\mathrm{Enc}_{\\mathrm{tea}}(\\mathcal{A}_1(\\boldsymbol{x})))$  and the student model  $z = \\tilde{\\mathcal{A}}_2(\\mathrm{Enc}_{\\mathrm{stu}}(\\mathcal{A}_2(\\boldsymbol{x})))$ . We extract pairs of features for every image in a batch, pool them along the spatial dimensions, and then normalize them along the batch dimension to ensure that each covariate has a mean of 0 over the batch. For simplicity, we overload  $z,\\tilde{z}$  to represent the normalized features, and drop the time index  $t$ .",
+        "bbox": [
+            496,
+            465,
+            893,
+            585
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We utilize feature-level consistency between the different views in order to accomplish feature space adaptation. This is achieved via a contrastive learning strategy which encourages the cross-correlation matrix between the outputs of the two networks to be as close to the identity matrix as possible [12],",
+        "bbox": [
+            496,
+            585,
+            893,
+            676
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {f e a t}} = \\sum_ {i} (1 - C _ {i i}) ^ {2} + \\gamma \\sum_ {i} \\sum_ {j \\neq i} C _ {i j} ^ {2}. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            560,
+            684,
+            893,
+            718
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We define  $\\gamma$  as a positive constant that balances the importance of the first and second terms of the loss function.  $C$  represents the cross-correlation matrix computed between the outputs of the student and teacher networks along the batch dimension:",
+        "bbox": [
+            496,
+            723,
+            893,
+            797
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nC _ {i j} = \\frac {\\sum_ {b = 1} ^ {| \\mathcal {B} |} \\tilde {z} _ {b , i} z _ {b , j}}{\\sqrt {\\sum_ {b = 1} ^ {| \\mathcal {B} |} \\left(\\tilde {z} _ {b , i}\\right) ^ {2}} \\sqrt {\\sum_ {b = 1} ^ {| \\mathcal {B} |} \\left(z _ {b , j}\\right) ^ {2}}}. \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            562,
+            801,
+            893,
+            849
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Here,  $b$  represents the batch index, while  $i$  and  $j$  index the feature dimensions of the networks' outputs. We set  $\\gamma = 5e - 3$  following [12].",
+        "bbox": [
+            496,
+            854,
+            893,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "14999",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The first term of  $\\mathcal{L}_{\\mathrm{feat}}$  encourages the consistency of pose features within the same image by equating the diagonal elements of the cross-correlation matrix to 1, effectively reversing the effects of augmentations. In contrast, the second term aims to decorrelate the different feature dimensions of the embedding by forcing the off-diagonal elements of the cross-correlation matrix to 0.",
+        "bbox": [
+            75,
+            90,
+            472,
+            196
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2. Regularization via Pose Prior",
+        "text_level": 1,
+        "bbox": [
+            76,
+            204,
+            339,
+            219
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Enhancing the performance of the pose estimation model on the target domain through joint adaptation over the output and feature space is undoubtedly valuable. However, this approach has its limitations, as it relies solely on general domain adaptation principles and overlooks the rich structural priors associated with human poses. To address this issue, we propose incorporating a parametric human pose prior to better adapt the pose estimation model to the target domain.",
+        "bbox": [
+            75,
+            227,
+            473,
+            349
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2.1 2D Human Pose Prior",
+        "text_level": 1,
+        "bbox": [
+            76,
+            366,
+            284,
+            378
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Building upon the work of [13], we propose a human pose prior modeled as a manifold consisting of plausible 2D poses. To represent the 2D poses while ignoring aspects such as size and scale, we use a set of 2D orientation vectors that connect pairs of joints in the human skeleton, denoted by  $\\mathcal{G} = \\{\\pmb {\\theta} = (\\theta_1,\\dots ,\\theta_L)\\mid \\theta_l\\in \\mathbb{R}^2,\\| \\theta_l\\| _2 = 1\\forall l\\in [L]\\}$ . We assume that plausible 2D human poses lie on a manifold embedded in this pose space  $\\mathcal{G}$ . We use a function  $g:\\mathcal{G}\\to$ $\\mathbb{R}^+$ , which maps a pose to a non-negative scalar, to represent the manifold of plausible poses as the zero-level set:",
+        "bbox": [
+            75,
+            388,
+            472,
+            541
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {P} = \\{\\boldsymbol {\\theta} \\in \\mathcal {G} \\mid g (\\boldsymbol {\\theta}) = 0 \\}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            184,
+            547,
+            470,
+            565
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $g$  represents the unsigned distance to the manifold. We construct this distance function by first encoding the pose using a hierarchical network  $g_{\\mathrm{enc}}$  that encodes the human pose based on its anatomical structure [42], and subsequently, use  $g_{\\mathrm{dec}}$  to predict the distance based on the pose representation.",
+        "bbox": [
+            75,
+            571,
+            472,
+            648
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Specifically, for a given pose  $\\theta$ , we encode it as follows,",
+        "bbox": [
+            94,
+            648,
+            465,
+            662
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {v} _ {1} = g _ {\\mathrm {e n c}} ^ {1} \\left(\\theta_ {1}\\right), \\quad \\mathbf {v} _ {l} = g _ {\\mathrm {e n c}} ^ {l} \\left(\\theta_ {l}, \\mathbf {v} _ {\\Omega (l)}\\right), l \\in \\{2, \\dots , L \\}. \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            84,
+            667,
+            468,
+            686
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Here,  $\\Omega(l)$  is a function that maps the index of each orientation vector to its parent orientation vector in the kinematic chain of the human skeleton. We obtain the overall pose encoding as  $\\mathbf{p} = [\\mathbf{v}_1|\\ldots|\\mathbf{v}_L]$  by concatenating all the individual orientation encodings. This pose encoding is then processed by  $g_{\\mathrm{dec}}: \\mathbb{R}^{d,L} \\to \\mathbb{R}^+$ , which predicts the unsigned distance for the given pose representation  $\\mathbf{p}$ . A lower distance value for a pose implies that the configuration of joints is more likely to be a plausible human pose.",
+        "bbox": [
+            75,
+            693,
+            472,
+            829
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2.2 Pose Prior Training",
+        "text_level": 1,
+        "bbox": [
+            76,
+            845,
+            269,
+            862
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To train the parametric prior  $g$ , we rely on an auxiliary dataset of  $M$  human poses  $\\mathcal{D}_{\\mathrm{A}} = \\{\\pmb{\\theta}^{i}\\}_{i=1}^{N_{\\mathrm{A}}}$ , where  $\\pmb{\\theta}^{i} =$",
+        "bbox": [
+            75,
+            869,
+            470,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "$(\\theta_1^i,\\dots ,\\theta_L^i)$ . Importantly, these poses are not associated with their corresponding RGB images, which preserves the privacy aspect of the method. Additionally, storing only the pose coordinates instead of entire images makes data storage much more efficient and feasible.",
+        "bbox": [
+            496,
+            90,
+            890,
+            165
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We adopt a supervised approach to train  $g$  to predict the  $L2$  distance to the plausible pose manifold for a given pose. To achieve this, we construct a dataset  $\\tilde{\\mathcal{D}} = \\{(\\pmb{\\theta}_i,d_i)\\}_{i=1}^M$ , consisting of pose and distance pairs, from  $\\mathcal{D}_{\\mathrm{A}}$ . As the poses from  $\\mathcal{D}_{\\mathrm{A}}$  lie on the desired manifold, we assign  $d = 0$  to all poses in the dataset. To diversify our training samples, we randomly generate negative samples with distance  $d > 0$  by perturbing the poses from  $\\mathcal{D}_{\\mathrm{A}}$  with noise.",
+        "bbox": [
+            496,
+            167,
+            893,
+            287
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We train the network with the standard  $L1$  loss,",
+        "bbox": [
+            517,
+            287,
+            834,
+            301
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t}} = \\sum_ {(\\boldsymbol {\\theta}, d) \\in \\tilde {D}} \\| g (\\boldsymbol {\\theta}) - d \\| _ {1}. \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            598,
+            306,
+            890,
+            340
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "More details on the training process of the prior are provided in Section 4.2.",
+        "bbox": [
+            498,
+            347,
+            893,
+            376
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2.3 Adaptation using Pose Prior",
+        "text_level": 1,
+        "bbox": [
+            500,
+            395,
+            751,
+            410
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We leverage the trained pose prior  $g$  to regularize the adaptation process by incentivizing the pose estimator to generate pseudo-labels that resemble plausible human poses.",
+        "bbox": [
+            496,
+            417,
+            893,
+            463
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Given the heatmaps  $\\{h_k^t\\}_{k=1}^K$  generated for each keypoint by the student model  $f_{\\mathrm{stu}}^t$  for a target image  $x$ , we calculate the corresponding orientation vectors in a differentiable manner to evaluate the plausibility of the predicted pose using the prior. First, we renormalize each heatmap to a probability distribution via spatial softmax and condense it to a point by computing the spatial expected value of the latter. For computational efficiency, we carry this out in a separable manner along the two spatial dimensions. Namely, assuming  $u = (u_1, u_2)$  to be the two components of each pixel coordinate, we set",
+        "bbox": [
+            496,
+            463,
+            893,
+            628
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nu _ {i} ^ {k} = \\frac {\\sum_ {u _ {i}} u _ {i} e ^ {h _ {k} ^ {t} (u _ {i})}}{\\sum_ {u _ {i}} e ^ {h _ {k} ^ {t} (u _ {i})}}, \\quad h _ {k} ^ {t} (u _ {i}) = \\sum_ {u _ {j}} h _ {k} ^ {t} (u _ {1}, u _ {2}), \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            519,
+            633,
+            890,
+            676
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $i = 1,2$  and  $j = 2,1$  respectively. Next, we use the pose coordinates to determine the orientation vectors between pairs of connected keypoints. Specifically, for every pair  $(a,b)$  of connected keypoints in the human skeleton (denoted by the set  $\\mathcal{E}$ ), we calculate the unit vector  $\\theta$  in the direction from  $u^a$  to  $u^b$ , where  $u^a$  and  $u^b$  are the estimated softmax coordinates of keypoints  $a$  and  $b$ , respectively, i.e.,",
+        "bbox": [
+            496,
+            681,
+            890,
+            787
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\theta = \\frac {u ^ {a} - u ^ {b}}{\\| u ^ {a} - u ^ {b} \\| _ {2}}, \\quad \\forall (a, b) \\in \\mathcal {E}. \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            584,
+            792,
+            890,
+            827
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Finally, we use the prior as a regularization term to minimize the distance of the current pose from our learned manifold,",
+        "bbox": [
+            498,
+            832,
+            890,
+            861
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {p r i o r}} = \\frac {1}{| \\mathcal {B} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {B}} g \\left(T \\left(f _ {\\mathrm {s t u}} ^ {t} (\\boldsymbol {x})\\right)\\right), \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            589,
+            867,
+            890,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "15000",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $T(\\cdot)$  converts the predicted heatmaps to the orientation vector format required as input by the prior.",
+        "bbox": [
+            76,
+            90,
+            468,
+            122
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.3. Overall Adaptation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            132,
+            263,
+            148
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The final training objective for the student model  $f_{\\mathrm{stu}}$  can be expressed as:",
+        "bbox": [
+            76,
+            155,
+            468,
+            186
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\min  _ {f _ {\\mathrm {s t u}}} \\mathcal {L} _ {\\text {o u t}} + \\lambda_ {1} \\mathcal {L} _ {\\text {f e a t}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p r i o r}}. \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            168,
+            199,
+            468,
+            223
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Here,  $\\lambda_{1}$  and  $\\lambda_{2}$  are hyper-parameters that control the influence of feature space adaptation and prior regularization, respectively. The teacher model  $f_{\\mathrm{tea}}$  is updated asynchronously by computing an exponential moving average of the student model weights as shown in Equation 1.",
+        "bbox": [
+            76,
+            234,
+            470,
+            310
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            325,
+            209,
+            343
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we demonstrate POST's ability to adapt a 2D human pose estimation model to a target domain using only unlabeled data from that domain. We conduct experiments on three domain adaptive scenarios and compare with state-of-the-art domain adaptation baselines that utilize source data during adaptation. We also conduct extensive analysis to analyze the contribution and interaction between each component in our framework.",
+        "bbox": [
+            76,
+            349,
+            468,
+            472
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Datasets",
+        "text_level": 1,
+        "bbox": [
+            76,
+            483,
+            179,
+            497
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "SURREAL: SURREAL [43] is a large-scale dataset of synthetically generated images of people rendered from 3D sequences of human motion capture data against indoor backgrounds. It contains over 6 million frames, making it one of the largest and most diverse datasets of its kind.",
+        "bbox": [
+            76,
+            506,
+            470,
+            580
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Human3.6M: Human3.6M [44] is a real-world video dataset captured in indoor environments, comprising 3.6 million frames. The dataset features human subjects performing various actions. In order to reduce redundancy and computational complexity, we down-sampled the videos from 50fps to 10fps as per the approach proposed in [5]. For training, we follow the standard protocol proposed in [6] and use 5 subjects (S1, S5, S6, S7, S8), while the remaining 2 subjects (S9, S11) are reserved for testing.",
+        "bbox": [
+            76,
+            587,
+            468,
+            723
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "LSP: Leeds Sports Pose (LSP) [45] is a real-world dataset that contains 2,000 images with annotated human body joint locations collected from sports activities. The images in LSP are captured in the wild, featuring a wide variety of human poses that are often challenging to detect.",
+        "bbox": [
+            76,
+            728,
+            468,
+            805
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "BRIAR: BRIAR [46] is a cutting-edge biometric dataset featuring a large-scale collection of videos of human subjects captured in extremely challenging conditions. The videos are recorded at varying distances, i.e., close range,  $100\\mathrm{m}$ ,  $200\\mathrm{m}$ ,  $400\\mathrm{m}$ ,  $500\\mathrm{m}$ , and unmanned aerial vehicles (UAV), with each video lasting around 90 seconds. We randomly",
+        "bbox": [
+            76,
+            809,
+            470,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "sample 20 frames from each sequence for each of the 158 subjects for our experiments.",
+        "bbox": [
+            500,
+            90,
+            890,
+            122
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Experiment Protocols",
+        "text_level": 1,
+        "bbox": [
+            500,
+            130,
+            702,
+            146
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Pose estimation model. We adopt the Simple Baseline [18] as our pose estimation model, with the ResNet-101 [47] as the backbone. To train the model, we use the Adam optimizer [48] with a base learning rate of  $1e - 4$ , scheduled to decrease to  $1e - 5$  after 5 epochs and to  $1e - 6$  after 20 epochs. The model is trained for 30 epochs. We use a batch size of 32 and run 500 iterations per epoch. To threshold the model predictions, we set the confidence thresholding ratio  $p$  to 0.5. To augment the images during training, we follow [6] and use rotation, translation, shear, and Gaussian blur. The hyper-parameters  $\\lambda_{1}$  and  $\\lambda_{2}$  are set to  $1e - 3$  and  $1e - 6$ , respectively.",
+        "bbox": [
+            498,
+            152,
+            890,
+            335
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Pose prior model. The training of the parametric prior follows a multi-stage approach that involves using different types of training samples. Initially, we use a combination of manifold poses  $\\theta_{m}$  and non-manifold poses  $\\theta_{nm}$  with a considerable distance from the desired manifold. Over the course of training, we gradually increase the number of non-manifold poses  $\\theta_{nm}$  with a small distance from the manifold. This enables our model to first learn a smooth surface and then gradually incorporate finer details as training progresses. We create these non-manifold poses  $\\theta_{nm}$  by injecting noise into the manifold poses  $\\theta_{m}$  obtained from the auxiliary dataset. Specifically, we sample directional noise from the Von-Mises distribution [49] and add it to the manifold poses in order to obtain the implausible poses.",
+        "bbox": [
+            498,
+            339,
+            893,
+            551
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The architecture for the encoder  $g_{\\mathrm{enc}}$  consists of a 2-layer MLP with an output feature size of  $d = 6$  for each orientation vector, similar to [42]. The distance field network  $g_{\\mathrm{dec}}$  is implemented as a 5-layer MLP. Given its large size and diverse poses, we train the prior for our primary experiments using the SURREAL dataset.",
+        "bbox": [
+            498,
+            551,
+            890,
+            642
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Comparison with Baselines",
+        "text_level": 1,
+        "bbox": [
+            500,
+            650,
+            745,
+            666
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Baselines. We evaluate the performance of our proposed method against several state-of-the-art domain adaptive frameworks. This includes adversarial learning-based feature alignment methods, such as DAN[31], DD[50], and RegDA[5]. Additionally, we consider approaches based on pseudo-labeling, namely CCSSL[38], and UDAPE [6]. It is worth noting that all these methods employ the source data during the adaptation process. To establish a comprehensive performance baseline, we report the results of two additional baselines: Oracle and Source only. The Oracle baseline represents the upper bound of the model's performance, achieved by training the model jointly with target 2D annotations. On the other hand, Source only represents the model's performance when it is directly applied to the target domain without any adaptation.",
+        "bbox": [
+            498,
+            674,
+            893,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "15001",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/8cc6403209faf3336274df573533ff128f80f2dfd107a47bc6f421cda02ba380.jpg",
+        "table_caption": [
+            "Table 1. PCK@0.05 on SURREAL  $\\rightarrow$  LSP. (Best value is in red color, while the second best value is in blue color.)"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>SF</td><td>Sld.</td><td>Elb.</td><td>Wrist</td><td>Hip</td><td>Knee</td><td>Ankle</td><td>Avg.</td></tr><tr><td>Source only</td><td>-</td><td>51.5</td><td>65.0</td><td>62.9</td><td>68.0</td><td>68.7</td><td>67.4</td><td>63.9</td></tr><tr><td>Oracle</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DAN</td><td>X</td><td>52.2</td><td>62.9</td><td>58.9</td><td>71.0</td><td>68.1</td><td>65.1</td><td>63.0</td></tr><tr><td>DD</td><td>X</td><td>28.4</td><td>65.9</td><td>56.8</td><td>75.0</td><td>74.3</td><td>73.9</td><td>62.4</td></tr><tr><td>RegDA</td><td>X</td><td>62.7</td><td>76.7</td><td>71.1</td><td>81.0</td><td>80.3</td><td>75.3</td><td>74.6</td></tr><tr><td>CCSSL</td><td>X</td><td>36.8</td><td>66.3</td><td>63.9</td><td>59.6</td><td>67.3</td><td>70.4</td><td>60.7</td></tr><tr><td>UDAPE</td><td>X</td><td>69.2</td><td>84.9</td><td>83.3</td><td>85.5</td><td>84.7</td><td>84.3</td><td>82.0</td></tr><tr><td>POST</td><td>✓</td><td>66.5</td><td>83.9</td><td>81.0</td><td>84.6</td><td>83.1</td><td>82.6</td><td>80.3</td></tr></table>",
+        "bbox": [
+            78,
+            121,
+            468,
+            268
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/a6fe4e1c5a1647be1f5939c567aa9b00152085a1219b1befdbf622759c82e948.jpg",
+        "table_caption": [
+            "Table 2. PCK@0.05 on SURREAL  $\\rightarrow$  Human3.6M."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>SF</td><td>Sld.</td><td>Elb.</td><td>Wrist</td><td>Hip</td><td>Knee</td><td>Ankle</td><td>Avg.</td></tr><tr><td>Source only</td><td>-</td><td>69.4</td><td>75.4</td><td>66.4</td><td>37.9</td><td>77.3</td><td>77.7</td><td>67.3</td></tr><tr><td>Oracle</td><td>-</td><td>95.3</td><td>91.8</td><td>86.9</td><td>95.6</td><td>94.1</td><td>93.6</td><td>92.9</td></tr><tr><td>DAN</td><td>X</td><td>68.1</td><td>77.5</td><td>62.3</td><td>30.4</td><td>78.4</td><td>79.4</td><td>66.0</td></tr><tr><td>DD</td><td>X</td><td>71.6</td><td>83.3</td><td>75.1</td><td>42.1</td><td>76.2</td><td>76.1</td><td>70.7</td></tr><tr><td>RegDA</td><td>X</td><td>73.3</td><td>86.4</td><td>72.8</td><td>54.8</td><td>82.0</td><td>84.4</td><td>75.6</td></tr><tr><td>CCSSL</td><td>X</td><td>44.3</td><td>68.5</td><td>55.2</td><td>22.2</td><td>62.3</td><td>57.8</td><td>51.7</td></tr><tr><td>UDAPE</td><td>X</td><td>78.1</td><td>89.6</td><td>81.1</td><td>52.6</td><td>85.3</td><td>87.1</td><td>79.0</td></tr><tr><td>POST</td><td>✓</td><td>81.3</td><td>88.5</td><td>77.4</td><td>46.1</td><td>83.4</td><td>83.4</td><td>76.7</td></tr></table>",
+        "bbox": [
+            78,
+            297,
+            468,
+            445
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Metrics. We adopt the evaluation metric of Percentage of Correct Keypoint (PCK) for all experiments and report PCK@0.05 that measures the ratio of correct prediction within a range of  $5\\%$  with respect to the image size.",
+        "bbox": [
+            75,
+            469,
+            468,
+            530
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4. Results",
+        "text_level": 1,
+        "bbox": [
+            76,
+            539,
+            171,
+            554
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Quantitative results. We evaluate POST in two adaptation scenarios: SURREAL  $\\rightarrow$  LSP and SURREAL  $\\rightarrow$  Human3.6M, and report the quantitative results in Table 1 and Table 2, respectively. Specifically, we report the PCK@0.05 on 16 keypoints of the human body, including shoulders (sld.), elbows (elb.), wrists, hips, knees, and ankles. Our method achieves comparable results to many recent approaches that leverage source data for adaptation. Among these methods, UDAPE [6] achieves the highest performance on both cases, with our framework achieving a close second and only 2 percentage points behind on average. Notably, we outperform every other method, including the recently proposed RegDA [5] approach, by a significant margin of up to 5.7 percentage points. It is worth noting that not only do these methods require source data, but also involve additional models such as discriminators or style transfer modules, and unstable adversarial training. In contrast, our framework is lightweight, only involves pseudo-label training, and utilizes a simple prior model that can be trained offline.",
+        "bbox": [
+            75,
+            561,
+            472,
+            851
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Qualitative results. In addition to quantitative results, we also present qualitative results on SURREAL  $\\rightarrow$  Human3.6M in Figure 3 and on SURREAL  $\\rightarrow$  LSP in Figure 4. Also,",
+        "bbox": [
+            75,
+            854,
+            472,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/9ca3dd4ca1a1e28750331b7cb8a5c701cc85b54cdf7f37b743deb37214ce2ba2.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            560,
+            89,
+            650,
+            157
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/6ddd8ed53a6b19b31f5e7fed23c697ea428f3503f8c8f9a11f0efe292d6a862a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            651,
+            89,
+            740,
+            157
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/06b8c705d828873c385eadc00468c9bc2fc87a4a57b2a58acd501b61a009979f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            741,
+            89,
+            831,
+            157
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/55035cceed5ab24a11f808a10e6e62344fe77e57fb571c064ec77e259f690089.jpg",
+        "image_caption": [
+            "Figure 3. Qualitative results on SURREAL  $\\rightarrow$  Human3.6M. We demonstrate sample results on the Human3.6M dataset. For each row, the leftmost shows the Source only prediction, the middle one shows the UDAPE [6] prediction, and the rightmost shows the prediction made POST."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            560,
+            159,
+            650,
+            228
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/bf638cc35de63688c3f809da13d6ecca3d074a6a5776fca8ba763b09bb5c3770.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            651,
+            159,
+            740,
+            228
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/1af0c514de7f1b2d09958562f8961b6c6cb0fc0159bb2d8a0d44fdaad1875b10.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            741,
+            159,
+            831,
+            228
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/3de0508bba699e9fb9ed70599bbfe639183e0de420765c2cb29b22922db41eeb.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            581,
+            321,
+            656,
+            380
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e88c80e6e37804c9a9155a709612e3849a0aaa276c55742ab55e4428bd64f1ea.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            658,
+            321,
+            733,
+            380
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/5800a5eff490007d9f72a9f3305ce806c7c576c1725fde8c20e35ff5cef02c7b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            735,
+            321,
+            810,
+            380
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/26f10a4c3df8781a5005c528fcf7cf87a1ea31a01fba7fe9144c0a13b57a9949.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            583,
+            386,
+            656,
+            443
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/87d6f34fb2bad67ef3179be8f212ff38d3cc26040310bce2f12f61dbaa69e179.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            658,
+            386,
+            733,
+            443
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/90ddfdaed2ee6ce2a5359bb92ed4d83fe00269159ce15ce6affd72347148e4ff.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            735,
+            386,
+            810,
+            443
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/30be3e3f739364914f6ac74358e4da78e6dc356ac72f5439de9a1d16fe5168af.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            583,
+            449,
+            656,
+            506
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/a24aa9696468f014d926a23875096da8c11ed3a8297a696ef0e21d4f5ab516a3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            660,
+            449,
+            733,
+            506
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/578aa8f70007e230899509d88df089d6989ca9c2fdd76f65830f0199b54453d9.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            735,
+            449,
+            810,
+            506
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2b917ce2bae4dc5d9ce79cef983cb4c7e5db69f2b175d78a10a9b48151f391f0.jpg",
+        "image_caption": [
+            "Figure 4. Qualitative results on SURREAL  $\\rightarrow$  LSP. We demonstrate sample results on the LSP dataset. For each row, the left-most shows the Source only prediction, the middle one shows the UDAPE [6] prediction, and the rightmost shows the prediction made POST."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            583,
+            513,
+            656,
+            570
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f45bf2fae60fa16dfdefd095aa7ee8e2514a82b33f48d654214f5c08b987128a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            658,
+            513,
+            733,
+            570
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f932f0b4aa506b25c209f2b2dc069513b86eee3321aba82c71b5d750428d5b9d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            735,
+            513,
+            810,
+            570
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "we present only visual results on the SURREAL  $\\rightarrow$  BRIAR adaptation scenario since pose annotations are absent in the BRIAR dataset. Figure 5 displays the predicted human poses on images taken from six different imaging ranges in the BRIAR dataset. While using the source model directly produces completely inaccurate poses, POST can accurately localize the keypoint locations, even in the presence of occlusions and atmospheric turbulence. Our approach can also accurately reconstruct poses even when the human is imaged from an elevated perspective, resulting in a high camera angle (Figure 5.f). Notably, our approach achieves this without using any source data for adaptation. Our results are comparable to those produced by UDAPE [6], which utilizes source data for adaptation. Additional qualitative results can be found in the supplementary document.",
+        "bbox": [
+            496,
+            672,
+            895,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "15002",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/b488637a570423b0c61b2b920d409367b8d37de12030b61d1cfa0a0d86a0fada.jpg",
+        "image_caption": [
+            "a)  $100\\mathrm{m}$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            163,
+            93,
+            464,
+            172
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6699f621e33ec7924a2e3e22e59af8268b17dc67c2b86f498bef502438207a34.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            93,
+            807,
+            172
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/7a74b7dbf63c3c52242e7f9de6d519c459d882828be198e5e53081e5072e7488.jpg",
+        "image_caption": [
+            "b)  $200\\mathrm{m}$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            163,
+            198,
+            464,
+            277
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/0e423e1f372fc6d2f448b59ecf1bf54c35a70d65eccaf60269c757cd1fe5f1bd.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            199,
+            807,
+            277
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/a2d47abb81dece636654ac0a86b4cca9356074298df4759fd5136a58f4212362.jpg",
+        "image_caption": [
+            "c)  $400\\mathrm{m}$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            163,
+            304,
+            464,
+            383
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/4e89899667fa45cde15a39a23c01f71b2bc367063323bb7caffea4f57f7f63e0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            508,
+            305,
+            807,
+            383
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/e1eb07fe7e882dfb82fce907089730185aeea9907de1c34c662f2eadacfe2082.jpg",
+        "image_caption": [
+            "d)  $500\\mathrm{m}$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            163,
+            410,
+            464,
+            488
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/117eb71616e26b7bb49299de2a2aa1031bf8907490a22a21fb563ffe508a9f3d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            508,
+            411,
+            807,
+            488
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ae26e585dc79fcadbe7beef7631175128cb92f7806c412598d808f049f8df802.jpg",
+        "image_caption": [
+            "e) close range"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            165,
+            516,
+            464,
+            594
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/20c06172a974e6383a36ad30d312c3833e0cf846cc3308210aceb24e8e687611.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            508,
+            516,
+            807,
+            594
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/f2aa48b85552239f56e0f04de58b137b6f5e654cda8884c106ba54f4b6e5fe1a.jpg",
+        "image_caption": [
+            "f) uav"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            165,
+            622,
+            464,
+            700
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/1c048a6e1e9309091c8b02dde1f23d8ed93272664bc72a57d8b8a03fc087ac8a.jpg",
+        "image_caption": [
+            "Figure 5. Qualitative results on SURREAL  $\\rightarrow$  BRIAR. We demonstrate sample results on BRIAR at all ranges. For each range, we display three images: the leftmost shows the Source only prediction, the middle one shows the UDAPE [6] prediction, and the rightmost shows the prediction using POST. Although POST does not use source data for adaptation, it is able to match the predictions produced by UDAPE, which uses source data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            508,
+            622,
+            807,
+            700
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.5. Ablation Studies",
+        "text_level": 1,
+        "bbox": [
+            76,
+            801,
+            240,
+            816
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of auxiliary dataset. We conduct experiments to evaluate the impact of the choice of the auxiliary dataset (used to train the prior) on downstream adaptation. The results are presented in Table 3. Our findings indicate that",
+        "bbox": [
+            75,
+            839,
+            470,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "POST is robust to the choice of the auxiliary dataset, with performance differences of  $\\sim 0.5$  percentage points.",
+        "bbox": [
+            498,
+            803,
+            890,
+            835
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of loss terms. We conduct an experiment to evaluate the performance of each component of our framework. The results on the SURREAL  $\\rightarrow$  Human3.6M adaptation scenario are presented in Table 4. Our findings indicate that in addi",
+        "bbox": [
+            498,
+            839,
+            893,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "15003",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/f33a9eab7d6b22371583b8a33afe270a2a874518ff5094082116c094d2114b6f.jpg",
+        "table_caption": [
+            "Table 3. Effect of auxiliary dataset. We evaluate the effect of auxiliary dataset on downstream adaptation tasks."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"8\">SURREAL→HUMAN3.6M</td></tr><tr><td>Aux. dataset</td><td>Sld.</td><td>Elb.</td><td>Wrist</td><td>Hip</td><td>Knee</td><td>Ankle</td><td>Avg.</td></tr><tr><td>SURREAL</td><td>81.3</td><td>88.5</td><td>77.4</td><td>46.1</td><td>83.4</td><td>83.4</td><td>76.7</td></tr><tr><td>Human3.6M</td><td>80.9</td><td>88.0</td><td>77.2</td><td>45.0</td><td>83.1</td><td>82.8</td><td>76.2</td></tr><tr><td colspan=\"8\">SURREAL→LSP</td></tr><tr><td>Aux. dataset</td><td>Sld.</td><td>Elb.</td><td>Wrist</td><td>Hip</td><td>Knee</td><td>Ankle</td><td>Avg.</td></tr><tr><td>SURREAL</td><td>66.5</td><td>83.9</td><td>81.0</td><td>84.6</td><td>83.1</td><td>82.6</td><td>80.3</td></tr><tr><td>Human3.6M</td><td>66.1</td><td>83.6</td><td>80.7</td><td>84.4</td><td>83.1</td><td>82.5</td><td>80.1</td></tr></table>",
+        "bbox": [
+            76,
+            119,
+            470,
+            267
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/8647424d859b191fa3cf852da3d10e0050e9733572ba9d8ee4f46abe481e298d.jpg",
+        "table_caption": [
+            "Table 4. Effect of each loss term. We evaluate the contribution of each loss term on SURREAL  $\\rightarrow$  Human3.6M."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Lout</td><td>Lfeat</td><td>Lprior</td><td>Sld.</td><td>Elb.</td><td>Wrist</td><td>Hip</td><td>Knee</td><td>Ankle</td><td>Avg.</td></tr><tr><td>X</td><td>X</td><td>X</td><td>69.4</td><td>75.4</td><td>66.4</td><td>37.9</td><td>77.3</td><td>77.7</td><td>67.3</td></tr><tr><td>✓</td><td>X</td><td>X</td><td>77.9</td><td>86.7</td><td>73.7</td><td>38.8</td><td>83.0</td><td>84.3</td><td>74.1</td></tr><tr><td>✓</td><td>✓</td><td>X</td><td>81.7</td><td>87.1</td><td>75.2</td><td>44.3</td><td>82.3</td><td>82.2</td><td>75.5</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>81.3</td><td>88.5</td><td>77.4</td><td>46.1</td><td>83.4</td><td>83.4</td><td>76.7</td></tr></table>",
+        "bbox": [
+            78,
+            313,
+            468,
+            393
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "tion to prediction space adaptation, feature space adaptation also plays a crucial role in enabling effective unsupervised learning from pseudo-labels. Moreover, we observed that the human pose prior brings additional improvements, thus validating our hypothesis that noisy pose pseudo-labels can be refined implicitly by a prior in the absence of source data. Overall, our results demonstrate the effectiveness of our framework and the importance of each of its components in achieving state-of-the-art performance.",
+        "bbox": [
+            75,
+            422,
+            468,
+            556
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Cross-dataset performance of prior. We evaluate the ability of the learned prior to handle distribution shift separately from the adaptation performance. Specifically, we assess the robustness of the prior by computing pose scores across datasets, and the results are presented in Figure 6. The plot demonstrates that our prior is effective in scoring plausible (real) poses with low scores and scoring implausible poses with higher scores. Note that the bell curve shape of the noisy pose scores is due to the Von-Mises noise added to create the noisy poses.",
+        "bbox": [
+            75,
+            564,
+            468,
+            715
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Effect of thresholding. We analyze the impact of the pseudo-label threshold  $\\tau$  on the adaptation performance in Table 5. The results on SURREAL  $\\rightarrow$  Human3.6M reveal that as we increase this ratio, the performance gradually decreases. This can be attributed to the fact that higher thresholding ratios tend to include lower confident predictions as pseudo-labels, which can negatively impact adaptation.",
+        "bbox": [
+            75,
+            720,
+            468,
+            827
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            76,
+            844,
+            194,
+            859
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We address the problem of adapting a pre-trained 2D human pose estimator to a new target domain with unlabeled",
+        "bbox": [
+            76,
+            869,
+            468,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/b26de1c511bd6d9ad94438716cb834d7a7533ad29ab73caad6ac905ee06cbb34.jpg",
+        "table_caption": [
+            "Table 5. Effect of  $\\tau$ . We evaluate adaptation performance on SURREAL  $\\rightarrow$  Human3.6M as  $\\tau$  is varied."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>τ = 0.1</td><td>τ = 0.3</td><td>τ = 0.5</td><td>τ = 0.7</td><td>τ = 0.9</td></tr><tr><td>76.4</td><td>76.5</td><td>76.7</td><td>75.3</td><td>73.8</td></tr></table>",
+        "bbox": [
+            540,
+            118,
+            851,
+            284
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/fe682b04ab13c49f01819e55fd9998b7720e2fa39836b2a08ac6227d112420df.jpg",
+        "image_caption": [
+            "(a) Human3.6M  $\\rightarrow$  SURREAL",
+            "(b) SURREAL  $\\rightarrow$  Human3.6M",
+            "Figure 6. Cross-dataset prior transfer. We plot the histogram of scores predicted by the prior. The prior can clearly demarcate plausible poses from implausible poses across datasets."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            540,
+            308,
+            846,
+            417
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "target data. To this end, we propose a self-training algorithm, POST, that leverages a Mean Teacher framework to enforce both prediction-level and feature-level consistency between a pair of student and teacher models. Our approach incorporates a human pose prior that captures the manifold of possible poses in a high-dimensional space, which helps to refine noisy pseudo-labels generated during self-training. We evaluate our method on three challenging adaptation scenarios and show that it achieves competitive performance compared to existing UDA methods that have access to source data.",
+        "bbox": [
+            496,
+            507,
+            893,
+            657
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements. This research is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via [2022-21102100007]. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein.",
+        "bbox": [
+            496,
+            666,
+            893,
+            818
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            500,
+            833,
+            596,
+            848
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "ref_text",
+        "text": "[1] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, 2019. 1, 2",
+        "bbox": [
+            506,
+            858,
+            893,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "15004",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[2] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 1",
+            "[3] Hai Liu, Tingting Liu, Zhaoli Zhang, Arun Kumar Sangaiah, Bing Yang, and Youfu Li. Arhpe: Asymmetric relation-aware representation learning for head pose estimation in industrial human-computer interaction. IEEE Transactions on Industrial Informatics, 18(10):7107-7117, 2022. 1",
+            "[4] Xiang Li, Yasushi Makihara, Chi Xu, Yasushi Yagi, Shiqi Yu, and Mingwu Ren. End-to-end model-based gait recognition. In ACCV, 2020. 1",
+            "[5] Junguang Jiang, Yifei Ji, Ximei Wang, Yufeng Liu, Jianmin Wang, and Mingsheng Long. Regressive domain adaptation for unsupervised keypoint detection. In CVPR, 2021. 1, 3, 6, 7",
+            "[6] Donghyun Kim, Kaihong Wang, Kate Saenko, Margrit Betke, and Stan Sclaroff. A unified framework for domain adaptive pose estimation. In ECCV, 2022. 1, 2, 3, 6, 7, 8",
+            "[7] Paul M Schwartz and Daniel J Solove. The pii problem: Privacy and a new concept of personally identifiable information. NYUL rev., 86:1814, 2011. 1",
+            "[8] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 2",
+            "[9] Sk Miraj Ahmed, Dripta S Raychaudhuri, Sujoy Paul, Samet Oymak, and Amit K Roy-Chowdhury. Unsupervised multisource domain adaptation without access to source data. In CVPR, 2021. 2",
+            "[10] Sujoy Paul, Ansh Khurana, and Gaurav Aggarwal. Unsupervised adaptation of semantic segmentation models without source data. arXiv preprint arXiv:2112.02359, 2021. 2",
+            "[11] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS, 2017. 2, 3",
+            "[12] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021. 2, 4",
+            "[13] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In ECCV, 2022. 2, 5",
+            "[14] Hao-Shu Fang, Shuqin Xie, Yu-Wing Tai, and Cewu Lu. *Rmpe: Regional multi-person pose estimation.* In ICCV, 2017. 2",
+            "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2",
+            "[16] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR, 2016. 2",
+            "[17] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In ECCV, 2016. 2, 4",
+            "[18] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In ECCV, 2018. 2, 4, 6"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[19] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multiperson pose estimation. In CVPR, 2018. 2",
+            "[20] Leonid Pishchulin, Eldar Insafutdinov, Siyu Tang, Bjoern Andres, Mykhaylo Andriluka, Peter V Gehler, and Bernt Schiele. Deepcut: Joint subset partition and labeling for multi-person pose estimation. In CVPR, 2016. 2",
+            "[21] Eldar Insafutdinov, Leonid Pishchulin, Bjoern Andres, Mykhaylo Andriluka, and Bernt Schiele. Deepercut: A deeper, stronger, and faster multi-person pose estimation model. In ECCV, 2016. 2",
+            "[22] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: realtime multi-person 2d pose estimation using part affinity fields. IEEE transactions on Pattern Analysis and Machine Intelligence, 43(1):172-186, 2019. 2",
+            "[23] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In CVPR, 2019. 2",
+            "[24] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In CVPR, 2020. 2",
+            "[25] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. arXiv preprint arXiv:2104.02300, 2021. 2",
+            "[26] Sheng Jin, Wentao Liu, Enze Xie, Wenhai Wang, Chen Qian, Wanli Ouyang, and Ping Luo. Differentiable hierarchical graph grouping for multi-person pose estimation. In ECCV, 2020. 2",
+            "[27] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In CVPR, 2017. 2",
+            "[28] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kihyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In CVPR, 2018. 2",
+            "[29] Han-Kai Hsu, Chun-Han Yao, Yi-Hsuan Tsai, Wei-Chih Hung, Hung-Yu Tseng, Maneesh Singh, and Ming-Hsuan Yang. Progressive domain adaptation for object detection. In WACV, 2020. 2",
+            "[30] Dripta S Raychaudhuri, Sujoy Paul, Jeroen Vanbaar, and Amit K Roy-Chowdhury. Cross-domain imitation from observations. In ICML, 2021. 2",
+            "[31] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, 2015. 2, 6",
+            "[32] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. In JMLR, 2016. 2"
+        ],
+        "bbox": [
+            501,
+            90,
+            893,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "15005",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[33] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei A. Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In ICML, 2018. 2",
+            "[34] Luan Tran, Kihyuk Sohn, Xiang Yu, Xiaoming Liu, and Manmohan Chandraker. Gotta adapt 'em all: Joint pixel and feature-level domain adaptation for recognition in the wild. In CVPR, 2019. 2",
+            "[35] Hao-Wei Yeh, Baoyao Yang, Pong C Yuen, and Tatsuya Harada. Sofa: Source-data-free feature alignment for unsupervised domain adaptation. In WACV, 2021. 2",
+            "[36] Vikash Kumar, Rohit Lal, Himanshu Patil, and Anirban Chakraborty. Conmix for source-free single and multi-target domain adaptation. In WACV, 2023. 2",
+            "[37] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In ICCV, 2021. 2",
+            "[38] Jiteng Mu, Weichao Qiu, Gregory D Hager, and Alan L Yuille. Learning from synthetic animals. In CVPR, 2020. 3, 6",
+            "[39] Chen Li and Gim Hee Lee. From synthetic to real: Unsupervised domain adaptation for animal pose estimation. In CVPR, 2021. 3",
+            "[40] Jonathan J. Tompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. NeurIPS, 2014. 3",
+            "[41] Luan Tran, Kihyuk Sohn, Xiang Yu, Xiaoming Liu, and Manmohan Chandraker. Gotta adapt' em all: Joint pixel and feature-level domain adaptation for recognition in the wild. In CVPR, 2019. 4",
+            "[42] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. Leap: Learning articulated occupancy of people. In CVPR, 2021. 5, 6",
+            "[43] Gül Varol, Javier Romero, Xavier Martin, Naureen Mahmood, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning from synthetic humans. In CVPR, 2017. 6",
+            "[44] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2013. 6",
+            "[45] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In BMVC, 2010. 6",
+            "[46] David Cornett, Joel Brogan, Nell Barber, Deniz Aykac, Seth Baird, Nicholas Burchfield, Carl Dukes, Andrew Duncan, Regina Ferrell, Jim Goddard, Gavin Jager, Matthew Larson, Bart Murphy, Christi Johnson, Ian Shelley, Nisha Srinivas, Brandon Stockwell, Leanne Thompson, Matthew Yohe, Robert Zhang, Scott Dolvin, Hector J. Santos-Villalobos, and David S. Bolme. Expanding accurate person recognition to new altitudes and ranges: The briar dataset. In WACV Workshops, 2023. 6"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            900
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[47] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6",
+            "[48] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6",
+            "[49] Riccardo Gatto and Sreenivasa Rao Jammalamadaka. The generalized von mises distribution. Statistical Methodology, 4(3):341-353, 2007. 6",
+            "[50] Yuchen Zhang, Tianle Liu, Mingsheng Long, and Michael Jordan. Bridging theory and algorithm for domain adaptation. In ICML, 2019. 6"
+        ],
+        "bbox": [
+            501,
+            90,
+            893,
+            272
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "15006",
+        "bbox": [
+            480,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2023/Prior-guided Source-free Domain Adaptation for Human Pose Estimation/67d12eb1-f1e4-4b90-aba2-4771e3bf7c39_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dec48c982809b54256b956a66444acd290c0921aa7a87eb2c1b87bba95c2a845
+size 2492773

2023/Prior-guided Source-free Domain Adaptation for Human Pose Estimation/full.md

@@ -0,0 +1,392 @@
+# Prior-guided Source-free Domain Adaptation for Human Pose Estimation
+
+Dripta S. Raychaudhuri $^{1,2*}$  Calvin-Khang Ta $^{1}$  Arindam Dutta $^{1}$  Rohit Lal $^{1}$  Amit K. Roy-Chowdhury $^{1}$ $^{1}$ University of California, Riverside  $^{2}$ AWS AI Labs
+
+{drayc001@, cta003@, adutt020@, rla1011@, amertc@ece.}ucr.edu
+
+# Abstract
+
+Domain adaptation methods for 2D human pose estimation typically require continuous access to the source data during adaptation, which can be challenging due to privacy, memory, or computational constraints. To address this limitation, we focus on the task of source-free domain adaptation for pose estimation, where a source model must adapt to a new target domain using only unlabeled target data. Although recent advances have introduced source-free methods for classification tasks, extending them to the regression task of pose estimation is non-trivial. In this paper, we present Prior-guided Self-training (POST), a pseudolabeling approach that builds on the popular Mean Teacher framework to compensate for the distribution shift. POST leverages prediction-level and feature-level consistency between a student and teacher model against certain image transformations. In the absence of source data, POST utilizes a human pose prior that regularizes the adaptation process by directing the model to generate more accurate and anatomically plausible pose pseudo-labels. Despite being simple and intuitive, our framework can deliver significant performance gains compared to applying the source model directly to the target data, as demonstrated in our extensive experiments and ablation studies. In fact, our approach achieves comparable performance to recent state-of-the-art methods that use source data for adaptation.
+
+# 1. Introduction
+
+Human pose estimation is a fundamental task in computer vision that involves determining the precise locations of keypoints, such as joints, on a human body in an image or video [1]. The growing need for pose estimation in various applications such as action recognition [2], human-computer interaction [3], and video surveillance [4] has driven the rapid development of highly accurate deep learning tech
+
+![](images/f269fd9a6d1df8c4a7d32d89b1b306e6b1a67e8a5f63ca42de4197fc02660e7b.jpg)  
+Figure 1. Problem setup. Existing UDA methods for pose estimation rely on a labeled source dataset while adapting to an unlabeled target dataset. However, privacy concerns surrounding the use of personally identifiable information in these labeled datasets, as well as the significant storage and computational requirements, can limit access to such data. Hence, our work focuses on source-free UDA of pose estimation models.
+
+niques. However, the challenge of obtaining large annotated datasets for training, compounded with the susceptibility to a performance decline in the face of data distribution shifts still poses limitations for current pose estimation models.
+
+To overcome these limitations, recent studies have focused on unsupervised domain adaptation (UDA) of pose estimators [5, 6]. UDA allows for transferring a pose estimation model trained on a source domain, where labeled data is available, to a target domain where labeled data is unavailable. Despite improved and robust pose estimation, the requirement of simultaneous access to both source and target domains during adaptation hinders real-world implementation. For instance, the labeled source data may not be accessible post-deployment due to privacy or proprietary issues. This is particularly relevant for human pose datasets, which contain personally identifiable information (PII) [7]. Furthermore, adaptation using the entire source data might
+
+be infeasible due to both memory and computational constraints. In light of these issues, we focus on source-free UDA of human pose estimation models.
+
+Concretely, our objective is to adapt a 2D human pose estimation model to a new target domain utilizing only its trained parameters and unlabeled target data. This presents a major challenge as the absence of source data for regularization can cause catastrophic forgetting. While recent advances have introduced methods to tackle this issue in classification tasks [8, 9, 10], extending them to the regression task of pose estimation is non-trivial. To address this challenge, we introduce Prior-guided Self-training (POST), a self-training regression framework that employs a human pose prior to effectively guide the adaptation process in the absence of source data. An overview of our problem setup is shown in Figure 1.
+
+Our approach builds on the Mean Teacher [11] framework, which uses consistency in the prediction space of a student and teacher model to produce trustworthy pseudo-labels and learn from the unlabeled target domain. To achieve this, we create two augmented views of each target image, varying in scale, spatial context, and color statistics. Aligned pose predictions from both models in both views are then obtained, and consistency between the predictions across the different views is encouraged to facilitate prediction space adaptation. However, our empirical results show that relying solely on consistency in the output space is insufficient when supervision from the source data is lacking. To address this, we also introduce feature space adaptation, which aims to encourage consistency across features extracted from the two separate views. We adopt the Barlow Twins [12] approach to accomplish this. Specifically, we seek to make the cross-correlation matrix calculated from a pair of feature embeddings as close to the identity matrix as possible.
+
+In addition to the adaptation across both outputs and features, we employ a human pose prior that models the full manifold of plausible poses in some high-dimensional pose space to refine possible noisy pseudo-labels that may arise during self-training. The plausible poses are represented as points on the manifold surface, with zero distance from it, while non-plausible poses are located outside the surface, with a non-zero distance from it. This manifold is learned using a high-dimensional neural field, similar to Pose-NDF [13]. The pose prior acts as a strong regularizer, directing the model to generate more accurate pose pseudolabels on the target data and leading to improved adaptation. The learning of this prior requires an auxiliary dataset of plausible human poses, but this does not compromise the privacy aspect of our framework as the prior does not make use of RGB images. In addition, it is worth noting that the prior can be trained offline, separately from the adaptation process. This not only saves computational resources but also reduces the amount of storage required. Compared to storing entire
+
+images, it is much more efficient to store pose coordinates, which requires approximately  $3000 \times$  less memory.
+
+Main contributions. To summarize, our primary contributions are as follows:
+
+- We address the problem of adapting a human pose estimation model to a target domain consisting of unlabeled data, without access to the original source dataset. This ameliorates the privacy concern associated with the current domain adaptive pose estimation methods.  
+- We introduce Prior-guided Self-training (POST), a simple source-free unsupervised adaptation algorithm. POST leverages both prediction-level and feature-level consistency, in addition to a human pose prior, to drive self-training for improved adaptation to the target domain.  
+- We evaluate our method qualitatively and quantitatively on three challenging domain adaptive scenarios, demonstrating comparable performance to existing UDA methods that have access to the source data.
+
+# 2. Related Works
+
+Pose Estimation. 2D human pose estimation aims to locate human anatomical keypoints, such as the elbow and knee. Prior works can be categorized into two primary frameworks: the top-down framework and the bottom-up framework. Top-down methods [14, 15, 16, 17, 18, 19, 1] first detect each person from the image and then perform single-person pose estimation on each bounding box independently. On the other hand, bottom-up methods [20, 21, 22, 23, 24, 25, 26] predict keypoints of each person directly in an end-to-end manner. Typical bottom-up methods consist of two steps: predicting keypoint heatmaps and grouping the detected keypoints into separate poses. In this work, we focus on the bottom-up framework for efficiency purposes and adopt the Simple Baseline [18] architecture following [6] to ensure fair comparisons with prior domain adaptation algorithms.
+
+Unsupervised Domain Adaptation. UDA methods have been extensively applied to a broad range of computer vision tasks, including image classification [27], semantic segmentation [28], object detection [29], and reinforcement learning [30] to tackle the issue of data distribution shift. Most approaches aim to align the source and target data distributions through techniques such as maximum mean discrepancy [31] and adversarial learning [32, 27]. Another line of research utilizes image translation methods to perform adaptation by transforming the source images into the target domain [33, 34]. More recently, there has been a surge of interest in adaptation using only a pre-trained source model due to privacy and memory storage concerns related to the source data. These include techniques such as information maximization [8, 9], pseudo-labeling [35, 36] and self-supervision [37]. Compared to other tasks, domain adaptation for regression tasks, such as pose estimation, remains
+
+relatively unexplored.
+
+Domain Adaptive Pose Estimation. UDA methods for pose estimation have explored various techniques for overcoming the domain gap, including adversarial feature alignment and pseudo-labeling. RegDA [5] estimates the domain discrepancy by evaluating false predictions on the target domain and minimizes it. Mu et al. [38] proposed consistency regularization with respect to transformations and temporal consistency learning within a video. Li et al. [39] proposed a refinement module and a self-feedback loop to obtain reliable pseudo-labels. Recently, Kim et al. [6] introduced a unified framework for both human and animal keypoint detection, which aligns representations using input-level and output-level cues. Typically, these methods require access to the source data, which may raise data privacy, memory and computation concerns. In contrast, our method addresses the domain adaptation problem in a source-free setting.
+
+# 3. Prior-guided Self-training
+
+Our work investigates source-free domain adaptation for 2D human pose estimation. In the pose estimation task, given an input image  $\pmb{x} \in \mathbb{R}^{H \times W \times 3}$ , the goal is to predict the corresponding  $y \in \mathbb{R}^{K \times 2}$  representing the 2D coordinates of  $K$  keypoints using a pose regression model  $f$ . In this paper, we assume access to a pre-trained model, denoted by  $f_{S}$ , as well as  $N$  unlabeled images  $\mathcal{D} = \{\pmb{x}_i\}_{i=1}^N$  from a target domain  $\mathcal{T}$ . Our goal is to adapt the source model to the target such that it performs better on images drawn from the target distribution than when directly using the source model on the target images.
+
+Overview. In the absence of source data, we propose to use self-training to adapt the source pose estimation model to the target domain. However, self-training methods are prone to error accumulation, particularly when labeled data is absent to act as regularization. Hence, we introduce POST, an enhanced self-training strategy that employs three essential ideas to prevent such errors:
+
+1. A weight-averaged teacher model is used to generate the pseudo-labels for self-training. This ensures better retention of the source knowledge within the teacher model by reducing the effect of updating the weights via noisy pseudo-labels (Section 3.1.1).  
+2. In addition to adaptation over the output space via pseudolabels, the model is adapted in the feature space as well. For each target image, two aligned predictive views are generated via data augmentation, and consistency across features extracted from the two separate views is encouraged via a contrastive learning strategy (Section 3.1.2).  
+3. A human pose prior is used to regularize the adaptation by directing the model to generate more accurate and anatomically plausible pose pseudo-labels (Section 3.2). An overview of our framework is presented in Figure 2.
+
+# 3.1. Self-training via Mean Teacher
+
+Motivated by research suggesting that weight-averaged models over training steps tend to perform better than the final model [6, 39], we utilize the Mean Teacher framework [11] to generate pseudo-labels for self-training. The framework involves creating two identical models, a teacher model  $f_{\mathrm{tea}}$  and a student model  $f_{\mathrm{stu}}$ , both of which are initialized with the pre-trained network  $f_{S}$  at time step  $t = 0$ . At each subsequent time step  $t$ , the student model parameters  $\theta$  are updated by backpropagating the supervisory signals provided by the teacher model. The parameters of the teacher model  $\tilde{\theta}$  are updated via an exponential moving average (EMA) of the student model parameters:
+
+$$
+\tilde {\theta} _ {t} = \alpha \tilde {\theta} _ {t - 1} + (1 - \alpha) \theta_ {t}, \tag {1}
+$$
+
+where  $\alpha$  denotes the smoothing coefficient which is set to 0.999 by default. The EMA update prevents the teacher model from overfitting to noisy pseudo-labels during the initial rounds of self-training, thereby, preserving the source knowledge. This is especially advantageous in our scenario where source data is unavailable to regularize the adaptation.
+
+In the following sections, we demonstrate how to adapt  $f_{\mathrm{stu}}$  using supervisory signals from  $f_{\mathrm{tea}}$  on both the feature space and the pose prediction space.
+
+# 3.1.1 Prediction Space Adaptation
+
+At each time step  $t$ , we apply two different data augmentations,  $\mathcal{A}_1$  and  $\mathcal{A}_2$ , to a target image  $\pmb{x}$  to generate two views. We then obtain the keypoint heatmap corresponding to the first transformed image  $\tilde{h}^t = f_{\mathrm{tea}}^t (\mathcal{A}_1(\pmb {x}))$ . Here  $\tilde{h}^t\in \mathbb{R}^{K\times H'\times W'}$  denotes the spatial likelihood of the  $K$  different keypoints on each channel. The pseudolabel from the teacher model is generated by obtaining the coordinates which produce the maximum activations  $\hat{y}^t = \operatorname{argmax}_u\tilde{h}_{[:,u]}^t$ .
+
+To reduce the influence of erroneous pseudo-labels in our training process, we utilize a confidence threshold to discard potentially unreliable labels. Specifically, we only retain the keypoint activations with the top  $p\%$  maximum values among all activations and discard the rest. We set the threshold  $\tau$  accordingly to reflect this, thereby ensuring that only the most confident labels are used for training.
+
+Following prior work on supervised pose estimation, we first convert the pseudo-labels to normalized Gaussian heatmaps [40] and then use the mean squared error (MSE) loss to update the student model over batches of target images sampled from  $\mathcal{D}$ :
+
+$$
+\mathcal {L} _ {\text {o u t}} = \frac {1}{| \mathcal {B} |} \sum_ {\mathbf {x} \in \mathcal {B}} \sum_ {k = 1} ^ {K} \mathbb {1} \left(\hat {h} _ {k} ^ {t} \geq \tau\right) \| \tilde {\mathcal {A}} _ {1} \left(\hat {h} _ {k} ^ {t}\right) - \tilde {\mathcal {A}} _ {2} \left(h _ {k} ^ {t}\right) \| _ {2}. \tag {2}
+$$
+
+![](images/c556debaba324cdf504fd1ec865cb5e134057372a5fa29e9517f0c4813dae355.jpg)  
+Figure 2. Framework overview. Our approach builds on the Mean Teacher framework and performs adaptation both in the pose prediction space using  $\mathcal{L}_{\mathrm{out}}$ , and the feature space using  $\mathcal{L}_{\mathrm{feat}}$ . This is supplemented by a human pose prior  $g$  that scores the predicted pseudo-labels in terms of plausibility. These scores are used to regularize the adaptation process in the absence of labeled source data via  $\mathcal{L}_{\mathrm{prior}}$ . The student model  $f_{\mathrm{stu}}$  is trained by the combination of the three losses, while the teacher model  $f_{\mathrm{tea}}$  is updated with the exponential moving average (EMA) of the weights of the student model.
+
+We denote the inverse functions of the chosen augmentations  $\mathcal{A}_1$  and  $\mathcal{A}_2$  as  $\tilde{\mathcal{A}}_1$  and  $\tilde{\mathcal{A}}_2$ , respectively. The heatmap predicted by the student model for the  $k$ -th keypoint is represented by  $h_k^t = f_{\mathrm{stu}}^t (\mathcal{A}_2(\pmb {x}))^k$ , while the heatmap generated from the pseudo-labels predicted by the teacher model is represented by  $\hat{h}_k^t = L(\hat{y}^t)^k$ . Here,  $L(\cdot)$  represents the heatmap generating function and  $\mathcal{B}$  denotes a batch of target images.
+
+# 3.1.2 Feature Apace Adaptation
+
+Pose estimation models primarily rely on a high-to-low-resolution feature encoder to generate low-resolution representations, which are then used to recover high-resolution pose heatmaps [18]. Since the upsampling process is inherently noisy, providing intermediate supervision to explicitly adapt the features, in addition to adapting the output pose keypoints as shown in the previous section, can be beneficial [17]. While previous work has demonstrated the benefits of joint adaptation across the output and feature space via adversarial learning in tasks such as semantic segmentation [41], this has been exclusively focused on scenarios where source data is available. Here, we propose an alternative way to accomplish feature space adaptation in the absence of source data via contrastive learning.
+
+We begin by creating two different views of each target image  $\mathbf{x}$  using a pair of sampled augmentations  $\mathcal{A}_1$  and  $\mathcal{A}_2$ , as previously shown. Next, considering the pose estimation model as a composition of a feature encoder and an output regressor, i.e.,  $f = \mathrm{Dec} \circ \mathrm{Enc}$ , we extract
+
+the augmentation reversed feature maps from the teacher model  $\tilde{z} = \tilde{\mathcal{A}}_1(\mathrm{Enc}_{\mathrm{tea}}(\mathcal{A}_1(\boldsymbol{x})))$  and the student model  $z = \tilde{\mathcal{A}}_2(\mathrm{Enc}_{\mathrm{stu}}(\mathcal{A}_2(\boldsymbol{x})))$ . We extract pairs of features for every image in a batch, pool them along the spatial dimensions, and then normalize them along the batch dimension to ensure that each covariate has a mean of 0 over the batch. For simplicity, we overload  $z,\tilde{z}$  to represent the normalized features, and drop the time index  $t$ .
+
+We utilize feature-level consistency between the different views in order to accomplish feature space adaptation. This is achieved via a contrastive learning strategy which encourages the cross-correlation matrix between the outputs of the two networks to be as close to the identity matrix as possible [12],
+
+$$
+\mathcal {L} _ {\text {f e a t}} = \sum_ {i} (1 - C _ {i i}) ^ {2} + \gamma \sum_ {i} \sum_ {j \neq i} C _ {i j} ^ {2}. \tag {3}
+$$
+
+We define  $\gamma$  as a positive constant that balances the importance of the first and second terms of the loss function.  $C$  represents the cross-correlation matrix computed between the outputs of the student and teacher networks along the batch dimension:
+
+$$
+C _ {i j} = \frac {\sum_ {b = 1} ^ {| \mathcal {B} |} \tilde {z} _ {b , i} z _ {b , j}}{\sqrt {\sum_ {b = 1} ^ {| \mathcal {B} |} \left(\tilde {z} _ {b , i}\right) ^ {2}} \sqrt {\sum_ {b = 1} ^ {| \mathcal {B} |} \left(z _ {b , j}\right) ^ {2}}}. \tag {4}
+$$
+
+Here,  $b$  represents the batch index, while  $i$  and  $j$  index the feature dimensions of the networks' outputs. We set  $\gamma = 5e - 3$  following [12].
+
+The first term of  $\mathcal{L}_{\mathrm{feat}}$  encourages the consistency of pose features within the same image by equating the diagonal elements of the cross-correlation matrix to 1, effectively reversing the effects of augmentations. In contrast, the second term aims to decorrelate the different feature dimensions of the embedding by forcing the off-diagonal elements of the cross-correlation matrix to 0.
+
+# 3.2. Regularization via Pose Prior
+
+Enhancing the performance of the pose estimation model on the target domain through joint adaptation over the output and feature space is undoubtedly valuable. However, this approach has its limitations, as it relies solely on general domain adaptation principles and overlooks the rich structural priors associated with human poses. To address this issue, we propose incorporating a parametric human pose prior to better adapt the pose estimation model to the target domain.
+
+# 3.2.1 2D Human Pose Prior
+
+Building upon the work of [13], we propose a human pose prior modeled as a manifold consisting of plausible 2D poses. To represent the 2D poses while ignoring aspects such as size and scale, we use a set of 2D orientation vectors that connect pairs of joints in the human skeleton, denoted by  $\mathcal{G} = \{\pmb {\theta} = (\theta_1,\dots ,\theta_L)\mid \theta_l\in \mathbb{R}^2,\| \theta_l\| _2 = 1\forall l\in [L]\}$ . We assume that plausible 2D human poses lie on a manifold embedded in this pose space  $\mathcal{G}$ . We use a function  $g:\mathcal{G}\to$ $\mathbb{R}^+$ , which maps a pose to a non-negative scalar, to represent the manifold of plausible poses as the zero-level set:
+
+$$
+\mathcal {P} = \{\boldsymbol {\theta} \in \mathcal {G} \mid g (\boldsymbol {\theta}) = 0 \}, \tag {5}
+$$
+
+where  $g$  represents the unsigned distance to the manifold. We construct this distance function by first encoding the pose using a hierarchical network  $g_{\mathrm{enc}}$  that encodes the human pose based on its anatomical structure [42], and subsequently, use  $g_{\mathrm{dec}}$  to predict the distance based on the pose representation.
+
+Specifically, for a given pose  $\theta$ , we encode it as follows,
+
+$$
+\mathbf {v} _ {1} = g _ {\mathrm {e n c}} ^ {1} \left(\theta_ {1}\right), \quad \mathbf {v} _ {l} = g _ {\mathrm {e n c}} ^ {l} \left(\theta_ {l}, \mathbf {v} _ {\Omega (l)}\right), l \in \{2, \dots , L \}. \tag {6}
+$$
+
+Here,  $\Omega(l)$  is a function that maps the index of each orientation vector to its parent orientation vector in the kinematic chain of the human skeleton. We obtain the overall pose encoding as  $\mathbf{p} = [\mathbf{v}_1|\ldots|\mathbf{v}_L]$  by concatenating all the individual orientation encodings. This pose encoding is then processed by  $g_{\mathrm{dec}}: \mathbb{R}^{d,L} \to \mathbb{R}^+$ , which predicts the unsigned distance for the given pose representation  $\mathbf{p}$ . A lower distance value for a pose implies that the configuration of joints is more likely to be a plausible human pose.
+
+# 3.2.2 Pose Prior Training
+
+To train the parametric prior  $g$ , we rely on an auxiliary dataset of  $M$  human poses  $\mathcal{D}_{\mathrm{A}} = \{\pmb{\theta}^{i}\}_{i=1}^{N_{\mathrm{A}}}$ , where  $\pmb{\theta}^{i} =$
+
+$(\theta_1^i,\dots ,\theta_L^i)$ . Importantly, these poses are not associated with their corresponding RGB images, which preserves the privacy aspect of the method. Additionally, storing only the pose coordinates instead of entire images makes data storage much more efficient and feasible.
+
+We adopt a supervised approach to train  $g$  to predict the  $L2$  distance to the plausible pose manifold for a given pose. To achieve this, we construct a dataset  $\tilde{\mathcal{D}} = \{(\pmb{\theta}_i,d_i)\}_{i=1}^M$ , consisting of pose and distance pairs, from  $\mathcal{D}_{\mathrm{A}}$ . As the poses from  $\mathcal{D}_{\mathrm{A}}$  lie on the desired manifold, we assign  $d = 0$  to all poses in the dataset. To diversify our training samples, we randomly generate negative samples with distance  $d > 0$  by perturbing the poses from  $\mathcal{D}_{\mathrm{A}}$  with noise.
+
+We train the network with the standard  $L1$  loss,
+
+$$
+\mathcal {L} _ {\text {d i s t}} = \sum_ {(\boldsymbol {\theta}, d) \in \tilde {D}} \| g (\boldsymbol {\theta}) - d \| _ {1}. \tag {7}
+$$
+
+More details on the training process of the prior are provided in Section 4.2.
+
+# 3.2.3 Adaptation using Pose Prior
+
+We leverage the trained pose prior  $g$  to regularize the adaptation process by incentivizing the pose estimator to generate pseudo-labels that resemble plausible human poses.
+
+Given the heatmaps  $\{h_k^t\}_{k=1}^K$  generated for each keypoint by the student model  $f_{\mathrm{stu}}^t$  for a target image  $x$ , we calculate the corresponding orientation vectors in a differentiable manner to evaluate the plausibility of the predicted pose using the prior. First, we renormalize each heatmap to a probability distribution via spatial softmax and condense it to a point by computing the spatial expected value of the latter. For computational efficiency, we carry this out in a separable manner along the two spatial dimensions. Namely, assuming  $u = (u_1, u_2)$  to be the two components of each pixel coordinate, we set
+
+$$
+u _ {i} ^ {k} = \frac {\sum_ {u _ {i}} u _ {i} e ^ {h _ {k} ^ {t} (u _ {i})}}{\sum_ {u _ {i}} e ^ {h _ {k} ^ {t} (u _ {i})}}, \quad h _ {k} ^ {t} (u _ {i}) = \sum_ {u _ {j}} h _ {k} ^ {t} (u _ {1}, u _ {2}), \tag {8}
+$$
+
+where  $i = 1,2$  and  $j = 2,1$  respectively. Next, we use the pose coordinates to determine the orientation vectors between pairs of connected keypoints. Specifically, for every pair  $(a,b)$  of connected keypoints in the human skeleton (denoted by the set  $\mathcal{E}$ ), we calculate the unit vector  $\theta$  in the direction from  $u^a$  to  $u^b$ , where  $u^a$  and  $u^b$  are the estimated softmax coordinates of keypoints  $a$  and  $b$ , respectively, i.e.,
+
+$$
+\theta = \frac {u ^ {a} - u ^ {b}}{\| u ^ {a} - u ^ {b} \| _ {2}}, \quad \forall (a, b) \in \mathcal {E}. \tag {9}
+$$
+
+Finally, we use the prior as a regularization term to minimize the distance of the current pose from our learned manifold,
+
+$$
+\mathcal {L} _ {\text {p r i o r}} = \frac {1}{| \mathcal {B} |} \sum_ {\boldsymbol {x} \in \mathcal {B}} g \left(T \left(f _ {\mathrm {s t u}} ^ {t} (\boldsymbol {x})\right)\right), \tag {10}
+$$
+
+where  $T(\cdot)$  converts the predicted heatmaps to the orientation vector format required as input by the prior.
