diff --git a/.gitattributes b/.gitattributes index 73a6f5c0f291892769b23145d7fb959cc2d138a1..ac8e1af9fcf2023ba88b3144c3ee96e497c3fe1d 100644 --- a/.gitattributes +++ b/.gitattributes @@ -4196,3 +4196,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/Move[[:space:]]as[[:space:]]You[[:space:]]Say[[:space:]]Interact[[:space:]]as[[:space:]]You[[:space:]]Can_[[:space:]]Language-guided[[:space:]]Human[[:space:]]Motion[[:space:]]Generation[[:space:]]with[[:space:]]Scene[[:space:]]Affordance/63d40f73-31a4-4c03-9015-7428005fa39b_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/MovieChat_[[:space:]]From[[:space:]]Dense[[:space:]]Token[[:space:]]to[[:space:]]Sparse[[:space:]]Memory[[:space:]]for[[:space:]]Long[[:space:]]Video[[:space:]]Understanding/87b1b3a8-af53-4dbf-8cc6-34b0b0d0a75a_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/MuGE_[[:space:]]Multiple[[:space:]]Granularity[[:space:]]Edge[[:space:]]Detection/a7e2ae2d-2070-4adb-a467-fe719afb4075_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/MuRF_[[:space:]]Multi-Baseline[[:space:]]Radiance[[:space:]]Fields/265b8c06-80fc-4e85-8d50-8ff0b6769d10_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Mudslide_[[:space:]]A[[:space:]]Universal[[:space:]]Nuclear[[:space:]]Instance[[:space:]]Segmentation[[:space:]]Method/5a64aa1e-710a-4a50-b851-b34aaaa3cbbb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-Attribute[[:space:]]Interactions[[:space:]]Matter[[:space:]]for[[:space:]]3D[[:space:]]Visual[[:space:]]Grounding/477f7931-e1b2-442e-a45b-11cf7d648a49_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-Level[[:space:]]Neural[[:space:]]Scene[[:space:]]Graphs[[:space:]]for[[:space:]]Dynamic[[:space:]]Urban[[:space:]]Environments/10ec9eda-ceb6-4b88-aa59-8a3b78a66785_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-Modal[[:space:]]Hallucination[[:space:]]Control[[:space:]]by[[:space:]]Visual[[:space:]]Information[[:space:]]Grounding/1e5cae18-a58b-4415-90f8-bc4858438f19_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-Modal[[:space:]]Proxy[[:space:]]Learning[[:space:]]Towards[[:space:]]Personalized[[:space:]]Visual[[:space:]]Multiple[[:space:]]Clustering/f9afe157-eaaf-440c-afcf-7604121ce960_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-Object[[:space:]]Tracking[[:space:]]in[[:space:]]the[[:space:]]Dark/4c20e9a2-1106-45cf-a36a-0161982e6015_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-Scale[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Anti-Aliased[[:space:]]Rendering/20977810-85e5-40f1-a4c6-cf52c533cece_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-Scale[[:space:]]Video[[:space:]]Anomaly[[:space:]]Detection[[:space:]]by[[:space:]]Multi-Grained[[:space:]]Spatio-Temporal[[:space:]]Representation[[:space:]]Learning/82eacc56-0492-4ac2-b935-f0e92d5461ce_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-Session[[:space:]]SLAM[[:space:]]with[[:space:]]Differentiable[[:space:]]Wide-Baseline[[:space:]]Pose[[:space:]]Optimization/04406ae4-0193-431d-9be9-c73afd625413_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-Space[[:space:]]Alignments[[:space:]]Towards[[:space:]]Universal[[:space:]]LiDAR[[:space:]]Segmentation/b0f5aa24-c6e7-4f07-a0f8-7432b1bda5de_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-Task[[:space:]]Dense[[:space:]]Prediction[[:space:]]via[[:space:]]Mixture[[:space:]]of[[:space:]]Low-Rank[[:space:]]Experts/4b27d7d8-e816-4974-8903-c5d6fd2b2750_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-View[[:space:]]Attentive[[:space:]]Contextualization[[:space:]]for[[:space:]]Multi-View[[:space:]]3D[[:space:]]Object[[:space:]]Detection/cb273d05-390f-4fb9-aa74-74b5a46402e1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-agent[[:space:]]Collaborative[[:space:]]Perception[[:space:]]via[[:space:]]Motion-aware[[:space:]]Robust[[:space:]]Communication[[:space:]]Network/8f321c57-13fe-4536-adc3-b4b1a5e6a84e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-agent[[:space:]]Long-term[[:space:]]3D[[:space:]]Human[[:space:]]Pose[[:space:]]Forecasting[[:space:]]via[[:space:]]Interaction-aware[[:space:]]Trajectory[[:space:]]Conditioning/12609e19-efed-4425-b5ad-32eb26a81347_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-criteria[[:space:]]Token[[:space:]]Fusion[[:space:]]with[[:space:]]One-step-ahead[[:space:]]Attention[[:space:]]for[[:space:]]Efficient[[:space:]]Vision[[:space:]]Transformers/c38e4100-b4f3-4d92-aff9-1fdadd7cfa0a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-modal[[:space:]]In-Context[[:space:]]Learning[[:space:]]Makes[[:space:]]an[[:space:]]Ego-evolving[[:space:]]Scene[[:space:]]Text[[:space:]]Recognizer/bf855ad8-e560-4eb0-9886-16b2ff2266f1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-modal[[:space:]]Instruction[[:space:]]Tuned[[:space:]]LLMs[[:space:]]with[[:space:]]Fine-grained[[:space:]]Visual[[:space:]]Perception/1060b587-378f-4876-81ef-a69e77b5ca70_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-modal[[:space:]]Learning[[:space:]]for[[:space:]]Geospatial[[:space:]]Vegetation[[:space:]]Forecasting/42a10fad-a12f-4501-8672-d7229d356fee_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-scale[[:space:]]Dynamic[[:space:]]and[[:space:]]Hierarchical[[:space:]]Relationship[[:space:]]Modeling[[:space:]]for[[:space:]]Facial[[:space:]]Action[[:space:]]Units[[:space:]]Recognition/75181bea-24ed-4d67-89b7-aa0610066eda_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-view[[:space:]]Aggregation[[:space:]]Network[[:space:]]for[[:space:]]Dichotomous[[:space:]]Image[[:space:]]Segmentation/cce93abc-7f7e-4582-a2d5-107b621a8f82_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/MultiDiff_[[:space:]]Consistent[[:space:]]Novel[[:space:]]View[[:space:]]Synthesis[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Image/2aa2e82a-dff4-4750-bb74-ff506ca72a62_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/MultiPLY_[[:space:]]A[[:space:]]Multisensory[[:space:]]Object-Centric[[:space:]]Embodied[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]in[[:space:]]3D[[:space:]]World/6cdd4677-266b-499a-a637-8d38b4cfc6f3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/MultiPhys_[[:space:]]Multi-Person[[:space:]]Physics-aware[[:space:]]3D[[:space:]]Motion[[:space:]]Estimation/8bf95065-c5b4-4afa-86d1-804b7f567a4b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/MultiPly_[[:space:]]Reconstruction[[:space:]]of[[:space:]]Multiple[[:space:]]People[[:space:]]from[[:space:]]Monocular[[:space:]]Video[[:space:]]in[[:space:]]the[[:space:]]Wild/9118ef01-a3a3-4d0e-a512-cab61af53519_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multiagent[[:space:]]Multitraversal[[:space:]]Multimodal[[:space:]]Self-Driving_[[:space:]]Open[[:space:]]MARS[[:space:]]Dataset/daf2a5e1-efd2-4f29-b37c-4c73f81b7002_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multimodal[[:space:]]Industrial[[:space:]]Anomaly[[:space:]]Detection[[:space:]]by[[:space:]]Crossmodal[[:space:]]Feature[[:space:]]Mapping/37da17b3-6824-4605-81f8-c01f9eebd49d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multimodal[[:space:]]Pathway_[[:space:]]Improve[[:space:]]Transformers[[:space:]]with[[:space:]]Irrelevant[[:space:]]Data[[:space:]]from[[:space:]]Other[[:space:]]Modalities/31541608-9780-4c92-8dfb-6ccd462a5a0b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multimodal[[:space:]]Prompt[[:space:]]Perceiver_[[:space:]]Empower[[:space:]]Adaptiveness[[:space:]]Generalizability[[:space:]]and[[:space:]]Fidelity[[:space:]]for[[:space:]]All-in-One[[:space:]]Image[[:space:]]Restoration/dd270751-648d-4a9b-b4c9-9b85d339c5c3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multimodal[[:space:]]Representation[[:space:]]Learning[[:space:]]by[[:space:]]Alternating[[:space:]]Unimodal[[:space:]]Adaptation/9f47df9d-5a19-4287-a41d-2d4ffe80f5db_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multimodal[[:space:]]Sense-Informed[[:space:]]Forecasting[[:space:]]of[[:space:]]3D[[:space:]]Human[[:space:]]Motions/b5ef95c4-a9a0-4681-a4a8-822e1b33a34a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multiplane[[:space:]]Prior[[:space:]]Guided[[:space:]]Few-Shot[[:space:]]Aerial[[:space:]]Scene[[:space:]]Rendering/15e4c819-6e34-4bd4-b379-6775e97abc28_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multiple[[:space:]]View[[:space:]]Geometry[[:space:]]Transformers[[:space:]]for[[:space:]]3D[[:space:]]Human[[:space:]]Pose[[:space:]]Estimation/07b64eb8-929d-4d4b-b3f7-dd15c8bf344f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multiscale[[:space:]]Vision[[:space:]]Transformers[[:space:]]Meet[[:space:]]Bipartite[[:space:]]Matching[[:space:]]for[[:space:]]Efficient[[:space:]]Single-stage[[:space:]]Action[[:space:]]Localization/eeb58ba1-b49b-44b5-b8bc-8e43b4f9d2bc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multiview[[:space:]]Aerial[[:space:]]Visual[[:space:]]RECognition[[:space:]](MAVREC)_[[:space:]]Can[[:space:]]Multi-view[[:space:]]Improve[[:space:]]Aerial[[:space:]]Visual[[:space:]]Perception_/5ed3fe81-108b-429b-8fc0-08881954a458_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multiway[[:space:]]Point[[:space:]]Cloud[[:space:]]Mosaicking[[:space:]]with[[:space:]]Diffusion[[:space:]]and[[:space:]]Global[[:space:]]Optimization/d99639eb-7fa3-4be0-a146-25a4bc301b9e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/MuseChat_[[:space:]]A[[:space:]]Conversational[[:space:]]Music[[:space:]]Recommendation[[:space:]]System[[:space:]]for[[:space:]]Videos/f0f6aab1-0b41-4348-8488-4140eb3d81e2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NAPGuard_[[:space:]]Towards[[:space:]]Detecting[[:space:]]Naturalistic[[:space:]]Adversarial[[:space:]]Patches/139c013e-88a7-4162-8220-0b4f45d02eeb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NARUTO_[[:space:]]Neural[[:space:]]Active[[:space:]]Reconstruction[[:space:]]from[[:space:]]Uncertain[[:space:]]Target[[:space:]]Observations/b788f3e6-255f-4cf1-93ae-ce6f871b55a1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NAYER_[[:space:]]Noisy[[:space:]]Layer[[:space:]]Data[[:space:]]Generation[[:space:]]for[[:space:]]Efficient[[:space:]]and[[:space:]]Effective[[:space:]]Data-free[[:space:]]Knowledge[[:space:]]Distillation/400ba431-dcaa-4545-88f2-8b0361ba2923_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NB-GTR_[[:space:]]Narrow-Band[[:space:]]Guided[[:space:]]Turbulence[[:space:]]Removal/c76eed54-1b6a-4a11-9752-38ffa0dd036b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NC-SDF_[[:space:]]Enhancing[[:space:]]Indoor[[:space:]]Scene[[:space:]]Reconstruction[[:space:]]Using[[:space:]]Neural[[:space:]]SDFs[[:space:]]with[[:space:]]View-Dependent[[:space:]]Normal[[:space:]]Compensation/d308c00a-4405-4e9c-a230-0021d3336660_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NC-TTT_[[:space:]]A[[:space:]]Noise[[:space:]]Constrastive[[:space:]]Approach[[:space:]]for[[:space:]]Test-Time[[:space:]]Training/bd059d62-00e8-44dd-afc5-8486d7ebc6ba_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NEAT_[[:space:]]Distilling[[:space:]]3D[[:space:]]Wireframes[[:space:]]from[[:space:]]Neural[[:space:]]Attraction[[:space:]]Fields/2ecd2fa3-e051-4f87-90aa-cbfb130cb94d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NECA_[[:space:]]Neural[[:space:]]Customizable[[:space:]]Human[[:space:]]Avatar/1fd371b3-1c77-4438-ace5-2f95f345e90e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NICE_[[:space:]]Neurogenesis[[:space:]]Inspired[[:space:]]Contextual[[:space:]]Encoding[[:space:]]for[[:space:]]Replay-free[[:space:]]Class[[:space:]]Incremental[[:space:]]Learning/e2ee9981-d8f0-4b73-ae4d-2cddbed7fe01_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NIFTY_[[:space:]]Neural[[:space:]]Object[[:space:]]Interaction[[:space:]]Fields[[:space:]]for[[:space:]]Guided[[:space:]]Human[[:space:]]Motion[[:space:]]Synthesis/a26a4987-a6c2-4a19-82a7-dc8df7105be2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NIVeL_[[:space:]]Neural[[:space:]]Implicit[[:space:]]Vector[[:space:]]Layers[[:space:]]for[[:space:]]Text-to-Vector[[:space:]]Generation/2b4a188b-d422-4fff-bac9-a5ee93dcab15_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NOPE_[[:space:]]Novel[[:space:]]Object[[:space:]]Pose[[:space:]]Estimation[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Image/eaddf8a4-618b-423e-9b47-d54ca4506fe6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NRDF_[[:space:]]Neural[[:space:]]Riemannian[[:space:]]Distance[[:space:]]Fields[[:space:]]for[[:space:]]Learning[[:space:]]Articulated[[:space:]]Pose[[:space:]]Priors/3c4744e1-d7a1-4bb7-9c52-5e6d82e2c55c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NTO3D_[[:space:]]Neural[[:space:]]Target[[:space:]]Object[[:space:]]3D[[:space:]]Reconstruction[[:space:]]with[[:space:]]Segment[[:space:]]Anything/d75ec439-fb66-44d1-8326-a3b002c61709_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NViST_[[:space:]]In[[:space:]]the[[:space:]]Wild[[:space:]]New[[:space:]]View[[:space:]]Synthesis[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Image[[:space:]]with[[:space:]]Transformers/ef8ae36f-fce2-4270-87ba-de527ac2b27e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Named[[:space:]]Entity[[:space:]]Driven[[:space:]]Zero-Shot[[:space:]]Image[[:space:]]Manipulation/29b9477a-ce33-450d-b6a0-e8383eaf57c8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Narrative[[:space:]]Action[[:space:]]Evaluation[[:space:]]with[[:space:]]Prompt-Guided[[:space:]]Multimodal[[:space:]]Interaction/9f79c1fa-6fc9-4b63-b0a8-abfc48152309_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Naturally[[:space:]]Supervised[[:space:]]3D[[:space:]]Visual[[:space:]]Grounding[[:space:]]with[[:space:]]Language-Regularized[[:space:]]Concept[[:space:]]Learners/098143a6-2834-44b7-90c6-1c263dd6e214_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Navigate[[:space:]]Beyond[[:space:]]Shortcuts_[[:space:]]Debiased[[:space:]]Learning[[:space:]]Through[[:space:]]the[[:space:]]Lens[[:space:]]of[[:space:]]Neural[[:space:]]Collapse/c60a89d3-50ed-43e1-9489-517c526d3ab0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Navigating[[:space:]]Beyond[[:space:]]Dropout_[[:space:]]An[[:space:]]Intriguing[[:space:]]Solution[[:space:]]towards[[:space:]]Generalizable[[:space:]]Image[[:space:]]Super[[:space:]]Resolution/306f09fc-f3eb-4f7a-8e1c-e7932830585d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NeISF_[[:space:]]Neural[[:space:]]Incident[[:space:]]Stokes[[:space:]]Field[[:space:]]for[[:space:]]Geometry[[:space:]]and[[:space:]]Material[[:space:]]Estimation/0ccf1001-f8bd-45cf-af5d-a6678647b774_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NeLF-Pro_[[:space:]]Neural[[:space:]]Light[[:space:]]Field[[:space:]]Probes[[:space:]]for[[:space:]]Multi-Scale[[:space:]]Novel[[:space:]]View[[:space:]]Synthesis/582c225d-f7ae-4dbf-b523-c3d7d61b55fc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NeRF[[:space:]]Analogies_[[:space:]]Example-Based[[:space:]]Visual[[:space:]]Attribute[[:space:]]Transfer[[:space:]]for[[:space:]]NeRFs/18430570-b9c7-4f0b-bc14-bce6d48fd341_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NeRF[[:space:]]Director_[[:space:]]Revisiting[[:space:]]View[[:space:]]Selection[[:space:]]in[[:space:]]Neural[[:space:]]Volume[[:space:]]Rendering/af270895-9751-4bb4-a280-a3905369976f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NeRF[[:space:]]On-the-go_[[:space:]]Exploiting[[:space:]]Uncertainty[[:space:]]for[[:space:]]Distractor-free[[:space:]]NeRFs[[:space:]]in[[:space:]]the[[:space:]]Wild/84919b4e-7dcd-414d-8d02-3749ba7af706_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NeRF-HuGS_[[:space:]]Improved[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields[[:space:]]in[[:space:]]Non-static[[:space:]]Scenes[[:space:]]Using[[:space:]]Heuristics-Guided[[:space:]]Segmentation/ddc8147a-c1a1-4504-bf06-6c58b9172cc6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/NeRFCodec_[[:space:]]Neural[[:space:]]Feature[[:space:]]Compression[[:space:]]Meets[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields[[:space:]]for[[:space:]]Memory-Efficient[[:space:]]Scene[[:space:]]Representation/db013d96-7fde-4b15-9ee1-0e7d5955b6d6_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/MuRF_ Multi-Baseline Radiance Fields/265b8c06-80fc-4e85-8d50-8ff0b6769d10_content_list.json b/2024/MuRF_ Multi-Baseline Radiance Fields/265b8c06-80fc-4e85-8d50-8ff0b6769d10_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..037e0f9c9909224015a82f3345422ddf2703fdc0 --- /dev/null +++ b/2024/MuRF_ Multi-Baseline Radiance Fields/265b8c06-80fc-4e85-8d50-8ff0b6769d10_content_list.json @@ -0,0 +1,1371 @@ +[ + { + "type": "text", + "text": "MuRF: Multi-Baseline Radiance Fields", + "text_level": 1, + "bbox": [ + 285, + 130, + 684, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haofei $\\mathrm{Xu}^{1,2}$ Anpei Chen $^{1,2}$ Yuedong Chen $^{3}$ Christos Sakaridis $^{1}$ Yulun Zhang $^{4}$ Marc Pollefeys $^{1,5}$ Andreas Geiger $^{2,\\dagger}$ Fisher $\\mathrm{Yu}^{1,\\dagger}$", + "bbox": [ + 135, + 178, + 833, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ETH Zurich $^{2}$ University of Tübingen, Tübingen AI Center $^{3}$ Monash University $^{4}$ Shanghai Jiao Tong University $^{5}$ Microsoft $^{\\dagger}$ Joint last author", + "bbox": [ + 161, + 219, + 810, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 290, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Multi-Baseline Radiance Fields (MuRF), a general feed-forward approach to solving sparse view synthesis under multiple different baseline settings (small and large baselines, and different number of input views). To render a target novel view, we discretize the 3D space into planes parallel to the target image plane, and accordingly construct a target view frustum volume. Such a target volume representation is spatially aligned with the target view, which effectively aggregates relevant information from the input views for high-quality rendering. It also facilitates subsequent radiance field regression with a convolutional network thanks to its axis-aligned nature. The 3D context modeled by the convolutional network enables our method to synthesis sharper scene structures than prior works. Our MuRF achieves state-of-the-art performance across multiple different baseline settings and diverse scenarios ranging from simple objects (DTU) to complex indoor and outdoor scenes (RealEstate10K and LLFF). We also show promising zero-shot generalization abilities on the Mip-NeRF 360 dataset, demonstrating the general applicability of MuRF.", + "bbox": [ + 75, + 321, + 473, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 632, + 209, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel view synthesis from sparse input views is a critical and practical problem in computer vision and graphics. However, typical optimization-based Neural Radiance Fields (NeRFs) [1, 2, 4, 18] cannot cope well with sparse views, and additional regularizations [19, 29, 33] are usually required to better constrain the optimization process.", + "bbox": [ + 75, + 657, + 468, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim at learning feed-forward NeRF models [3, 32, 38] that are able to perform feed-forward inference on new unseen data, without requiring per-scene optimization. Moreover, inductive biases could be acquired from data and better results can potentially be attained.", + "bbox": [ + 75, + 750, + 470, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To achieve this, existing sparse view methods can be broadly classified into methods that take small baseline images as input and methods that focus on large baseline settings. However, none of the existing methods works well across different baselines (see Fig. 1a and Fig. 1b). In par", + "bbox": [ + 75, + 824, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5b16c347522680150641c89b3ef6c8996a33d53409241d46a6ab3f594fa890dd.jpg", + "image_caption": [ + "(a) MuRF outperforms previous state-of-the-art small baseline method ENeRF. The larger baseline, the larger performance gap." + ], + "image_footnote": [], + "bbox": [ + 501, + 290, + 694, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f74c0de3814f50ad9bc2ad0220474aa18c0732453508e8df039727a59f94f39.jpg", + "image_caption": [ + "(b) MuRF outperforms previous state-of-the-art large baseline method AttnRend. The smaller baseline, the larger performance gap." + ], + "image_footnote": [], + "bbox": [ + 699, + 290, + 890, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0d31ace908e061cc66813723fd604c73a8c4c2c0eeb837c007eb17ebec57785d.jpg", + "image_caption": [ + "Figure 1. MuRF supports multiple different baseline settings. Previous methods are specifically designed for either small (e.g., ENeRF [16]) or large (e.g., AttnRend1 [8]) baselines. However, no existing method performs well on both (see Table 1)." + ], + "image_footnote": [], + "bbox": [ + 506, + 455, + 880, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ticular, small baseline methods [3, 5, 6, 15-17, 32] rely heavily on feature matching to extract relevant information from input views and suffer significantly when there is insufficient image overlap or occlusion occurs. In contrast, large baseline methods [8, 24] are mostly data-driven, and often discard matching cues. Instead, they resort to large-scale datasets and generic architectures (e.g., Transformers [7, 30]) to learn geometric inductive biases implicitly. However, due to their correspondence-unaware and data-driven nature, even state-of-the-art methods [8, 24] tend to produce blurry renderings and their generalization ability is still limited, as we show in Fig. 3 and Table 6.", + "bbox": [ + 496, + 662, + 892, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our goal is to address sparse view synthesis under both", + "bbox": [ + 517, + 845, + 890, + 861 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1We name the previously unnamed method of Du et al. [8] as AttnRend, according its GitHub repository name.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20041", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "small and large camera baselines, different number of input views, and diverse scenarios ranging from simple objects to complex scenes. To this end, we introduce Multi-Baseline Radiance Fields (MuRF), a method for sparse view synthesis from multiple different baseline settings (Fig. 1).", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More specifically, to render a target novel view, we discretize the 3D space with planes parallel to the target image plane, and accordingly construct a target view frustum volume that is spatially aligned with the target view to render. This is a crucial difference to previous volume-based methods like MVSNeRF [3] and GeoNeRF [15], where the volumes are constructed in a pre-defined reference input view by straightforwardly following the standard practice in multi-view stereo (MVS). However, a key difference between MVS and novel view synthesis (NVS) is that MVS estimates the depth of a reference input view, while NVS aims to render a novel view. Taking this difference into account is crucial, since the reference view volume will not be able to effectively aggregate information from input views when the overlap between the reference and target view is small, which results in failure for large baselines (Fig. 5).", + "bbox": [ + 75, + 167, + 470, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our target view frustum volume is constructed by sampling multi-view input image colors and features, which provide appearance information to the prediction of the 3D point's color. We also compute the cosine similarities between sampled features to provide multi-view consistency cues to aid the prediction of volume density [5]. The sampled colors and features, and the computed cosine similarities are concatenated together to provide relevant information for both small and large baseline scenarios.", + "bbox": [ + 75, + 411, + 468, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Equipped with this axis-aligned target volume representation, we further propose to use a convolutional neural network (CNN) to reconstruct the radiance field from this volume. Thanks to the context modeling abilities of CNNs, our method yields more accurate scene structures than previous (MLP-based) point-wise [6, 8] and (Ray Transformer-based) ray-wise methods [5, 27, 32] (see Fig. 6).", + "bbox": [ + 75, + 549, + 468, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In practice, we perform $8 \\times$ subsampling in the spatial dimension when constructing the target volume, which enables our method to efficiently handle high-resolution images. The full resolution radiance field is finally obtained with a lightweight $8 \\times$ upsampler [25]. We also use a computationally more efficient 3D CNN alternative by factorizing the $3\\mathrm{D}3 \\times 3 \\times 3$ convolution to a $2\\mathrm{D}3 \\times 3 \\times 1$ convolution on the spatial dimension and a $1\\mathrm{D}1 \\times 1 \\times 3$ convolution on the depth dimension, i.e., $(2 + 1)\\mathrm{D}$ CNN, a popular strategy in video recognition works [9, 28].", + "bbox": [ + 75, + 657, + 468, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments to demonstrate the effectiveness of our proposed target view frustum volume and 3D context-aware radiance field decoder. More specifically, we outperform previous specialized models on both small (DTU [14], LLFF [18]) and large (RealEstate10K [40]) baselines, and achieve promising zero-short generalization", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "abilities on the Mip-NeRF 360 [2] dataset, indicating the general applicability of our method.", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 131, + 640, + 147 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Baseline. The concept of \"multi-baseline\" dates back to multi-baseline stereo depth estimation [10, 20, 36, 37], where multiple stereo pairs with various baselines are used to surpass matching ambiguities and thus achieve improved precision. In this paper, we aim at designing a generally applicable approach to solving novel view synthesis from multiple different sparse view settings. When considering multiple images as input, our problem setup is somewhat similar to multi-baseline stereo, while with the radiance field as the output for volumetric