+
+# 3.3. Overall Adaptation
+
+The final training objective for the student model  $f_{\mathrm{stu}}$  can be expressed as:
+
+$$
+\min  _ {f _ {\mathrm {s t u}}} \mathcal {L} _ {\text {o u t}} + \lambda_ {1} \mathcal {L} _ {\text {f e a t}} + \lambda_ {2} \mathcal {L} _ {\text {p r i o r}}. \tag {11}
+$$
+
+Here,  $\lambda_{1}$  and  $\lambda_{2}$  are hyper-parameters that control the influence of feature space adaptation and prior regularization, respectively. The teacher model  $f_{\mathrm{tea}}$  is updated asynchronously by computing an exponential moving average of the student model weights as shown in Equation 1.
+
+# 4. Experiments
+
+In this section, we demonstrate POST's ability to adapt a 2D human pose estimation model to a target domain using only unlabeled data from that domain. We conduct experiments on three domain adaptive scenarios and compare with state-of-the-art domain adaptation baselines that utilize source data during adaptation. We also conduct extensive analysis to analyze the contribution and interaction between each component in our framework.
+
+# 4.1. Datasets
+
+SURREAL: SURREAL [43] is a large-scale dataset of synthetically generated images of people rendered from 3D sequences of human motion capture data against indoor backgrounds. It contains over 6 million frames, making it one of the largest and most diverse datasets of its kind.
+
+Human3.6M: Human3.6M [44] is a real-world video dataset captured in indoor environments, comprising 3.6 million frames. The dataset features human subjects performing various actions. In order to reduce redundancy and computational complexity, we down-sampled the videos from 50fps to 10fps as per the approach proposed in [5]. For training, we follow the standard protocol proposed in [6] and use 5 subjects (S1, S5, S6, S7, S8), while the remaining 2 subjects (S9, S11) are reserved for testing.
+
+LSP: Leeds Sports Pose (LSP) [45] is a real-world dataset that contains 2,000 images with annotated human body joint locations collected from sports activities. The images in LSP are captured in the wild, featuring a wide variety of human poses that are often challenging to detect.
+
+BRIAR: BRIAR [46] is a cutting-edge biometric dataset featuring a large-scale collection of videos of human subjects captured in extremely challenging conditions. The videos are recorded at varying distances, i.e., close range,  $100\mathrm{m}$ ,  $200\mathrm{m}$ ,  $400\mathrm{m}$ ,  $500\mathrm{m}$ , and unmanned aerial vehicles (UAV), with each video lasting around 90 seconds. We randomly
+
+sample 20 frames from each sequence for each of the 158 subjects for our experiments.
+
+# 4.2. Experiment Protocols
+
+Pose estimation model. We adopt the Simple Baseline [18] as our pose estimation model, with the ResNet-101 [47] as the backbone. To train the model, we use the Adam optimizer [48] with a base learning rate of  $1e - 4$ , scheduled to decrease to  $1e - 5$  after 5 epochs and to  $1e - 6$  after 20 epochs. The model is trained for 30 epochs. We use a batch size of 32 and run 500 iterations per epoch. To threshold the model predictions, we set the confidence thresholding ratio  $p$  to 0.5. To augment the images during training, we follow [6] and use rotation, translation, shear, and Gaussian blur. The hyper-parameters  $\lambda_{1}$  and  $\lambda_{2}$  are set to  $1e - 3$  and  $1e - 6$ , respectively.
+
+Pose prior model. The training of the parametric prior follows a multi-stage approach that involves using different types of training samples. Initially, we use a combination of manifold poses  $\theta_{m}$  and non-manifold poses  $\theta_{nm}$  with a considerable distance from the desired manifold. Over the course of training, we gradually increase the number of non-manifold poses  $\theta_{nm}$  with a small distance from the manifold. This enables our model to first learn a smooth surface and then gradually incorporate finer details as training progresses. We create these non-manifold poses  $\theta_{nm}$  by injecting noise into the manifold poses  $\theta_{m}$  obtained from the auxiliary dataset. Specifically, we sample directional noise from the Von-Mises distribution [49] and add it to the manifold poses in order to obtain the implausible poses.
+
+The architecture for the encoder  $g_{\mathrm{enc}}$  consists of a 2-layer MLP with an output feature size of  $d = 6$  for each orientation vector, similar to [42]. The distance field network  $g_{\mathrm{dec}}$  is implemented as a 5-layer MLP. Given its large size and diverse poses, we train the prior for our primary experiments using the SURREAL dataset.
+
+# 4.3. Comparison with Baselines
+
+Baselines. We evaluate the performance of our proposed method against several state-of-the-art domain adaptive frameworks. This includes adversarial learning-based feature alignment methods, such as DAN[31], DD[50], and RegDA[5]. Additionally, we consider approaches based on pseudo-labeling, namely CCSSL[38], and UDAPE [6]. It is worth noting that all these methods employ the source data during the adaptation process. To establish a comprehensive performance baseline, we report the results of two additional baselines: Oracle and Source only. The Oracle baseline represents the upper bound of the model's performance, achieved by training the model jointly with target 2D annotations. On the other hand, Source only represents the model's performance when it is directly applied to the target domain without any adaptation.
+
+Table 1. PCK@0.05 on SURREAL  $\rightarrow$  LSP. (Best value is in red color, while the second best value is in blue color.)  
+
+<table><tr><td>Method</td><td>SF</td><td>Sld.</td><td>Elb.</td><td>Wrist</td><td>Hip</td><td>Knee</td><td>Ankle</td><td>Avg.</td></tr><tr><td>Source only</td><td>-</td><td>51.5</td><td>65.0</td><td>62.9</td><td>68.0</td><td>68.7</td><td>67.4</td><td>63.9</td></tr><tr><td>Oracle</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DAN</td><td>X</td><td>52.2</td><td>62.9</td><td>58.9</td><td>71.0</td><td>68.1</td><td>65.1</td><td>63.0</td></tr><tr><td>DD</td><td>X</td><td>28.4</td><td>65.9</td><td>56.8</td><td>75.0</td><td>74.3</td><td>73.9</td><td>62.4</td></tr><tr><td>RegDA</td><td>X</td><td>62.7</td><td>76.7</td><td>71.1</td><td>81.0</td><td>80.3</td><td>75.3</td><td>74.6</td></tr><tr><td>CCSSL</td><td>X</td><td>36.8</td><td>66.3</td><td>63.9</td><td>59.6</td><td>67.3</td><td>70.4</td><td>60.7</td></tr><tr><td>UDAPE</td><td>X</td><td>69.2</td><td>84.9</td><td>83.3</td><td>85.5</td><td>84.7</td><td>84.3</td><td>82.0</td></tr><tr><td>POST</td><td>✓</td><td>66.5</td><td>83.9</td><td>81.0</td><td>84.6</td><td>83.1</td><td>82.6</td><td>80.3</td></tr></table>
+
+Table 2. PCK@0.05 on SURREAL  $\rightarrow$  Human3.6M.  
+
+<table><tr><td>Method</td><td>SF</td><td>Sld.</td><td>Elb.</td><td>Wrist</td><td>Hip</td><td>Knee</td><td>Ankle</td><td>Avg.</td></tr><tr><td>Source only</td><td>-</td><td>69.4</td><td>75.4</td><td>66.4</td><td>37.9</td><td>77.3</td><td>77.7</td><td>67.3</td></tr><tr><td>Oracle</td><td>-</td><td>95.3</td><td>91.8</td><td>86.9</td><td>95.6</td><td>94.1</td><td>93.6</td><td>92.9</td></tr><tr><td>DAN</td><td>X</td><td>68.1</td><td>77.5</td><td>62.3</td><td>30.4</td><td>78.4</td><td>79.4</td><td>66.0</td></tr><tr><td>DD</td><td>X</td><td>71.6</td><td>83.3</td><td>75.1</td><td>42.1</td><td>76.2</td><td>76.1</td><td>70.7</td></tr><tr><td>RegDA</td><td>X</td><td>73.3</td><td>86.4</td><td>72.8</td><td>54.8</td><td>82.0</td><td>84.4</td><td>75.6</td></tr><tr><td>CCSSL</td><td>X</td><td>44.3</td><td>68.5</td><td>55.2</td><td>22.2</td><td>62.3</td><td>57.8</td><td>51.7</td></tr><tr><td>UDAPE</td><td>X</td><td>78.1</td><td>89.6</td><td>81.1</td><td>52.6</td><td>85.3</td><td>87.1</td><td>79.0</td></tr><tr><td>POST</td><td>✓</td><td>81.3</td><td>88.5</td><td>77.4</td><td>46.1</td><td>83.4</td><td>83.4</td><td>76.7</td></tr></table>
+
+Metrics. We adopt the evaluation metric of Percentage of Correct Keypoint (PCK) for all experiments and report PCK@0.05 that measures the ratio of correct prediction within a range of  $5\%$  with respect to the image size.
+
+# 4.4. Results
+
+Quantitative results. We evaluate POST in two adaptation scenarios: SURREAL  $\rightarrow$  LSP and SURREAL  $\rightarrow$  Human3.6M, and report the quantitative results in Table 1 and Table 2, respectively. Specifically, we report the PCK@0.05 on 16 keypoints of the human body, including shoulders (sld.), elbows (elb.), wrists, hips, knees, and ankles. Our method achieves comparable results to many recent approaches that leverage source data for adaptation. Among these methods, UDAPE [6] achieves the highest performance on both cases, with our framework achieving a close second and only 2 percentage points behind on average. Notably, we outperform every other method, including the recently proposed RegDA [5] approach, by a significant margin of up to 5.7 percentage points. It is worth noting that not only do these methods require source data, but also involve additional models such as discriminators or style transfer modules, and unstable adversarial training. In contrast, our framework is lightweight, only involves pseudo-label training, and utilizes a simple prior model that can be trained offline.
+
+Qualitative results. In addition to quantitative results, we also present qualitative results on SURREAL  $\rightarrow$  Human3.6M in Figure 3 and on SURREAL  $\rightarrow$  LSP in Figure 4. Also,
+
+![](images/9ca3dd4ca1a1e28750331b7cb8a5c701cc85b54cdf7f37b743deb37214ce2ba2.jpg)
+
+![](images/6ddd8ed53a6b19b31f5e7fed23c697ea428f3503f8c8f9a11f0efe292d6a862a.jpg)
+
+![](images/06b8c705d828873c385eadc00468c9bc2fc87a4a57b2a58acd501b61a009979f.jpg)
+
+![](images/55035cceed5ab24a11f808a10e6e62344fe77e57fb571c064ec77e259f690089.jpg)  
+Figure 3. Qualitative results on SURREAL  $\rightarrow$  Human3.6M. We demonstrate sample results on the Human3.6M dataset. For each row, the leftmost shows the Source only prediction, the middle one shows the UDAPE [6] prediction, and the rightmost shows the prediction made POST.
+
+![](images/bf638cc35de63688c3f809da13d6ecca3d074a6a5776fca8ba763b09bb5c3770.jpg)
+
+![](images/1af0c514de7f1b2d09958562f8961b6c6cb0fc0159bb2d8a0d44fdaad1875b10.jpg)
+
+![](images/3de0508bba699e9fb9ed70599bbfe639183e0de420765c2cb29b22922db41eeb.jpg)
+
+![](images/e88c80e6e37804c9a9155a709612e3849a0aaa276c55742ab55e4428bd64f1ea.jpg)
+
+![](images/5800a5eff490007d9f72a9f3305ce806c7c576c1725fde8c20e35ff5cef02c7b.jpg)
+
+![](images/26f10a4c3df8781a5005c528fcf7cf87a1ea31a01fba7fe9144c0a13b57a9949.jpg)
+
+![](images/87d6f34fb2bad67ef3179be8f212ff38d3cc26040310bce2f12f61dbaa69e179.jpg)
+
+![](images/90ddfdaed2ee6ce2a5359bb92ed4d83fe00269159ce15ce6affd72347148e4ff.jpg)
+
+![](images/30be3e3f739364914f6ac74358e4da78e6dc356ac72f5439de9a1d16fe5168af.jpg)
+
+![](images/a24aa9696468f014d926a23875096da8c11ed3a8297a696ef0e21d4f5ab516a3.jpg)
+
+![](images/578aa8f70007e230899509d88df089d6989ca9c2fdd76f65830f0199b54453d9.jpg)
+
+![](images/2b917ce2bae4dc5d9ce79cef983cb4c7e5db69f2b175d78a10a9b48151f391f0.jpg)  
+Figure 4. Qualitative results on SURREAL  $\rightarrow$  LSP. We demonstrate sample results on the LSP dataset. For each row, the left-most shows the Source only prediction, the middle one shows the UDAPE [6] prediction, and the rightmost shows the prediction made POST.
+
+![](images/f45bf2fae60fa16dfdefd095aa7ee8e2514a82b33f48d654214f5c08b987128a.jpg)
+
+![](images/f932f0b4aa506b25c209f2b2dc069513b86eee3321aba82c71b5d750428d5b9d.jpg)
+
+we present only visual results on the SURREAL  $\rightarrow$  BRIAR adaptation scenario since pose annotations are absent in the BRIAR dataset. Figure 5 displays the predicted human poses on images taken from six different imaging ranges in the BRIAR dataset. While using the source model directly produces completely inaccurate poses, POST can accurately localize the keypoint locations, even in the presence of occlusions and atmospheric turbulence. Our approach can also accurately reconstruct poses even when the human is imaged from an elevated perspective, resulting in a high camera angle (Figure 5.f). Notably, our approach achieves this without using any source data for adaptation. Our results are comparable to those produced by UDAPE [6], which utilizes source data for adaptation. Additional qualitative results can be found in the supplementary document.
+
+![](images/b488637a570423b0c61b2b920d409367b8d37de12030b61d1cfa0a0d86a0fada.jpg)  
+a)  $100\mathrm{m}$
+
+![](images/6699f621e33ec7924a2e3e22e59af8268b17dc67c2b86f498bef502438207a34.jpg)
+
+![](images/7a74b7dbf63c3c52242e7f9de6d519c459d882828be198e5e53081e5072e7488.jpg)  
+b)  $200\mathrm{m}$
+
+![](images/0e423e1f372fc6d2f448b59ecf1bf54c35a70d65eccaf60269c757cd1fe5f1bd.jpg)
+
+![](images/a2d47abb81dece636654ac0a86b4cca9356074298df4759fd5136a58f4212362.jpg)  
+c)  $400\mathrm{m}$
+
+![](images/4e89899667fa45cde15a39a23c01f71b2bc367063323bb7caffea4f57f7f63e0.jpg)
+
+![](images/e1eb07fe7e882dfb82fce907089730185aeea9907de1c34c662f2eadacfe2082.jpg)  
+d)  $500\mathrm{m}$
+
+![](images/117eb71616e26b7bb49299de2a2aa1031bf8907490a22a21fb563ffe508a9f3d.jpg)
+
+![](images/ae26e585dc79fcadbe7beef7631175128cb92f7806c412598d808f049f8df802.jpg)  
+e) close range
+
+![](images/20c06172a974e6383a36ad30d312c3833e0cf846cc3308210aceb24e8e687611.jpg)
+
+![](images/f2aa48b85552239f56e0f04de58b137b6f5e654cda8884c106ba54f4b6e5fe1a.jpg)  
+f) uav
+
+![](images/1c048a6e1e9309091c8b02dde1f23d8ed93272664bc72a57d8b8a03fc087ac8a.jpg)  
+Figure 5. Qualitative results on SURREAL  $\rightarrow$  BRIAR. We demonstrate sample results on BRIAR at all ranges. For each range, we display three images: the leftmost shows the Source only prediction, the middle one shows the UDAPE [6] prediction, and the rightmost shows the prediction using POST. Although POST does not use source data for adaptation, it is able to match the predictions produced by UDAPE, which uses source data.
+
+# 4.5. Ablation Studies
+
+Effect of auxiliary dataset. We conduct experiments to evaluate the impact of the choice of the auxiliary dataset (used to train the prior) on downstream adaptation. The results are presented in Table 3. Our findings indicate that
+
+POST is robust to the choice of the auxiliary dataset, with performance differences of  $\sim 0.5$  percentage points.
+
+Effect of loss terms. We conduct an experiment to evaluate the performance of each component of our framework. The results on the SURREAL  $\rightarrow$  Human3.6M adaptation scenario are presented in Table 4. Our findings indicate that in addi
+
+Table 3. Effect of auxiliary dataset. We evaluate the effect of auxiliary dataset on downstream adaptation tasks.  
+
+<table><tr><td colspan="8">SURREAL→HUMAN3.6M</td></tr><tr><td>Aux. dataset</td><td>Sld.</td><td>Elb.</td><td>Wrist</td><td>Hip</td><td>Knee</td><td>Ankle</td><td>Avg.</td></tr><tr><td>SURREAL</td><td>81.3</td><td>88.5</td><td>77.4</td><td>46.1</td><td>83.4</td><td>83.4</td><td>76.7</td></tr><tr><td>Human3.6M</td><td>80.9</td><td>88.0</td><td>77.2</td><td>45.0</td><td>83.1</td><td>82.8</td><td>76.2</td></tr><tr><td colspan="8">SURREAL→LSP</td></tr><tr><td>Aux. dataset</td><td>Sld.</td><td>Elb.</td><td>Wrist</td><td>Hip</td><td>Knee</td><td>Ankle</td><td>Avg.</td></tr><tr><td>SURREAL</td><td>66.5</td><td>83.9</td><td>81.0</td><td>84.6</td><td>83.1</td><td>82.6</td><td>80.3</td></tr><tr><td>Human3.6M</td><td>66.1</td><td>83.6</td><td>80.7</td><td>84.4</td><td>83.1</td><td>82.5</td><td>80.1</td></tr></table>
+
+Table 4. Effect of each loss term. We evaluate the contribution of each loss term on SURREAL  $\rightarrow$  Human3.6M.  
+
+<table><tr><td>Lout</td><td>Lfeat</td><td>Lprior</td><td>Sld.</td><td>Elb.</td><td>Wrist</td><td>Hip</td><td>Knee</td><td>Ankle</td><td>Avg.</td></tr><tr><td>X</td><td>X</td><td>X</td><td>69.4</td><td>75.4</td><td>66.4</td><td>37.9</td><td>77.3</td><td>77.7</td><td>67.3</td></tr><tr><td>✓</td><td>X</td><td>X</td><td>77.9</td><td>86.7</td><td>73.7</td><td>38.8</td><td>83.0</td><td>84.3</td><td>74.1</td></tr><tr><td>✓</td><td>✓</td><td>X</td><td>81.7</td><td>87.1</td><td>75.2</td><td>44.3</td><td>82.3</td><td>82.2</td><td>75.5</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>81.3</td><td>88.5</td><td>77.4</td><td>46.1</td><td>83.4</td><td>83.4</td><td>76.7</td></tr></table>
+
+tion to prediction space adaptation, feature space adaptation also plays a crucial role in enabling effective unsupervised learning from pseudo-labels. Moreover, we observed that the human pose prior brings additional improvements, thus validating our hypothesis that noisy pose pseudo-labels can be refined implicitly by a prior in the absence of source data. Overall, our results demonstrate the effectiveness of our framework and the importance of each of its components in achieving state-of-the-art performance.
+
+Cross-dataset performance of prior. We evaluate the ability of the learned prior to handle distribution shift separately from the adaptation performance. Specifically, we assess the robustness of the prior by computing pose scores across datasets, and the results are presented in Figure 6. The plot demonstrates that our prior is effective in scoring plausible (real) poses with low scores and scoring implausible poses with higher scores. Note that the bell curve shape of the noisy pose scores is due to the Von-Mises noise added to create the noisy poses.
+
+Effect of thresholding. We analyze the impact of the pseudo-label threshold  $\tau$  on the adaptation performance in Table 5. The results on SURREAL  $\rightarrow$  Human3.6M reveal that as we increase this ratio, the performance gradually decreases. This can be attributed to the fact that higher thresholding ratios tend to include lower confident predictions as pseudo-labels, which can negatively impact adaptation.
+
+# 5. Conclusion
+
+We address the problem of adapting a pre-trained 2D human pose estimator to a new target domain with unlabeled
+
+Table 5. Effect of  $\tau$ . We evaluate adaptation performance on SURREAL  $\rightarrow$  Human3.6M as  $\tau$  is varied.  
+
+<table><tr><td>τ = 0.1</td><td>τ = 0.3</td><td>τ = 0.5</td><td>τ = 0.7</td><td>τ = 0.9</td></tr><tr><td>76.4</td><td>76.5</td><td>76.7</td><td>75.3</td><td>73.8</td></tr></table>
+
+![](images/fe682b04ab13c49f01819e55fd9998b7720e2fa39836b2a08ac6227d112420df.jpg)  
+(a) Human3.6M  $\rightarrow$  SURREAL  
+(b) SURREAL  $\rightarrow$  Human3.6M  
+Figure 6. Cross-dataset prior transfer. We plot the histogram of scores predicted by the prior. The prior can clearly demarcate plausible poses from implausible poses across datasets.
+
+target data. To this end, we propose a self-training algorithm, POST, that leverages a Mean Teacher framework to enforce both prediction-level and feature-level consistency between a pair of student and teacher models. Our approach incorporates a human pose prior that captures the manifold of possible poses in a high-dimensional space, which helps to refine noisy pseudo-labels generated during self-training. We evaluate our method on three challenging adaptation scenarios and show that it achieves competitive performance compared to existing UDA methods that have access to source data.
+
+Acknowledgements. This research is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via [2022-21102100007]. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein.
+
+# References
+
+[1] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, 2019. 1, 2
+
+[2] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 1  
+[3] Hai Liu, Tingting Liu, Zhaoli Zhang, Arun Kumar Sangaiah, Bing Yang, and Youfu Li. Arhpe: Asymmetric relation-aware representation learning for head pose estimation in industrial human-computer interaction. IEEE Transactions on Industrial Informatics, 18(10):7107-7117, 2022. 1  
+[4] Xiang Li, Yasushi Makihara, Chi Xu, Yasushi Yagi, Shiqi Yu, and Mingwu Ren. End-to-end model-based gait recognition. In ACCV, 2020. 1  
+[5] Junguang Jiang, Yifei Ji, Ximei Wang, Yufeng Liu, Jianmin Wang, and Mingsheng Long. Regressive domain adaptation for unsupervised keypoint detection. In CVPR, 2021. 1, 3, 6, 7  
+[6] Donghyun Kim, Kaihong Wang, Kate Saenko, Margrit Betke, and Stan Sclaroff. A unified framework for domain adaptive pose estimation. In ECCV, 2022. 1, 2, 3, 6, 7, 8  
+[7] Paul M Schwartz and Daniel J Solove. The pii problem: Privacy and a new concept of personally identifiable information. NYUL rev., 86:1814, 2011. 1  
+[8] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 2  
+[9] Sk Miraj Ahmed, Dripta S Raychaudhuri, Sujoy Paul, Samet Oymak, and Amit K Roy-Chowdhury. Unsupervised multisource domain adaptation without access to source data. In CVPR, 2021. 2  
+[10] Sujoy Paul, Ansh Khurana, and Gaurav Aggarwal. Unsupervised adaptation of semantic segmentation models without source data. arXiv preprint arXiv:2112.02359, 2021. 2  
+[11] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS, 2017. 2, 3  
+[12] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021. 2, 4  
+[13] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In ECCV, 2022. 2, 5  
+[14] Hao-Shu Fang, Shuqin Xie, Yu-Wing Tai, and Cewu Lu. *Rmpe: Regional multi-person pose estimation.* In ICCV, 2017. 2  
+[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2  
+[16] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR, 2016. 2  
+[17] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In ECCV, 2016. 2, 4  
+[18] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In ECCV, 2018. 2, 4, 6
+
+[19] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multiperson pose estimation. In CVPR, 2018. 2  
+[20] Leonid Pishchulin, Eldar Insafutdinov, Siyu Tang, Bjoern Andres, Mykhaylo Andriluka, Peter V Gehler, and Bernt Schiele. Deepcut: Joint subset partition and labeling for multi-person pose estimation. In CVPR, 2016. 2  
+[21] Eldar Insafutdinov, Leonid Pishchulin, Bjoern Andres, Mykhaylo Andriluka, and Bernt Schiele. Deepercut: A deeper, stronger, and faster multi-person pose estimation model. In ECCV, 2016. 2  
+[22] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: realtime multi-person 2d pose estimation using part affinity fields. IEEE transactions on Pattern Analysis and Machine Intelligence, 43(1):172-186, 2019. 2  
+[23] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In CVPR, 2019. 2  
+[24] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In CVPR, 2020. 2  
+[25] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. arXiv preprint arXiv:2104.02300, 2021. 2  
+[26] Sheng Jin, Wentao Liu, Enze Xie, Wenhai Wang, Chen Qian, Wanli Ouyang, and Ping Luo. Differentiable hierarchical graph grouping for multi-person pose estimation. In ECCV, 2020. 2  
+[27] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In CVPR, 2017. 2  
+[28] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kihyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In CVPR, 2018. 2  
+[29] Han-Kai Hsu, Chun-Han Yao, Yi-Hsuan Tsai, Wei-Chih Hung, Hung-Yu Tseng, Maneesh Singh, and Ming-Hsuan Yang. Progressive domain adaptation for object detection. In WACV, 2020. 2  
+[30] Dripta S Raychaudhuri, Sujoy Paul, Jeroen Vanbaar, and Amit K Roy-Chowdhury. Cross-domain imitation from observations. In ICML, 2021. 2  
+[31] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, 2015. 2, 6  
+[32] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. In JMLR, 2016. 2
+
+[33] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei A. Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In ICML, 2018. 2  
+[34] Luan Tran, Kihyuk Sohn, Xiang Yu, Xiaoming Liu, and Manmohan Chandraker. Gotta adapt 'em all: Joint pixel and feature-level domain adaptation for recognition in the wild. In CVPR, 2019. 2  
+[35] Hao-Wei Yeh, Baoyao Yang, Pong C Yuen, and Tatsuya Harada. Sofa: Source-data-free feature alignment for unsupervised domain adaptation. In WACV, 2021. 2  
+[36] Vikash Kumar, Rohit Lal, Himanshu Patil, and Anirban Chakraborty. Conmix for source-free single and multi-target domain adaptation. In WACV, 2023. 2  
+[37] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In ICCV, 2021. 2  
+[38] Jiteng Mu, Weichao Qiu, Gregory D Hager, and Alan L Yuille. Learning from synthetic animals. In CVPR, 2020. 3, 6  
+[39] Chen Li and Gim Hee Lee. From synthetic to real: Unsupervised domain adaptation for animal pose estimation. In CVPR, 2021. 3  
+[40] Jonathan J. Tompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. NeurIPS, 2014. 3  
+[41] Luan Tran, Kihyuk Sohn, Xiang Yu, Xiaoming Liu, and Manmohan Chandraker. Gotta adapt' em all: Joint pixel and feature-level domain adaptation for recognition in the wild. In CVPR, 2019. 4  
+[42] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. Leap: Learning articulated occupancy of people. In CVPR, 2021. 5, 6  
+[43] Gül Varol, Javier Romero, Xavier Martin, Naureen Mahmood, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning from synthetic humans. In CVPR, 2017. 6  
+[44] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2013. 6  
+[45] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In BMVC, 2010. 6  
+[46] David Cornett, Joel Brogan, Nell Barber, Deniz Aykac, Seth Baird, Nicholas Burchfield, Carl Dukes, Andrew Duncan, Regina Ferrell, Jim Goddard, Gavin Jager, Matthew Larson, Bart Murphy, Christi Johnson, Ian Shelley, Nisha Srinivas, Brandon Stockwell, Leanne Thompson, Matthew Yohe, Robert Zhang, Scott Dolvin, Hector J. Santos-Villalobos, and David S. Bolme. Expanding accurate person recognition to new altitudes and ranges: The briar dataset. In WACV Workshops, 2023. 6
+
+[47] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6  
+[48] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6  
+[49] Riccardo Gatto and Sreenivasa Rao Jammalamadaka. The generalized von mises distribution. Statistical Methodology, 4(3):341-353, 2007. 6  
+[50] Yuchen Zhang, Tianle Liu, Mingsheng Long, and Michael Jordan. Bridging theory and algorithm for domain adaptation. In ICML, 2019. 6

2023/Prior-guided Source-free Domain Adaptation for Human Pose Estimation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65474a6f316313e4e71e7aa108f7e1f437159ca963906f4eaec285eabac506a4
+size 594032

2023/Priority-Centric Human Motion Generation in Discrete Latent Space/58c7e478-f288-424e-a702-051324564633_content_list.json

@@ -0,0 +1,1733 @@
+[
+    {
+        "type": "text",
+        "text": "Priority-Centric Human Motion Generation in Discrete Latent Space",
+        "text_level": 1,
+        "bbox": [
+            135,
+            130,
+            833,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Hanyang Kong $^{1}$ , Kehong Gong $^{1}$ , Dongze Lian $^{1}$ , Michael Bi Mi $^{2}$ , Xinchao Wang $^{1*}$",
+        "bbox": [
+            158,
+            179,
+            810,
+            200
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ National University of Singapore  $^{2}$ Huawei International Pte Ltd",
+        "bbox": [
+            218,
+            204,
+            750,
+            223
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{hanyang.k, gongkehong}@u.nus.edu {dongze, xinchao}@nus.edu.sg",
+        "bbox": [
+            205,
+            224,
+            764,
+            241
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            275,
+            313,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Text-to-motion generation is a formidable task, aiming to produce human motions that align with the input text while also adhering to human capabilities and physical laws. While there have been advancements in diffusion models, their application in discrete spaces remains underexplored. Current methods often overlook the varying significance of different motions, treating them uniformly. It is essential to recognize that not all motions hold the same relevance to a particular textual description. Some motions, being more salient and informative, should be given precedence during generation. In response, we introduce a Priority-Centric Motion Discrete Diffusion Model (M2DM), which utilizes a Transformer-based VQ-VAE to derive a concise, discrete motion representation, incorporating a global self-attention mechanism and a regularization term to counteract code collapse. We also present a motion discrete diffusion model that employs an innovative noise schedule, determined by the significance of each motion token within the entire motion sequence. This approach retains the most salient motions during the reverse diffusion process, leading to more semantically rich and varied motions. Additionally, we formulate two strategies to gauge the importance of motion tokens, drawing from both textual and visual indicators. Comprehensive experiments on the HumanML3D and KIT-ML datasets confirm that our model surpasses existing techniques in fidelity and diversity, particularly for intricate textual descriptions.",
+        "bbox": [
+            76,
+            306,
+            473,
+            717
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            744,
+            209,
+            762
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Generating realistic and diverse 3D human motions under various conditions e.g., action labels [34, 18], natural language descriptions [55, 17, 5, 56], and musical cues [27, 28, 29], presents a significant challenge across multiple domains, including gaming, filmmaking, and robotic animation. Notably, motion generation based on language descriptions has garnered substantial interest, given its promise",
+        "bbox": [
+            75,
+            770,
+            470,
+            878
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "for enhancing realism and broadening practical applications.",
+        "bbox": [
+            500,
+            277,
+            893,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, generating a high-quality motion is not trivial due to the inherent modality gap and the complex mapping between text and motion modalities. Previous works [35, 43] align the latent feature space between text and motion modalities. For instance, TEMOS [35] aligns the text and motion feature by learning text and motion encoders. MotionClip [43] aligns the human motion manifold to CLIP [37] space for infusing semantic knowledge of CLIP into the motion extractor. Despite these advancements, they might encounter performance degradation when dealing with complex textual descriptions.",
+        "bbox": [
+            496,
+            292,
+            893,
+            459
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To process these complex textual descriptions, some text-to-motion methods are proposed. State-of-the-art approaches such as TM2T [17] and T2M-GPT [55] use vector-quantized autoencoder (VQ-VAE) [46] to learn a discrete and compact motion representation, followed by a translation model [38, 47] to map the text modality to the motion modality. With the popularity and the superior performance in the generation tasks of diffusion models [20], MDM [44] and MLD [5] are proposed to learn conditioned diffusion models on the raw motion representation space and the latent feature space, respectively.",
+        "bbox": [
+            496,
+            460,
+            895,
+            626
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "While there have been promising advancements, two primary issues remain unresolved: i) The aforementioned diffusion methods, namely MDM [44] and MLD [5], predominantly address the latent feature within a continuous space. Although VQ-VAE-inspired architectures have made considerable strides in motion generation [55, 17], particularly with the support of discrete and compact motion representations, the integration of the diffusion model into a discrete space remains inadequately explored. ii) Discrete diffusion models employed in prior studies [15, 14, 22] treat all tokens uniformly. This approach presupposes that every token conveys an equivalent amount of information, neglecting the inherent disparities among tokens within a sequence. A more intuitive generative approach for humans would involve a progressive hierarchy, commencing with overarching concepts and gradually delving into finer details.",
+        "bbox": [
+            496,
+            627,
+            895,
+            868
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address the aforementioned challenges, we introduce a priority-centric motion discrete diffusion model (M2DM)",
+        "bbox": [
+            500,
+            869,
+            893,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            805,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author: xinchao@nus.edu.sg",
+        "bbox": [
+            94,
+            886,
+            336,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "14806",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/b439b8314da8a8ceb3326548da576200317184eb052eb4f23a161258ea0bf35f.jpg",
+        "image_caption": [
+            "a person steps forward with left leg walking in a circle counterclockwise motion swinging arms and returns back."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            90,
+            267,
+            198
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/9d0c251455da47bd3d24f66b71118aebf0d37946e80a8daacde1e958f2e04810.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            276,
+            90,
+            439,
+            198
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/c8ea4c1f7282c3d12e82f475d439ce5e00a7a9482b33fc27749fca5bf60e559b.jpg",
+        "image_caption": [
+            "a person runs forward, then turns around and walks in the opposite direction while swinging their right arm."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            83,
+            233,
+            251,
+            345
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/c859db1c1545f793b8cdf537f4e5ba72913f0744a729ef102d8e27a747e06912.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            269,
+            236,
+            447,
+            347
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/ed8e60537fcd7457573289571d364e02360b7d0c102a0ab97a985dc78d08cfc3.jpg",
+        "image_caption": [
+            "a man gets up from the ground, walks around in a circle and then sits down on the ground again."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            88,
+            875,
+            200
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/85861fed1c799c75fb9e0e963334011b3432907a77306bb750bfcea593abbd37.jpg",
+        "image_caption": [
+            "a man postures his arms like holding a dance partner and dances the waltz from the right hand side to the left hand side."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            532,
+            247,
+            694,
+            347
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/a8b75794bd13b135797c0134ed741dcf543f0c5d0c7087ad0abad9fa0f0f1441.jpg",
+        "image_caption": [
+            "Figure 1: Our Priority-Centric Motion Discrete Diffusion Model (M2DM) generates diverse and precise human motions given complex textual descriptions. The color of human mesh goes from light to dark over time."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            709,
+            244,
+            861,
+            348
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "designed to generate motion sequences from textual descriptions, progressing in a primary to secondary manner. Initially, we employ a Transformer-based VQ-VAE, which is adept at learning a concise, discrete motion representation through the global self-attention mechanism. To circumvent code collapse and guarantee the optimal utilization of each motion token within the codebook, a regularization term is incorporated during the VQ-VAE training phase. Furthermore, we craft a noise schedule wherein individual tokens within a sequence are allocated varying corruption probabilities, contingent on their respective priorities during the forward process. Specifically, tokens of higher priority are slated for corruption towards the latter stages of the forward process. This ensures that the subsequent learnable reverse process adheres to a primary-to-secondary sequence, with top-priority tokens being reinstated foremost.",
+        "bbox": [
+            75,
+            446,
+            472,
+            688
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To discern the significance of motion tokens within a sequence, we introduce two evaluative strategies: static assessment and dynamic assessment. 1) Static assessment: Drawing inspiration from the neural language domain, where the significance of individual words is gauged by their entropy across datasets, we calculate the entropy of each motion token across the entire motion dataset. 2) Dynamic assessment: We cultivate an agent specifically to dynamically gauge the importance of tokens within a sequence. Given a discrete motion token sequence, the agent masks one token at each iteration. These masked token sequences are then fed into the VQ decoder for continuous motion reconstruction. The agent's objective is to curtail the cumulative reconstruction discrepancy between the original and the reconstructed mo",
+        "bbox": [
+            75,
+            688,
+            472,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "tions at every phase. This is achieved using a reinforcement learning [23] (RL) strategy, facilitating the agent's identification of motion tokens that convey minimal information within a sequence—a process we term dynamic analysis. Our priority-centric discrete diffusion model has showcased commendable generative prowess, especially when dealing with intricate textual descriptions. Through rigorous testing, our method has proven its mettle, consistently matching or surpassing the performance of prevailing text-to-motion generation techniques on the HumanML3D and KIT-ML datasets.",
+        "bbox": [
+            496,
+            446,
+            893,
+            611
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We summarize our contributions as follows:",
+        "bbox": [
+            519,
+            613,
+            812,
+            627
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- To capture long-range dependencies in the motion sequences, we apply Transformer as the architecture of VQ-VAE. Besides, a regularization term is applied to increase the usage of tokens in the codebook.",
+            "- We design a priority-centric scheme for human motion generation in the discrete latent space. To enhance the performance given complex descriptions, we design a novel priority-centric scheme for the discrete diffusion model.",
+            "- Our proposed priority-centric M2DM achieves state-of-the-art performance on the HumanML3D [16] and KIT-ML [36] datasets."
+        ],
+        "bbox": [
+            519,
+            635,
+            893,
+            832
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            844,
+            640,
+            859
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "VQ-VAE. Vector Quantized Variational Autoencoder (VQ-VAE) [46] aims to minimize reconstruction error with dis",
+        "bbox": [
+            500,
+            869,
+            893,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "14807",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "crete representations by training a variational autoencoder. VQ-VAE achieves promising performance on generative tasks, such as text-to-image generation [39, 52], music generation [7], unconditional image synthesis [9], etc. For motion generation tasks, most of the methods [17, 55] tokenize the continuous motion representation by the 1D-CNN network. Since the motion style varies from person to person and the number of action combinations is almost infinite, it is essential to extract discriminative and distinctive motion features from the raw motion representations. Moreover, raw motion representation is much more redundant compared with neural language, it is feasible to represent a motion sequence using few motion tokens from a global perspective. To this end, we introduce a transformer-based VQ-VAE and several techniques are applied to balance the usage of each motion token in the codebook and prevent the codebook from collapsing, i.e., only a small proportion of codes in the codebook are activated.",
+        "bbox": [
+            75,
+            90,
+            472,
+            363
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Diffusion models. Diffusion generative models are first proposed in [42] and achieve promising results on image generation [8, 33, 21, 20, 50, 48, 53, 11, 10]. Discrete diffusion model is first proposed by [42] and several works [22, 14] apply the discrete diffusion model for text generation. D3PM [3] then applies the discrete diffusion model to the image generation task. For motion generation tasks, several works [24, 56, 44, 5] introduce diffusion models to generate diverse motion sequences from textual descriptions. In this work, we first introduce the discrete diffusion model for text-to-motion generation tasks which achieves promising results, especially for long text descriptions with several action commands.",
+        "bbox": [
+            75,
+            382,
+            472,
+            580
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Conditional human motion synthesis Compared with unconditional Human motion synthesis, human motion synthesis conditioned on various conditions, such as textual description [55, 17, 5, 56], music [27, 28, 29, 13], and actions [34, 18] is more challenging due to the huge gap between two different domains. A common strategy of conditioned human motion synthesis is to employ generative models, for instance, conditional VAE [26], and learn a latent motion representation. Text-to-motion task is more challenging because neural language is a type of information highly refined by human beings, whereas human motion is much more redundant. Most recently, several promising works [5, 16, 17, 56, 55] are proposed for text-to-motion tasks. T2M-GPT [55] and TM2T [17] regard the text-to-motion task as a translation problem. They first train a VQ-VAE [46] to tokenize the motion sequences into motion tokens, and Recurrent Neural Networks (RNN) [32], Transformer [47, 51, 49, 41, 40, 54], or GPT-like [38] models are further applied to 'translate' textual descriptions to motions. Moreover, diffusion-based [20] models are introduced for",
+        "bbox": [
+            75,
+            598,
+            472,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "text-to-motion generation by [56, 6, 44, 5] in the contentious space. In this work, to the best of our knowledge, we are the first ones to diffuse the motion representation in the discrete latent space. Moreover, we notice that different motion clips in a whole motion sequence play various roles, for instance, stepping forward is more important than swinging arms for a running motion. We further design a diffuse schedule such that the generation process follows a from-primary-to-secondary manner.",
+        "bbox": [
+            496,
+            90,
+            893,
+            227
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            500,
+            242,
+            591,
+            258
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Motion Tokenizer",
+        "text_level": 1,
+        "bbox": [
+            500,
+            267,
+            671,
+            282
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The motion tokenizer aims to learn a codebook that can represent a wide range of motion sequences with the priority awareness of motion tokens. To achieve this, we propose a transformer-based VQ-VAE that leverages self-attention to capture long-range dependencies between motion frames and learns a stable and diverse codebook.",
+        "bbox": [
+            496,
+            291,
+            890,
+            381
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Fig. 2 shows the overview of our motion tokenizer. Given a motion sequence  $X \\in \\mathbb{R}^{T \\times d}$  with  $T$  frames and  $d$  dimensional motion representation, we reconstruct the motion sequence using a transformer autoencoder and learn an informative codebook  $Z = \\{z_k\\}_{k=1}^K$  with  $K$  entries. We use the transformer encoder  $E$  to compute the latent motion features  $\\boldsymbol{b} = \\{b_t\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}$ , where  $d$  is the dimension of the latent features. Then for each feature  $b_t$ , we query the codebook  $Z$  by measuring the distance between feature  $b_t$  and each entry in the codebook  $Z$ . To improve the codebook usage and avoid dead codes, we first normalize  $b_t$  and  $z_t$  and then measure the Euclidean distance between two normalized vectors in the unit sphere coordinate, which is formulated as:",
+        "bbox": [
+            496,
+            383,
+            893,
+            580
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {b} _ {t} = \\underset {b _ {t} \\in \\boldsymbol {b}} {\\arg \\min } \\| \\ell_ {2} (b _ {t}) - \\ell_ {2} \\left(z _ {k}\\right) \\| _ {2} ^ {2}, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            578,
+            594,
+            890,
+            622
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\ell_2$  is the  $l_{2}$  normalized function.",
+        "bbox": [
+            498,
+            628,
+            756,
+            643
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The standard optimization goal consists of three parts: the reconstruction loss, the embedding loss, and the commitment loss, which is formulated as",
+        "bbox": [
+            496,
+            645,
+            893,
+            690
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {v q} = \\left\\| X - \\tilde {X} \\right\\| _ {2} ^ {2} + \\left\\| Z - s g [ \\hat {Z} ] \\right\\| _ {2} ^ {2} \\tag {2} \\\\ + \\eta | | s g [ Z ] - \\hat {Z} | | _ {2} ^ {2}. \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            573,
+            700,
+            890,
+            752
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Moreover, to achieve an informative codebook with diverse and high usage of its entries, we apply an orthogonal regularization. Specifically, we calculate the distance between each pair of codebook entries to enforce orthogonality among all the codebook entries:",
+        "bbox": [
+            496,
+            763,
+            893,
+            839
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {o r t h}} (Z) = \\left\\| \\ell_ {2} (Z) ^ {\\top} \\ell_ {2} (Z) - I _ {K} \\right\\| _ {2} ^ {2}, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            568,
+            851,
+            890,
+            875
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{L}_2$  is the  $l_{2}$  normalized function.",
+        "bbox": [
+            500,
+            885,
+            761,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "14808",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/5b14b9f440f563c742bd7350990b93a1daf121fbfd30f65ec94ca918acd80905.jpg",
+        "image_caption": [
+            "Figure 2: Overview of our framework for our Motion Discrete Diffusion Model (M2DM). M2DM consists of a Transformer-based VQ-VAE (Sec. 3.1) and a discrete diffusion model (Sec. 3.2). Firstly, the Transformer-based motion tokenizer is learned to discretize motion sequences into tokens and reconstruct original motion sequences from tokens queried from the learnable codebook. Then the discrete diffusion model is learned to generate diverse motions conditioned on the textual descriptions. Two priority assessment strategies (Sec. 3.3) are applied to mask or replace motion tokens with less information at the beginning of the forward diffusion process and the most important motion tokens are corrupted in the last few diffusion steps. The diffusion model learns to restore the original motion token index conditioned on the given textual descriptions."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            98,
+            88,
+            880,
+            308
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Therefore, the overall training objective of vector quantization is defined as follows:",
+        "bbox": [
+            76,
+            450,
+            472,
+            479
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {v q} = \\left\\| X - \\tilde {X} \\right\\| _ {2} ^ {2} + \\| Z - s g [ \\hat {Z} ] \\| _ {2} ^ {2} \\tag {4} \\\\ + \\eta | | s g [ Z ] - \\hat {Z} | | _ {2} ^ {2} + \\delta \\mathcal {L} _ {\\text {o r t h}} (Z), \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            147,
+            502,
+            468,
+            551
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\tilde{X} = D(\\pmb {b})$ $D$  is the decoder,  $sg[\\cdot ]$  denotes the stopgradient operation,  $\\eta$  and  $\\delta$  are the weighting scalars.",
+        "bbox": [
+            76,
+            580,
+            470,
+            612
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Discrete Diffusion Models",
+        "text_level": 1,
+        "bbox": [
+            76,
+            643,
+            312,
+            657
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Generally speaking, the forward discrete diffusion process corrupts each element  $x_{t}$  at timestep  $t$  with  $K$  categories via fixed Markov chain  $q(x_{t}|x_{t - 1})$ , for instance, replace some tokens of  $x_{t - 1}$  or mask the token directly. After a fixed number of  $T$  timesteps, the forward process yields a sequence of increasingly noisy discrete tokens  $z_{1},\\ldots ,z_{T}$  of the same dimension of  $z_{0}$  and  $z_{T}$  becomes pure noise token. For the denoising procedure, the real  $x_0$  is restored sequentially based on the pure noisy token  $z_{T}$ . To be specific, for scalar random variable with  $K$  categories  $x_{t},x_{t - 1}\\in 1,\\dots K$ , the forward transition probabilities can be represented by the transition matrices  $[Q_{t}]_{ij} = q(x_{t} = i|x_{t - 1} = j)\\in \\mathbb{R}^{(K + 1)\\times (K + 1)}$ .  $[Q_{t}]_{i,j}$  is the Markov transition matrix from state  $t$  to state  $t - 1$  that is applied to each token in the sequence independently, which",
+        "bbox": [
+            76,
+            674,
+            472,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "can be written as:",
+        "bbox": [
+            500,
+            450,
+            619,
+            463
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\boldsymbol {Q} _ {t} = \\left[ \\begin{array}{c c c c c} \\alpha_ {t} + \\beta_ {t} & \\beta_ {t} & \\beta_ {t} & \\dots & 0 \\\\ \\beta_ {t} & \\alpha_ {t} + \\beta_ {t} & \\beta_ {t} & \\dots & 0 \\\\ \\beta_ {t} & \\beta_ {t} & \\alpha_ {t} + \\beta_ {t} & \\dots & 0 \\\\ \\vdots & \\vdots & \\vdots & & \\vdots \\\\ \\gamma_ {t} & \\gamma_ {t} & \\gamma_ {t} & \\dots & 1 \\end{array} \\right], \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            534,
+            472,
+            892,
+            561
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\alpha_{t} \\in [0,1]$  is the probability of retaining the token,  $\\gamma_{t}$  is the probability of replacing the original token to [MASK] token, leaving the probability  $\\beta_{t} = (1 - \\alpha_{t} - \\gamma_{t}) / K$  to be diffused. The forward Markov diffusion process for the whole token sequence is written as,",
+        "bbox": [
+            496,
+            571,
+            892,
+            647
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nq \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right) = \\operatorname {C a t} (\\boldsymbol {x}; \\boldsymbol {p} = \\boldsymbol {x} _ {t - 1} \\boldsymbol {Q} _ {t}) = \\boldsymbol {x} _ {t} \\boldsymbol {Q} _ {t} \\boldsymbol {x} _ {t - 1}, \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            516,
+            659,
+            892,
+            676
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\pmb{x}$  is the one-hot row vector identifying the token index and  $\\operatorname{Cat}(\\pmb{x};\\pmb{p})$  is a categorical distribution over the one-hot vector  $\\pmb{x}_t$  with probabilities  $\\pmb{p}$ .",
+        "bbox": [
+            496,
+            686,
+            890,
+            733
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Starting from the initial step  $x_0$ , the posterior diffusion process is formulated as:",
+        "bbox": [
+            496,
+            733,
+            890,
+            762
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} q \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}\\right) = \\frac {q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1} , \\boldsymbol {x} _ {0}\\right) q \\left(\\boldsymbol {x} _ {t - 1} , \\boldsymbol {x} _ {0}\\right)}{q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right)} (7) \\\\ = \\frac {\\left(\\boldsymbol {x} ^ {\\top} \\boldsymbol {Q} _ {t} \\boldsymbol {x} _ {t - 1}\\right) \\left(\\boldsymbol {x} _ {t - 1} ^ {T} \\bar {\\boldsymbol {Q}} _ {t - 1} \\boldsymbol {x} _ {0}\\right)}{\\boldsymbol {x} _ {t} ^ {\\top} \\bar {\\boldsymbol {Q}} _ {t} \\boldsymbol {x} _ {0}} (7) \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            535,
+            773,
+            890,
+            844
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "with  $\\bar{Q} = Q_{1}Q_{2}\\dots Q_{t}$",
+        "bbox": [
+            498,
+            854,
+            663,
+            869
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Recording to Markov chain, the intermediate step can be marginalized out and the derivation of the probability of  $\\boldsymbol{x}_t$",
+        "bbox": [
+            498,
+            869,
+            890,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "14809",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "at arbitrary timestep directly from  $\\pmb{x}_0$  can be formulated by",
+        "bbox": [
+            76,
+            90,
+            465,
+            107
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nq \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right) = \\boldsymbol {x} _ {t} ^ {\\top} \\bar {\\boldsymbol {Q}} \\boldsymbol {x} _ {0}. \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            199,
+            116,
+            468,
+            135
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The cumulative transition matrix  $\\bar{Q}$  and the probability  $q(\\pmb{x}_t|\\pmb{x}_0)$  are then computed as",
+        "bbox": [
+            76,
+            146,
+            468,
+            179
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\bar {Q} \\boldsymbol {x} _ {0} = \\bar {\\alpha} _ {t} \\boldsymbol {x} _ {0} + (\\bar {\\gamma} - \\bar {\\beta} _ {t}) \\boldsymbol {m} + \\bar {\\beta} _ {t} \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            161,
+            186,
+            468,
+            207
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $m$  is the one-hot vector for [MASK] and [PAD],  $\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_i$ ,  $\\bar{\\gamma}_t = 1 - \\prod_{i=1}^t (1 - \\gamma_i)$ , and  $\\bar{\\beta}_t = (1 - \\bar{\\alpha}_t - \\bar{\\gamma}_t) / K$ .",
+        "bbox": [
+            76,
+            217,
+            468,
+            250
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Regarding the reverse diffusion process, we train a transformer-like [47] denoising network  $p_{\\theta}(\\pmb{x}_{t-1}|\\pmb{x}_t,\\pmb{y})$  to estimate the posterior transition distribution  $q(\\pmb{x}_{t-1}|\\pmb{x}_t,\\pmb{x}_0)$ . The network is trained to minimize the variational lower bound (VLB) [42]:",
+        "bbox": [
+            75,
+            250,
+            470,
+            325
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {V L B} = \\mathbb {E} _ {q} \\left[ D _ {K L} \\left(q \\left(\\boldsymbol {x} _ {T} \\mid \\boldsymbol {x} _ {0}\\right) \\| p _ {\\theta} \\left(\\boldsymbol {x} _ {T}\\right)\\right) \\right] \\\\ + \\mathbb {E} _ {q} [ \\sum_ {t = 2} ^ {T} D _ {K L} (q (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}) \\| p _ {\\theta} (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t}, t)) ] \\\\ - \\log p _ {\\theta} \\left(\\boldsymbol {x} _ {0} \\mid \\boldsymbol {x} _ {1}\\right) \\tag {10} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            76,
+            334,
+            473,
+            428
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbb{E}_q[\\cdot ]$  denotes the expectation over the joint distribution  $q(\\pmb{x}_{0:T})$ .",
+        "bbox": [
+            75,
+            429,
+            470,
+            460
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3. Priority-Centric Denoising Process",
+        "text_level": 1,
+        "bbox": [
+            76,
+            468,
+            380,
+            484
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The designation of the noise schedule in the continuous domain, for instance, the linear schedule [20] and the cosine schedule [33], achieves the excellent performance of diffusion models. The noise can be easily controlled by the variation of Gaussian noise. For the discrete diffusion models, several noise schedules [22, 3] have been explored to control the data corruption and denoise the reverse process by choosing the transition matrix  $Q_{t}$ . For text-to-motion generation, however, such a schedule assumes all motion tokens carry the same amount of information and do not consider the difference among the discrete tokens extracted from the continuous motion sequence. Intuitively, the motion generation process would be in a progressive manner, where the most important motion tokens should appear in earlier steps for laying a solid foundation when denoising the noise in the discrete domain.",
+        "bbox": [
+            73,
+            492,
+            468,
+            733
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we apply the mask-and-replace strategy in Eq. (5) to corrupt each ordinary discrete token with a probability of  $\\gamma_{t}$  for masking [MASK] token and a probability of  $\\beta_{t}$  for uniform diffusion to a random token  $x_{k}$ , resulting in a remaining probability of  $\\alpha_{t} = 1 - K\\beta_{t} - \\gamma_{t}$  for retention. When corrupting across the forward process in the discrete domain, we expect that  $\\bar{\\alpha}_{t}^{i} < \\bar{\\alpha}_{t}^{j}$  if token  $x_{i}$  is more informative than  $x_{j}$  such that the tokens with high information emerge earlier than less informative motion tokens in the reverse process. Instead of using a normal corruption strategy with linearly increasing  $\\bar{\\gamma}_{t}$  and  $\\bar{\\beta}_{t}$ , we propose a",
+        "bbox": [
+            75,
+            734,
+            470,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "priority-score function  $F$  to assess the relative importance of information for a motion token sequence, which enables the recovery of the most informative tokens during the reverse process.",
+        "bbox": [
+            496,
+            90,
+            892,
+            151
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Given an original motion token sequences  $\\pmb{x}_0$  at timestep 0 with length  $N$ , we aim to analyze the importance of each token  $x_0^i$  in  $\\pmb{x}_0$ . Based on the information entropy theory, we calculate the information entropy for each motion token  $x_0^i$  as the importance score",
+        "bbox": [
+            496,
+            152,
+            893,
+            227
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF \\left(x _ {0} ^ {i}\\right) = \\frac {N H \\left(x _ {0} ^ {i}\\right)}{\\sum_ {j = 1} ^ {N} H \\left(x _ {0} ^ {j}\\right)}, \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            612,
+            239,
+            892,
+            280
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where",
+        "bbox": [
+            500,
+            291,
+            544,
+            303
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nH (x) = - \\sum_ {i = 1} ^ {K} p \\left(x _ {i}\\right) \\log p \\left(x _ {i}\\right), \\tag {12}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            591,
+            304,
+            892,
+            345
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "and  $K$  denotes the number of categories belonging to  $x_{i}$",
+        "bbox": [
+            498,
+            353,
+            874,
+            369
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After obtaining the importance score function and setting linearly increasing schedule  $\\bar{\\gamma}_t^i$  and  $\\bar{\\beta}_t^i$  for all tokens,  $\\bar{\\gamma}_t^i$  and  $\\bar{\\beta}_t^i$  for specific token  $x_0^i$  can be further updated by",
+        "bbox": [
+            498,
+            369,
+            890,
+            415
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\bar {\\gamma} _ {t} ^ {i} \\Leftarrow \\bar {\\gamma} _ {t} ^ {i} \\cdot \\sin \\frac {t \\pi}{T} \\cdot F \\left(x _ {0} ^ {i}\\right), \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            607,
+            426,
+            892,
+            455
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "and",
+        "bbox": [
+            500,
+            468,
+            529,
+            479
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\bar {\\beta} ^ {i} \\Leftarrow \\bar {\\beta} ^ {i} \\cdot \\sin \\frac {t \\pi}{T} \\cdot F \\left(x _ {0} ^ {i}\\right), \\tag {14}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            607,
+            479,
+            892,
+            508
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $x_0^i$  is the  $i$ -th token at timestep 0.",
+        "bbox": [
+            498,
+            517,
+            764,
+            532
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Now we introduce two solutions for obtaining  $p(x_{i})$ .",
+        "bbox": [
+            517,
+            532,
+            867,
+            549
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Static assessment. Analogously to computing the entropy for each word in natural language processing, we estimate the probability  $p(x_{i})$  for each motion token by counting its frequency in the quantized motion sequences over the entire dataset.",
+        "bbox": [
+            496,
+            571,
+            892,
+            646
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/3d0befa6afe7389516bbd25840dc6cd2e93ada0b6bc000589e1a52883fd2441f.jpg",
+        "image_caption": [
+            "Figure 3: Illustration for dynamic assessment strategy. The gray blocks are masked tokens and other colorful blocks are motion tokens."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            566,
+            660,
+            826,
+            823
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "14810",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Dynamic assessment. Rather than counting motion token frequency and calculating the importance of each token, we further propose to learn an agent to estimate the importance of each token in the token sequences by reinforcement learning strategy. As shown in Fig. 3, given a series of complete motion tokens  $\\boldsymbol{x} = \\{x_{n}\\}_{n=1}^{N}$  that are queried from the codebook at current state  $s_{t} \\in S$ , the scorer samples the motion tokens  $\\tilde{\\boldsymbol{x}} = \\{x_{n}\\}_{n=1}^{N'}$  and tries to minimize the reconstruction error between the original motion sequences  $\\boldsymbol{m}$  and the motion sequences decoded from the sampled tokens  $\\tilde{\\boldsymbol{m}}$ .",
+        "bbox": [
+            75,
+            90,
+            472,
+            244
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "State includes the currently selected motion tokens  $\\tilde{\\pmb{x}} = \\{x_{n}\\}_{n = 1}^{N^{\\prime}}$ , the complete motion tokens  $\\pmb {x} = \\{x_{n}\\}_{n = 1}^{N}$ , and the corresponding reconstructed continuous motions  $\\tilde{m}$  and  $m$ . The reconstructed continuous motion sequences  $\\tilde{m}$  and  $m$  are obtained by decoding the motion tokens by VQ decoder. The action is to sample one motion token from the complete motion token sequence at each stage. To be specific, given a motion token sequence, the agent  $\\pi_{\\theta}(a_t|s_t)$  tries to sample a motion token that can recover the continuous motion sequences with the minimum reconstruction error. The reward measures the reconstruction difference between  $\\tilde{m}$  and  $m$  with the minimum sample times.",
+        "bbox": [
+            75,
+            251,
+            472,
+            435
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/64487833a06942ff58a25558356f6c19cec3d49745c206c3d6a61b15fa11e306.jpg",
+        "image_caption": [
+            "Figure 4: Priority-centric denoising schedule for motion tokens."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            89,
+            460,
+            455,
+            588
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "After the agent is well-trained, we calculate and restore the sampling order for each motion token sequence by the agent  $\\pi_{\\theta}(a_t|s_t)$  and compute the noise schedule for each motion token in the codebook by Eq. (13) and (14). Fig. 4 shows the noise schedule for different motion tokens w.r.t. the corresponding priorities. As shown in Fig. 4, high-priority tokens (orange and green curves) are encouraged to be corrupted at the end of the forward process such that the learnable reverse process follows a primary-to-secondary manner and high-priority tokens will be restored first. It is worth noting that both strategies are applied to design noise schedules for each token, which can be calculated and stored in advance such that the agent will not join the forward and reverse diffusion processes. Thus, there is no extra computational cost for the diffusion model during the training and inference stages.",
+        "bbox": [
+            75,
+            659,
+            472,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            90,
+            633,
+            107
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Datasets and Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            500,
+            114,
+            785,
+            130
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Our experiments are conducted on two standard text-to-motion datasets for text-to-motion generation: HumanML3D [16] and KIT-ML [36]. We follow the evaluation metrics provided in [16].",
+        "bbox": [
+            498,
+            138,
+            893,
+            199
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "HumanML3D. HumanML3D [16] is currently the largest 3D human motion dataset that covers a broad range of daily human actions, for instance, exercising and dancing. The dataset contains 14,616 motion sequences and 44970 text descriptions. The motion sequences are processed to 20 frame-per-second (FPS) and cropped to 10 seconds if the motion sequences are longer than 10 seconds, resulting in duration ranges from 2 to 10 seconds. For each motion clip, the corresponding number of descriptions is at least three.",
+        "bbox": [
+            496,
+            219,
+            893,
+            354
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "KIT Motion-Language (KIT-ML). KIT-ML dataset [36] contains 3,911 3D human motion clips with 6,278 text descriptions. For each motion clip, one to four textual descriptions are provided. The motion sequences are collected from the KIT dataset [31] and the CMU dataset [1] with the down-sampled 12.5 FPS.",
+        "bbox": [
+            496,
+            375,
+            893,
+            465
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Both of the aforementioned datasets are split into training, validation, and testing datasets with the proportions of  $80\\%$ ,  $5\\%$ , and  $15\\%$ .",
+        "bbox": [
+            498,
+            465,
+            893,
+            511
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation details. Both our VQ-encoder and decoder consist of 4 transformer layers with 8 heads and the dimension is 512. For the HumanML3D [16] and KIT-ML [36] datasets, the motion sequences are cropped to 64 frames for training. We apply Adam with  $\\beta_{1} = 0.9$ ,  $\\beta_{2} = 0.999$ , weight decay is 1e-4. The learning rate is 1e-4 after warmup linearly for 5 epochs. The number of tokens in the codebook is 8192. We train our VQ-VAE with a batch size of 1024 across 8 Tesla T4 for a total 100 training epochs for 8 hours.",
+        "bbox": [
+            496,
+            532,
+            893,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For the discrete diffusion model, we set timesteps  $T = 100$  and the network is trained using Adam [25] with  $\\beta_{1} = 0.9$ ,  $\\beta_{2} = 0.999$ . The learning rate reaches 2e-4 after 2000 iterations of a linear warmup schedule. The transformer-based discrete diffusion model consists of 12 transformer layers with 8 heads, and the dimension is 512. The text features are extracted by pre-trained CLIP [37] model. The diffusion model is trained with a batch size of 64 across 8 Tesla T4 for 100k iterations for 36 hours.",
+        "bbox": [
+            496,
+            667,
+            893,
+            804
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation metrics. Following the evaluation protocols provided in previous works [16, 55, 44, 5], we evaluate the correspondences between motion and language using deep multimodal features with the pre-trained models in [16] by the following metrics: 1) Generation diversity: We randomly",
+        "bbox": [
+            496,
+            825,
+            893,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "14811",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/f9cd13004f2232ed68f65266461c6c6dc29584485240cf341a67533677a4ea85.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "${}^{8}$  reports results using ground-truth motion length."