rendering. Besides, we are also interested in novel view synthesis from very sparse (as few as 2) input views with different baselines. In this sense, our proposed multi-baseline radiance field not only draws connections to classic multi-baseline stereo, but also extends it to view synthesis from 2 input views with both small and large baselines.", + "bbox": [ + 496, + 155, + 890, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sparse Input Novel View Synthesis. Applying NeRF to novel view synthesis with sparse input views has garnered considerable interest in recent years. Several works improve the optimization-based pipeline via the incorporation of additional training regularizations, e.g., depth [19], correspondence [29] and diffusion model [33]. An alternative area of research turns to the feed-forward designs by learning reliable representations from data, and accordingly removing the need of per-scene optimization. Among them, MVSNeRF [3], GeoNeRF [15] and NeuRay [17] follow the spirit of multi-view stereo (MVS) to construct cost volumes at pre-defined reference viewpoints. However, the reference volume will not be able to effectively aggregate the information for input views when the overlap between the reference and target novel view is small, which accordingly results in failure for large baselines. Besides, ENeRF [16] essentially relies on an MVS network to guide NeRF's sampling process with the estimated depth. However, the depth quality will be unreliable for large baselines and thus severely affects the subsequent rendering process.", + "bbox": [ + 496, + 400, + 890, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another line of works such as SRF [6], IBRNet [32], GPNR [26], GNT [27] and MatchNeRF [5] mostly perform ray-based rendering, where each ray is rendered independently and no explicit 3D context between different rays is modeled. In contrast, our target view volume representation effectively encodes the scene geometry and naturally enables 3D context modeling with a convolutional network, yielding better scene structures. Despite various motivations and implementations, the above models are all designed and experimented only on small baseline settings, limiting their applications in real-world scenarios. Only a few existing approaches focus on the large baseline settings. Nevertheless, these works [8, 24] resort to large-scale", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20042", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b3439f4e5879f86ff2b7aa67e614d9600a2bff2123cf7c85447371c9202d55a1.jpg", + "image_caption": [ + "Figure 2. Overview. Given multiple input images, we first extract multi-view image features with a multi-view Transformer. To render a target image of resolution $H \\times W$ , we construct a target view frustum volume by performing $8 \\times$ subsampling in the spatial dimension while casting rays and sampling $D$ equidistant points on each ray. For each 3D point, we sample feature and color information from the extracted feature maps and input images, which consists of the elements of the target volume $z \\in \\mathbb{R}^{\\frac{H}{8} \\times \\frac{W}{8} \\times D \\times C}$ . Here, $C$ denotes the channel dimension after aggregating sampled features and colors. To reconstruct the radiance field from the volume, we model the context information in the decoder with a (2+1)D CNN operating on low resolution and subsequently obtain the full-resolution radiance field with a lightweight $8 \\times$ upsampler. The target image is finally rendered with volumetric rendering." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 893, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "datasets and generic architectures, without employing explicit 3D representations, which ultimately leads to blurry rendering and limited generalization. As a comparison, our MuRF is developed with both geometry-aware target view frustum volume representation and 3D context-aware convolutional network, which make it excel at both small and large baseline settings on different datasets.", + "bbox": [ + 75, + 369, + 472, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 486, + 184, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our MuRF (Multi-Baseline Radiance Fields) is an encoder-decoder architecture (Fig. 2), where the encoder maps the multi-view input images to features that are used to construct a volume at the target view's camera frustum, and the decoder regresses the radiance field from this volume representation. The target image is finally obtained using volume rendering [18]. The key components are introduced below.", + "bbox": [ + 75, + 507, + 468, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multi-View Feature Encoder", + "text_level": 1, + "bbox": [ + 76, + 622, + 333, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To aggregate essential information required as input for learning feed-forward NeRF models, we first extract features $\\{F_k\\}_{k=1}^K$ from $K$ input images $\\{I_k\\}_{k=1}^K$ . Our feature encoder consists of a weight-sharing per-image 2D CNN and a joint multi-view Transformer. The CNN consists of 6 residual blocks [12] and every 2 residual blocks include a stride-2 convolutional layer to achieve $2 \\times$ downsampling. Accordingly, we obtain features at $1/2$ , $1/4$ and $1/8$ resolutions. The $1/8$ convolutional features are then fed into a joint multi-view Transformer to obtain features at $1/8$ resolution. Our Transformer is built on GMFlow [34, 35]'s 2-view Transformer architecture, where we extend the 2-view cross-attention to $K(K \\geq 2)$ input views by performing cross-attention for all $K$ views simultaneously. This facilitates more efficient processing of additional input views without significantly increasing memory footprint compared to the pair-wise architecture [5]. The multi-scale", + "bbox": [ + 75, + 643, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "features are sampled in next volume reconstruction step and the sampled features are concatenated together.", + "bbox": [ + 498, + 369, + 890, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Target View Frustum Volume", + "text_level": 1, + "bbox": [ + 498, + 411, + 764, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To render an image for a target novel view, our key difference with previous methods [3, 15, 17] is that we construct our volume aligned with the target view frustum, instead of a pre-defined reference input view. Such a spatially-aligned target view frustum volume representation naturally enables effective aggregation of information from input images. This is in contrast to the reference volume construction approach which has a high risk missing the information outside the epipolar lines of the reference view. This phenomenon is more pronounced for large baselines where the scene overlap is small, as illustrated in Fig. 5.", + "bbox": [ + 496, + 431, + 890, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The volume elements are sampled from the input images and features to provide necessary cues to aid the prediction of color and density for volume rendering. More specifically, to render a target image of resolution $H \\times W$ , we perform $8 \\times$ subsampling in the spatial dimension while casting rays and uniformly sample $D$ points on each ray. The $8 \\times$ subsampling enables our method to maintain an acceptable volume resolution for high-resolution images, and we finally obtain the full resolution volume with a lightweight upsampler [25]. Next, we introduce the sampling process.", + "bbox": [ + 496, + 599, + 890, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Color sampling within a window. Due to the $8 \\times$ subsampling, we might risk losing some information if we only sample a single point compared to full resolution sampling. To remedy this, we instead sample a $9 \\times 9$ window centered at the 2D projection of the 3D point and thus obtain a color vector of $\\{\\tilde{c}_k^{\\mathrm{w}}\\}_{k=1}^{K}$ for $K$ input views, where $\\tilde{c}_k^{\\mathrm{w}}$ denotes the concatenation of colors within the window.", + "bbox": [ + 496, + 760, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature sampling and matching. We also sample image features $\\{\\pmb{f}_i\\}_{i=1}^K$ from $K$ feature maps. To obtain geometric", + "bbox": [ + 500, + 869, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20043", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cues, we quantify the multi-view consistency of these features by computing their pair-wise cosine similarities. We then use the cosine feature similarities to provide geometric cues for volume density prediction, based on the observation that the multi-view features of a surface point tend to exhibit high multi-view consistency [5]. All the pair-wise cosine similarities are aggregated with a learned weighted average, which can be expressed as", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {s}} = w _ {i j} \\sum_ {i < j} \\cos \\left(\\boldsymbol {f} _ {i}, \\boldsymbol {f} _ {j}\\right), \\quad i, j = 1, 2, \\dots , K \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 217, + 468, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s_{ij}$ denotes the cosine similarity between pair $(\\pmb{f}_i, \\pmb{f}_j)$ , and $(i,j)$ iterates over all possible $K(K - 1)/2$ non-repeated pairs. Following previous work [39], $w_{ij}$ are normalized visibility weights learned from the entropy of the depth distribution on each ray.", + "bbox": [ + 76, + 253, + 468, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It's worth noting that the pair-wise cosine similarities $\\cos (\\pmb {f}_i,\\pmb {f}_j)$ in Eq. (1) can be computed in parallel with a single matrix multiplication. More specifically, we first normalize the features to unit vectors and then collect them as a $K\\times M$ matrix, where $M$ is the feature dimension. Finally all $K(K - 1) / 2$ pair-wise cosine similarities can be obtained by extracting the upper triangle of the multiplication of this matrix with its transpose. In practice, we use group-wise cosine similarities [5, 11] for further improved expressiveness, where the same computation still applies.", + "bbox": [ + 75, + 330, + 468, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-view aggregation. Next, the sampled colors and features from $K$ input views are aggregated with a learned weighted average. More specifically, the aggregated color and features are", + "bbox": [ + 76, + 484, + 468, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {c}} = w _ {i} \\sum_ {i = 1} ^ {K} \\tilde {\\boldsymbol {c}} _ {i} ^ {\\mathrm {w}}, \\quad \\hat {\\boldsymbol {f}} = w _ {i} \\sum_ {i = 1} ^ {K} \\boldsymbol {f} _ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 547, + 468, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{i}$ are normalized learned weights [16]. In particular, we use a small MLP network to predict $w_{i}$ from the concatenation of sampled features and view direction difference between the source and target views.", + "bbox": [ + 75, + 593, + 468, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The cosine similarity $\\hat{s}$ in Eq. (1), colors $\\hat{c}$ and features $\\hat{f}$ in Eq. (2) are then concatenated and projected to a $C$ -dimensional vector with a linear layer. Accordingly we obtain a target volume $z \\in \\mathbb{R}^{\\frac{H}{8} \\times \\frac{W}{8} \\times D \\times C}$ for all $\\frac{H}{8} \\times \\frac{W}{8} \\times D$ 3D points. This volume encodes appearance and geometry information from multi-view images and features, which is next fed as input to the decoder for radiance field prediction.", + "bbox": [ + 76, + 654, + 468, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Context-aware Radiance Field Decoder", + "text_level": 1, + "bbox": [ + 76, + 770, + 413, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the target view frustum volume $\\mathbf{z} \\in \\mathbb{R}^{\\frac{H}{8} \\times \\frac{W}{8} \\times D \\times C}$ , our decoder learns to predict the 4-dimensional (color and density) radiance field $\\mathbf{R} \\in \\mathbb{R}^{H \\times W \\times D \\times 4}$ of all $H \\times W \\times D$ 3D points. In order to perform well on both small and large baseline input views, we observe that it's crucial to model the 3D context between different 3D points. The context information can help learn useful inductive biases from data", + "bbox": [ + 75, + 791, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and accordingly leads to better scene structures.", + "bbox": [ + 498, + 90, + 815, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve this, we model the 3D context with a convolutional network. A straightforward approach would be using a 3D CNN. In this paper, we explore an alternative for better memory and parameter efficiency, while maintaining similar performance (Fig. 7). More specifically, we factorize the 3D $(3\\times 3\\times 3)$ convolution to a 2D $(3\\times 3\\times 1)$ convolution in the spatial dimension and a 1D $(1\\times 1\\times 3)$ convolution in the depth dimension, i.e., $(2 + 1)\\mathrm{D}$ CNN, which is a popular strategy in video recognition works [9, 28].", + "bbox": [ + 496, + 106, + 890, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The full decoder consists of two major components: a $(2 + 1)\\mathrm{D}$ CNN which operates on the low resolution volume and a lightweight upsampler which achieves $8\\times$ upsampling and thus outputs a full-resolution radiance filed. The low-resolution $(2 + 1)\\mathrm{D}$ CNN architecture is composed of 12 stacked $(2 + 1)\\mathrm{D}$ residual blocks [12]. The final color and density predictions are obtained using two linear layers, with output channels of 3 and 1, respectively. More architectural details are presented in the supplementary material.", + "bbox": [ + 496, + 243, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Hierarchical Volume Sampling", + "text_level": 1, + "bbox": [ + 498, + 390, + 772, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Like other NeRF methods [1, 18, 32], our model also supports hierarchical volume sampling for further improved performance. When hierarchical sampling is used, the model presented before can be viewed as a coarse model, and the hierarchical stage as a fine model. The fine model has a very similar overall architecture as the coarse stage, with the ray sampling process as the key difference. More specifically, given the density prediction from the coarse model, we first compute the Probability Distribution Function (PDF) [18] on each ray by normalizing the color composition weights in the volume rendering equation. Next, we sample a new set of points on each ray according to this distribution. Thanks to the coarse geometry predicted by the coarse network, the fine model requires sampling fewer points since the influence of empty space or occluded regions can be removed. In our implementation, we uniformly sample 64 points on each ray in the coarse model, and only sample 16 points in the fine stage. Due to the smaller number of sampling points, we are able to directly construct a target view frustum volume at the full resolution. We also reduce the volume's channel dimension from 128 to 16, which makes it more efficient for subsequent regression. Thus the volume at the fine stage is more compact. We use a lightweight 3D U-Net [22, 23] to predict the color and density from this volume, and the final color prediction is similarly obtained using volume rendering [18]. More details are presented in the supplementary material.", + "bbox": [ + 496, + 414, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Training Loss", + "text_level": 1, + "bbox": [ + 500, + 833, + 642, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use random crops from the full image for training, the training loss is an addition of $\\ell_1$ , SSIM and LPIPS losses between rendered and ground truth image color.", + "bbox": [ + 496, + 857, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "footer", + "text": "20044", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 89, + 209, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We implement our method using PyTorch [21] and we adopt a two-stage training process, where we first train the coarse model only, and then train the fine model with the coarse model frozen. For all the experiments, we sample 64 points in the coarse model, and 16 points in the fine model. For experiments on the RealEstate10K [40] dataset, we follow AttnRend [8]'s setting to train and evaluate on the $256 \\times 256$ resolution. On this resolution, our method can easily use $4 \\times$ subsampling when constructing the volume, unlike the default $8 \\times$ subsampling. We also reduce the number of channels of the volume from 128 to 64 when $4 \\times$ subsampling is used. We don't include the additional hierarchical sampling stage for experiments on the RealEstate10K dataset since the single scale model already performs very competitively with $4 \\times$ subsampling. Our code is available at https://github.com/autonomousvision/murf.", + "bbox": [ + 75, + 113, + 472, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation settings. Our evaluations include both small and large baselines, different number of views, and diverse scenarios, including object-centric, indoor and unbounded outdoor scenes, to test the model's general applicability.", + "bbox": [ + 75, + 375, + 470, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 76, + 446, + 215, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both small & large baselines. To compare with existing methods on both small and large baselines, we choose previous state-of-the-art small baseline method ENeRF [16] and large baseline method AttnRend [8] as two representative approaches. The evaluations are conducted on DTU [14] and RealEstate10K [40] datasets. Since ENeRF only reported its results on DTU and AttnRend only on RealEstate10K, not both, we re-train ENeRF on RealEstate10K and AttnRend on DTU, which allows for more comprehensive comparisons. The results in Table 1 demonstrate that previous state-of-the-art methods are specialized to either small or large baselines, but they can not work well on both. In contrast, our MuRF performs consistently better on both small and large baselines. Next, we conduct comparisons in different specialized settings.", + "bbox": [ + 75, + 468, + 468, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3-view small baseline on DTU. In this setting, to render a target view, 3 nearest views are selected from all source views based on the distance to the target camera location. Thus this constitutes a small baseline setting. As shown in Table 2, we achieve more than 1dB PSNR improvement compared to previous best method ENeRF [16]. The visual comparisons are shown in Fig. 3, where our method produces significantly better scene structures.", + "bbox": [ + 75, + 699, + 468, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2-view large baseline on RealEstate10K. The RealEstate10K is a very large dataset, which consists of more than 66K training scenes and more than 7K testing scenes. We follow the evaluation setting of AttnRend [8] for comparisons. More specifically, the two input views are", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1b380ad9ff852756ebd0667abc707328a50c1e3bcb5ddec0d4d9c2711d98cacd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDTU (small baseline)RealEstate10K (large baseline)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
ENeRF [16]27.610.9560.09119.520.7390.341
AttnRend[8]18.570.7320.41921.380.8390.262
MuRF28.760.9610.07724.200.8650.170
", + "bbox": [ + 504, + 93, + 888, + 167 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f812cf1861ad1665082cead986f4c69f1a2d88db53fa75687b27f31c2d10f990.jpg", + "table_caption": [ + "Table 1. Comparison on both small and large baselines. Previous state-of-the-art methods are specialized to either small (ENeRF) or large (AttnRend) baselines, but can not work well on both." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓
PixelNeRF [38]19.310.7890.382
SRF [6]22.120.8450.292
IBRNet [32]26.040.9170.190
MVSNeRF [3]26.630.9310.168
GeoNeRF [15]26.760.8930.150
MatchNeRF [5]26.910.9340.159
ENeRF [16]27.610.9560.091
MuRF28.760.9610.077
", + "bbox": [ + 542, + 223, + 849, + 363 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4144e479bb43dd97b5ae92ab25c784d8b9c8b49709d8d8a83829916b0b1b1aae.jpg", + "table_caption": [ + "Table 2. DTU 3-view small baseline." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓
PixelNeRF [38]13.910.4600.591
SRF [6]15.400.4860.604
GeoNeRF [15]16.650.5110.541
IBRNet [32]15.990.4840.532
GPNR [26]18.550.7480.459
AttnRend [8]21.380.8390.262
MuRF24.200.8650.170
", + "bbox": [ + 542, + 386, + 849, + 513 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1f9e83f1d0a3b034393cdb1da4f26d457ba2fc521af8153cb42c01471a88586a.jpg", + "table_caption": [ + "Table 3. RealEstate10K 2-view large baseline." + ], + "table_footnote": [], + "table_body": "
GTENeRFMuRF
GTAttnRenderMuRF
GTGPNRMuRF
", + "bbox": [ + 501, + 529, + 890, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3. Visual comparisons with previous best methods on DTU, RealEstate10K and LLFF datasets.", + "bbox": [ + 498, + 821, + 890, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "selected from a video with a distance of 128 frames, and the target view to synthesis is an intermediate frame. This is a challenging setting since the overlap between two input", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20045", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/26e6407cf1d0206b6763cf0622714398e2fa1a7e54c53c715eaca2e47116f8b5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSetting3-view6-view9-view
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Mip-NeRF [1]8.680.5710.35316.650.7410.19823.580.8790.092
DietNeRF [13]11.850.6330.31420.630.7780.20123.830.8230.173
RegNeRF [19]per-scene optimization18.890.7450.19022.200.8410.11724.930.8840.089
DiffusioNeRF [33]16.200.6980.20720.340.8180.13925.180.8830.095
SPARF [29]21.010.8700.100------
SRF [6]15.320.6710.30417.540.7300.25018.350.7520.232
PixelNeRF [38]16.820.6950.27019.110.7450.23220.400.7680.220
MVSNeRF [3]feed-forward inference18.630.7690.19720.700.8230.15622.400.8530.135
ENeRF† [16]19.840.8560.17121.820.8940.13123.490.9190.106
MuRF21.310.8850.12723.740.9210.09525.280.9360.084
", + "bbox": [ + 174, + 88, + 797, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "views is usually small, useful priors should be acquired from the model and the data in order to obtain reliable synthesis results. In this large baseline setting, we achieve significant improvement ( $\\sim$ 3dB PSNR) than previous best method AttnRend [8], as shown in Table 3. The visual comparison in Fig. 3 indicates that our method produces clearer rendering results than AttnRend [8]. We attribute such large improvements to our target view frustum volume and the convolutional radiance field decoder, which are two key missing components in AttnRend [8]. The target view frustum volume effectively encodes the geometric structure of the view synthesis task and the convolutional decoder enables context-modeling between neighbouring 3D points, both significantly contributing to our final performance. Their effects are more thoroughly analyzed in Sec. 4.2.", + "bbox": [ + 75, + 295, + 472, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different number of input views on DTU and LLFF. We also evaluate on different number of input views to further understand the effect of different baselines. Firstly, we follow RegNeRF [19]'s setting of 3, 6 and 9 input views on DTU. Different from the aforementioned 3-view small baseline setting on DTU, the baseline of this setting is larger. More specifically, for each test scene, different numbers of views (3, 6 and 9) are sampled from a fixed set of 9 views. Each scene has 25 test views and they all share the same input views. This is different from the small baseline DTU setting before where the nearest views (instead of shared fixed views for all test views in each scene) are selected for each test view. Thus this setting is more close to the large baseline setting, especially when the number of input views is small. In this setting, our MuRF significantly outperforms previous representative methods like PixelNeRF [38] and MVSNeRF [3], as shown in Table 4.", + "bbox": [ + 75, + 532, + 468, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To better understand the performance of different methods on different baselines, we re-train ENeRF [16], the previous state-of-the-art method on DTU's small baseline setting, on this large baseline setting with their official code. Since the key component of ENeRF is to use depth-guided sampling to skip empty space in NeRF's volume rendering process, its performance heavily relies on the quality of the", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d8326132edc8756001db669a21eb32f7a5c333c64ac7bcceeee43f3f79fb0960.jpg", + "table_caption": [ + "Table 4. DTU 3, 6 and 9 input views. ${}^{ \\dagger }$ Enhanced ENeRF baseline by doubling the number of sampling points on each ray, otherwise we were not able to obtain meaningful results. We again outperform ENeRF by $1 \\sim 2\\mathrm{{dB}}$ PSNR even compared to this enhanced baseline." + ], + "table_footnote": [], + "table_body": "
Method#viewsPSNR↑SSIM↑LPIPS↓
PixelNeRF [38]1018.660.5880.463
MVSNeRF [3]1021.180.6910.301
IBRNet [32]1025.170.8130.200
NeuRay [17]1025.350.8180.198
GPNR [26]1025.720.8800.175
GNT [31]1025.530.8850.218
425.950.8970.149
MuRF626.040.9000.153
1026.490.9090.143