+        ],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"3\">R-Precision ↑</td><td rowspan=\"2\">FID ↓</td><td rowspan=\"2\">MM-Dist ↓</td><td rowspan=\"2\">Diversity →</td><td rowspan=\"2\">MModality ↑</td></tr><tr><td>Top-1</td><td>Top-2</td><td>Top-3</td></tr><tr><td>Real motion</td><td>0.511±.003</td><td>0.703±.003</td><td>0.797±.002</td><td>0.002±.000</td><td>2.974±.008</td><td>9.503±.065</td><td>-</td></tr><tr><td>Our VQ-VAE</td><td>0.508±.002</td><td>0.691±.002</td><td>0.791±.003</td><td>0.63±.001</td><td>3.015±.010</td><td>9.577±.081</td><td>-</td></tr><tr><td>Our VQ-VAE w/o l2</td><td>0.494±.003</td><td>0.685±.002</td><td>0.772±.003</td><td>0.70±.001</td><td>3.251±.008</td><td>9.525±.092</td><td>-</td></tr><tr><td>Our VQ-VAE w/o l2, Lorth</td><td>0.485±.002</td><td>0.671±.003</td><td>0.752±.003</td><td>0.79±.001</td><td>3.378±.012</td><td>9.511±.095</td><td>-</td></tr><tr><td>Seq2Seq [30]</td><td>0.180±.002</td><td>0.300±.002</td><td>0.396±.002</td><td>11.75±.035</td><td>5.529±.007</td><td>6.223±.061</td><td>-</td></tr><tr><td>Language2Pose [2]</td><td>0.246±.002</td><td>0.387±.002</td><td>0.486±.002</td><td>11.02±.046</td><td>5.296±.008</td><td>7.676±.058</td><td>-</td></tr><tr><td>Text2Gesture [4]</td><td>0.165±.001</td><td>0.267±.002</td><td>0.345±.002</td><td>5.012±.030</td><td>6.030±.008</td><td>6.409±.071</td><td>-</td></tr><tr><td>Hier [12]</td><td>0.301±.002</td><td>0.425±.002</td><td>0.552±.004</td><td>6.532±.024</td><td>5.012±.018</td><td>8.332±.042</td><td>-</td></tr><tr><td>MoCoGAN [45]</td><td>0.037±.000</td><td>0.072±.001</td><td>0.106±.001</td><td>94.41±.021</td><td>9.643±.006</td><td>0.462±.008</td><td>0.019±.000</td></tr><tr><td>Dance2Music [27]</td><td>0.033±.000</td><td>0.065±.001</td><td>0.097±.001</td><td>66.98±.016</td><td>8.116±.006</td><td>0.725±.011</td><td>0.043±.001</td></tr><tr><td>TM2T [17]</td><td>0.424±.003</td><td>0.618±.003</td><td>0.729±.002</td><td>1.501±.017</td><td>3.467±.011</td><td>8.589±.076</td><td>2.424±.093</td></tr><tr><td>Guo et al. [16]</td><td>0.455±.003</td><td>0.636±.003</td><td>0.736±.002</td><td>1.087±.021</td><td>3.347±.008</td><td>9.175±.083</td><td>2.219±.074</td></tr><tr><td>MDM [44]$</td><td>-</td><td>-</td><td>0.611±.007</td><td>0.544±.044</td><td>5.566±.027</td><td>9.559±.086</td><td>2.799±.072</td></tr><tr><td>MotionDiffuse [56]$</td><td>0.491±.001</td><td>0.681±.001</td><td>0.782±.001</td><td>0.630±.001</td><td>3.113±.001</td><td>9.410±.049</td><td>1.553±.042</td></tr><tr><td>T2M-GPT [55]</td><td>0.492±.003</td><td>0.679±.002</td><td>0.775±.002</td><td>0.141±.005</td><td>3.121±.009</td><td>9.722±.082</td><td>1.831±.048</td></tr><tr><td>MLD [5]</td><td>0.481±.003</td><td>0.673±.003</td><td>0.772±.002</td><td>0.473±.013</td><td>3.196±.010</td><td>9.724±.082</td><td>2.413±.079</td></tr><tr><td>M2DM (Ours) w. linear schedule</td><td>0.452±.003</td><td>0.678±.002</td><td>0.752±.003</td><td>0.417±.004</td><td>3.167±.008</td><td>9.972±.089</td><td>3.562±.071</td></tr><tr><td>M2DM (Ours) w. static assessment</td><td>0.492±.003</td><td>0.671±.003</td><td>0.775±.003</td><td>0.395±.005</td><td>3.116±.008</td><td>9.937±.075</td><td>3.413±.035</td></tr><tr><td>M2DM (Ours) w. dynamic assessment</td><td>0.497±.003</td><td>0.682±.002</td><td>0.763±.003</td><td>0.352±.005</td><td>3.134±.010</td><td>9.926±.073</td><td>3.587±.072</td></tr></table>",
+        "bbox": [
+            84,
+            88,
+            885,
+            383
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Comparison with the state-of-the-art methods on HumanML3D [16] test set. The evaluation metrics are evaluated by the motion encoder from Guo et al. [16]. The right row  $\\rightarrow$  means the closer to the real motion the better.",
+        "bbox": [
+            75,
+            397,
+            890,
+            428
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "sample 300 pairs of motions from the motion pool and extracted the selected motion features by the extractor provided by [16]. Euclidean distances of the motion pairs are computed for measuring the motion diversity. 2) R-Precision: Provided one motion representation and 32 textual descriptions, the Euclidean distances between the description feature and each motion feature in the candidates are calculated and ranked. Top-k  $(k = 1,2,3)$  average accuracy of motion-to-text retrieval is reported. 3) Frechet Inception Distance (FID) [19]: FID is the principal metric to evaluate the feature descriptions between the generated motions and the ground-truth motions by the feature extractor [16]. 4) Multimodal Distance (MM-Dist): MM-Dist measures the average Euclidean distances between each text feature and the generated motion feature. 5) MModality measures the generation diversity within the same textual descriptions. Please refer to the supplementary material for a detailed introduction.",
+        "bbox": [
+            75,
+            454,
+            472,
+            710
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2. Comparisons with state-of-the-art approaches",
+        "text_level": 1,
+        "bbox": [
+            76,
+            720,
+            465,
+            738
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We compare our methods with other state-of-the-art methods [30, 2, 4, 12, 45, 27, 17, 16, 44, 56, 55, 5] on the HumanML3D [16] and KIT-ML [36] datasets.",
+        "bbox": [
+            75,
+            744,
+            470,
+            789
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Quantitative comparisons. For each metric, we repeat the evaluation 20 times and report the average with  $95\\%$  confidence interval, followed by [16]. Most of the results are borrowed from [55]. Tab. 1 and Tab. 2 summarize the comparison results on the HumanML3D [16] and the KIT-ML [36] datasets, respectively. Firstly, the metrics scores for",
+        "bbox": [
+            75,
+            810,
+            470,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "our reconstruction motion achieve close performance compared with real motion, which suggests the distinctive and high-quality discrete motion tokens in the codebook learned by our Transformer-based VQ-VAE. The performance of our VQ-VAE degrades when the normalized function  $\\ell_2$  and the orthogonal regularization term  $\\mathcal{L}_{\\mathrm{orth}}$  are removed. For text-to-motion generation, our approach achieves compatible performance (R-Precision, FID, and MModality) compared to other state-of-the-art methods. Moreover, compared with MDM [44] and MotionDiffuse [56] which evaluate their performance with the ground-truth motion lengths, our model can generate motion sequences with arbitrary lengths conditioned on the given textual descriptions since the [MASK] and [PAD] tokens are introduced in the discrete diffusion process.",
+        "bbox": [
+            496,
+            454,
+            893,
+            680
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, we conduct more experiments to evaluate the generation performance conditioned on different lengths of textual descriptions. The testing dataset is split into three parts according to the length of the textual descriptions: less than 15 words, between 15 words and 30 words, and more than 30 words. Tab. 3 shows the quantitative results on the aforementioned three sub-sets. Both TM2T [17] and MLD [5] suffer degradation when the textual descriptions are long and complex, especially the descriptions are more than 30 words. Our method significantly outperforms the other two methods with various metrics and there is not much degradation when the descriptions are more complex. The quantitative results demonstrate the effectiveness of our proposed primary-to-secondary diffusion manner.",
+        "bbox": [
+            496,
+            689,
+            893,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "14812",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/ae6150a8e1c2a9bdf849573f0306b72d6a23f0def0aaa4aaa7b5b1caea5b8038.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "${}^{8}$  reports results using ground-truth motion length."
+        ],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"3\">R-Precision ↑</td><td rowspan=\"2\">FID ↓</td><td rowspan=\"2\">MM-Dist ↓</td><td rowspan=\"2\">Diversity →</td><td rowspan=\"2\">MModality ↑</td></tr><tr><td>Top-1</td><td>Top-2</td><td>Top-3</td></tr><tr><td>Real motion</td><td>0.424±.005</td><td>0.649±.006</td><td>0.779±.006</td><td>0.031±.004</td><td>2.788±.012</td><td>11.08±.097</td><td>-</td></tr><tr><td>Our VQ-VAE (Recons.)</td><td>0.417±.004</td><td>0.621±.003</td><td>0.741±.006</td><td>0.413±.009</td><td>2.772±.018</td><td>10.851±.105</td><td>-</td></tr><tr><td>Seq2Seq [30]</td><td>0.103±.003</td><td>0.178±.005</td><td>0.241±.006</td><td>24.86±.348</td><td>7.960±.031</td><td>6.744±.106</td><td>-</td></tr><tr><td>Language2Pose [2]</td><td>0.221±.005</td><td>0.373±.004</td><td>0.483±.005</td><td>6.545±.072</td><td>5.147±.030</td><td>9.073±.100</td><td>-</td></tr><tr><td>Text2Gesture [4]</td><td>0.156±.004</td><td>0.255±.004</td><td>0.338±.005</td><td>12.12±.183</td><td>6.964±.029</td><td>9.334±.079</td><td>-</td></tr><tr><td>Hier [12]</td><td>0.255±.006</td><td>0.432±.007</td><td>0.531±.007</td><td>5.203±.107</td><td>4.986±.027</td><td>9.563±.072</td><td>-</td></tr><tr><td>MoCoGAN [45]</td><td>0.022±.002</td><td>0.042±.003</td><td>0.063±.003</td><td>82.69±.242</td><td>10.47±.012</td><td>3.091±.043</td><td>0.250±.009</td></tr><tr><td>Dance2Music [27]</td><td>0.031±.002</td><td>0.058±.002</td><td>0.086±.003</td><td>115.4±.240</td><td>10.40±.016</td><td>0.241±.004</td><td>0.062±.002</td></tr><tr><td>TM2T [17]</td><td>0.280±.005</td><td>0.463±.006</td><td>0.587±.005</td><td>3.599±.153</td><td>4.591±.026</td><td>9.473±.117</td><td>3.292±.081</td></tr><tr><td>Guo et al. [16]</td><td>0.361±.006</td><td>0.559±.007</td><td>0.681±.007</td><td>3.022±.107</td><td>3.488±.028</td><td>10.72±.145</td><td>2.052±.107</td></tr><tr><td>MDM [44]$</td><td>-</td><td>-</td><td>0.396±.004</td><td>0.497±.021</td><td>9.191±.022</td><td>10.847±.109</td><td>1.907±.214</td></tr><tr><td>MotionDiffuse [56]$</td><td>0.417±.004</td><td>0.621±.004</td><td>0.739±.004</td><td>1.954±.062</td><td>2.958±.005</td><td>11.10±.143</td><td>0.730±.013</td></tr><tr><td>T2M-GPT [55]</td><td>0.416±.006</td><td>0.627±.006</td><td>0.745±.006</td><td>0.514±.029</td><td>3.007±.023</td><td>10.921±.108</td><td>1.570±.039</td></tr><tr><td>MLD [5]</td><td>0.390±.003</td><td>0.609±.003</td><td>0.734±.002</td><td>0.404±.013</td><td>3.204±.010</td><td>10.8±.082</td><td>2.192±.079</td></tr><tr><td>M2DM (Ours) w. linear schedule</td><td>0.405±.003</td><td>0.629±.005</td><td>0.739±.004</td><td>0.502±.049</td><td>3.012±.015</td><td>11.375±.079</td><td>3.273±.045</td></tr><tr><td>M2DM (Ours) w. static assessment</td><td>0.417±.006</td><td>0.625±.003</td><td>0.741±.006</td><td>0.521±.041</td><td>3.024±.018</td><td>11.373±.081</td><td>3.317±.031</td></tr><tr><td>M2DM (Ours) w. dynamic assessment</td><td>0.416±.004</td><td>0.628±.004</td><td>0.743±.004</td><td>0.515±.029</td><td>3.015±.017</td><td>11.417±.97</td><td>3.325±.37</td></tr></table>",
+        "bbox": [
+            84,
+            88,
+            885,
+            358
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/df198118ee40c61ad1c53f774873c14103d6ce10cad67b2ec089f41744a4eff8.jpg",
+        "table_caption": [
+            "Table 2: Comparison with the state-of-the-art methods on KIT-ML [36] test set. The evaluation metrics are evaluated by the motion encoder from Guo et al. [16]. The right row  $\\rightarrow$  means the closer to the real motion the better."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>TM2T</td><td>MDM</td><td>T2M-GPT</td><td>MLD</td><td>Ours</td></tr><tr><td colspan=\"5\">a person grabs something, turns around, walks forward, squats, stands up, walks forward, squats and stands up again.</td></tr><tr><td></td><td></td><td></td><td></td><td></td></tr><tr><td colspan=\"5\">a person puts both hands by their face in a defensive stance, then kicks with their right foot.</td></tr><tr><td></td><td></td><td></td><td></td><td></td></tr><tr><td colspan=\"5\">a person jumps and turns around in the air to face behind him.</td></tr><tr><td></td><td></td><td></td><td></td><td></td></tr><tr><td colspan=\"5\">a person walks forward briskly, crouches and lifts an object with both hands, puts the object down in a higher position, then stands up and turn right.</td></tr><tr><td></td><td></td><td></td><td></td><td></td></tr><tr><td colspan=\"5\">a person takes something, put it down and throws something.</td></tr><tr><td></td><td></td><td></td><td></td><td></td></tr></table>",
+        "bbox": [
+            84,
+            417,
+            888,
+            777
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 5: Qualitative comparison of the state-of-the-art methods on the HumanML3D [16] dataset. We compare our generations with TM2T [17], MDM [44], T2M-GPT [55], and MLD [5]. The color of human mesh goes from light to dark over time.",
+        "bbox": [
+            75,
+            789,
+            892,
+            821
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Qualitative comparisons. Fig. 5 qualitatively compares the generated motions from the same textual descriptions. Most of the methods suffer from motion freezing if the tex",
+        "bbox": [
+            75,
+            845,
+            472,
+            892
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "tual descriptions consist of several complex commands. For example, the first row in Fig. 5 shows the generation results from a long textual description with various duplicate action",
+        "bbox": [
+            498,
+            845,
+            893,
+            892
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "14813",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/420cc7244abe45a33681b07f0d5cdf88593415861b117205df0de0cf0fd1f147.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Text Length</td><td rowspan=\"2\">Methods</td><td colspan=\"3\">R-Precision ↑</td><td rowspan=\"2\">FID ↓</td><td rowspan=\"2\">MM-Dist ↓</td><td rowspan=\"2\">Diversity →</td><td rowspan=\"2\">MModality ↑</td></tr><tr><td>Top-1</td><td>Top-2</td><td>Top-3</td></tr><tr><td rowspan=\"3\">All</td><td>TM2T [17]</td><td>0.424±0.003</td><td>0.618±0.003</td><td>0.729±0.002</td><td>1.501±0.017</td><td>3.467±0.011</td><td>8.580±0.076</td><td>2.424±0.093</td></tr><tr><td>MLD [5]</td><td>0.481±0.003</td><td>0.673±0.003</td><td>0.772±0.002</td><td>0.473±0.013</td><td>3.196±0.010</td><td>9.724±0.082</td><td>2.413±0.079</td></tr><tr><td>M2DM (Ours)</td><td>0.497±0.003</td><td>0.689±0.002</td><td>0.763±0.003</td><td>0.352±0.005</td><td>3.134±0.010</td><td>9.926±0.073</td><td>3.587±0.072</td></tr><tr><td rowspan=\"3\">&lt; 15 words</td><td>TM2T [17]</td><td>0.433±0.003</td><td>0.627±0.003</td><td>0.738±0.003</td><td>1.592±0.017</td><td>3.446±0.008</td><td>8.677±0.077</td><td>2.520±0.051</td></tr><tr><td>MLD [5]</td><td>0.492±0.003</td><td>0.680±0.004</td><td>0.779±0.004</td><td>0.469±0.014</td><td>3.191±0.011</td><td>9.770±0.010</td><td>2.586±0.086</td></tr><tr><td>M2DM (Ours)</td><td>0.508±0.003</td><td>0.685±0.003</td><td>0.767±0.004</td><td>0.347±0.015</td><td>3.129±0.013</td><td>9.915±0.068</td><td>3.594±0.077</td></tr><tr><td rowspan=\"3\">15-30 words</td><td>TM2T [17]</td><td>0.406±0.003</td><td>0.596±0.002</td><td>0.708±0.002</td><td>1.425±0.014</td><td>3.528±0.007</td><td>8.000±0.051</td><td>2.477±0.057</td></tr><tr><td>MLD [5]</td><td>0.406±0.005</td><td>0.592±0.005</td><td>0.701±0.006</td><td>0.480±0.020</td><td>3.591±0.020</td><td>9.089±0.083</td><td>2.727±0.094</td></tr><tr><td>M2DM (Ours)</td><td>0.493±0.003</td><td>0.681±0.004</td><td>0.761±0.004</td><td>0.353±0.031</td><td>3.135±0.014</td><td>9.926±0.073</td><td>3.588±0.081</td></tr><tr><td rowspan=\"3\">&gt;30 words</td><td>TM2T [17]</td><td>0.353±0.009</td><td>0.536±0.009</td><td>0.650±0.007</td><td>1.779±0.061</td><td>3.577±0.024</td><td>7.411±0.071</td><td>2.612±0.055</td></tr><tr><td>MLD [5]</td><td>0.289±0.010</td><td>0.466±0.012</td><td>0.581±0.015</td><td>0.947±0.066</td><td>3.870±0.039</td><td>8.347±0.100</td><td>2.916±0.063</td></tr><tr><td>M2DM (Ours)</td><td>0.491±0.004</td><td>0.677±0.004</td><td>0.758±0.004</td><td>0.356±0.038</td><td>3.138±0.025</td><td>9.930±0.088</td><td>3.585±0.071</td></tr></table>",
+        "bbox": [
+            80,
+            88,
+            480,
+            202
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Comparison with the state-of-the-art methods on HumanML3D [16] test set. The evaluation metrics are evaluated by the motion encoder from Guo et al. [16]. The right row  $\\rightarrow$  means the closer to the real motion the better. The first three rows are the evaluation results on the whole testing dataset. The rest are the evaluation results from textual descriptions with different lengths.",
+        "bbox": [
+            75,
+            212,
+            473,
+            319
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "commands. MDM [44] and MLD [5] generate fewer semantic motions compared with the given description. With the help of compact discrete motion tokens, TM2T [17] and T2M-GPT [55] achieve better results compared with MDM [44] and MLD [5]. Our motion generation results conform to the textual descriptions and exhibit a high degree of diversity compared with other state-of-the-art methods, which validates the efficacy of our proposed method.",
+        "bbox": [
+            75,
+            344,
+            472,
+            467
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "4.3. Ablation studies",
+        "text_level": 1,
+        "bbox": [
+            76,
+            474,
+            238,
+            491
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We delved into the significance of various quantization techniques, as delineated in the initial four rows of Tab.1. The findings indicate that a rudimentary Transformer-based VQ-VAE struggles to reconstruct credible motion sequences, primarily due to the limited efficacy of the codebook. However, the incorporation of the  $\\ell_2$  norm and  $\\mathcal{L}_{\\mathrm{orth}}$  ensures that each motion token acquires a distinct motion representation. We further scrutinized the utilization of each motion token on the HumanML3D [16] test set, with the outcomes depicted in Fig.6. As evidenced by Fig.6a and 6b, both the vanilla CNN-based VQ-VAE and Transformer-based VQ-VAE grapple with sub-optimal codebook usage. This is because the tokens in the codebook aren't inherently driven to develop a distinguishing motion representation. Yet, with the aid of the  $\\ell_2$  norm and  $\\mathcal{L}_{\\mathrm{orth}}$ , there's a marked surge in the codebook token utilization. Moreover, when juxtaposing the codebook usage in Fig.6c with that in Fig.6d, the latter exhibits a more equitably distributed frequency, attributable to the orthogonal regularization term  $\\mathcal{L}_{\\mathrm{orth}}$ .",
+        "bbox": [
+            75,
+            498,
+            473,
+            787
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            76,
+            799,
+            196,
+            814
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this study, we introduced the priority-centric motion discrete diffusion model (M2DM) tailored for text-to-motion generation. Our model excels in producing varied and life-like human motions, aligning seamlessly with the input text. This is achieved by mastering a discrete motion representa",
+        "bbox": [
+            75,
+            824,
+            472,
+            902
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/fc329d554b712ebab8445885af79b588165483f118280db747c3031a33d97614.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            88,
+            694,
+            183
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/28883f5084cde44709e4afc09a8ae2903a48ef63c332b77913d0e163e9082e7b.jpg",
+        "image_caption": [
+            "(a) 1D-CNN VQ-VAE.",
+            "Figure 6: Codebook usage for different training strategies on HumanML3D [16] test set. We train our VQ-VAE (Reconstruction) with different training strategy on the HumanML3D [16] dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            205,
+            694,
+            299
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/c0e1017151f31bbe46a5760b61b3892cd922ab459b8f8b4271862dbc78c61cd1.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            94,
+            890,
+            183
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/e99bc9909e6b52f64dded8cc3908bfbee41524111947c4e675d4ffad96c19899.jpg",
+        "image_caption": [
+            "(c) Our VQ-VAE w.  $\\ell_2$  norm.",
+            "(b) Our VQ-VAE.",
+            "(d) Our VQ-VAE w.  $\\ell_2$  and  $\\mathcal{L}_{\\mathrm{orth}}$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            205,
+            890,
+            297
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "tion through a Transformer-based VQ-VAE and implementing a priority-informed noise schedule within the motion discrete diffusion framework. Additionally, we unveiled two innovative techniques for gauging the priority or significance of each motion token. Experimental evaluations across two datasets underscore our model's competitive edge, outpacing existing methodologies in R-Precision and diversity, particularly with intricate textual narratives.",
+        "bbox": [
+            496,
+            419,
+            895,
+            541
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            500,
+            556,
+            660,
+            574
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This project is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award Number: MOE-T2EP20122-0006).",
+        "bbox": [
+            498,
+            584,
+            893,
+            628
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            500,
+            645,
+            596,
+            660
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Cmu graphics lab motion capture database. http://mocap.cs.cmu.edu/.6",
+            "[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 7, 8",
+            "[3] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34:17981-17993, 2021. 3, 5",
+            "[4] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE virtual reality and 3D user interfaces (VR), pages 1-10. IEEE, 2021. 7, 8"
+        ],
+        "bbox": [
+            506,
+            671,
+            895,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "14814",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[5] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. arXiv preprint arXiv:2212.04048, 2022. 1, 3, 6, 7, 8, 9",
+            "[6] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. arXiv preprint arXiv:2212.04495, 2022. 3",
+            "[7] Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341, 2020. 3",
+            "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3",
+            "[9] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3",
+            "[10] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16091-16101, 2023. 3",
+            "[11] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. arXiv preprint arXiv:2305.10924, 2023. 3",
+            "[12] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1396-1406, 2021. 7, 8",
+            "[13] Kehong Gong, Dongze Lian, Heng Chang, Chuan Guo, Zihang Jiang, Xinxin Zuo, Michael Bi Mi, and Xinchao Wang. TM2D: Bimodality Driven 3D Dance Generation via Music-Text Integration. In IEEE/CVF International Conference on Computer Vision, 2023. 3",
+            "[14] Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and LingPeng Kong. Diffuseq: Sequence to sequence text generation with diffusion models. arXiv preprint arXiv:2210.08933, 2022. 1, 3",
+            "[15] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 1",
+            "[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 3, 6, 7, 8, 9",
+            "[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 580-597. Springer, 2022. 1, 3, 7, 8, 9"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[18] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 1, 3",
+            "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7",
+            "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1, 3, 5",
+            "[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23(47):1-33, 2022. 3",
+            "[22] Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forre, and Max Welling. Argmax flows and multinomial diffusion: Towards non-autoregressive language models. arXiv preprint arXiv:2102.05379, 3(4):5, 2021. 1, 3, 5",
+            "[23] Leslie Pack Kaelbling, Michael L Littman, and Andrew W Moore. Reinforcement learning: A survey. Journal of artificial intelligence research, 4:237-285, 1996. 2",
+            "[24] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. arXiv preprint arXiv:2209.00349, 2022. 3",
+            "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6",
+            "[26] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3",
+            "[27] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. Advances in neural information processing systems, 32, 2019. 1, 3, 7, 8",
+            "[28] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1272-1279, 2022. 1, 3",
+            "[29] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 1, 3",
+            "[30] Angela S. Lin, Lemeng Wu, Rodolfo Corona, Kevin Tai, Qixing Huang, and Raymond J. Mooney. Generating animated videos of human activities from natural language descriptions. In Proceedings of the Visually Grounded Interaction and Language Workshop at NeurIPS 2018, December 2018. 7, 8",
+            "[31] Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. The kit whole-body human motion database. In 2015 International Conference on Advanced Robotics (ICAR), pages 329-336. IEEE, 2015. 6",
+            "[32] Larry R Medsker and LC Jain. Recurrent neural networks. Design and Applications, 5:64-67, 2001. 3"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "14815",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 3, 5",
+            "[34] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 1, 3",
+            "[35] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXII, pages 480–497. Springer, 2022. 1",
+            "[36] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236–252, 2016. 2, 6, 7, 8",
+            "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 6",
+            "[38] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 1, 3",
+            "[39] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3",
+            "[40] Sucheng Ren, Xingyi Yang, Songhua Liu, and Xinchao Wang. SG-Former: Self-guided Transformer with Evolving Token Reallocation. In IEEE International Conference on Computer Vision, 2023. 3",
+            "[41] Sucheng Ren, Daquan Zhou, Shengfeng He, Jiashi Feng, and Xinchao Wang. Shunted Self-Attention via Multi-Scale Token Aggregation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3",
+            "[42] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3, 5",
+            "[43] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXII, pages 358–374. Springer, 2022. 1",
+            "[44] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 1, 3, 6, 7, 8, 9",
+            "[45] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 7, 8"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[46] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 2, 3",
+            "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 3, 5",
+            "[48] Xingyi Yang and Xinchao Wang. Diffusion Model as Representation Learner. In IEEE/CVF International Conference on Computer Vision, 2023. 3",
+            "[49] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In European Conference on Computer Vision, 2022. 3",
+            "[50] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion Probabilistic Model Made Slim. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3",
+            "[51] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In Advances in neural information processing systems, 2022. 3",
+            "[52] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 3",
+            "[53] Runpeng Yu, Songhua Liu, Xingyi Yang, and Xinchao Wang. Distribution Shift Inversion for Out-of-Distribution Prediction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3",
+            "[54] Weihao Yu, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, and Xinchao Wang. Metaformer Baselines for Vision. arXiv preprint arXiv:2210.13452, 2022. 3",
+            "[55] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023. 1, 3, 6, 7, 8, 9",
+            "[56] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 1, 3, 7, 8"
+        ],
+        "bbox": [
+            501,
+            92,
+            892,
+            686
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "14816",
+        "bbox": [
+            480,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2023/Priority-Centric Human Motion Generation in Discrete Latent Space/58c7e478-f288-424e-a702-051324564633_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fae5046f9be915ad0f6e7b3fb9813137b4249af6a135f5d498ad59b8a90e2724
+size 5824223

2023/Priority-Centric Human Motion Generation in Discrete Latent Space/full.md

@@ -0,0 +1,355 @@
+# Priority-Centric Human Motion Generation in Discrete Latent Space
+
+Hanyang Kong $^{1}$ , Kehong Gong $^{1}$ , Dongze Lian $^{1}$ , Michael Bi Mi $^{2}$ , Xinchao Wang $^{1*}$
+
+$^{1}$ National University of Singapore  $^{2}$ Huawei International Pte Ltd
+
+{hanyang.k, gongkehong}@u.nus.edu {dongze, xinchao}@nus.edu.sg
+
+# Abstract
+
+Text-to-motion generation is a formidable task, aiming to produce human motions that align with the input text while also adhering to human capabilities and physical laws. While there have been advancements in diffusion models, their application in discrete spaces remains underexplored. Current methods often overlook the varying significance of different motions, treating them uniformly. It is essential to recognize that not all motions hold the same relevance to a particular textual description. Some motions, being more salient and informative, should be given precedence during generation. In response, we introduce a Priority-Centric Motion Discrete Diffusion Model (M2DM), which utilizes a Transformer-based VQ-VAE to derive a concise, discrete motion representation, incorporating a global self-attention mechanism and a regularization term to counteract code collapse. We also present a motion discrete diffusion model that employs an innovative noise schedule, determined by the significance of each motion token within the entire motion sequence. This approach retains the most salient motions during the reverse diffusion process, leading to more semantically rich and varied motions. Additionally, we formulate two strategies to gauge the importance of motion tokens, drawing from both textual and visual indicators. Comprehensive experiments on the HumanML3D and KIT-ML datasets confirm that our model surpasses existing techniques in fidelity and diversity, particularly for intricate textual descriptions.
+
+# 1. Introduction
+
+Generating realistic and diverse 3D human motions under various conditions e.g., action labels [34, 18], natural language descriptions [55, 17, 5, 56], and musical cues [27, 28, 29], presents a significant challenge across multiple domains, including gaming, filmmaking, and robotic animation. Notably, motion generation based on language descriptions has garnered substantial interest, given its promise
+
+for enhancing realism and broadening practical applications.
+
+However, generating a high-quality motion is not trivial due to the inherent modality gap and the complex mapping between text and motion modalities. Previous works [35, 43] align the latent feature space between text and motion modalities. For instance, TEMOS [35] aligns the text and motion feature by learning text and motion encoders. MotionClip [43] aligns the human motion manifold to CLIP [37] space for infusing semantic knowledge of CLIP into the motion extractor. Despite these advancements, they might encounter performance degradation when dealing with complex textual descriptions.
+
+To process these complex textual descriptions, some text-to-motion methods are proposed. State-of-the-art approaches such as TM2T [17] and T2M-GPT [55] use vector-quantized autoencoder (VQ-VAE) [46] to learn a discrete and compact motion representation, followed by a translation model [38, 47] to map the text modality to the motion modality. With the popularity and the superior performance in the generation tasks of diffusion models [20], MDM [44] and MLD [5] are proposed to learn conditioned diffusion models on the raw motion representation space and the latent feature space, respectively.
+
+While there have been promising advancements, two primary issues remain unresolved: i) The aforementioned diffusion methods, namely MDM [44] and MLD [5], predominantly address the latent feature within a continuous space. Although VQ-VAE-inspired architectures have made considerable strides in motion generation [55, 17], particularly with the support of discrete and compact motion representations, the integration of the diffusion model into a discrete space remains inadequately explored. ii) Discrete diffusion models employed in prior studies [15, 14, 22] treat all tokens uniformly. This approach presupposes that every token conveys an equivalent amount of information, neglecting the inherent disparities among tokens within a sequence. A more intuitive generative approach for humans would involve a progressive hierarchy, commencing with overarching concepts and gradually delving into finer details.
+
+To address the aforementioned challenges, we introduce a priority-centric motion discrete diffusion model (M2DM)
+
+![](images/b439b8314da8a8ceb3326548da576200317184eb052eb4f23a161258ea0bf35f.jpg)  
+a person steps forward with left leg walking in a circle counterclockwise motion swinging arms and returns back.
+
+![](images/9d0c251455da47bd3d24f66b71118aebf0d37946e80a8daacde1e958f2e04810.jpg)
+
+![](images/c8ea4c1f7282c3d12e82f475d439ce5e00a7a9482b33fc27749fca5bf60e559b.jpg)  
+a person runs forward, then turns around and walks in the opposite direction while swinging their right arm.
+
+![](images/c859db1c1545f793b8cdf537f4e5ba72913f0744a729ef102d8e27a747e06912.jpg)
+
+![](images/ed8e60537fcd7457573289571d364e02360b7d0c102a0ab97a985dc78d08cfc3.jpg)  
+a man gets up from the ground, walks around in a circle and then sits down on the ground again.
+
+![](images/85861fed1c799c75fb9e0e963334011b3432907a77306bb750bfcea593abbd37.jpg)  
+a man postures his arms like holding a dance partner and dances the waltz from the right hand side to the left hand side.
+
+![](images/a8b75794bd13b135797c0134ed741dcf543f0c5d0c7087ad0abad9fa0f0f1441.jpg)  
+Figure 1: Our Priority-Centric Motion Discrete Diffusion Model (M2DM) generates diverse and precise human motions given complex textual descriptions. The color of human mesh goes from light to dark over time.
+
+designed to generate motion sequences from textual descriptions, progressing in a primary to secondary manner. Initially, we employ a Transformer-based VQ-VAE, which is adept at learning a concise, discrete motion representation through the global self-attention mechanism. To circumvent code collapse and guarantee the optimal utilization of each motion token within the codebook, a regularization term is incorporated during the VQ-VAE training phase. Furthermore, we craft a noise schedule wherein individual tokens within a sequence are allocated varying corruption probabilities, contingent on their respective priorities during the forward process. Specifically, tokens of higher priority are slated for corruption towards the latter stages of the forward process. This ensures that the subsequent learnable reverse process adheres to a primary-to-secondary sequence, with top-priority tokens being reinstated foremost.
+
+To discern the significance of motion tokens within a sequence, we introduce two evaluative strategies: static assessment and dynamic assessment. 1) Static assessment: Drawing inspiration from the neural language domain, where the significance of individual words is gauged by their entropy across datasets, we calculate the entropy of each motion token across the entire motion dataset. 2) Dynamic assessment: We cultivate an agent specifically to dynamically gauge the importance of tokens within a sequence. Given a discrete motion token sequence, the agent masks one token at each iteration. These masked token sequences are then fed into the VQ decoder for continuous motion reconstruction. The agent's objective is to curtail the cumulative reconstruction discrepancy between the original and the reconstructed mo
+
+tions at every phase. This is achieved using a reinforcement learning [23] (RL) strategy, facilitating the agent's identification of motion tokens that convey minimal information within a sequence—a process we term dynamic analysis. Our priority-centric discrete diffusion model has showcased commendable generative prowess, especially when dealing with intricate textual descriptions. Through rigorous testing, our method has proven its mettle, consistently matching or surpassing the performance of prevailing text-to-motion generation techniques on the HumanML3D and KIT-ML datasets.
+
+We summarize our contributions as follows:
+
+- To capture long-range dependencies in the motion sequences, we apply Transformer as the architecture of VQ-VAE. Besides, a regularization term is applied to increase the usage of tokens in the codebook.  
+- We design a priority-centric scheme for human motion generation in the discrete latent space. To enhance the performance given complex descriptions, we design a novel priority-centric scheme for the discrete diffusion model.  
+- Our proposed priority-centric M2DM achieves state-of-the-art performance on the HumanML3D [16] and KIT-ML [36] datasets.
+
+# 2. Related Work
+
+VQ-VAE. Vector Quantized Variational Autoencoder (VQ-VAE) [46] aims to minimize reconstruction error with dis
+
+crete representations by training a variational autoencoder. VQ-VAE achieves promising performance on generative tasks, such as text-to-image generation [39, 52], music generation [7], unconditional image synthesis [9], etc. For motion generation tasks, most of the methods [17, 55] tokenize the continuous motion representation by the 1D-CNN network. Since the motion style varies from person to person and the number of action combinations is almost infinite, it is essential to extract discriminative and distinctive motion features from the raw motion representations. Moreover, raw motion representation is much more redundant compared with neural language, it is feasible to represent a motion sequence using few motion tokens from a global perspective. To this end, we introduce a transformer-based VQ-VAE and several techniques are applied to balance the usage of each motion token in the codebook and prevent the codebook from collapsing, i.e., only a small proportion of codes in the codebook are activated.
+
+Diffusion models. Diffusion generative models are first proposed in [42] and achieve promising results on image generation [8, 33, 21, 20, 50, 48, 53, 11, 10]. Discrete diffusion model is first proposed by [42] and several works [22, 14] apply the discrete diffusion model for text generation. D3PM [3] then applies the discrete diffusion model to the image generation task. For motion generation tasks, several works [24, 56, 44, 5] introduce diffusion models to generate diverse motion sequences from textual descriptions. In this work, we first introduce the discrete diffusion model for text-to-motion generation tasks which achieves promising results, especially for long text descriptions with several action commands.
+
+Conditional human motion synthesis Compared with unconditional Human motion synthesis, human motion synthesis conditioned on various conditions, such as textual description [55, 17, 5, 56], music [27, 28, 29, 13], and actions [34, 18] is more challenging due to the huge gap between two different domains. A common strategy of conditioned human motion synthesis is to employ generative models, for instance, conditional VAE [26], and learn a latent motion representation. Text-to-motion task is more challenging because neural language is a type of information highly refined by human beings, whereas human motion is much more redundant. Most recently, several promising works [5, 16, 17, 56, 55] are proposed for text-to-motion tasks. T2M-GPT [55] and TM2T [17] regard the text-to-motion task as a translation problem. They first train a VQ-VAE [46] to tokenize the motion sequences into motion tokens, and Recurrent Neural Networks (RNN) [32], Transformer [47, 51, 49, 41, 40, 54], or GPT-like [38] models are further applied to 'translate' textual descriptions to motions. Moreover, diffusion-based [20] models are introduced for
+
+text-to-motion generation by [56, 6, 44, 5] in the contentious space. In this work, to the best of our knowledge, we are the first ones to diffuse the motion representation in the discrete latent space. Moreover, we notice that different motion clips in a whole motion sequence play various roles, for instance, stepping forward is more important than swinging arms for a running motion. We further design a diffuse schedule such that the generation process follows a from-primary-to-secondary manner.
+
+# 3. Method
+
+# 3.1. Motion Tokenizer
+
+The motion tokenizer aims to learn a codebook that can represent a wide range of motion sequences with the priority awareness of motion tokens. To achieve this, we propose a transformer-based VQ-VAE that leverages self-attention to capture long-range dependencies between motion frames and learns a stable and diverse codebook.
+
+Fig. 2 shows the overview of our motion tokenizer. Given a motion sequence  $X \in \mathbb{R}^{T \times d}$  with  $T$  frames and  $d$  dimensional motion representation, we reconstruct the motion sequence using a transformer autoencoder and learn an informative codebook  $Z = \{z_k\}_{k=1}^K$  with  $K$  entries. We use the transformer encoder  $E$  to compute the latent motion features  $\boldsymbol{b} = \{b_t\}_{t=1}^T \in \mathbb{R}^{T \times d}$ , where  $d$  is the dimension of the latent features. Then for each feature  $b_t$ , we query the codebook  $Z$  by measuring the distance between feature  $b_t$  and each entry in the codebook  $Z$ . To improve the codebook usage and avoid dead codes, we first normalize  $b_t$  and  $z_t$  and then measure the Euclidean distance between two normalized vectors in the unit sphere coordinate, which is formulated as:
+
+$$
+\hat {b} _ {t} = \underset {b _ {t} \in \boldsymbol {b}} {\arg \min } \| \ell_ {2} (b _ {t}) - \ell_ {2} \left(z _ {k}\right) \| _ {2} ^ {2}, \tag {1}
+$$
+
+where  $\ell_2$  is the  $l_{2}$  normalized function.
+
+The standard optimization goal consists of three parts: the reconstruction loss, the embedding loss, and the commitment loss, which is formulated as
+
+$$
+\begin{array}{l} \mathcal {L} _ {v q} = \left\| X - \tilde {X} \right\| _ {2} ^ {2} + \left\| Z - s g [ \hat {Z} ] \right\| _ {2} ^ {2} \tag {2} \\ + \eta | | s g [ Z ] - \hat {Z} | | _ {2} ^ {2}. \\ \end{array}
+$$
+
+Moreover, to achieve an informative codebook with diverse and high usage of its entries, we apply an orthogonal regularization. Specifically, we calculate the distance between each pair of codebook entries to enforce orthogonality among all the codebook entries:
+
+$$
+\mathcal {L} _ {\text {o r t h}} (Z) = \left\| \ell_ {2} (Z) ^ {\top} \ell_ {2} (Z) - I _ {K} \right\| _ {2} ^ {2}, \tag {3}
+$$
+
+where  $\mathcal{L}_2$  is the  $l_{2}$  normalized function.
+
+![](images/5b14b9f440f563c742bd7350990b93a1daf121fbfd30f65ec94ca918acd80905.jpg)  
+Figure 2: Overview of our framework for our Motion Discrete Diffusion Model (M2DM). M2DM consists of a Transformer-based VQ-VAE (Sec. 3.1) and a discrete diffusion model (Sec. 3.2). Firstly, the Transformer-based motion tokenizer is learned to discretize motion sequences into tokens and reconstruct original motion sequences from tokens queried from the learnable codebook. Then the discrete diffusion model is learned to generate diverse motions conditioned on the textual descriptions. Two priority assessment strategies (Sec. 3.3) are applied to mask or replace motion tokens with less information at the beginning of the forward diffusion process and the most important motion tokens are corrupted in the last few diffusion steps. The diffusion model learns to restore the original motion token index conditioned on the given textual descriptions.
+
+Therefore, the overall training objective of vector quantization is defined as follows:
+
+$$
+\begin{array}{l} \mathcal {L} _ {v q} = \left\| X - \tilde {X} \right\| _ {2} ^ {2} + \| Z - s g [ \hat {Z} ] \| _ {2} ^ {2} \tag {4} \\ + \eta | | s g [ Z ] - \hat {Z} | | _ {2} ^ {2} + \delta \mathcal {L} _ {\text {o r t h}} (Z), \\ \end{array}
+$$
+
+where  $\tilde{X} = D(\pmb {b})$ $D$  is the decoder,  $sg[\cdot ]$  denotes the stopgradient operation,  $\eta$  and  $\delta$  are the weighting scalars.
+
+# 3.2. Discrete Diffusion Models
+
+Generally speaking, the forward discrete diffusion process corrupts each element  $x_{t}$  at timestep  $t$  with  $K$  categories via fixed Markov chain  $q(x_{t}|x_{t - 1})$ , for instance, replace some tokens of  $x_{t - 1}$  or mask the token directly. After a fixed number of  $T$  timesteps, the forward process yields a sequence of increasingly noisy discrete tokens  $z_{1},\ldots ,z_{T}$  of the same dimension of  $z_{0}$  and  $z_{T}$  becomes pure noise token. For the denoising procedure, the real  $x_0$  is restored sequentially based on the pure noisy token  $z_{T}$ . To be specific, for scalar random variable with  $K$  categories  $x_{t},x_{t - 1}\in 1,\dots K$ , the forward transition probabilities can be represented by the transition matrices  $[Q_{t}]_{ij} = q(x_{t} = i|x_{t - 1} = j)\in \mathbb{R}^{(K + 1)\times (K + 1)}$ .  $[Q_{t}]_{i,j}$  is the Markov transition matrix from state  $t$  to state  $t - 1$  that is applied to each token in the sequence independently, which
+
+can be written as:
+
+$$
+\boldsymbol {Q} _ {t} = \left[ \begin{array}{c c c c c} \alpha_ {t} + \beta_ {t} & \beta_ {t} & \beta_ {t} & \dots & 0 \\ \beta_ {t} & \alpha_ {t} + \beta_ {t} & \beta_ {t} & \dots & 0 \\ \beta_ {t} & \beta_ {t} & \alpha_ {t} + \beta_ {t} & \dots & 0 \\ \vdots & \vdots & \vdots & & \vdots \\ \gamma_ {t} & \gamma_ {t} & \gamma_ {t} & \dots & 1 \end{array} \right], \tag {5}
+$$
+
+where  $\alpha_{t} \in [0,1]$  is the probability of retaining the token,  $\gamma_{t}$  is the probability of replacing the original token to [MASK] token, leaving the probability  $\beta_{t} = (1 - \alpha_{t} - \gamma_{t}) / K$  to be diffused. The forward Markov diffusion process for the whole token sequence is written as,
+
+$$
+q \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {t - 1}\right) = \operatorname {C a t} (\boldsymbol {x}; \boldsymbol {p} = \boldsymbol {x} _ {t - 1} \boldsymbol {Q} _ {t}) = \boldsymbol {x} _ {t} \boldsymbol {Q} _ {t} \boldsymbol {x} _ {t - 1}, \tag {6}
+$$
+
+where  $\pmb{x}$  is the one-hot row vector identifying the token index and  $\operatorname{Cat}(\pmb{x};\pmb{p})$  is a categorical distribution over the one-hot vector  $\pmb{x}_t$  with probabilities  $\pmb{p}$ .
+
+Starting from the initial step  $x_0$ , the posterior diffusion process is formulated as:
+
+$$
+\begin{array}{l} q \left(\boldsymbol {x} _ {t - 1} \mid \boldsymbol {x} _ {t}, \boldsymbol {x} _ {0}\right) = \frac {q \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {t - 1} , \boldsymbol {x} _ {0}\right) q \left(\boldsymbol {x} _ {t - 1} , \boldsymbol {x} _ {0}\right)}{q \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {0}\right)} (7) \\ = \frac {\left(\boldsymbol {x} ^ {\top} \boldsymbol {Q} _ {t} \boldsymbol {x} _ {t - 1}\right) \left(\boldsymbol {x} _ {t - 1} ^ {T} \bar {\boldsymbol {Q}} _ {t - 1} \boldsymbol {x} _ {0}\right)}{\boldsymbol {x} _ {t} ^ {\top} \bar {\boldsymbol {Q}} _ {t} \boldsymbol {x} _ {0}} (7) \\ \end{array}
+$$
+
+with  $\bar{Q} = Q_{1}Q_{2}\dots Q_{t}$
+
+Recording to Markov chain, the intermediate step can be marginalized out and the derivation of the probability of  $\boldsymbol{x}_t$
+
+at arbitrary timestep directly from  $\pmb{x}_0$  can be formulated by
+
+$$
+q \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {0}\right) = \boldsymbol {x} _ {t} ^ {\top} \bar {\boldsymbol {Q}} \boldsymbol {x} _ {0}. \tag {8}
+$$
+
+The cumulative transition matrix  $\bar{Q}$  and the probability  $q(\pmb{x}_t|\pmb{x}_0)$  are then computed as
+
+$$
+\bar {Q} \boldsymbol {x} _ {0} = \bar {\alpha} _ {t} \boldsymbol {x} _ {0} + (\bar {\gamma} - \bar {\beta} _ {t}) \boldsymbol {m} + \bar {\beta} _ {t} \tag {9}
+$$
+
+where  $m$  is the one-hot vector for [MASK] and [PAD],  $\bar{\alpha}_t = \prod_{i=1}^t \alpha_i$ ,  $\bar{\gamma}_t = 1 - \prod_{i=1}^t (1 - \gamma_i)$ , and  $\bar{\beta}_t = (1 - \bar{\alpha}_t - \bar{\gamma}_t) / K$ .
+
+Regarding the reverse diffusion process, we train a transformer-like [47] denoising network  $p_{\theta}(\pmb{x}_{t-1}|\pmb{x}_t,\pmb{y})$  to estimate the posterior transition distribution  $q(\pmb{x}_{t-1}|\pmb{x}_t,\pmb{x}_0)$ . The network is trained to minimize the variational lower bound (VLB) [42]:
+
+$$
+\begin{array}{l} \mathcal {L} _ {V L B} = \mathbb {E} _ {q} \left[ D _ {K L} \left(q \left(\boldsymbol {x} _ {T} \mid \boldsymbol {x} _ {0}\right) \| p _ {\theta} \left(\boldsymbol {x} _ {T}\right)\right) \right] \\ + \mathbb {E} _ {q} [ \sum_ {t = 2} ^ {T} D _ {K L} (q (\boldsymbol {x} _ {t - 1} | \boldsymbol {x} _ {t}, \boldsymbol {x} _ {0}) \| p _ {\theta} (\boldsymbol {x} _ {t - 1} | \boldsymbol {x} _ {t}, t)) ] \\ - \log p _ {\theta} \left(\boldsymbol {x} _ {0} \mid \boldsymbol {x} _ {1}\right) \tag {10} \\ \end{array}
+$$
+
+where  $\mathbb{E}_q[\cdot ]$  denotes the expectation over the joint distribution  $q(\pmb{x}_{0:T})$ .
+
+# 3.3. Priority-Centric Denoising Process
+
+The designation of the noise schedule in the continuous domain, for instance, the linear schedule [20] and the cosine schedule [33], achieves the excellent performance of diffusion models. The noise can be easily controlled by the variation of Gaussian noise. For the discrete diffusion models, several noise schedules [22, 3] have been explored to control the data corruption and denoise the reverse process by choosing the transition matrix  $Q_{t}$ . For text-to-motion generation, however, such a schedule assumes all motion tokens carry the same amount of information and do not consider the difference among the discrete tokens extracted from the continuous motion sequence. Intuitively, the motion generation process would be in a progressive manner, where the most important motion tokens should appear in earlier steps for laying a solid foundation when denoising the noise in the discrete domain.
+
+Specifically, we apply the mask-and-replace strategy in Eq. (5) to corrupt each ordinary discrete token with a probability of  $\gamma_{t}$  for masking [MASK] token and a probability of  $\beta_{t}$  for uniform diffusion to a random token  $x_{k}$ , resulting in a remaining probability of  $\alpha_{t} = 1 - K\beta_{t} - \gamma_{t}$  for retention. When corrupting across the forward process in the discrete domain, we expect that  $\bar{\alpha}_{t}^{i} < \bar{\alpha}_{t}^{j}$  if token  $x_{i}$  is more informative than  $x_{j}$  such that the tokens with high information emerge earlier than less informative motion tokens in the reverse process. Instead of using a normal corruption strategy with linearly increasing  $\bar{\gamma}_{t}$  and  $\bar{\beta}_{t}$ , we propose a
+
+priority-score function  $F$  to assess the relative importance of information for a motion token sequence, which enables the recovery of the most informative tokens during the reverse process.
+
+Given an original motion token sequences  $\pmb{x}_0$  at timestep 0 with length  $N$ , we aim to analyze the importance of each token  $x_0^i$  in  $\pmb{x}_0$ . Based on the information entropy theory, we calculate the information entropy for each motion token  $x_0^i$  as the importance score
+
+$$
+F \left(x _ {0} ^ {i}\right) = \frac {N H \left(x _ {0} ^ {i}\right)}{\sum_ {j = 1} ^ {N} H \left(x _ {0} ^ {j}\right)}, \tag {11}
+$$
+
+where
+
+$$
+H (x) = - \sum_ {i = 1} ^ {K} p \left(x _ {i}\right) \log p \left(x _ {i}\right), \tag {12}
+$$
+
+and  $K$  denotes the number of categories belonging to  $x_{i}$
+
+After obtaining the importance score function and setting linearly increasing schedule  $\bar{\gamma}_t^i$  and  $\bar{\beta}_t^i$  for all tokens,  $\bar{\gamma}_t^i$  and  $\bar{\beta}_t^i$  for specific token  $x_0^i$  can be further updated by
+
+$$
+\bar {\gamma} _ {t} ^ {i} \Leftarrow \bar {\gamma} _ {t} ^ {i} \cdot \sin \frac {t \pi}{T} \cdot F \left(x _ {0} ^ {i}\right), \tag {13}
+$$
+
+and
+
+$$
+\bar {\beta} ^ {i} \Leftarrow \bar {\beta} ^ {i} \cdot \sin \frac {t \pi}{T} \cdot F \left(x _ {0} ^ {i}\right), \tag {14}
+$$
+
+where  $x_0^i$  is the  $i$ -th token at timestep 0.
+
+Now we introduce two solutions for obtaining  $p(x_{i})$ .
+
+Static assessment. Analogously to computing the entropy for each word in natural language processing, we estimate the probability  $p(x_{i})$  for each motion token by counting its frequency in the quantized motion sequences over the entire dataset.
+
+![](images/3d0befa6afe7389516bbd25840dc6cd2e93ada0b6bc000589e1a52883fd2441f.jpg)  
+Figure 3: Illustration for dynamic assessment strategy. The gray blocks are masked tokens and other colorful blocks are motion tokens.
+
+Dynamic assessment. Rather than counting motion token frequency and calculating the importance of each token, we further propose to learn an agent to estimate the importance of each token in the token sequences by reinforcement learning strategy. As shown in Fig. 3, given a series of complete motion tokens  $\boldsymbol{x} = \{x_{n}\}_{n=1}^{N}$  that are queried from the codebook at current state  $s_{t} \in S$ , the scorer samples the motion tokens  $\tilde{\boldsymbol{x}} = \{x_{n}\}_{n=1}^{N'}$  and tries to minimize the reconstruction error between the original motion sequences  $\boldsymbol{m}$  and the motion sequences decoded from the sampled tokens  $\tilde{\boldsymbol{m}}$ .
+
+State includes the currently selected motion tokens  $\tilde{\pmb{x}} = \{x_{n}\}_{n = 1}^{N^{\prime}}$ , the complete motion tokens  $\pmb {x} = \{x_{n}\}_{n = 1}^{N}$ , and the corresponding reconstructed continuous motions  $\tilde{m}$  and  $m$ . The reconstructed continuous motion sequences  $\tilde{m}$  and  $m$  are obtained by decoding the motion tokens by VQ decoder. The action is to sample one motion token from the complete motion token sequence at each stage. To be specific, given a motion token sequence, the agent  $\pi_{\theta}(a_t|s_t)$  tries to sample a motion token that can recover the continuous motion sequences with the minimum reconstruction error. The reward measures the reconstruction difference between  $\tilde{m}$  and  $m$  with the minimum sample times.
+
+![](images/64487833a06942ff58a25558356f6c19cec3d49745c206c3d6a61b15fa11e306.jpg)  
+Figure 4: Priority-centric denoising schedule for motion tokens.
+
+After the agent is well-trained, we calculate and restore the sampling order for each motion token sequence by the agent  $\pi_{\theta}(a_t|s_t)$  and compute the noise schedule for each motion token in the codebook by Eq. (13) and (14). Fig. 4 shows the noise schedule for different motion tokens w.r.t. the corresponding priorities. As shown in Fig. 4, high-priority tokens (orange and green curves) are encouraged to be corrupted at the end of the forward process such that the learnable reverse process follows a primary-to-secondary manner and high-priority tokens will be restored first. It is worth noting that both strategies are applied to design noise schedules for each token, which can be calculated and stored in advance such that the agent will not join the forward and reverse diffusion processes. Thus, there is no extra computational cost for the diffusion model during the training and inference stages.
+
+# 4. Experiments
+
+# 4.1. Datasets and Evaluation Metrics
+
+Our experiments are conducted on two standard text-to-motion datasets for text-to-motion generation: HumanML3D [16] and KIT-ML [36]. We follow the evaluation metrics provided in [16].
+
+HumanML3D. HumanML3D [16] is currently the largest 3D human motion dataset that covers a broad range of daily human actions, for instance, exercising and dancing. The dataset contains 14,616 motion sequences and 44970 text descriptions. The motion sequences are processed to 20 frame-per-second (FPS) and cropped to 10 seconds if the motion sequences are longer than 10 seconds, resulting in duration ranges from 2 to 10 seconds. For each motion clip, the corresponding number of descriptions is at least three.
+
+KIT Motion-Language (KIT-ML). KIT-ML dataset [36] contains 3,911 3D human motion clips with 6,278 text descriptions. For each motion clip, one to four textual descriptions are provided. The motion sequences are collected from the KIT dataset [31] and the CMU dataset [1] with the down-sampled 12.5 FPS.
+
+Both of the aforementioned datasets are split into training, validation, and testing datasets with the proportions of  $80\%$ ,  $5\%$ , and  $15\%$ .
+
+Implementation details. Both our VQ-encoder and decoder consist of 4 transformer layers with 8 heads and the dimension is 512. For the HumanML3D [16] and KIT-ML [36] datasets, the motion sequences are cropped to 64 frames for training. We apply Adam with  $\beta_{1} = 0.9$ ,  $\beta_{2} = 0.999$ , weight decay is 1e-4. The learning rate is 1e-4 after warmup linearly for 5 epochs. The number of tokens in the codebook is 8192. We train our VQ-VAE with a batch size of 1024 across 8 Tesla T4 for a total 100 training epochs for 8 hours.
+
+For the discrete diffusion model, we set timesteps  $T = 100$  and the network is trained using Adam [25] with  $\beta_{1} = 0.9$ ,  $\beta_{2} = 0.999$ . The learning rate reaches 2e-4 after 2000 iterations of a linear warmup schedule. The transformer-based discrete diffusion model consists of 12 transformer layers with 8 heads, and the dimension is 512. The text features are extracted by pre-trained CLIP [37] model. The diffusion model is trained with a batch size of 64 across 8 Tesla T4 for 100k iterations for 36 hours.
+
+Evaluation metrics. Following the evaluation protocols provided in previous works [16, 55, 44, 5], we evaluate the correspondences between motion and language using deep multimodal features with the pre-trained models in [16] by the following metrics: 1) Generation diversity: We randomly
+
+<table><tr><td rowspan="2">Methods</td><td colspan="3">R-Precision ↑</td><td rowspan="2">FID ↓</td><td rowspan="2">MM-Dist ↓</td><td rowspan="2">Diversity →</td><td rowspan="2">MModality ↑</td></tr><tr><td>Top-1</td><td>Top-2</td><td>Top-3</td></tr><tr><td>Real motion</td><td>0.511±.003</td><td>0.703±.003</td><td>0.797±.002</td><td>0.002±.000</td><td>2.974±.008</td><td>9.503±.065</td><td>-</td></tr><tr><td>Our VQ-VAE</td><td>0.508±.002</td><td>0.691±.002</td><td>0.791±.003</td><td>0.63±.001</td><td>3.015±.010</td><td>9.577±.081</td><td>-</td></tr><tr><td>Our VQ-VAE w/o l2</td><td>0.494±.003</td><td>0.685±.002</td><td>0.772±.003</td><td>0.70±.001</td><td>3.251±.008</td><td>9.525±.092</td><td>-</td></tr><tr><td>Our VQ-VAE w/o l2, Lorth</td><td>0.485±.002</td><td>0.671±.003</td><td>0.752±.003</td><td>0.79±.001</td><td>3.378±.012</td><td>9.511±.095</td><td>-</td></tr><tr><td>Seq2Seq [30]</td><td>0.180±.002</td><td>0.300±.002</td><td>0.396±.002</td><td>11.75±.035</td><td>5.529±.007</td><td>6.223±.061</td><td>-</td></tr><tr><td>Language2Pose [2]</td><td>0.246±.002</td><td>0.387±.002</td><td>0.486±.002</td><td>11.02±.046</td><td>5.296±.008</td><td>7.676±.058</td><td>-</td></tr><tr><td>Text2Gesture [4]</td><td>0.165±.001</td><td>0.267±.002</td><td>0.345±.002</td><td>5.012±.030</td><td>6.030±.008</td><td>6.409±.071</td><td>-</td></tr><tr><td>Hier [12]</td><td>0.301±.002</td><td>0.425±.002</td><td>0.552±.004</td><td>6.532±.024</td><td>5.012±.018</td><td>8.332±.042</td><td>-</td></tr><tr><td>MoCoGAN [45]</td><td>0.037±.000</td><td>0.072±.001</td><td>0.106±.001</td><td>94.41±.021</td><td>9.643±.006</td><td>0.462±.008</td><td>0.019±.000</td></tr><tr><td>Dance2Music [27]</td><td>0.033±.000</td><td>0.065±.001</td><td>0.097±.001</td><td>66.98±.016</td><td>8.116±.006</td><td>0.725±.011</td><td>0.043±.001</td></tr><tr><td>TM2T [17]</td><td>0.424±.003</td><td>0.618±.003</td><td>0.729±.002</td><td>1.501±.017</td><td>3.467±.011</td><td>8.589±.076</td><td>2.424±.093</td></tr><tr><td>Guo et al. [16]</td><td>0.455±.003</td><td>0.636±.003</td><td>0.736±.002</td><td>1.087±.021</td><td>3.347±.008</td><td>9.175±.083</td><td>2.219±.074</td></tr><tr><td>MDM [44]$</td><td>-</td><td>-</td><td>0.611±.007</td><td>0.544±.044</td><td>5.566±.027</td><td>9.559±.086</td><td>2.799±.072</td></tr><tr><td>MotionDiffuse [56]$</td><td>0.491±.001</td><td>0.681±.001</td><td>0.782±.001</td><td>0.630±.001</td><td>3.113±.001</td><td>9.410±.049</td><td>1.553±.042</td></tr><tr><td>T2M-GPT [55]</td><td>0.492±.003</td><td>0.679±.002</td><td>0.775±.002</td><td>0.141±.005</td><td>3.121±.009</td><td>9.722±.082</td><td>1.831±.048</td></tr><tr><td>MLD [5]</td><td>0.481±.003</td><td>0.673±.003</td><td>0.772±.002</td><td>0.473±.013</td><td>3.196±.010</td><td>9.724±.082</td><td>2.413±.079</td></tr><tr><td>M2DM (Ours) w. linear schedule</td><td>0.452±.003</td><td>0.678±.002</td><td>0.752±.003</td><td>0.417±.004</td><td>3.167±.008</td><td>9.972±.089</td><td>3.562±.071</td></tr><tr><td>M2DM (Ours) w. static assessment</td><td>0.492±.003</td><td>0.671±.003</td><td>0.775±.003</td><td>0.395±.005</td><td>3.116±.008</td><td>9.937±.075</td><td>3.413±.035</td></tr><tr><td>M2DM (Ours) w. dynamic assessment</td><td>0.497±.003</td><td>0.682±.002</td><td>0.763±.003</td><td>0.352±.005</td><td>3.134±.010</td><td>9.926±.073</td><td>3.587±.072</td></tr></table>
+
+${}^{8}$  reports results using ground-truth motion length.
+
+Table 1: Comparison with the state-of-the-art methods on HumanML3D [16] test set. The evaluation metrics are evaluated by the motion encoder from Guo et al. [16]. The right row  $\rightarrow$  means the closer to the real motion the better.
+
+sample 300 pairs of motions from the motion pool and extracted the selected motion features by the extractor provided by [16]. Euclidean distances of the motion pairs are computed for measuring the motion diversity. 2) R-Precision: Provided one motion representation and 32 textual descriptions, the Euclidean distances between the description feature and each motion feature in the candidates are calculated and ranked. Top-k  $(k = 1,2,3)$  average accuracy of motion-to-text retrieval is reported. 3) Frechet Inception Distance (FID) [19]: FID is the principal metric to evaluate the feature descriptions between the generated motions and the ground-truth motions by the feature extractor [16]. 4) Multimodal Distance (MM-Dist): MM-Dist measures the average Euclidean distances between each text feature and the generated motion feature. 5) MModality measures the generation diversity within the same textual descriptions. Please refer to the supplementary material for a detailed introduction.
+
+# 4.2. Comparisons with state-of-the-art approaches
+
+We compare our methods with other state-of-the-art methods [30, 2, 4, 12, 45, 27, 17, 16, 44, 56, 55, 5] on the HumanML3D [16] and KIT-ML [36] datasets.
+
+Quantitative comparisons. For each metric, we repeat the evaluation 20 times and report the average with  $95\%$  confidence interval, followed by [16]. Most of the results are borrowed from [55]. Tab. 1 and Tab. 2 summarize the comparison results on the HumanML3D [16] and the KIT-ML [36] datasets, respectively. Firstly, the metrics scores for
+
+our reconstruction motion achieve close performance compared with real motion, which suggests the distinctive and high-quality discrete motion tokens in the codebook learned by our Transformer-based VQ-VAE. The performance of our VQ-VAE degrades when the normalized function  $\ell_2$  and the orthogonal regularization term  $\mathcal{L}_{\mathrm{orth}}$  are removed. For text-to-motion generation, our approach achieves compatible performance (R-Precision, FID, and MModality) compared to other state-of-the-art methods. Moreover, compared with MDM [44] and MotionDiffuse [56] which evaluate their performance with the ground-truth motion lengths, our model can generate motion sequences with arbitrary lengths conditioned on the given textual descriptions since the [MASK] and [PAD] tokens are introduced in the discrete diffusion process.
+
+Furthermore, we conduct more experiments to evaluate the generation performance conditioned on different lengths of textual descriptions. The testing dataset is split into three parts according to the length of the textual descriptions: less than 15 words, between 15 words and 30 words, and more than 30 words. Tab. 3 shows the quantitative results on the aforementioned three sub-sets. Both TM2T [17] and MLD [5] suffer degradation when the textual descriptions are long and complex, especially the descriptions are more than 30 words. Our method significantly outperforms the other two methods with various metrics and there is not much degradation when the descriptions are more complex. The quantitative results demonstrate the effectiveness of our proposed primary-to-secondary diffusion manner.
+
+<table><tr><td rowspan="2">Methods</td><td colspan="3">R-Precision ↑</td><td rowspan="2">FID ↓</td><td rowspan="2">MM-Dist ↓</td><td rowspan="2">Diversity →</td><td rowspan="2">MModality ↑</td></tr><tr><td>Top-1</td><td>Top-2</td><td>Top-3</td></tr><tr><td>Real motion</td><td>0.424±.005</td><td>0.649±.006</td><td>0.779±.006</td><td>0.031±.004</td><td>2.788±.012</td><td>11.08±.097</td><td>-</td></tr><tr><td>Our VQ-VAE (Recons.)</td><td>0.417±.004</td><td>0.621±.003</td><td>0.741±.006</td><td>0.413±.009</td><td>2.772±.018</td><td>10.851±.105</td><td>-</td></tr><tr><td>Seq2Seq [30]</td><td>0.103±.003</td><td>0.178±.005</td><td>0.241±.006</td><td>24.86±.348</td><td>7.960±.031</td><td>6.744±.106</td><td>-</td></tr><tr><td>Language2Pose [2]</td><td>0.221±.005</td><td>0.373±.004</td><td>0.483±.005</td><td>6.545±.072</td><td>5.147±.030</td><td>9.073±.100</td><td>-</td></tr><tr><td>Text2Gesture [4]</td><td>0.156±.004</td><td>0.255±.004</td><td>0.338±.005</td><td>12.12±.183</td><td>6.964±.029</td><td>9.334±.079</td><td>-</td></tr><tr><td>Hier [12]</td><td>0.255±.006</td><td>0.432±.007</td><td>0.531±.007</td><td>5.203±.107</td><td>4.986±.027</td><td>9.563±.072</td><td>-</td></tr><tr><td>MoCoGAN [45]</td><td>0.022±.002</td><td>0.042±.003</td><td>0.063±.003</td><td>82.69±.242</td><td>10.47±.012</td><td>3.091±.043</td><td>0.250±.009</td></tr><tr><td>Dance2Music [27]</td><td>0.031±.002</td><td>0.058±.002</td><td>0.086±.003</td><td>115.4±.240</td><td>10.40±.016</td><td>0.241±.004</td><td>0.062±.002</td></tr><tr><td>TM2T [17]</td><td>0.280±.005</td><td>0.463±.006</td><td>0.587±.005</td><td>3.599±.153</td><td>4.591±.026</td><td>9.473±.117</td><td>3.292±.081</td></tr><tr><td>Guo et al. [16]</td><td>0.361±.006</td><td>0.559±.007</td><td>0.681±.007</td><td>3.022±.107</td><td>3.488±.028</td><td>10.72±.145</td><td>2.052±.107</td></tr><tr><td>MDM [44]$</td><td>-</td><td>-</td><td>0.396±.004</td><td>0.497±.021</td><td>9.191±.022</td><td>10.847±.109</td><td>1.907±.214</td></tr><tr><td>MotionDiffuse [56]$</td><td>0.417±.004</td><td>0.621±.004</td><td>0.739±.004</td><td>1.954±.062</td><td>2.958±.005</td><td>11.10±.143</td><td>0.730±.013</td></tr><tr><td>T2M-GPT [55]</td><td>0.416±.006</td><td>0.627±.006</td><td>0.745±.006</td><td>0.514±.029</td><td>3.007±.023</td><td>10.921±.108</td><td>1.570±.039</td></tr><tr><td>MLD [5]</td><td>0.390±.003</td><td>0.609±.003</td><td>0.734±.002</td><td>0.404±.013</td><td>3.204±.010</td><td>10.8±.082</td><td>2.192±.079</td></tr><tr><td>M2DM (Ours) w. linear schedule</td><td>0.405±.003</td><td>0.629±.005</td><td>0.739±.004</td><td>0.502±.049</td><td>3.012±.015</td><td>11.375±.079</td><td>3.273±.045</td></tr><tr><td>M2DM (Ours) w. static assessment</td><td>0.417±.006</td><td>0.625±.003</td><td>0.741±.006</td><td>0.521±.041</td><td>3.024±.018</td><td>11.373±.081</td><td>3.317±.031</td></tr><tr><td>M2DM (Ours) w. dynamic assessment</td><td>0.416±.004</td><td>0.628±.004</td><td>0.743±.004</td><td>0.515±.029</td><td>3.015±.017</td><td>11.417±.97</td><td>3.325±.37</td></tr></table>
+
+${}^{8}$  reports results using ground-truth motion length.
+
+Table 2: Comparison with the state-of-the-art methods on KIT-ML [36] test set. The evaluation metrics are evaluated by the motion encoder from Guo et al. [16]. The right row  $\rightarrow$  means the closer to the real motion the better.  
+
+<table><tr><td>TM2T</td><td>MDM</td><td>T2M-GPT</td><td>MLD</td><td>Ours</td></tr><tr><td colspan="5">a person grabs something, turns around, walks forward, squats, stands up, walks forward, squats and stands up again.</td></tr><tr><td></td><td></td><td></td><td></td><td></td></tr><tr><td colspan="5">a person puts both hands by their face in a defensive stance, then kicks with their right foot.</td></tr><tr><td></td><td></td><td></td><td></td><td></td></tr><tr><td colspan="5">a person jumps and turns around in the air to face behind him.</td></tr><tr><td></td><td></td><td></td><td></td><td></td></tr><tr><td colspan="5">a person walks forward briskly, crouches and lifts an object with both hands, puts the object down in a higher position, then stands up and turn right.</td></tr><tr><td></td><td></td><td></td><td></td><td></td></tr><tr><td colspan="5">a person takes something, put it down and throws something.</td></tr><tr><td></td><td></td><td></td><td></td><td></td></tr></table>
+
+Figure 5: Qualitative comparison of the state-of-the-art methods on the HumanML3D [16] dataset. We compare our generations with TM2T [17], MDM [44], T2M-GPT [55], and MLD [5]. The color of human mesh goes from light to dark over time.