", + "bbox": [ + 526, + 292, + 867, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. LLFF. Our 4-view model already outperforms previous 10-view methods. It improves further with more views.", + "bbox": [ + 500, + 436, + 890, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "estimated depth. In this large baseline setting, our re-trained ENeRF is not able to produce reasonable results with their original hyper-parameters since the small overlap between input views makes the depth estimation quality lower and thus negatively affect NeRF's rendering process. Thus we train an enhanced ENeRF baseline by doubling the number of sampling points (from original 8 and 2 to 16 and 4 for 2-stage NeRFs) on each ray such that the model is more tolerant to the depth estimation error. The results are shown in Table 4. Our MuRF again outperforms ENeRF by $1 \\sim 2\\mathrm{dB}$ PSNR in different number of input views.", + "bbox": [ + 496, + 472, + 890, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared with per-scene optimization methods, our MuRF achieves similar performance with the state-of-the-art 3-view method SPARF [29] in the 3-view setting. It's worth noting that SPARF specifically focuses on 3 input views and they didn't report the performance on 6 and 9 views. In contrast, we purse thorough evaluation on different number of views and we show clear improvement than previous per-scene optimization methods with 6 and 9 input views, despite without using any per-scene optimization.", + "bbox": [ + 496, + 641, + 890, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also compare with previous methods on the LLFF dataset [18]. Since this is a forward-facing dataset, the baselines between different images are typically small. Previous methods usually evaluate on this dataset with 10 input views following IBRNet [32]'s setting. Surprisingly, our MuRF trained with 4 input views already outperforms all previous 10-view methods (Table 5), suggesting our model is also more data-efficient. Our performance gets further improve", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20046", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d71efff9ad2a57d60f219caefe0497b39e898f15980b9ef5d076e48c5d5060d4.jpg", + "image_caption": [ + "Figure 4. Zero-shot generalization on Mip-NeRF 360 dataset." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e5d0051de422dbf8636620ce746498071839e7966ff4989658b5a34acb40d85.jpg", + "image_caption": [ + "Figure 5. Our target view volume vs. reference (first image) view volume. Constructing the volume at the pre-defined reference view space might miss relevant information (e.g., red arrows regions) in other views since such information could be far away from the reference view's epipolar lines and thus hard to be sampled. In contrast, we construct the volume in the target view, which more effectively aggregates information from all input views and thus maximizes the information usage." + ], + "image_footnote": [], + "bbox": [ + 78, + 294, + 895, + 488 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5f171706def1c0d54c417257730618b138ad0755fccbb17aae9f70671b1eedb1.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Zero-shot generalization after trained on RealEstate10K. Our method outperforms AttnRend [8] by significantly large margins." + ], + "table_body": "
MethodDTUMip-NeRF 360 Dataset
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
AttnRend [8]11.350.5670.65114.000.4740.712
MuRF22.190.8940.21123.980.8000.293
", + "bbox": [ + 81, + 559, + 467, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ment with more input views. The visual comparison with previous state-of-the-art method GPNR [26] in Fig. 3 shows that our rendering has better structures.", + "bbox": [ + 75, + 680, + 468, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization on DTU and Mip-NeRF 360 dataset. For feed-forward models, we are also interested in their zero-shot generalization abilities on unseen datasets, which is a practical problem setting. For this evaluation, we compare with AttnRend [8] using their released model trained on the same RealEstate10K dataset. The results are reported on the object-centric dataset DTU and scene-level Mip-NeRF 360 dataset with 2 input views. From Table 6, we observe our MuRF generalizes significantly better than AttnRend. The lack of geometric inductive biases makes AttnRend more data-hungry and prone to overfitting to the training data.", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further explore the zero-shot performance limit on Mip-NeRF 360 dataset, we conduct additional fine-tuning with our RealEstate10K pre-trained model on a mixed dataset collected by IBRNet [32]. We achieve further improvement and the promising visual results shown in Fig. 4 indicate the general applicability of our method.", + "bbox": [ + 496, + 556, + 893, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation and Analysis", + "text_level": 1, + "bbox": [ + 498, + 660, + 705, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For fast ablation experiments, we only train coarse models on both DTU and RealEstate10K datasets. Since the full RealEstate10K dataset is very large, which would take significant compute to train all the ablations, we use the subset provided by AttnRend [8] for training and evaluation.", + "bbox": [ + 496, + 678, + 893, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Volume Orientation. One key difference of our method with previous approaches is that we construct the volume in the target view frustum, unlike the popular reference view-based volume in multi-view stereo related methods [3, 15]. Such a volume representation essentially aggregates information from multi-view input images, where the reference view volume would suffer from information loss since the regions outside its epipolar lines are less likely to be sampled. This issue would be more pronounced for large base", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20047", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c8e778c65c0d205ee9648b89f76348283687a59aae3265894138c825e3dc8ced.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModuleMethodLarge Baseline (RealEstate10K)Small Baseline (DTU)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Volume Orientationtarget view21.290.8080.24626.890.9250.119
reference view20.200.7760.30926.500.9220.138
Volume Elementscosine & feature & color21.290.8080.24626.890.9250.119
w/o cosine21.060.8000.25526.680.9190.124
w/o feature20.910.7940.27526.790.9230.132
w/o color20.040.7890.27525.800.9160.133
Radiance Decoder(2+1)D CNN (spatial + depth)21.290.8080.24626.890.9250.119
3D CNN (spatial & depth)21.390.8090.24326.860.9240.121
2D CNN (spatial only)21.060.8010.25626.670.9210.124
1D CNN (depth only)20.780.7880.27226.370.9200.124
Ray Transformer (depth only)20.570.7800.28326.220.9120.129
MLP (point only)20.410.7780.29826.240.9180.126
", + "bbox": [ + 150, + 88, + 821, + 284 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablations. Settings used in our final model are underlined. The 3D CNN model has more than $50\\%$ parameters (16.6M vs. 10.4M) than our factorized $(2 + 1)\\mathrm{D}$ CNN model, and the performance is similar, thus we choose to use the $(2 + 1)\\mathrm{D}$ CNN.", + "bbox": [ + 75, + 287, + 892, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/01b0132d8d5ee1bdf46e49e4864c3d9c6d942634541c5b7ba974ca0c205bc82a.jpg", + "image_caption": [ + "Figure 6. CNN vs. MLP vs. Ray Transformer. Our CNN-based decoder yields sharper structures thanks to its context modeling abilities in both spatial and depth dimensions." + ], + "image_footnote": [], + "bbox": [ + 80, + 327, + 468, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lines, as illustrated in Fig. 5. In Table 7, we can observe a 1dB PSNR performance drop by replacing our target view volume with the reference view volume on the large baseline RealEstate10K dataset. Even on the small baseline DTU dataset, our target view volume is still better since it maximizes the information usage from input views.", + "bbox": [ + 75, + 553, + 468, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Volume Elements. We construct the volume using information from the sampled colors, features, and features' cosine similarities, they are concatenated together as the volume elements. We ablate different inputs in Table 7 to understand their roles. It's interesting to see that for the large baseline setting, the features are more important than the cosine similarities, while cosine similarities become more important for small baselines. This could be expected since the feature matching information would be less reliable for large baselines due to insufficient scene overlap. It's also worth noting that for both small and large baselines, the sampled colors also contribute to the final performance. The image color provides valuable cues for predicting the 3D points' color, thus leading to improved performance.", + "bbox": [ + 75, + 651, + 468, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Radiance Decoder. We compare different approaches to predict the radiance field from the target view frustum vol", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ume in Table 7. It can be observed that the spatial context modeled by the 2D CNN is more important to the final performance than that on the depth dimension. Previous methods like IBRNet [32] and GNT [27] use Ray Transformer to only model the context information on the depth dimension. However, we observe in our experiments that the Ray Transformer not always produces better results than MLP (see Fig. 6), while our $(2 + 1)\\mathrm{D}$ CNN is consistently better. We also compare with the straightforward 3D CNN decoder and the performance is similar. However, 3D CNN has more than $50\\%$ more parameters than our $(2 + 1)\\mathrm{D}$ CNN, and thus we choose to use the $(2 + 1)\\mathrm{D}$ CNN for better efficiency. The visual comparison in Fig. 6 illustrates that our CNN-based decoder produces clearly better structures.", + "bbox": [ + 496, + 325, + 893, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 551, + 619, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present MuRF, a feed-forward method to sparse view synthesis from multiple different baseline settings. Key to our approach are the proposed target view frustum volume and CNN-based radiance field decoder. We achieve state-of-the-art performance on various evaluation settings, demonstrating the general applicability of our method.", + "bbox": [ + 496, + 578, + 890, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation and Discussion. Currently, large-scale scene-level datasets are still scarce, we expect more diverse training data would improve our performance further. Besides, our model currently assumes known camera parameters and static scenes, extending our method to pose-free scenarios and dynamic scenes could be interesting future directions.", + "bbox": [ + 496, + 678, + 890, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We thank Shaofei Wang, Zehao Yu and Stefano Esposito for the insightful comments on the early draft of this work. We thank Tobias Fischer and Kashyap Chitta for the constructive discussions. We thank Takeru Miyato for the help with the RealEstate10K dataset. Andreas Geiger was supported by the ERC Starting Grant LEGO-3D (850533) and the DFG EXC number 2064/1 - project number 390727645.", + "bbox": [ + 496, + 779, + 892, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20048", + "bbox": [ + 478, + 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] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1, 4, 6", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1, 2", + "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 1, 2, 3, 5, 6, 7", + "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 1", + "[5] Yuedong Chen, Haofei Xu, Qianyi Wu, Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Explicit correspondence matching for generalizable neural radiance fields. In arXiv, 2023. 1, 2, 3, 4, 5", + "[6] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In CVPR, 2021. 1, 2, 5, 6", + "[7] 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", + "[8] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In CVPR, 2023. 1, 2, 5, 6, 7", + "[9] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 2, 4", + "[10] David Gallup, Jan-Michael Frahm, Philippos Mordohai, and Marc Pollefeys. Variable baseline/resolution stereo. In CVPR, 2008. 2", + "[11] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In CVPR, 2019. 4", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 4", + "[13] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In ICCV, 2021. 6", + "[14] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2, 5", + "[15] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In CVPR, 2022. 1, 2, 3, 5, 7", + "[16] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2, 4, 5, 6" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, 2022. 1, 2, 3, 6", + "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 4, 6", + "[19] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 1, 2, 6", + "[20] Masatoshi Okutomi and Takeo Kanade. A multiple-baseline stereo. TPAMI, 15(4):353-363, 1993. 2", + "[21] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5", + "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4", + "[23] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4", + "[24] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In CVPR, 2022. 1, 2", + "[25] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 2, 3", + "[26] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In ECCV. Springer, 2022. 