+
+Qualitative comparisons. Fig. 5 qualitatively compares the generated motions from the same textual descriptions. Most of the methods suffer from motion freezing if the tex
+
+tual descriptions consist of several complex commands. For example, the first row in Fig. 5 shows the generation results from a long textual description with various duplicate action
+
+<table><tr><td rowspan="2">Text Length</td><td rowspan="2">Methods</td><td colspan="3">R-Precision ↑</td><td rowspan="2">FID ↓</td><td rowspan="2">MM-Dist ↓</td><td rowspan="2">Diversity →</td><td rowspan="2">MModality ↑</td></tr><tr><td>Top-1</td><td>Top-2</td><td>Top-3</td></tr><tr><td rowspan="3">All</td><td>TM2T [17]</td><td>0.424±0.003</td><td>0.618±0.003</td><td>0.729±0.002</td><td>1.501±0.017</td><td>3.467±0.011</td><td>8.580±0.076</td><td>2.424±0.093</td></tr><tr><td>MLD [5]</td><td>0.481±0.003</td><td>0.673±0.003</td><td>0.772±0.002</td><td>0.473±0.013</td><td>3.196±0.010</td><td>9.724±0.082</td><td>2.413±0.079</td></tr><tr><td>M2DM (Ours)</td><td>0.497±0.003</td><td>0.689±0.002</td><td>0.763±0.003</td><td>0.352±0.005</td><td>3.134±0.010</td><td>9.926±0.073</td><td>3.587±0.072</td></tr><tr><td rowspan="3">&lt; 15 words</td><td>TM2T [17]</td><td>0.433±0.003</td><td>0.627±0.003</td><td>0.738±0.003</td><td>1.592±0.017</td><td>3.446±0.008</td><td>8.677±0.077</td><td>2.520±0.051</td></tr><tr><td>MLD [5]</td><td>0.492±0.003</td><td>0.680±0.004</td><td>0.779±0.004</td><td>0.469±0.014</td><td>3.191±0.011</td><td>9.770±0.010</td><td>2.586±0.086</td></tr><tr><td>M2DM (Ours)</td><td>0.508±0.003</td><td>0.685±0.003</td><td>0.767±0.004</td><td>0.347±0.015</td><td>3.129±0.013</td><td>9.915±0.068</td><td>3.594±0.077</td></tr><tr><td rowspan="3">15-30 words</td><td>TM2T [17]</td><td>0.406±0.003</td><td>0.596±0.002</td><td>0.708±0.002</td><td>1.425±0.014</td><td>3.528±0.007</td><td>8.000±0.051</td><td>2.477±0.057</td></tr><tr><td>MLD [5]</td><td>0.406±0.005</td><td>0.592±0.005</td><td>0.701±0.006</td><td>0.480±0.020</td><td>3.591±0.020</td><td>9.089±0.083</td><td>2.727±0.094</td></tr><tr><td>M2DM (Ours)</td><td>0.493±0.003</td><td>0.681±0.004</td><td>0.761±0.004</td><td>0.353±0.031</td><td>3.135±0.014</td><td>9.926±0.073</td><td>3.588±0.081</td></tr><tr><td rowspan="3">&gt;30 words</td><td>TM2T [17]</td><td>0.353±0.009</td><td>0.536±0.009</td><td>0.650±0.007</td><td>1.779±0.061</td><td>3.577±0.024</td><td>7.411±0.071</td><td>2.612±0.055</td></tr><tr><td>MLD [5]</td><td>0.289±0.010</td><td>0.466±0.012</td><td>0.581±0.015</td><td>0.947±0.066</td><td>3.870±0.039</td><td>8.347±0.100</td><td>2.916±0.063</td></tr><tr><td>M2DM (Ours)</td><td>0.491±0.004</td><td>0.677±0.004</td><td>0.758±0.004</td><td>0.356±0.038</td><td>3.138±0.025</td><td>9.930±0.088</td><td>3.585±0.071</td></tr></table>
+
+Table 3: Comparison with the state-of-the-art methods on HumanML3D [16] test set. The evaluation metrics are evaluated by the motion encoder from Guo et al. [16]. The right row  $\rightarrow$  means the closer to the real motion the better. The first three rows are the evaluation results on the whole testing dataset. The rest are the evaluation results from textual descriptions with different lengths.
+
+commands. MDM [44] and MLD [5] generate fewer semantic motions compared with the given description. With the help of compact discrete motion tokens, TM2T [17] and T2M-GPT [55] achieve better results compared with MDM [44] and MLD [5]. Our motion generation results conform to the textual descriptions and exhibit a high degree of diversity compared with other state-of-the-art methods, which validates the efficacy of our proposed method.
+
+# 4.3. Ablation studies
+
+We delved into the significance of various quantization techniques, as delineated in the initial four rows of Tab.1. The findings indicate that a rudimentary Transformer-based VQ-VAE struggles to reconstruct credible motion sequences, primarily due to the limited efficacy of the codebook. However, the incorporation of the  $\ell_2$  norm and  $\mathcal{L}_{\mathrm{orth}}$  ensures that each motion token acquires a distinct motion representation. We further scrutinized the utilization of each motion token on the HumanML3D [16] test set, with the outcomes depicted in Fig.6. As evidenced by Fig.6a and 6b, both the vanilla CNN-based VQ-VAE and Transformer-based VQ-VAE grapple with sub-optimal codebook usage. This is because the tokens in the codebook aren't inherently driven to develop a distinguishing motion representation. Yet, with the aid of the  $\ell_2$  norm and  $\mathcal{L}_{\mathrm{orth}}$ , there's a marked surge in the codebook token utilization. Moreover, when juxtaposing the codebook usage in Fig.6c with that in Fig.6d, the latter exhibits a more equitably distributed frequency, attributable to the orthogonal regularization term  $\mathcal{L}_{\mathrm{orth}}$ .
+
+# 5. Conclusion
+
+In this study, we introduced the priority-centric motion discrete diffusion model (M2DM) tailored for text-to-motion generation. Our model excels in producing varied and life-like human motions, aligning seamlessly with the input text. This is achieved by mastering a discrete motion representa
+
+![](images/fc329d554b712ebab8445885af79b588165483f118280db747c3031a33d97614.jpg)
+
+![](images/28883f5084cde44709e4afc09a8ae2903a48ef63c332b77913d0e163e9082e7b.jpg)  
+(a) 1D-CNN VQ-VAE.  
+Figure 6: Codebook usage for different training strategies on HumanML3D [16] test set. We train our VQ-VAE (Reconstruction) with different training strategy on the HumanML3D [16] dataset.
+
+![](images/c0e1017151f31bbe46a5760b61b3892cd922ab459b8f8b4271862dbc78c61cd1.jpg)
+
+![](images/e99bc9909e6b52f64dded8cc3908bfbee41524111947c4e675d4ffad96c19899.jpg)  
+(c) Our VQ-VAE w.  $\ell_2$  norm.  
+(b) Our VQ-VAE.  
+(d) Our VQ-VAE w.  $\ell_2$  and  $\mathcal{L}_{\mathrm{orth}}$
+
+tion through a Transformer-based VQ-VAE and implementing a priority-informed noise schedule within the motion discrete diffusion framework. Additionally, we unveiled two innovative techniques for gauging the priority or significance of each motion token. Experimental evaluations across two datasets underscore our model's competitive edge, outpacing existing methodologies in R-Precision and diversity, particularly with intricate textual narratives.
+
+# Acknowledgement
+
+This project is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award Number: MOE-T2EP20122-0006).
+
+# References
+
+[1] Cmu graphics lab motion capture database. http://mocap.cs.cmu.edu/.6  
+[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 7, 8  
+[3] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34:17981-17993, 2021. 3, 5  
+[4] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE virtual reality and 3D user interfaces (VR), pages 1-10. IEEE, 2021. 7, 8
+
+[5] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. arXiv preprint arXiv:2212.04048, 2022. 1, 3, 6, 7, 8, 9  
+[6] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. arXiv preprint arXiv:2212.04495, 2022. 3  
+[7] Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341, 2020. 3  
+[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3  
+[9] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3  
+[10] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16091-16101, 2023. 3  
+[11] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. arXiv preprint arXiv:2305.10924, 2023. 3  
+[12] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1396-1406, 2021. 7, 8  
+[13] Kehong Gong, Dongze Lian, Heng Chang, Chuan Guo, Zihang Jiang, Xinxin Zuo, Michael Bi Mi, and Xinchao Wang. TM2D: Bimodality Driven 3D Dance Generation via Music-Text Integration. In IEEE/CVF International Conference on Computer Vision, 2023. 3  
+[14] Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and LingPeng Kong. Diffuseq: Sequence to sequence text generation with diffusion models. arXiv preprint arXiv:2210.08933, 2022. 1, 3  
+[15] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 1  
+[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 3, 6, 7, 8, 9  
+[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 580-597. Springer, 2022. 1, 3, 7, 8, 9
+
+[18] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 1, 3  
+[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7  
+[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1, 3, 5  
+[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23(47):1-33, 2022. 3  
+[22] Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forre, and Max Welling. Argmax flows and multinomial diffusion: Towards non-autoregressive language models. arXiv preprint arXiv:2102.05379, 3(4):5, 2021. 1, 3, 5  
+[23] Leslie Pack Kaelbling, Michael L Littman, and Andrew W Moore. Reinforcement learning: A survey. Journal of artificial intelligence research, 4:237-285, 1996. 2  
+[24] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. arXiv preprint arXiv:2209.00349, 2022. 3  
+[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6  
+[26] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3  
+[27] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. Advances in neural information processing systems, 32, 2019. 1, 3, 7, 8  
+[28] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1272-1279, 2022. 1, 3  
+[29] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 1, 3  
+[30] Angela S. Lin, Lemeng Wu, Rodolfo Corona, Kevin Tai, Qixing Huang, and Raymond J. Mooney. Generating animated videos of human activities from natural language descriptions. In Proceedings of the Visually Grounded Interaction and Language Workshop at NeurIPS 2018, December 2018. 7, 8  
+[31] Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. The kit whole-body human motion database. In 2015 International Conference on Advanced Robotics (ICAR), pages 329-336. IEEE, 2015. 6  
+[32] Larry R Medsker and LC Jain. Recurrent neural networks. Design and Applications, 5:64-67, 2001. 3
+
+[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 3, 5  
+[34] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 1, 3  
+[35] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXII, pages 480–497. Springer, 2022. 1  
+[36] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236–252, 2016. 2, 6, 7, 8  
+[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 6  
+[38] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 1, 3  
+[39] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3  
+[40] Sucheng Ren, Xingyi Yang, Songhua Liu, and Xinchao Wang. SG-Former: Self-guided Transformer with Evolving Token Reallocation. In IEEE International Conference on Computer Vision, 2023. 3  
+[41] Sucheng Ren, Daquan Zhou, Shengfeng He, Jiashi Feng, and Xinchao Wang. Shunted Self-Attention via Multi-Scale Token Aggregation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3  
+[42] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3, 5  
+[43] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXII, pages 358–374. Springer, 2022. 1  
+[44] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 1, 3, 6, 7, 8, 9  
+[45] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 7, 8
+
+[46] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 2, 3  
+[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 3, 5  
+[48] Xingyi Yang and Xinchao Wang. Diffusion Model as Representation Learner. In IEEE/CVF International Conference on Computer Vision, 2023. 3  
+[49] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In European Conference on Computer Vision, 2022. 3  
+[50] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion Probabilistic Model Made Slim. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3  
+[51] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In Advances in neural information processing systems, 2022. 3  
+[52] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 3  
+[53] Runpeng Yu, Songhua Liu, Xingyi Yang, and Xinchao Wang. Distribution Shift Inversion for Out-of-Distribution Prediction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3  
+[54] Weihao Yu, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, and Xinchao Wang. Metaformer Baselines for Vision. arXiv preprint arXiv:2210.13452, 2022. 3  
+[55] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023. 1, 3, 6, 7, 8, 9  
+[56] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 1, 3, 7, 8

2023/Priority-Centric Human Motion Generation in Discrete Latent Space/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4e724f4f6e517ae72a4402bea866272999e20389bfb57145bfdbf5eb4d43751
+size 814208

2023/Privacy Preserving Localization via Coordinate Permutations/f992eeb4-2e60-42f9-9e28-abc98f428b2a_content_list.json

@@ -0,0 +1,1754 @@
+[
+    {
+        "type": "text",
+        "text": "Privacy Preserving Localization via Coordinate Permutations",
+        "text_level": 1,
+        "bbox": [
+            174,
+            130,
+            795,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Linfei Pan<sup>1</sup> Johannes L. Schonberger<sup>2</sup> Viktor Larsson<sup>3</sup> Marc Pollefeys<sup>1,2</sup> ETH Zurich Microsoft Lund University",
+        "bbox": [
+            163,
+            179,
+            844,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1{linfei.pan,marc.pollefeys}@inf.ethz.ch,2joschonb@microsoft.com,3viktor.larsson@math.lth.se",
+        "bbox": [
+            84,
+            215,
+            919,
+            234
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            268,
+            313,
+            284
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent methods on privacy-preserving image-based localization use a random line parameterization to protect the privacy of query images and database maps. The lifting of points to lines effectively drops one of the two geometric constraints traditionally used with point-to-point correspondences in structure-based localization. This leads to a significant loss of accuracy for the privacy-preserving methods. In this paper, we overcome this limitation by devising a coordinate permutation scheme that allows for recovering the original point positions during pose estimation. The recovered points provide the full 2D geometric constraints and enable us to close the gap between privacy-preserving and traditional methods in terms of accuracy. Another limitation of random line methods is their vulnerability to density based 3D line cloud inversion attacks. Our method not only provides better accuracy than the original random line based approach but also provides stronger privacy guarantees against these recently proposed attacks. Extensive experiments on standard benchmark datasets demonstrate these improvements consistently across both scenarios of protecting the privacy of query images as well as the database map.",
+        "bbox": [
+            75,
+            301,
+            473,
+            633
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            662,
+            209,
+            679
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The interest in commercial solutions for cloud-based localization and mapping in mixed reality and robotics (e.g., Google VPS [34], Microsoft Azure Spatial Anchors [21], or Facebook LiveMaps [1]) has raised a host of privacy concerns [24, 31, 35, 51]. Recently, several approaches have been introduced by the research community to address these concerns [14, 17-19, 25, 45-47]. The common theme underlying these methods is based on the principle of lifting traditional point-based features to randomly oriented lines to conceal the appearance of query images and database maps.",
+        "bbox": [
+            75,
+            688,
+            468,
+            840
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "While this family of approaches is successful in enabling privacy preserving image-based localization and mapping, it comes at significant trade-offs in terms of accuracy and recall. These trade-offs are primarily caused by the fact that",
+        "bbox": [
+            75,
+            839,
+            470,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/524c6b154c2a14c46cacf05b0c2a3aa503f66a41bac75c032d1d12660794369b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            270,
+            696,
+            416
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/d2399fcb7658fe3c79a7ec4e74dd914c3cacf990cf989a1d6588aca9083aa944.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            431,
+            696,
+            536
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/e1b5aada20d4f28e8c3633c703f0a32d355ad8acbd868d1aef5abaa15185b738.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            277,
+            888,
+            415
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/f95403eda44190c1f50465ad34d41ad45cafe1ca676559edecc618a2a2c298e6.jpg",
+        "image_caption": [
+            "Figure 1: Privacy-preserving localization. Top Left: The client extracts keypoints and descriptors from the query image. Top Right: The client secretly pairs keypoints and permutes their coordinates; hiding their true position from the server. Bottom Left: Server performs localization using the permuted points and recovers the original 2D positions for points that exist in both query and map. Bottom Right: Feature inversion on the set of recovered points only reveals objects that are already present in the map."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            420,
+            887,
+            566
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "the privacy preserving line-to-point or point-to-line correspondences provide weaker geometric constraints than the traditional point-to-point correspondences [46]. When concealing the content of images or maps, the number of constraints is effectively halved, as each point-to-line or line-to-point reprojection only contributes a 1D geometric constraint compared to the 2D constraint from a point-to-point reprojection. Especially for the hard to localize images, where only few correspondences can be recovered during feature matching, this limitation makes a significant difference in terms of accuracy and recall.",
+        "bbox": [
+            496,
+            734,
+            893,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "18174",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, without careful construction of random line maps, density based attacks have been shown to be effective in uncovering information from the privacy preserving representations, which was already mentioned as a limitation by Speciale et al. [46]. The main idea behind density based attacks exploits the fact that pairs of randomly oriented lines have a relatively higher chance of intersecting near the true surface as compared to empty or unobserved space. As such, a naive attack could simply gather statistics from pairwise intersections of random lines and assume the true surface in dense regions of the intersected space. More sophisticated attacks based on this idea can yield realistic image renderings from 3D line maps when attacking naively constructed line maps [9].",
+        "bbox": [
+            75,
+            90,
+            472,
+            301
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this work, we revisit the approach of lifting points to lines in order to improve upon these two main limitations. In contrast to the original approach, we introduce a coordinate permutation scheme by which we effectively construct axis-aligned lines. During the pose estimation routine, the permutation scheme enables us to recover the original points for a large fraction of the inlier correspondences. Therefore, we can use the same 2D geometric constraints as in the traditional setup. This significantly improves the accuracy of the localization results. Moreover, when applying our approach for hiding the 3D map, we drastically increase the combinatorial complexity for density based attacks as well as reduce the chance for line intersections due to constructing axis-aligned lines. Counter-intuitively, our proposed method therefore not only leads to better accuracy but also provides a stronger level of privacy preservation. We experimentally demonstrate these improvements on standard benchmark datasets and by attacking our axis-aligned line maps using the approach by Chelani et al. [9]. In summary, we make the following contributions:",
+        "bbox": [
+            75,
+            301,
+            472,
+            604
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Propose a new privacy-preserving lifting approach, applicable to protecting the map and query.",
+            "- Attacking the lifted representation has a higher combinatorial complexity compared to random lifting, rendering previous density based attacks ineffective.",
+            "- In contrast to previous work, we can recover the original 2D or 3D points, allowing for stronger geometric constraints and more accurate poses.",
+            "- Experimentally, we show significant improvements over previous privacy preserving work, closing the gap to traditional point-based non-private methods."
+        ],
+        "bbox": [
+            96,
+            613,
+            467,
+            808
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            829,
+            218,
+            845
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Image-based Localization Methods in image-based localization have made tremendous progress. Over the years, the methods have been scaled for robust localization in",
+        "bbox": [
+            75,
+            854,
+            468,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "large-scale scenes [27,38,40,54] with compressed map representations [8,15], while some of the methods can even run in real-time on mobile devices [4,20,22,27,28,39]. Some of the recent works focus on improving the robustness to drastic environmental changes [3,42,50] or enabling crossdevice localization and mapping scenarios [13]. The developed approaches are generally divided into structure-based methods [20, 38-40] using on an explicit geometric map representation or learning-based methods [7,22,44] with an implicit map representation.",
+        "bbox": [
+            496,
+            90,
+            890,
+            243
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Privacy Risks Image-based localization is subject to privacy risks due to requiring images as an input [24, 31, 35, 51]. For example, in a cloud-based localization system, users may be concerned about sharing images of private spaces with the service provider. In a practical setting, the clients could only send local image features to the cloud to avoid these risks. However, recent inversion attacks have proven extremely effective in recovering the original images from only sparse local image features [12, 30, 32]. Recently, Chelani et al. proposed a method to query a localization service with a database of objects to recover information about scene contents [10].",
+        "bbox": [
+            496,
+            261,
+            890,
+            443
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Privacy Preservation To protect against these risks, several methods have been developed recently. The first work on this topic [46, 47] proposed a solution to preserve the privacy of 2D image or 3D map points in image-based localization. Their main idea was to lift 2D or 3D points to random lines by dropping one of the two geometric constraints during the pose estimation problem. Based on the same idea, a number of works followed up to address the problems of structure-from-motion [18, 19] and real-time SLAM [45], or to directly protect the high-dimensional feature representations instead of the geometry [14].",
+        "bbox": [
+            496,
+            462,
+            890,
+            628
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Underlying all these approaches, there are two fundamental limitations that were already observed by the original work [47]. First, by dropping one geometric constraint per correspondence for pose estimation, the accuracy of these approaches is significantly lowered, especially in the challenging cases with few matches between query and database. Second, without careful sparsification of the 3D map, density based attacks can recover information from 3D line clouds [9]. Sparsification of the map is an effective protection against density based attacks but comes with a further reduction of accuracy as well as recall.",
+        "bbox": [
+            496,
+            628,
+            890,
+            794
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To protect in a more principled manner against these attacks, Geppert et al. [17] proposed to decompose the full 3D map into 1D partial maps, which are distributed across different servers. Each server then solves a partial localization problem and sends back the answer to the client, which can assemble a full 6-DoF localization result from the partial answers. Despite being effective against density based at",
+        "bbox": [
+            496,
+            795,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "18175",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/4cf0f0d1a72c8c2e4aa121a4a6dd209e9abbe199ab036ac03a486d34397c0373.jpg",
+        "image_caption": [
+            "Figure 2: Method overview for localization with private queries. The client extracts keypoints in the query image and secretly pairs them to swap one of their coordinates. The server only receives the permuted points, which can be considered as axis-aligned lines. The secretly paired points form rectangles, which can be leveraged during RANSAC to efficiently recover the original point positions for pairs that are both inliers. The scoring of candidate poses and the final non-linear pose refinement jointly leverage the stronger 2D constraints for recovered points and otherwise the weaker 1D constraints."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            87,
+            410,
+            215
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/9723d42c230ae8c976a9a35e81051ea175c704f4c571dc3ced2864ac61d4fe8e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            416,
+            88,
+            846,
+            215
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "tacks, the system is complex and it comes with further accuracy and recall trade-offs. In contrast, our method does not require a complex multi-server setup and we improve upon the two main limitations of the random lifting approaches simultaneously. Our method provides both stronger privacy guarantees while also closing the accuracy gap to traditional localization approaches.",
+        "bbox": [
+            75,
+            310,
+            472,
+            416
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            76,
+            440,
+            168,
+            455
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We now present our method for privacy-preserving localization. We address both the scenario of protecting the client's 2D image points in the localization query (Section 3.2) as well as protecting the service's 3D points in the database map (Section 3.3).",
+        "bbox": [
+            75,
+            470,
+            468,
+            546
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our approach follows the main idea in Speciale et al. [46, 47], where the exact 2D or 3D point positions are hidden by introducing ambiguity in the coordinates. In the original works, this is done by lifting each point to a random line. The line representation still enables localization, using line-to-point instead of point-to-point correspondences, but makes it difficult to identify potentially sensitive image or map content. Similar to the original idea [46], we also obscure the true position of the point. However, instead of replacing it with a random line, we select a subset of the point coordinates and corrupt them. This is done by secretly forming pairs of points and swapping their coordinates at random. Since the list of pairs is not known to an attacker, it is practically intractable to recover the original point positions. As we will later show, for point-pairs that exist in both the query and the map, we can devise a scheme to recover the true positions, allowing us to use the full 2D-3D geometric constraints for accurate pose refinement.",
+        "bbox": [
+            75,
+            549,
+            470,
+            821
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "An overview of our pipeline can be found in Figure 2 and we detail the steps of our method in the following sections. For ease of presentation, we first review the traditional setup and then explain the 2D setting that hides the query, but the same principle generalizes to the 3D setting as well.",
+        "bbox": [
+            75,
+            824,
+            468,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Review of Traditional Localization",
+        "text_level": 1,
+        "bbox": [
+            498,
+            309,
+            802,
+            324
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Traditionally, localization is performed by establishing a set of tentative 2D-3D point correspondences between the query image and the database map. Each correspondence  $(\\pmb{x},\\pmb{X})\\in \\mathbb{R}^2\\times \\mathbb{R}^3$  yields two geometric constraints on the unknown camera pose  $(R,t)\\in \\mathrm{SE}(3)$  as",
+        "bbox": [
+            496,
+            332,
+            890,
+            409
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\lambda \\bar {\\boldsymbol {x}} = R \\boldsymbol {X} + \\boldsymbol {t}, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            637,
+            417,
+            890,
+            431
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where we let  $\\bar{x}$  denote the homogeneous 2D point in the normalized image plane (after removing the intrinsic calibration), and  $\\lambda$  represents the unknown depth. To identify outlier matches and estimate an initial pose, the standard approach is to use hypothesize-and-test frameworks, such as RANSAC [16], in order to generate multiple pose candidates by fitting  $(R,t)$  to minimal subsets of the points, followed by measuring the consensus among the rest of the matches. The consistency between a candidate pose and a 2D-3D match is computed using the 2D reprojection error",
+        "bbox": [
+            496,
+            440,
+            892,
+            592
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nd _ {p} = \\| \\boldsymbol {x} - \\pi (R \\boldsymbol {X} + t) \\|, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            604,
+            599,
+            890,
+            618
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\pi : \\mathbb{R}^3 \\to \\mathbb{R}^2$  is the projection mapping. Once a candidate pose, together with the inlier matches, have been established, it is refined by minimizing the reprojection errors for all inliers jointly.",
+        "bbox": [
+            496,
+            625,
+            890,
+            686
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Localization with Private Queries",
+        "text_level": 1,
+        "bbox": [
+            500,
+            693,
+            794,
+            709
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The traditional method relies on the availability of the original 2D image points, and is thus not privacy preserving. To preserve the privacy of the image, Speciale et al. [47] proposed to lift 2D image points into random 2D lines. Our method also builds upon 2D lines but aligns them with the coordinate axes through a random coordinate permutation scheme. The following paragraphs first establish notation by reviewing (random) line constraints for localization and then explain our permutation scheme in detail.",
+        "bbox": [
+            496,
+            715,
+            890,
+            852
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Random Line Constraints In Speciale et al. [47], each 2D line  $\\ell$  back-projects into a 3D plane  $\\Pi_i = P^\\top \\ell_i$  passing",
+        "bbox": [
+            500,
+            869,
+            892,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "18176",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/1bbdcd6bee67e2cee3b5d12cf251658bcbefee1565373d3daa6b45f8cfb3caa5.jpg",
+        "image_caption": [
+            "Figure 3: Illustration of geometric constraints induced by coordinate permutation and point recovery mechanism in the scenario of private queries. Red image points are secretly permuted. Black points are 3D map points. Blue points are the recovered, original image points obtained by back-swapping and form a rectangle together with their respective permuted points."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            155,
+            87,
+            395,
+            218
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "through camera projection center and the line in the image plane. The corresponding 3D point  $\\mathbf{X}$  is expected to lie on this plane, which constrains the pose as",
+        "bbox": [
+            75,
+            345,
+            470,
+            392
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\boldsymbol {\\ell} ^ {\\top} \\left(R \\boldsymbol {X} + \\boldsymbol {t}\\right) = 0. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            202,
+            402,
+            468,
+            422
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Note that this only yields a single geometric constraint instead of two as in Equation (1). In this case, it is still possible to estimate the pose, though the minimal problem now requires 6 instead of 3 correspondences. The minimal solver for this problem is known as  $p6LP$  [33] and was used in [47]. Further, to validate a match, we can only measure the distance between the 2D line and the projected 3D point",
+        "bbox": [
+            75,
+            435,
+            468,
+            541
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nd _ {l} = \\frac {\\left| \\boldsymbol {\\ell} ^ {\\top} \\bar {\\pi} (\\boldsymbol {R} \\boldsymbol {X} + \\boldsymbol {t}) \\right|}{\\sqrt {l _ {1} ^ {2} + l _ {2} ^ {2}}} , \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            553,
+            468,
+            592
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\ell = [l_1, l_2, l_3]^\\top$  and  $\\bar{\\pi}$  is the homogeneous projected 2D point. This is a weaker constraint compared to the 2D error measured in Equation (2) and thus explains the observed loss in accuracy by this approach.",
+        "bbox": [
+            75,
+            604,
+            468,
+            666
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Coordinate Permutation Constraints In our case, we have 2D image points, where we know that only one of the two coordinates is correct because of the a priori secret permutation (illustrated in Figures 2 and 3). In the context of the line-based framework from Speciale et al. [47], this can be considered as, instead of having a random line, having two axis-aligned lines passing through the given point. The original point will share one of the coordinates, and thus lie on one of these two lines. Mathematically, each point  $\\pmb{x} = [x_1,x_2]^\\top$  gives rise to two axis-aligned lines  $\\ell_1,\\ell_2$ , where  $\\ell_{1} = [1,0, - x_{1}]^{\\top}$ $\\ell_{2} = [0,1, - x_{2}]^{\\top}$ . Each correspondence then gives us the following constraint",
+        "bbox": [
+            75,
+            689,
+            470,
+            872
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\ell_ {1} ^ {\\top} (R \\boldsymbol {X} + \\boldsymbol {t}) = 0 \\quad \\vee \\quad \\ell_ {2} ^ {\\top} (R \\boldsymbol {X} + \\boldsymbol {t}) = 0, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            116,
+            883,
+            468,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where we only know that one of the two constraints should be satisfied. For  $n$  point correspondences, we thus have  $2^{n}$  possible combinations of constraints to choose from.",
+        "bbox": [
+            496,
+            90,
+            890,
+            135
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "While considering all  $n$  points leads to a combinatorial explosion and would be computationally intractable, we only need to consider small subsets of matches to hypothesize candidate poses inside RANSAC. For a minimal sample of 6 correspondences, there exist  $2^6 = 64$  different combinations of lines in Equation (5). One of these combinations will correspond to having the true (uncorrupted) coordinate in all 6 lines, and should give a correct camera pose for an all-inlier minimal sample. So, to find a camera pose, we can simply iterate exhaustively over all  $2^6 = 64$  combinations. Each of these combinations corresponds to a  $p6LP$  minimal estimation problem, which can be efficiently solved. If gravity is known, we can efficiently solve the minimal  $up4LP$  problem [33] from  $2^4 = 16$  combinations.",
+        "bbox": [
+            496,
+            137,
+            892,
+            348
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The key observation is now that, if we have a candidate pose, then for any inlier correspondence, we can easily resolve which of the two axis-aligned lines (i.e., coordinates) is the correct one by simply projecting the 3D points to the image and taking the one with the smaller 1D residual  $d_{l}$ . By determining the inlier coordinates  $\\mathcal{I}$  within the threshold (i.e.,  $d_{l} < \\varepsilon$ ), we can score candidate poses and thereby embed our method into a standard RANSAC loop to robustly estimate a camera pose. This approach still uses the weaker 1D constraints. The next paragraph explains how we can efficiently recover the original 2D points in this framework to benefit from stronger 2D constraints.",
+        "bbox": [
+            496,
+            349,
+            893,
+            530
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Original Point Recovery After finding a candidate pose using RANSAC, we devise an efficient scheme for recovering the original points  $\\pmb{x}$  from the set of inliers  $\\mathcal{I}$ . Note that our scheme can only recover the positions for points that are visible in both the query and the map. As such, recovering the original points is not a concern for privacy, as they are known to both client and server.",
+        "bbox": [
+            496,
+            551,
+            890,
+            656
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Let's recall that we randomly picked pairs of points and swapped one of their coordinates to preserve their privacy. To recover the original points, we need to rely on the projected points and the property that the swapped back points should be close to the projected points (within  $\\varepsilon$  distance) if it is an inlier. As for outliers points, there is no proper mechanism to distinguish a correct pair from an incorrect one, thus they remain unrecoverable.",
+        "bbox": [
+            496,
+            657,
+            890,
+            779
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "One way of recovering points is to iterate through all pairwise combinations of inlier correspondences and choose a configuration with the most swapped back points close to the projected points. However, this method has a squared complexity in the number of correspondences and is thus computationally expensive to execute inside a RANSAC loop. To achieve a better complexity, it is helpful to notice that each pair of points, that are selected for",
+        "bbox": [
+            496,
+            780,
+            892,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "18177",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "coordinate permutation, their original points and the permuted points form a 2D rectangle in the image (c.f. Figures 2 and 3). Using this observation, we can devise a more efficient recovery algorithm. For bookkeeping of recovered point coordinates, we first define the two sets",
+        "bbox": [
+            75,
+            90,
+            468,
+            167
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS _ {j} = \\left\\{\\left(x _ {k j}, k\\right) \\mid \\hat {\\boldsymbol {x}} _ {k} = \\left[ x _ {k 1} \\quad x _ {k 2} \\right] ^ {\\top} \\right\\}, \\quad j = \\{1, 2 \\} \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            89,
+            176,
+            468,
+            196
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Each element in the sets consists of a single coordinate  $(x_{1}$  or  $x_{2})$  and its index  $k$  in the ordered list of points. The index uniquely identifies coordinates to enable swapping of point pairs. For example, if the query image has 3 points  $\\{\\hat{x}_1 = [0,10]^\\top ,\\hat{x}_2 = [20,30]^\\top ,\\hat{x}_3 = [40,50]^\\top \\}$ , then  $S_{1} = \\{(0,1),(20,2),(40,3)\\}$ ,  $S_{2} = \\{(10,1),(30,2),(50,3)\\}$ . For each projected image point  $\\tilde{x}$ , we first decide which coordinate could be the original one by comparing  $\\tilde{x}$  and the permuted image point  $\\hat{x}$ . We label coordinate  $j$  as correct if the  $j$ -th coordinate is within  $\\varepsilon$  from the projected point, that is  $|\\hat{x}_j - \\tilde{x}_j| < \\varepsilon$ . Denoting the other coordinate as  $j'$ , we collect the possible swapped point candidates  $C$  by looking in the set  $S_{j'}$ , and finding points with the respective coordinate close to the projected point",
+        "bbox": [
+            75,
+            205,
+            470,
+            417
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nC = \\left\\{k \\mid \\left| x _ {k j ^ {\\prime}} - \\tilde {x} _ {j ^ {\\prime}} \\right| <   \\varepsilon , \\left(x _ {k j ^ {\\prime}}, k\\right) \\in S _ {j ^ {\\prime}} \\right\\} \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            125,
+            428,
+            468,
+            446
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Since this set likely contains multiple candidates for each point (having a similar  $x_{1}$  or  $x_{2}$  coordinate), and the candidate points after swapping back can be distant from their projected points, we use a symmetric swapping error to filter out the coincident candidates",
+        "bbox": [
+            75,
+            455,
+            468,
+            531
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\varepsilon_ {s y m} = \\left\\| \\tilde {\\boldsymbol {x}} - \\left[ \\begin{array}{l l} x _ {j} & x _ {k j ^ {\\prime}} \\end{array} \\right] ^ {\\top} \\right\\| _ {2} + \\left\\| \\tilde {\\boldsymbol {x}} _ {k} - \\left[ \\begin{array}{l l} x _ {k j} & x _ {j ^ {\\prime}} \\end{array} \\right] ^ {\\top} \\right\\| _ {2} (8)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            84,
+            541,
+            468,
+            561
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We only consider a candidate to be valid, if both points are close to the respective reprojected image points after swapping them back. If multiple candidates are valid, we keep the one with the smallest  $\\varepsilon_{sym}$ . Notice that, with such a selection mechanism, the recovered original points may not be symmetric. However, we have not found any drawbacks from this asymmetry in our experiments.",
+        "bbox": [
+            75,
+            571,
+            468,
+            676
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "By implementing the sets  $S_{j}$  using hashing, we achieve amortized linear runtime complexity on average w.r.t. the number of correspondences for our recovery scheme. More details can be found in the suppl. material.",
+        "bbox": [
+            75,
+            676,
+            468,
+            738
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Pose Refinement To further improve the accuracy of our pose estimates, we conduct a non-linear refinement over all inliers using the stronger 2D constraints from Equation (2) for recovered points  $\\mathcal{I}_p$  and the weaker 1D constraints from Equation (4) for coordinates  $\\mathcal{I}_l$  that could not be swapped back. The overall cost function is defined as",
+        "bbox": [
+            75,
+            755,
+            468,
+            845
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\sum_ {k \\in \\mathcal {I} _ {p}} \\rho \\left(d _ {p} \\left(\\boldsymbol {x} _ {k}, P \\bar {\\boldsymbol {X}} _ {k}\\right)\\right) + \\sum_ {k \\in \\mathcal {I} _ {\\ell}} \\rho \\left(\\underset {i \\in \\{1, 2 \\}} {\\operatorname {s o f t m i n}} \\left(d _ {l} \\left(\\boldsymbol {\\ell} _ {k i}, P \\bar {\\boldsymbol {X}} _ {k}\\right)\\right)\\right) \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            86,
+            852,
+            468,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/2dbf1298f0c8dbce52b528f0c2f52a9a9c516458f56de08ce214d7bce1591ee1.jpg",
+        "image_caption": [
+            "Figure 4: Illustration of point recovery mechanism in the 3D scenario with private maps, where the permuted red points form a cuboid in 3D. The reprojection of the 3D points is used for thresholding of errors in image space."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            576,
+            87,
+            818,
+            210
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\rho$  is a robustifier. To smoothly select between the two possible axis-aligned lines, we use the (Boltzmann) soft-min operator [5], defined as  $\\operatorname*{softmin}_{i}\\{d_{i}\\} = (\\sum_{i}d_{i}e^{-d_{i}}) / (\\sum_{i}e^{-d_{i}})$ . We incorporate this non-linear refinement into LO-RANSAC [11] using the Arctan loss as  $\\rho$  inside the RANSAC loop and Cauchy as the robust loss for the final pose refinement.",
+        "bbox": [
+            498,
+            292,
+            892,
+            404
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Degenerate Configurations Note that when all  $\\ell_{i}$  lines are parallel (i.e., originating from the same coordinate), the configuration becomes degenerate as the full 6-DoF of the pose cannot be recovered. This happens for 2 out of 64 configurations that we can trivially discard. For the two degenerate cases, it would be possible to solve for the rotation and two translation parameters using 5 correspondences. We initially explored this direction as well yet found the solutions to be less stable, as they provide less accurate pose estimates, so we do not include them in our final method.",
+        "bbox": [
+            498,
+            422,
+            890,
+            571
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3. Localization with Private Maps",
+        "text_level": 1,
+        "bbox": [
+            500,
+            580,
+            777,
+            598
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the previous section, we described our coordinate permutation scheme for 2D image points to preserve the privacy of localization queries. Similar to the 2D case, we can also permute coordinates of 3D points to hide the contents of private maps. This is a scenario that was originally addressed by Speciale et al. [46] using lifting to random 3D lines. We denote a random 3D line  $L$  as",
+        "bbox": [
+            498,
+            604,
+            890,
+            709
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\boldsymbol {L} = \\left[ \\begin{array}{l} \\boldsymbol {p} \\\\ \\boldsymbol {d} \\end{array} \\right] \\in \\mathbb {R} ^ {3} \\times \\mathbb {P} ^ {2}, \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            617,
+            717,
+            890,
+            751
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $p$  and  $q$  are 3D points on  $L$  subject to  $d = \\frac{p - q}{\\|p - q\\|_2}$  and such that the line passes through the original 3D point  $X$ . The reprojection of the 3D line to the image defines a 1D geometric constraint",
+        "bbox": [
+            498,
+            761,
+            890,
+            833
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n0 = \\hat {\\boldsymbol {L}} _ {i} ^ {\\top} \\bar {\\boldsymbol {x}} = \\left(\\left[ \\boldsymbol {R} \\boldsymbol {d} _ {i} \\right] _ {\\times} \\boldsymbol {P} \\bar {\\boldsymbol {p}} _ {i}\\right) ^ {\\top} \\bar {\\boldsymbol {x}} \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            591,
+            840,
+            890,
+            861
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "that can be used for privacy-preserving pose estimation. To solve for the full 6-DoF pose, at least 6 correspondences are",
+        "bbox": [
+            498,
+            869,
+            890,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "18178",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "needed to solve the minimal  $p6L$  problem [48] in RANSAC. If gravity is known, only 4 correspondences are needed to solve the up4L problem [49].",
+        "bbox": [
+            75,
+            90,
+            468,
+            136
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Instead of lifting to random 3D lines, we construct axis-aligned 3D lines by coordinate permutation of randomly selected pairs of 3D points. Analogous to the 2D case, we can hypothesize pose candidates in RANSAC by generating the different  $3^{6} = 729$  combinations of line constraints, and we can also recover the original 3D point positions by back-swapping with sets  $S_{j=1,2,3}$ . Notice that the original 3D point may lie on any of the three axis-aligned lines and the randomly selected point pairs form a cuboid in 3D (see Figure 4). To robustify the back-swapping algorithm against non-isotropic noise on the (triangulated) 3D points, we define the error  $\\varepsilon_{sym}$  in image space through reprojection. Moreover, if runtime is a concern for the final application, we can also choose to expose one of the 3 coordinates, so the number of line combinations would be  $2^{6} = 64$ .",
+        "bbox": [
+            75,
+            136,
+            470,
+            363
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Pose Refinement For pose refinement in the 3D case, we again add constraints from two sources: the ones from inlier lines  $\\mathcal{I}_L$  and the ones from recovered inlier points  $\\mathcal{I}_p$ . The overall cost function is defined as",
+        "bbox": [
+            75,
+            380,
+            468,
+            439
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\sum_ {k \\in \\mathcal {I} _ {p}} \\rho (d _ {p} (\\boldsymbol {x} _ {k}, P \\bar {\\boldsymbol {X}} _ {k})) + \\sum_ {k \\in \\mathcal {I} _ {L}} \\rho (\\underset {i \\in \\{1, 2 \\}} {\\operatorname {s o f t m i n}} (d _ {L} (\\boldsymbol {x} _ {k}, \\hat {\\boldsymbol {L}} _ {k i}))\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            86,
+            441,
+            454,
+            478
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\text {w h e r e} d _ {L} = \\frac {\\left| \\hat {\\boldsymbol {L}} _ {k i} ^ {\\top} \\bar {\\boldsymbol {x}} _ {k} \\right|}{\\sqrt {\\hat {L} _ {k i 1} ^ {2} + \\hat {L} _ {k i 2} ^ {2}}} \\tag {12}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            179,
+            481,
+            468,
+            532
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "and we use the same setup of LO-RANSAC and choice of robustifiers as in the 2D scenario.",
+        "bbox": [
+            75,
+            532,
+            468,
+            563
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            575,
+            209,
+            590
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In our experiments, we first focus on demonstrating the effectiveness of our proposed point recovery mechanism using an ablation study. We then compare our method against the traditional non-privacy preserving approach (PnP) and the original privacy-preserving approach based on random lines. Results are reported on existing benchmark datasets [23,46,52] for image-based localization, evaluated for the two scenarios of private localization queries (PnLP) and maps (PnL). All results assume unknown gravity and we show further results with known gravity in the supplementary material. Finally, we show image inversion results using the density-based attack by Chelani et al. [9].",
+        "bbox": [
+            75,
+            599,
+            470,
+            782
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details We embed our method into LO-RANSAC [11] with Levenberg-Marquardt [2, 26] as nonlinear optimizer for pose refinement. Inlier reprojection thresholds on 2D points are set to  $12px$  for the Cambridge dataset [23] and to  $4px$  otherwise. Reprojection thresholds on lines are chosen as  $\\frac{12px}{\\sqrt{2}}$  and  $\\frac{4px}{\\sqrt{2}}$ , respectively.",
+        "bbox": [
+            75,
+            799,
+            468,
+            904
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/fcfdda0c24f7add157c62919900c928caae6e6cf651a9c77f77fd29735bad04c.jpg",
+        "image_caption": [
+            "Figure 5: Ablation study on King's college sequence [23] w.r.t. the number of correspondences. Our method closes the performance gap, that is especially pronounced in difficult cases with few correspondences."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            90,
+            700,
+            217
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/f9b3f6097d3455aa7522fe3b6bc4d97e2303d939e895b7637f10002ab5974189.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            704,
+            90,
+            888,
+            217
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We re-implement all methods in the same framework for a fair comparison of accuracy, recall, and runtime. All point clouds are triangulated using COLMAP [41, 43].",
+        "bbox": [
+            496,
+            295,
+            892,
+            340
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            498,
+            351,
+            653,
+            368
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1.1 Impact of Number of Correspondences",
+        "text_level": 1,
+        "bbox": [
+            498,
+            375,
+            825,
+            391
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As argued in previous sections, the reduced accuracy and recall in privacy-preserving localization methods based on random lines is largely attributed to the weaker geometric constraints. To showcase this performance gap, we conduct an experiment on the King's College sequence in the Cambridge dataset [23]. Localization performance is tested on varying number of correspondences per query image by randomly dropping input matches. Cases with few correspondences typically depict the challenging localization scenarios, where feature matching only yields few matches and only few constraints can be used for pose estimation. The results of this experiment are summarized in Figure 5. First, we observe that the performance gap between the privacy-preserving and traditional approach is especially pronounced for low number of matches, which in turn suggests that the impact of the number of constraints on the localization accuracy saturates eventually. We also see that our method closes the accuracy gap to the traditional approach across the board.",
+        "bbox": [
+            496,
+            400,
+            890,
+            688
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1.2 Impact of Point Recovery Mechanism",
+        "text_level": 1,
+        "bbox": [
+            498,
+            709,
+            815,
+            724
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To analyze the effectiveness of the proposed point recovery mechanism, we conduct experiments on HoloLens data [47] with our method in different settings: with only line constraints, with recovered point constraints, and with the true recovered points using an oracle that has access to groundtruth information. This dataset contains images with a large variance on the inlier ratio, which makes it suitable for this ablation study, as the point recovery mechanism is strongly influenced by it. Results are summarized in Figure 6, which shows a significant improvement in accuracy for our method when using the stronger 2D constraints from",
+        "bbox": [
+            496,
+            734,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "18179",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/14eebd814c76fc5b5200e557d5bdc68f0f072e276edf43984cfcf95cb849b390.jpg",
+        "image_caption": [
+            "Figure 6: Ablation study on HoloLens data from [47] to measure the effectiveness of our point recovery mechanism."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            89,
+            285,
+            218
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2e74e1d0727393db298b76a34c2d0c4f550394adde9bc3a584292f974f9ea90a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            290,
+            89,
+            467,
+            217
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/03bb19e9f7286bd9f392c3f6c0693956d0e5f30e5362b5516d9af07430396d66.jpg",
+        "image_caption": [
+            "Figure 7: Ablation study on HoloLens data from [47] that shows effective point recovery across all outlier ratios and overall low error rates on the recovered points."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            270,
+            269,
+            387
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/ebb9e667a9b4211c02f21788eac833da46a54bbd88fcd7d8e77b84786fb0ad3e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            274,
+            272,
+            464,
+            388
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/d5705481af6e29edd51ab27014749c570972495691838be58e84959d8286c4cd.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>MED (°)</td><td>AUC@1°</td><td>MED (cm)</td><td>AUC@10cm</td><td>(1°, 10 cm)</td><td>Time (s)</td></tr><tr><td>PnLP [47]</td><td>0.08</td><td>84.79</td><td>0.63</td><td>88.88</td><td>97.80</td><td>0.07</td></tr><tr><td>Proposed</td><td>0.07</td><td>86.69</td><td>0.56</td><td>90.12</td><td>98.41</td><td>0.36</td></tr><tr><td>PnP</td><td>0.07</td><td>87.92</td><td>0.51</td><td>90.85</td><td>98.80</td><td>0.05</td></tr></table>",
+        "bbox": [
+            78,
+            457,
+            467,
+            507
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "recovered points. In comparison to the oracle that defines a theoretical upper bound for our method, there is no significant difference in performance.",
+        "bbox": [
+            75,
+            561,
+            468,
+            606
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, we also report the correlation between the outlier ratio and the ratio of recovered points in Figure 7. The expected number is calculated as  $p^2$  where  $p$  is the inlier ratio. From the figure we can see that the ratios of correctly recovered points matches the expected number across all outlier ratios. The ratio of mismatched points stays low in this dataset even when the inlier ratio is as low as  $20\\%$ . Those observations combined provide strong evidence that the recovering mechanism is effective.",
+        "bbox": [
+            75,
+            608,
+            468,
+            743
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2. Localization with Private Queries",
+        "text_level": 1,
+        "bbox": [
+            76,
+            755,
+            372,
+            771
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "HoloLens [47] This dataset features mostly small-scale scenes and unconstrained motion trajectories from a headworn augmented reality device with VGA-resolution cameras. Equivalent to the original dataset, we use COLMAP for reconstruction of the full image sequence while limiting the number of SIFT features per image to at most 500. There are 3325 query images in total and each query has on average 320 correspondences. After reconstructing the full",
+        "bbox": [
+            75,
+            779,
+            470,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/fa3f599bcb1979fcf78caaa3e9c785abeea0f2cdb230f41a39f55d48ff6e1205.jpg",
+        "table_caption": [
+            "Table 1: Localization performance on HoloLens dataset [47] with indoor and outdoor scenes."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>MED (°)</td><td>AUC@1°</td><td>MED (cm)</td><td>AUC@10cm</td><td>(1°, 10 cm)</td><td>Time (s)</td></tr><tr><td>PnLP [47]</td><td>0.18</td><td>71.59</td><td>12.13</td><td>18.38</td><td>41.29</td><td>0.36</td></tr><tr><td>Proposed</td><td>0.18</td><td>72.61</td><td>11.73</td><td>20.03</td><td>44.26</td><td>2.31</td></tr><tr><td>PnP</td><td>0.17</td><td>73.09</td><td>11.66</td><td>20.22</td><td>43.95</td><td>0.20</td></tr></table>",
+        "bbox": [
+            501,
+            88,
+            890,
+            137
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/11ca6fa186ad75b9cb2b80c4a3d785898f2c78cd5836473dd96becbd430700fa.jpg",
+        "table_caption": [
+            "Table 2: Localization performance on Cambridge dataset [23] with handheld videos in an outdoor scene."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>MED (°)</td><td>AUC@1°</td><td>MED (cm)</td><td>AUC@10cm</td><td>(1°, 10 cm)</td><td>Time (s)</td></tr><tr><td>PnLP [47]</td><td>0.02</td><td>88.32</td><td>2.24</td><td>59.57</td><td>78.23</td><td>0.09</td></tr><tr><td>Proposed</td><td>0.02</td><td>89.67</td><td>1.92</td><td>63.62</td><td>81.33</td><td>0.68</td></tr><tr><td>PnP</td><td>0.02</td><td>90.74</td><td>1.85</td><td>64.58</td><td>82.62</td><td>0.08</td></tr></table>",
+        "bbox": [
+            503,
+            186,
+            890,
+            237
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/40d5dc88f55da7c587fdc4c15498dd877fd8fcbf0ac336b1401460227a8cf8fa.jpg",
+        "table_caption": [
+            "Table 3: Localization performance on crowd-sourced Internet photo collections [29, 53]."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>MED (°)</td><td>AUC@1°</td><td>MED (cm)</td><td>AUC@10cm</td><td>(1°, 10 cm)</td><td>Time (s)</td></tr><tr><td>PnL [46]</td><td>0.08</td><td>84.71</td><td>0.64</td><td>88.66</td><td>97.71</td><td>1.16</td></tr><tr><td>Proposed</td><td>0.07</td><td>86.31</td><td>0.56</td><td>89.71</td><td>97.92</td><td>6.61</td></tr><tr><td>PnP</td><td>0.07</td><td>87.85</td><td>0.51</td><td>90.84</td><td>98.83</td><td>0.07</td></tr></table>",
+        "bbox": [
+            501,
+            287,
+            890,
+            335
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Localization performance on HoloLens dataset [47] in the private map scenario.",
+        "bbox": [
+            500,
+            337,
+            890,
+            367
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "sequence, we hold out every 5-th image as a query, remove any point correspondences associated to these images from the reconstruction, and retriangulate the remaining database images. The pose of the query images in the full reconstruction serve as the groundtruth for evaluation. The results are summarized in Table 1, where we can see that our method performs almost on par with the traditional non-privacy preserving case in terms of accuracy and recall. The additional combinatorial overhead of evaluating more pose candidates leads to a roughly 5-fold runtime increase for our method.",
+        "bbox": [
+            496,
+            378,
+            890,
+            531
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Cambridge [23] This localization benchmark dataset contains handheld HD-resolution videos taken around Cambridge University in an outdoor setting. We use the same reference reconstruction as in [6] for the database map and extract SuperPoint+SuperGlue feature correspondences using the hloc toolbox [36, 37]. The dataset has a total of 1918 query images and an average of around 2850 2D-3D correspondences per image. Results are summarized in Table 2 and the performance trend for accuracy, recall, and runtime are all consistent with the results on the lower-resolution HoloLens dataset. This underlines the robustness of our approach against different sensors, scene types, as well as feature detection and matching algorithm characteristics.",
+        "bbox": [
+            496,
+            549,
+            892,
+            746
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Internet Photo Collections [29, 53] This collection of datasets is crowd-sourced from the internet and comprises a large variety in image resolution and quality, lighting conditions, etc. We conduct experiments on the Dubrovnik6k, Gendarmenmarkt, Roman Forum, and Trafalgar locations with equivalent settings as used by Speciale et al. [47] for generating pseudo-groundtruth and establishing feature correspondences using COLMAP. On average, there are around 1000 SIFT correspondences per image with a total",
+        "bbox": [
+            496,
+            763,
+            893,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "18180",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/b2d44ac88d4801cda4a69e67eff0099aa6ac971d574fc65992836e440b62a3e1.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            76,
+            88,
+            285,
+            207
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/9e551d266bf5c6131ff9f1420b864b965c16cf4c23f9347fe6f2c9c3737b0751.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            285,
+            92,
+            486,
+            207
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5ce5c8a2a75b9c85a500d9449ddf849627a70ba93664f8742a97889ad1cdab2e.jpg",
+        "image_caption": [
+            "(a) Original"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            209,
+            179,
+            285
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ef8e893c68bd4aca50b8e99e1d847e6ee9d40e13e87894723d44644cffd77ef9.jpg",
+        "image_caption": [
+            "Figure 8: Privacy attack on (private) maps using a combination of 3D image inversion [32] and the density based attack method Line2Point [9]. The top row shows the 3D map that is recovered using Line2Point for all but the leftmost result. The 3D map is used as an input for rendering synthesized images shown at two different viewpoints in the bottom row. From left to right: (a) Results on original, non-privacy preserving point cloud. (b) Results on random line cloud. (c) Results on our method with an attacker that gained access to swapped coordinates. (d) Results on our method with permuted coordinates."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            179,
+            209,
+            279,
+            285
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/0233f679175491a3c125b02efcaf46f3165233349eaf8f6c03ec55f701754edd.jpg",
+        "image_caption": [
+            "(b) Random line"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            285,
+            209,
+            383,
+            285
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/c59a2479938f7529e748de30739078938d93956484356f4f46a8c62e3aedb6b8.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            383,
+            209,
+            482,
+            285
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ca60b5e169f291d2dc35afd49444e32767fdeea66416909f0708ad60724ac263.jpg",
+        "image_caption": [
+            "(c) Oracle"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            486,
+            209,
+            586,
+            285
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/38e85b2d1cf80a6ef1f57c80a35ff301f74873b4b134a5eba57d8db2b52ca315.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            586,
+            209,
+            684,
+            285
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/fb1986a0c95d80e0536062c8f1af82e57bb9d9ad03280206dd6acdc9175ef71b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            687,
+            88,
+            890,
+            207
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/a1d8eb92aad51fb1caf5f23b5f0ea870cc03274335a0a97f00fe4940b677823a.jpg",
+        "image_caption": [
+            "(d) Proposed"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            689,
+            209,
+            787,
+            285
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/fb2b36e85cb9bceaf814a9c5b6ebfc224acad267e2d8ef4bcd0ab15679a7b6a4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            790,
+            209,
+            888,
+            285
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "of 2416 query images. Results are summarized in Table 3. The accuracy of all methods on this dataset is surprisingly high, yet we can still observe significant improvements.",
+        "bbox": [
+            75,
+            393,
+            468,
+            441
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.3. Localization with a Private Map",
+        "text_level": 1,
+        "bbox": [
+            76,
+            450,
+            359,
+            465
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To evaluate localization scenario with a private map, we conduct experiments on HoloLens data [47] using the same setup as in the private query scenario described in the previous section. We adopt the setting of exposing one correct coordinate in Table 4 and provide results for the computationally more expensive case with 729 combinations in the supplementary material. Like in the private query scenario, we are able to significantly close the accuracy gap but increase run-time by a factor of around six.",
+        "bbox": [
+            75,
+            474,
+            468,
+            609
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.4. Inversion Attack on Line Cloud",
+        "text_level": 1,
+        "bbox": [
+            76,
+            619,
+            354,
+            633
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this section, we analyze and compare the level of privacy protection for the map representation. We perform this analysis by attacking the 3D point maps using image inversion techniques, as proposed by Pittaluga et al. [32]. To render images from the privacy preserving 3D line maps, we first execute a density based attack on the map representations using the Line2Point [9] method and then run image inversion on the recovered point cloud. Figure 8 shows results for this experiment, where inverted images from the non-privacy preserving map are nearly indistinguishable from original images of the scene. This again highlights the importance of research on privacy preserving methods for image-based localization. Images rendered from naively constructed random line clouds without sparsification closely resemble the original images as well and thus are not sufficient to protect private information in the map. The results of the attack on our method in column (d)",
+        "bbox": [
+            75,
+            643,
+            468,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "shows mostly recognizable images and demonstrates the increased level of privacy that our method provides. Even if an attacker had access to an oracle that knows about which of the coordinates is correct, the rendered images (c) are less recognizable than the ones from a random line cloud. It is important to note that this information is not available to an attacker in practice, but it allows us to understand the relative impact on the increased level of privacy. On the one hand, our method constructs axis-aligned lines, which, by construction, cannot be intersected in a density based attack. On the other hand, we increase the combinatorial complexity of executing a density based attack, as significantly more coordinate pairs must be tried to find a valid intersection. In the supplementary material we provide additional inversion results and more discussion regarding potential attacks.",
+        "bbox": [
+            496,
+            393,
+            890,
+            621
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            638,
+            617,
+            654
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we present a significant improvement over existing privacy-preserving image-based localization approaches. Our method improves upon the accuracy limitation of random line based approaches and closes the gap to traditional non-privacy preserving methods. At the same time, our method proves as robust against recently presented density attacks on random 3D line clouds. An interesting avenue for future work could be in addressing the increased runtime of our method. One promising idea involves voting for correct coordinate swaps during RANSAC and progressively sampling from more likely lines.",
+        "bbox": [
+            496,
+            666,
+            890,
+            832
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments We would like to thank Marcel Geppert for reviewing this paper, for the insightful discussions, and his help in the supp. material, as well as Torsten Sattler and Kunal Chelani for the code and data of line cloud inversion experiments. Viktor Larsson was supported by ELLIIT.",
+        "bbox": [
+            496,
+            840,
+            890,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "18181",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            174,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Inside Facebook Reality Labs: Research updates and the future of social connection. https://tech.fb.com/inside-facebook-reality-labs-research-updates-and-the-future-of-social-connection/, 2019. 1",
+            "[2] Sameer Agarwal, Keir Mierle, and The Ceres Solver Team. Ceres Solver, 3 2022. 6",
+            "[3] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN architecture for weakly supervised place recognition. In Computer Vision and Pattern Recognition (CVPR), 2016. 2",
+            "[4] Clemens Arth, Daniel Wagner, Manfred Klopschitz, Arnold Irschara, and Dieter Schmalstieg. Wide area localization on mobile phones. In International Symposium on Mixed and Augmented Reality (ISMAR), 2009. 2",
+            "[5] Kavosh Asadi and Michael L Littman. An alternative softmax operator for reinforcement learning. In International Conference on Machine Learning, pages 243-252. PMLR, 2017. 5",
+            "[6] Eric Brachmann, Martin Humenberger, Carsten Rother, and Torsten Sattler. On the limits of pseudo ground truth in visual camera re-localisation. In International Conference on Computer Vision (ICCV), 2021. 7",
+            "[7] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. DSAC-differentiable RANSAC for camera localization. In Computer Vision and Pattern Recognition (CVPR), 2017. 2",
+            "[8] Song Cao and Noah Snavely. Minimal scene descriptions from structure from motion models. In Computer Vision and Pattern Recognition (CVPR), 2014. 2",
+            "[9] Kunal Chelani, Fredrik Kahl, and Torsten Sattler. How privacy-preserving are line clouds? recovering scene details from 3d lines. In Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6, 8",
+            "[10] Kunal Chelani, Torsten Sattler, Fredrik Kahl, and Zuzana Kukelova. Privacy-preserving representations are not enough: Recovering scene content from camera poses. In CVPR, 2023. 2",
+            "[11] Ondrej Chum, Jiří Matas, and Josef Kittler. Locally optimized ransac. In Joint Pattern Recognition Symposium, 2003. 5, 6",
+            "[12] Alexey Dosovitskiy and Thomas Brox. Inverting visual representations with convolutional networks. In Computer Vision and Pattern Recognition (CVPR), 2016. 2",
+            "[13] Mihai Dusmanu, Ondrej Miksik, Johannes Lutz Schonberger, and Marc Pollefeys. Cross-Descriptor Visual Localization and Mapping. In International Conference on Computer Vision (ICCV), 2021. 2",
+            "[14] Mihai Dusmanu, Johannes Lutz Schonberger, Sudipta Sinha, and Marc Pollefeys. Privacy-Preserving Image Features via Adversarial Affine Subspace Embeddings. In Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2",
+            "[15] Marcin Dymczyk, Simon Lynen, Michael Bosse, and Roland Siegwart. Keep it brief: Scalable creation of compressed localization maps. In International Conference on Intelligent Robots and Systems (IROS), 2015. 2"
+        ],
+        "bbox": [
+            78,
+            114,
+            468,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[16] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM (CACM), 1981. 3",
+            "[17] Marcel Geppert, Viktor Larsson, Johannes Lutz Schonberger, and Marc Pollefeys. Privacy Preserving Partial Localization. In Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2",
+            "[18] Marcel Geppert, Viktor Larsson, Pablo Speciale, Johannes Lutz Schonberger, and Marc Pollefeys. Privacy Preserving Structure-from-Motion. In European Conference on Computer Vision (ECCV), 2020. 1, 2",
+            "[19] Marcel Geppert, Viktor Larsson, Pablo Speciale, Johannes L. Schonberger, and Marc Pollefeys. Privacy preserving localization and mapping from uncalibrated cameras. In Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2",
+            "[20] Arnold Irschara, Christopher Zach, Jan-Michael Frahm, and Horst Bischof. From structure-from-motion point clouds to fast location recognition. In Computer Vision and Pattern Recognition (CVPR), 2009. 2",
+            "[21] Neena Kamath. Announcing Azure Spatial Anchors for collaborative, cross-platform mixed reality apps. https://azure.microsoft.com/en-us/blog/announcing-azure-spatial-anchors-for-collaborative-cross-platform-mixed-reality-apps/, 2019.1",
+            "[22] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In International Conference on Computer Vision (ICCV), 2015. 2",
+            "[23] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Proceedings of the IEEE international conference on computer vision, pages 2938-2946, 2015. 6, 7",
+            "[24] Alex Kipman. Azure Spatial Anchors approach to privacy and ethical design. https://www.linkedin.com/pulse/azure-spatial-anchors-approach-privacy-ethical-design-alex-kipman, 2019. 1, 2",
+            "[25] Chunghwan Lee, Jaihoon Kim, Chanhyuk Yun, and Je Hyeong Hong. Paired-point lifting for enhanced privacy-preserving visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17266-17275, 2023. 1",
+            "[26] Kenneth Levenberg. A method for the solution of certain non-linear problems in least squares. Quarterly of applied mathematics, 2(2):164-168, 1944. 6",
+            "[27] Yunpeng Li, Noah Snavely, Dan Huttenlocher, and Pascal Fua. Worldwide pose estimation using 3d point clouds. In European Conference on Computer Vision (ECCV), 2012. 2",
+            "[28] Yunpeng Li, Noah Snavely, and Daniel P Huttenlocher. Location recognition using prioritized feature matching. In European Conference on Computer Vision (ECCV), 2010. 2",
+            "[29] Yunpeng Li, Noah Snavely, and Daniel P Huttenlocher. Location recognition using prioritized feature matching. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part II 11, pages 791-804. Springer, 2010. 7"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "18182",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[30] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In Computer Vision and Pattern Recognition (CVPR), 2015. 2",
+            "[31] Mary Lynne Nielsen. Augmented Reality and its Impact on the Internet, Security, and Privacy. https://beyondstandards.ieee.org/augmented-reality/augmented-reality-and-its-impact-on-the-internet-security-and-privacy/, 2015. 1, 2",
+            "[32] Francesco Pittaluga, Sanjeev J Koppal, Sing Bing Kang, and Sudipta N Sinha. Revealing scenes by inverting structure from motion reconstructions. In Computer Vision and Pattern Recognition (CVPR), 2019. 2, 8",
+            "[33] Srikumar Ramalingam and Yuichi Taguchi. A theory of minimal 3d point to 3d plane registration and its generalization. International Conference on Computer Vision (ICCV), 2013. 4",
+            "[34] Tilman Reinhardt. Google Visual Positioning Service. https://ai.googleblog.com/2019/02/using-global-localization-to-improve.html, 2019. 1",
+            "[35] Franziska Roesner. Who Is Thinking About Security and Privacy for Augmented Reality? https://www.technologyreview.com/s/609143/who-is-thinking-about-security-and-privacy-for-augmented-reality/, 2017. 1, 2",
+            "[36] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Computer Vision and Pattern Recognition (CVPR), 2019. 7",
+            "[37] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. 7",
+            "[38] Torsten Sattler, Michal Havlena, Filip Radenovic, Konrad Schindler, and Marc Pollefeys. Hyperpoints and fine vocabularies for large-scale location recognition. In International Conference on Computer Vision (ICCV), 2015. 2",
+            "[39] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Fast image-based localization using direct 2d-to-3d matching. In International Conference on Computer Vision (ICCV), 2011. 2",
+            "[40] T. Sattler, B. Leibe, and L. Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. Trans. Pattern Analysis and Machine Intelligence (PAMI), 2017. 2",
+            "[41] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Computer Vision and Pattern Recognition (CVPR), 2016. 6",
+            "[42] Johannes L Schonberger, Marc Pollefeys, Andreas Geiger, and Torsten Sattler. Semantic visual localization. Computer Vision and Pattern Recognition (CVPR), 2018. 2",
+            "[43] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise View Selection for Unstructured Multi-View Stereo. In European Conference on Computer Vision (ECCV), 2016. 6",
+            "[44] Paul Hongsuck Seo, Tobias Weyand, Jack Sim, and Bohyung Han. Cplanet: Enhancing image geolocation by combinatorial partitioning of maps. In European Conference on Computer Vision (ECCV), 2018. 2"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[45] Mikiya Shibuya, Shinya Sumikura, and Ken Sakurada. Privacy preserving visual SLAM. In European Conference on Computer Vision (ECCV), 2020. 1, 2",
+            "[46] Pablo Speciale, Johannes L. Schonberger, Sing Bing Kang, Sudipta Sinha, and Marc Pollefeys. Privacy Preserving Image-Based Localization. In Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 3, 5, 6, 7",
+            "[47] Pablo Speciale, Johannes L. Schonberger, Sudipta N. Sinha, and Marc Pollefeys. Privacy preserving image queries for camera localization. In International Conference on Computer Vision (ICCV), 2019. 1, 2, 3, 4, 6, 7, 8",
+            "[48] Henrik Stewénius, David Nister, Magnus Oskarsson, and Kalle Astrom. Solutions to minimal generalized relative pose problems. In Workshop on omnidirectional vision, 2005. 6",
+            "[49] Chris Sweeney, John Flynn, and Matthew Turk. Solving for relative pose with a partially known rotation is a quadratic eigenvalue problem. In International Conference on 3D Vision (3DV), 2014. 6",
+            "[50] Carl Toft, Will Maddern, Akihiko Torii, Lars Hammarstrand, Erik Stenberg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, et al. Long-term visual localization revisited. Trans. Pattern Analysis and Machine Intelligence (PAMI), 2020. 2",
+            "[51] Jan-Erik Vinje. Privacy Manifesto for AR Cloud Solutions. https://medium.com/openarcloud/privacy-manifesto-for-ar-cloud-solutions-9507543f50b6, 2018. 1, 2",
+            "[52] Kyle Wilson and Noah Snavely. Robust global translations with 1DSFM. In European Conference on Computer Vision (ECCV), 2014. 6",
+            "[53] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In European Conference on Computer Vision (ECCV), 2014. 7",
+            "[54] Bernhard Zeisl, Torsten Sattler, and Marc Pollefeys. Camera pose voting for large-scale image-based localization. In International Conference on Computer Vision (ICCV), 2015. 2"
+        ],
+        "bbox": [
+            501,
+            90,
+            890,
+            613
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "18183",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 9
+    }
+]