2, 5, 6, 7", + "[27] Mukund Varma T, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that neRF needs? In ICLR, 2023. 2, 8", + "[28] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 2, 4", + "[29] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In CVPR, 2023. 1, 2, 6", + "[30] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NIPS, 2017. 1", + "[31] Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, Zhangyang Wang, et al. Is attention all nerf needs? arXiv preprint arXiv:2207.13298, 2022. 6", + "[32] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20049", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 1, 2, 4, 5, 6, 7, 8", + "[33] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In CVPR, 2023. 1, 2, 6", + "[34] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In CVPR, 2022. 3", + "[35] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. PAMI, 2023. 3", + "[36] Ruigang Yang and Marc Pollefeys. Multi-resolution real-time stereo on commodity graphics hardware. In CVPR, 2003. 2", + "[37] Ruigang Yang, Greg Welch, and Gary Bishop. Real-time consensus-based scene reconstruction using commodity graphics hardware. In 10th Pacific Conference on Computer Graphics and Applications, 2002. 2", + "[38] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 1, 5, 6", + "[39] Jingyang Zhang, Yao Yao, Shiwei Li, Zixin Luo, and Tian Fang. Visibility-aware multi-view stereo network. arXiv preprint arXiv:2008.07928, 2020. 4", + "[40] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view synthesis using multiplane images. ACM Trans. Graph., 37(4): 65, 2018. 2, 5" + ], + "bbox": [ + 78, + 90, + 468, + 501 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20050", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/MuRF_ Multi-Baseline Radiance Fields/265b8c06-80fc-4e85-8d50-8ff0b6769d10_model.json b/2024/MuRF_ Multi-Baseline Radiance Fields/265b8c06-80fc-4e85-8d50-8ff0b6769d10_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3988ab9f0c7554a53d753bba13a17544bc9a652e --- /dev/null +++ b/2024/MuRF_ Multi-Baseline Radiance Fields/265b8c06-80fc-4e85-8d50-8ff0b6769d10_model.json @@ -0,0 +1,1848 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.286, + 0.131, + 0.685, + 0.151 + ], + "angle": 0, + "content": "MuRF: Multi-Baseline Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.179, + 0.834, + 0.218 + ], + "angle": 0, + "content": "Haofei \\(\\mathrm{Xu}^{1,2}\\) Anpei Chen\\(^{1,2}\\) Yuedong Chen\\(^{3}\\) Christos Sakaridis\\(^{1}\\) Yulun Zhang\\(^{4}\\) Marc Pollefeys\\(^{1,5}\\) Andreas Geiger\\(^{2,\\dagger}\\) Fisher \\(\\mathrm{Yu}^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.22, + 0.812, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)ETH Zurich \\(^{2}\\)University of Tübingen, Tübingen AI Center \\(^{3}\\)Monash University \\(^{4}\\)Shanghai Jiao Tong University \\(^{5}\\)Microsoft \\(^{\\dagger}\\)Joint last author" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.291, + 0.314, + 0.307 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.474, + 0.626 + ], + "angle": 0, + "content": "We present Multi-Baseline Radiance Fields (MuRF), a general feed-forward approach to solving sparse view synthesis under multiple different baseline settings (small and large baselines, and different number of input views). To render a target novel view, we discretize the 3D space into planes parallel to the target image plane, and accordingly construct a target view frustum volume. Such a target volume representation is spatially aligned with the target view, which effectively aggregates relevant information from the input views for high-quality rendering. It also facilitates subsequent radiance field regression with a convolutional network thanks to its axis-aligned nature. The 3D context modeled by the convolutional network enables our method to synthesis sharper scene structures than prior works. Our MuRF achieves state-of-the-art performance across multiple different baseline settings and diverse scenarios ranging from simple objects (DTU) to complex indoor and outdoor scenes (RealEstate10K and LLFF). We also show promising zero-shot generalization abilities on the Mip-NeRF 360 dataset, demonstrating the general applicability of MuRF." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.633, + 0.21, + 0.648 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Novel view synthesis from sparse input views is a critical and practical problem in computer vision and graphics. However, typical optimization-based Neural Radiance Fields (NeRFs) [1, 2, 4, 18] cannot cope well with sparse views, and additional regularizations [19, 29, 33] are usually required to better constrain the optimization process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.825 + ], + "angle": 0, + "content": "In this paper, we aim at learning feed-forward NeRF models [3, 32, 38] that are able to perform feed-forward inference on new unseen data, without requiring per-scene optimization. Moreover, inductive biases could be acquired from data and better results can potentially be attained." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To achieve this, existing sparse view methods can be broadly classified into methods that take small baseline images as input and methods that focus on large baseline settings. However, none of the existing methods works well across different baselines (see Fig. 1a and Fig. 1b). In par" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.291, + 0.695, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.408, + 0.695, + 0.451 + ], + "angle": 0, + "content": "(a) MuRF outperforms previous state-of-the-art small baseline method ENeRF. The larger baseline, the larger performance gap." + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.291, + 0.891, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.408, + 0.892, + 0.45 + ], + "angle": 0, + "content": "(b) MuRF outperforms previous state-of-the-art large baseline method AttnRend. The smaller baseline, the larger performance gap." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.456, + 0.882, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.654 + ], + "angle": 0, + "content": "Figure 1. MuRF supports multiple different baseline settings. Previous methods are specifically designed for either small (e.g., ENeRF [16]) or large (e.g., AttnRend1 [8]) baselines. However, no existing method performs well on both (see Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.893, + 0.846 + ], + "angle": 0, + "content": "ticular, small baseline methods [3, 5, 6, 15-17, 32] rely heavily on feature matching to extract relevant information from input views and suffer significantly when there is insufficient image overlap or occlusion occurs. In contrast, large baseline methods [8, 24] are mostly data-driven, and often discard matching cues. Instead, they resort to large-scale datasets and generic architectures (e.g., Transformers [7, 30]) to learn geometric inductive biases implicitly. However, due to their correspondence-unaware and data-driven nature, even state-of-the-art methods [8, 24] tend to produce blurry renderings and their generalization ability is still limited, as we show in Fig. 3 and Table 6." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.847, + 0.892, + 0.862 + ], + "angle": 0, + "content": "Our goal is to address sparse view synthesis under both" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1We name the previously unnamed method of Du et al. [8] as AttnRend, according its GitHub repository name." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20041" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "small and large camera baselines, different number of input views, and diverse scenarios ranging from simple objects to complex scenes. To this end, we introduce Multi-Baseline Radiance Fields (MuRF), a method for sparse view synthesis from multiple different baseline settings (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.471, + 0.412 + ], + "angle": 0, + "content": "More specifically, to render a target novel view, we discretize the 3D space with planes parallel to the target image plane, and accordingly construct a target view frustum volume that is spatially aligned with the target view to render. This is a crucial difference to previous volume-based methods like MVSNeRF [3] and GeoNeRF [15], where the volumes are constructed in a pre-defined reference input view by straightforwardly following the standard practice in multi-view stereo (MVS). However, a key difference between MVS and novel view synthesis (NVS) is that MVS estimates the depth of a reference input view, while NVS aims to render a novel view. Taking this difference into account is crucial, since the reference view volume will not be able to effectively aggregate information from input views when the overlap between the reference and target view is small, which results in failure for large baselines (Fig. 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.47, + 0.548 + ], + "angle": 0, + "content": "Our target view frustum volume is constructed by sampling multi-view input image colors and features, which provide appearance information to the prediction of the 3D point's color. We also compute the cosine similarities between sampled features to provide multi-view consistency cues to aid the prediction of volume density [5]. The sampled colors and features, and the computed cosine similarities are concatenated together to provide relevant information for both small and large baseline scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.469, + 0.656 + ], + "angle": 0, + "content": "Equipped with this axis-aligned target volume representation, we further propose to use a convolutional neural network (CNN) to reconstruct the radiance field from this volume. Thanks to the context modeling abilities of CNNs, our method yields more accurate scene structures than previous (MLP-based) point-wise [6, 8] and (Ray Transformer-based) ray-wise methods [5, 27, 32] (see Fig. 6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.469, + 0.809 + ], + "angle": 0, + "content": "In practice, we perform \\(8 \\times\\) subsampling in the spatial dimension when constructing the target volume, which enables our method to efficiently handle high-resolution images. The full resolution radiance field is finally obtained with a lightweight \\(8 \\times\\) upsampler [25]. We also use a computationally more efficient 3D CNN alternative by factorizing the \\(3\\mathrm{D}3 \\times 3 \\times 3\\) convolution to a \\(2\\mathrm{D}3 \\times 3 \\times 1\\) convolution on the spatial dimension and a \\(1\\mathrm{D}1 \\times 1 \\times 3\\) convolution on the depth dimension, i.e., \\((2 + 1)\\mathrm{D}\\) CNN, a popular strategy in video recognition works [9, 28]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We conduct extensive experiments to demonstrate the effectiveness of our proposed target view frustum volume and 3D context-aware radiance field decoder. More specifically, we outperform previous specialized models on both small (DTU [14], LLFF [18]) and large (RealEstate10K [40]) baselines, and achieve promising zero-short generalization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.123 + ], + "angle": 0, + "content": "abilities on the Mip-NeRF 360 [2] dataset, indicating the general applicability of our method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.132, + 0.642, + 0.148 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.892, + 0.399 + ], + "angle": 0, + "content": "Multi-Baseline. The concept of \"multi-baseline\" dates back to multi-baseline stereo depth estimation [10, 20, 36, 37], where multiple stereo pairs with various baselines are used to surpass matching ambiguities and thus achieve improved precision. In this paper, we aim at designing a generally applicable approach to solving novel view synthesis from multiple different sparse view settings. When considering multiple images as input, our problem setup is somewhat similar to multi-baseline stereo, while with the radiance field as the output for volumetric rendering. Besides, we are also interested in novel view synthesis from very sparse (as few as 2) input views with different baselines. In this sense, our proposed multi-baseline radiance field not only draws connections to classic multi-baseline stereo, but also extends it to view synthesis from 2 input views with both small and large baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Sparse Input Novel View Synthesis. Applying NeRF to novel view synthesis with sparse input views has garnered considerable interest in recent years. Several works improve the optimization-based pipeline via the incorporation of additional training regularizations, e.g., depth [19], correspondence [29] and diffusion model [33]. An alternative area of research turns to the feed-forward designs by learning reliable representations from data, and accordingly removing the need of per-scene optimization. Among them, MVSNeRF [3], GeoNeRF [15] and NeuRay [17] follow the spirit of multi-view stereo (MVS) to construct cost volumes at pre-defined reference viewpoints. However, the reference volume will not be able to effectively aggregate the information for input views when the overlap between the reference and target novel view is small, which accordingly results in failure for large baselines. Besides, ENeRF [16] essentially relies on an MVS network to guide NeRF's sampling process with the estimated depth. However, the depth quality will be unreliable for large baselines and thus severely affects the subsequent rendering process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Another line of works such as SRF [6], IBRNet [32], GPNR [26], GNT [27] and MatchNeRF [5] mostly perform ray-based rendering, where each ray is rendered independently and no explicit 3D context between different rays is modeled. In contrast, our target view volume representation effectively encodes the scene geometry and naturally enables 3D context modeling with a convolutional network, yielding better scene structures. Despite various motivations and implementations, the above models are all designed and experimented only on small baseline settings, limiting their applications in real-world scenarios. Only a few existing approaches focus on the large baseline settings. Nevertheless, these works [8, 24] resort to large-scale" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20042" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.895, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.261, + 0.895, + 0.361 + ], + "angle": 0, + "content": "Figure 2. Overview. Given multiple input images, we first extract multi-view image features with a multi-view Transformer. To render a target image of resolution \\( H \\times W \\), we construct a target view frustum volume by performing \\( 8 \\times \\) subsampling in the spatial dimension while casting rays and sampling \\( D \\) equidistant points on each ray. For each 3D point, we sample feature and color information from the extracted feature maps and input images, which consists of the elements of the target volume \\( z \\in \\mathbb{R}^{\\frac{H}{8} \\times \\frac{W}{8} \\times D \\times C} \\). Here, \\( C \\) denotes the channel dimension after aggregating sampled features and colors. To reconstruct the radiance field from the volume, we model the context information in the decoder with a (2+1)D CNN operating on low resolution and subsequently obtain the full-resolution radiance field with a lightweight \\( 8 \\times \\) upsampler. The target image is finally rendered with volumetric rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.473, + 0.476 + ], + "angle": 0, + "content": "datasets and generic architectures, without employing explicit 3D representations, which ultimately leads to blurry rendering and limited generalization. As a comparison, our MuRF is developed with both geometry-aware target view frustum volume representation and 3D context-aware convolutional network, which make it excel at both small and large baseline settings on different datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.487, + 0.186, + 0.503 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.614 + ], + "angle": 0, + "content": "Our MuRF (Multi-Baseline Radiance Fields) is an encoder-decoder architecture (Fig. 2), where the encoder maps the multi-view input images to features that are used to construct a volume at the target view's camera frustum, and the decoder regresses the radiance field from this volume representation. The target image is finally obtained using volume rendering [18]. The key components are introduced below." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.623, + 0.334, + 0.637 + ], + "angle": 0, + "content": "3.1. Multi-View Feature Encoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.901 + ], + "angle": 0, + "content": "To aggregate essential information required as input for learning feed-forward NeRF models, we first extract features \\(\\{F_k\\}_{k=1}^K\\) from \\(K\\) input images \\(\\{I_k\\}_{k=1}^K\\). Our feature encoder consists of a weight-sharing per-image 2D CNN and a joint multi-view Transformer. The CNN consists of 6 residual blocks [12] and every 2 residual blocks include a stride-2 convolutional layer to achieve \\(2 \\times\\) downsampling. Accordingly, we obtain features at \\(1/2\\), \\(1/4\\) and \\(1/8\\) resolutions. The \\(1/8\\) convolutional features are then fed into a joint multi-view Transformer to obtain features at \\(1/8\\) resolution. Our Transformer is built on GMFlow [34, 35]'s 2-view Transformer architecture, where we extend the 2-view cross-attention to \\(K(K \\geq 2)\\) input views by performing cross-attention for all \\(K\\) views simultaneously. This facilitates more efficient processing of additional input views without significantly increasing memory footprint compared to the pair-wise architecture [5]. The multi-scale" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.371, + 0.892, + 0.401 + ], + "angle": 0, + "content": "features are sampled in next volume reconstruction step and the sampled features are concatenated together." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.412, + 0.765, + 0.427 + ], + "angle": 0, + "content": "3.2. Target View Frustum Volume" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.6 + ], + "angle": 0, + "content": "To render an image for a target novel view, our key difference with previous methods [3, 15, 17] is that we construct our volume aligned with the target view frustum, instead of a pre-defined reference input view. Such a spatially-aligned target view frustum volume representation naturally enables effective aggregation of information from input images. This is in contrast to the reference volume construction approach which has a high risk missing the information outside the epipolar lines of the reference view. This phenomenon is more pronounced for large baselines where the scene overlap is small, as illustrated in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.892, + 0.753 + ], + "angle": 0, + "content": "The volume elements are sampled from the input images and features to provide necessary cues to aid the prediction of color and density for volume rendering. More specifically, to render a target image of resolution \\( H \\times W \\), we perform \\( 8 \\times \\) subsampling in the spatial dimension while casting rays and uniformly sample \\( D \\) points on each ray. The \\( 8 \\times \\) subsampling enables our method to maintain an acceptable volume resolution for high-resolution images, and we finally obtain the full resolution volume with a lightweight upsampler [25]. Next, we introduce the sampling process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.892, + 0.866 + ], + "angle": 0, + "content": "Color sampling within a window. Due to the \\(8 \\times\\) subsampling, we might risk losing some information if we only sample a single point compared to full resolution sampling. To remedy this, we instead sample a \\(9 \\times 9\\) window centered at the 2D projection of the 3D point and thus obtain a color vector of \\(\\{\\tilde{c}_k^{\\mathrm{w}}\\}_{k=1}^{K}\\) for \\(K\\) input views, where \\(\\tilde{c}_k^{\\mathrm{w}}\\) denotes the concatenation of colors within the window." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Feature sampling and matching. We also sample image features \\(\\{\\pmb{f}_i\\}_{i=1}^K\\) from \\(K\\) feature maps. To obtain geometric" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20043" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "cues, we quantify the multi-view consistency of these features by computing their pair-wise cosine similarities. We then use the cosine feature similarities to provide geometric cues for volume density prediction, based on the observation that the multi-view features of a surface point tend to exhibit high multi-view consistency [5]. All the pair-wise cosine similarities are aggregated with a learned weighted average, which can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.218, + 0.47, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {s}} = w _ {i j} \\sum_ {i < j} \\cos \\left(\\boldsymbol {f} _ {i}, \\boldsymbol {f} _ {j}\\right), \\quad i, j = 1, 2, \\dots , K \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.255, + 0.47, + 0.33 + ], + "angle": 0, + "content": "where \\( s_{ij} \\) denotes the cosine similarity between pair \\( (\\pmb{f}_i, \\pmb{f}_j) \\), and \\( (i,j) \\) iterates over all possible \\( K(K - 1)/2 \\) non-repeated pairs. Following previous work [39], \\( w_{ij} \\) are normalized visibility weights learned from the entropy of the depth distribution on each ray." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.331, + 0.469, + 0.481 + ], + "angle": 0, + "content": "It's worth noting that the pair-wise cosine similarities \\(\\cos (\\pmb {f}_i,\\pmb {f}_j)\\) in Eq. (1) can be computed in parallel with a single matrix multiplication. More specifically, we first normalize the features to unit vectors and then collect them as a \\(K\\times M\\) matrix, where \\(M\\) is the feature dimension. Finally all \\(K(K - 1) / 2\\) pair-wise cosine similarities can be obtained by extracting the upper triangle of the multiplication of this matrix with its transpose. In practice, we use group-wise cosine similarities [5, 11] for further improved expressiveness, where the same computation still applies." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.486, + 0.47, + 0.546 + ], + "angle": 0, + "content": "Multi-view aggregation. Next, the sampled colors and features from \\( K \\) input views are aggregated with a learned weighted average. More specifically, the aggregated color and features are" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.549, + 0.469, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {c}} = w _ {i} \\sum_ {i = 1} ^ {K} \\tilde {\\boldsymbol {c}} _ {i} ^ {\\mathrm {w}}, \\quad \\hat {\\boldsymbol {f}} = w _ {i} \\sum_ {i = 1} ^ {K} \\boldsymbol {f} _ {i}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.469, + 0.655 + ], + "angle": 0, + "content": "where \\(w_{i}\\) are normalized learned weights [16]. In particular, we use a small MLP network to predict \\(w_{i}\\) from the concatenation of sampled features and view direction difference between the source and target views." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.655, + 0.469, + 0.76 + ], + "angle": 0, + "content": "The cosine similarity \\(\\hat{s}\\) in Eq. (1), colors \\(\\hat{c}\\) and features \\(\\hat{f}\\) in Eq. (2) are then concatenated and projected to a \\(C\\)-dimensional vector with a linear layer. Accordingly we obtain a target volume \\(z \\in \\mathbb{R}^{\\frac{H}{8} \\times \\frac{W}{8} \\times D \\times C}\\) for all \\(\\frac{H}{8} \\times \\frac{W}{8} \\times D\\) 3D points. This volume encodes appearance and geometry information from multi-view images and features, which is next fed as input to the decoder for radiance field prediction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.415, + 0.787 + ], + "angle": 0, + "content": "3.3. Context-aware Radiance Field Decoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Given the target view frustum volume \\( \\mathbf{z} \\in \\mathbb{R}^{\\frac{H}{8} \\times \\frac{W}{8} \\times D \\times C} \\), our decoder learns to predict the 4-dimensional (color and density) radiance field \\( \\mathbf{R} \\in \\mathbb{R}^{H \\times W \\times D \\times 4} \\) of all \\( H \\times W \\times D \\) 3D points. In order to perform well on both small and large baseline input views, we observe that it's crucial to model the 3D context between different 3D points. The context information can help learn useful inductive biases from data" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.816, + 0.106 + ], + "angle": 0, + "content": "and accordingly leads to better scene structures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.243 + ], + "angle": 0, + "content": "To achieve this, we model the 3D context with a convolutional network. A straightforward approach would be using a 3D CNN. In this paper, we explore an alternative for better memory and parameter efficiency, while maintaining similar performance (Fig. 7). More specifically, we factorize the 3D \\((3\\times 3\\times 3)\\) convolution to a 2D \\((3\\times 3\\times 1)\\) convolution in the spatial dimension and a 1D \\((1\\times 1\\times 3)\\) convolution in the depth dimension, i.e., \\((2 + 1)\\mathrm{D}\\) CNN, which is a popular strategy in video recognition works [9, 28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.38 + ], + "angle": 0, + "content": "The full decoder consists of two major components: a \\((2 + 1)\\mathrm{D}\\) CNN which operates on the low resolution volume and a lightweight upsampler which achieves \\(8\\times\\) upsampling and thus outputs a full-resolution radiance filed. The low-resolution \\((2 + 1)\\mathrm{D}\\) CNN architecture is composed of 12 stacked \\((2 + 1)\\mathrm{D}\\) residual blocks [12]. The final color and density predictions are obtained using two linear layers, with output channels of 3 and 1, respectively. More architectural details are presented in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.391, + 0.774, + 0.407 + ], + "angle": 0, + "content": "3.4. Hierarchical Volume Sampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Like other NeRF methods [1, 18, 32], our model also supports hierarchical volume sampling for further improved performance. When hierarchical sampling is used, the model presented before can be viewed as a coarse model, and the hierarchical stage as a fine model. The fine model has a very similar overall architecture as the coarse stage, with the ray sampling process as the key difference. More specifically, given the density prediction from the coarse model, we first compute the Probability Distribution Function (PDF) [18] on each ray by normalizing the color composition weights in the volume rendering equation. Next, we sample a new set of points on each ray according to this distribution. Thanks to the coarse geometry predicted by the coarse network, the fine model requires sampling fewer points since the influence of empty space or occluded regions can be removed. In our implementation, we uniformly sample 64 points on each ray in the coarse model, and only sample 16 points in the fine stage. Due to the smaller number of sampling points, we are able to directly construct a target view frustum volume at the full resolution. We also reduce the volume's channel dimension from 128 to 16, which makes it more efficient for subsequent regression. Thus the volume at the fine stage is more compact. We use a lightweight 3D U-Net [22, 23] to predict the color and density from this volume, and the final color prediction is similarly obtained using volume rendering [18]. More details are presented in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.834, + 0.643, + 0.85 + ], + "angle": 0, + "content": "3.5. Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.858, + 0.892, + 0.904 + ], + "angle": 0, + "content": "We use random crops from the full image for training, the training loss is an addition of \\(\\ell_1\\), SSIM and LPIPS losses between rendered and ground truth image color." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20044" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.21, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.473, + 0.371 + ], + "angle": 0, + "content": "Implementation details. We implement our method using PyTorch [21] and we adopt a two-stage training process, where we first train the coarse model only, and then train the fine model with the coarse model frozen. For all the experiments, we sample 64 points in the coarse model, and 16 points in the fine model. For experiments on the RealEstate10K [40] dataset, we follow AttnRend [8]'s setting to train and evaluate on the \\(256 \\times 256\\) resolution. On this resolution, our method can easily use \\(4 \\times\\) subsampling when constructing the volume, unlike the default \\(8 \\times\\) subsampling. We also reduce the number of channels of the volume from 128 to 64 when \\(4 \\times\\) subsampling is used. We don't include the additional hierarchical sampling stage for experiments on the RealEstate10K dataset since the single scale model already performs very competitively with \\(4 \\times\\) subsampling. Our code is available at https://github.com/autonomousvision/murf." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.471, + 0.437 + ], + "angle": 0, + "content": "Evaluation settings. Our evaluations include both small and large baselines, different number of views, and diverse scenarios, including object-centric, indoor and unbounded outdoor scenes, to test the model's general applicability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.448, + 0.216, + 0.463 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.697 + ], + "angle": 0, + "content": "Both small & large baselines. To compare with existing methods on both small and large baselines, we choose previous state-of-the-art small baseline method ENeRF [16] and large baseline method AttnRend [8] as two representative approaches. The evaluations are conducted on DTU [14] and RealEstate10K [40] datasets. Since ENeRF only reported its results on DTU and AttnRend only on RealEstate10K, not both, we re-train ENeRF on RealEstate10K and AttnRend on DTU, which allows for more comprehensive comparisons. The results in Table 1 demonstrate that previous state-of-the-art methods are specialized to either small or large baselines, but they can not work well on both. In contrast, our MuRF performs consistently better on both small and large baselines. Next, we conduct comparisons in different specialized settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.47, + 0.821 + ], + "angle": 0, + "content": "3-view small baseline on DTU. In this setting, to render a target view, 3 nearest views are selected from all source views based on the distance to the target camera location. Thus this constitutes a small baseline setting. As shown in Table 2, we achieve more than 1dB PSNR improvement compared to previous best method ENeRF [16]. The visual comparisons are shown in Fig. 3, where our method produces significantly better scene structures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "2-view large baseline on RealEstate10K. The RealEstate10K is a very large dataset, which consists of more than 66K training scenes and more than 7K testing scenes. We follow the evaluation setting of AttnRend [8] for comparisons. More specifically, the two input views are" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.094, + 0.89, + 0.169 + ], + "angle": 0, + "content": "
MethodDTU (small baseline)RealEstate10K (large baseline)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
ENeRF [16]27.610.9560.09119.520.7390.341
AttnRend[8]18.570.7320.41921.380.8390.262
MuRF28.760.9610.07724.200.8650.170
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.171, + 0.892, + 0.213 + ], + "angle": 0, + "content": "Table 1. Comparison on both small and large baselines. Previous state-of-the-art methods are specialized to either small (ENeRF) or large (AttnRend) baselines, but can not work well on both." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.224, + 0.85, + 0.364 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓
PixelNeRF [38]19.310.7890.382
SRF [6]22.120.8450.292
IBRNet [32]26.040.9170.190
MVSNeRF [3]26.630.9310.168
GeoNeRF [15]26.760.8930.150
MatchNeRF [5]26.910.9340.159
ENeRF [16]27.610.9560.091
MuRF28.760.9610.077
" + }, + { + "type": "table_caption", + "bbox": [ + 0.583, + 0.367, + 0.808, + 0.38 + ], + "angle": 0, + "content": "Table 2. DTU 3-view small baseline." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.387, + 0.85, + 0.514 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓
PixelNeRF [38]13.910.4600.591
SRF [6]15.400.4860.604
GeoNeRF [15]16.650.5110.541
IBRNet [32]15.990.4840.532
GPNR [26]18.550.7480.459
AttnRend [8]21.380.8390.262
MuRF24.200.8650.170
" + }, + { + "type": "table_caption", + "bbox": [ + 0.553, + 0.516, + 0.838, + 0.529 + ], + "angle": 0, + "content": "Table 3. RealEstate10K 2-view large baseline." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.53, + 0.892, + 0.821 + ], + "angle": 0, + "content": "
GTENeRFMuRF
GTAttnRenderMuRF
GTGPNRMuRF
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.822, + 0.892, + 0.849 + ], + "angle": 0, + "content": "Figure 3. Visual comparisons with previous best methods on DTU, RealEstate10K and LLFF datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "selected from a video with a distance of 128 frames, and the target view to synthesis is an intermediate frame. This is a challenging setting since the overlap between two input" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20045" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.175, + 0.089, + 0.798, + 0.254 + ], + "angle": 0, + "content": "
MethodSetting3-view6-view9-view
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Mip-NeRF [1]8.680.5710.35316.650.7410.19823.580.8790.092
DietNeRF [13]11.850.6330.31420.630.7780.20123.830.8230.173
RegNeRF [19]per-scene optimization18.890.7450.19022.200.8410.11724.930.8840.089
DiffusioNeRF [33]16.200.6980.20720.340.8180.13925.180.8830.095
SPARF [29]21.010.8700.100------
SRF [6]15.320.6710.30417.540.7300.25018.350.7520.232
PixelNeRF [38]16.820.6950.27019.110.7450.23220.400.7680.220
MVSNeRF [3]feed-forward inference18.630.7690.19720.700.8230.15622.400.8530.135
ENeRF† [16]19.840.8560.17121.820.8940.13123.490.9190.106
MuRF21.310.8850.12723.740.9210.09525.280.9360.084
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.257, + 0.892, + 0.286 + ], + "angle": 0, + "content": "Table 4. DTU 3, 6 and 9 input views. \\( {}^{ \\dagger } \\) Enhanced ENeRF baseline by doubling the number of sampling points on each ray, otherwise we were not able to obtain meaningful results. We again outperform ENeRF by \\( 1 \\sim 2\\mathrm{{dB}} \\) PSNR even compared to this enhanced baseline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.473, + 0.524 + ], + "angle": 0, + "content": "views is usually small, useful priors should be acquired from the model and the data in order to obtain reliable synthesis results. In this large baseline setting, we achieve significant improvement (\\(\\sim\\)3dB PSNR) than previous best method AttnRend [8], as shown in Table 3. The visual comparison in Fig. 3 indicates that our method produces clearer rendering results than AttnRend [8]. We attribute such large improvements to our target view frustum volume and the convolutional radiance field decoder, which are two key missing components in AttnRend [8]. The target view frustum volume effectively encodes the geometric structure of the view synthesis task and the convolutional decoder enables context-modeling between neighbouring 3D points, both significantly contributing to our final performance. Their effects are more thoroughly analyzed in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.79 + ], + "angle": 0, + "content": "Different number of input views on DTU and LLFF. We also evaluate on different number of input views to further understand the effect of different baselines. Firstly, we follow RegNeRF [19]'s setting of 3, 6 and 9 input views on DTU. Different from the aforementioned 3-view small baseline setting on DTU, the baseline of this setting is larger. More specifically, for each test scene, different numbers of views (3, 6 and 9) are sampled from a fixed set of 9 views. Each scene has 25 test views and they all share the same input views. This is different from the small baseline DTU setting before where the nearest views (instead of shared fixed views for all test views in each scene) are selected for each test view. Thus this setting is more close to the large baseline setting, especially when the number of input views is small. In this setting, our MuRF significantly outperforms previous representative methods like PixelNeRF [38] and MVSNeRF [3], as shown in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To better understand the performance of different methods on different baselines, we re-train ENeRF [16], the previous state-of-the-art method on DTU's small baseline setting, on this large baseline setting with their official code. Since the key component of ENeRF is to use depth-guided sampling to skip empty space in NeRF's volume rendering process, its performance heavily relies on the quality of the" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.293, + 0.868, + 0.436 + ], + "angle": 0, + "content": "
Method#viewsPSNR↑SSIM↑LPIPS↓
PixelNeRF [38]1018.660.5880.463
MVSNeRF [3]1021.180.6910.301
IBRNet [32]1025.170.8130.200
NeuRay [17]1025.350.8180.198
GPNR [26]1025.720.8800.175
GNT [31]1025.530.8850.218
425.950.8970.149
MuRF626.040.9000.153
1026.490.9090.143
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.438, + 0.892, + 0.465 + ], + "angle": 0, + "content": "Table 5. LLFF. Our 4-view model already outperforms previous 10-view methods. It improves further with more views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.639 + ], + "angle": 0, + "content": "estimated depth. In this large baseline setting, our re-trained ENeRF is not able to produce reasonable results with their original hyper-parameters since the small overlap between input views makes the depth estimation quality lower and thus negatively affect NeRF's rendering process. Thus we train an enhanced ENeRF baseline by doubling the number of sampling points (from original 8 and 2 to 16 and 4 for 2-stage NeRFs) on each ray such that the model is more tolerant to the depth estimation error. The results are shown in Table 4. Our MuRF again outperforms ENeRF by \\(1 \\sim 2\\mathrm{dB}\\) PSNR in different number of input views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.778 + ], + "angle": 0, + "content": "Compared with per-scene optimization methods, our MuRF achieves similar performance with the state-of-the-art 3-view method SPARF [29] in the 3-view setting. It's worth noting that SPARF specifically focuses on 3 input views and they didn't report the performance on 6 and 9 views. In contrast, we purse thorough evaluation on different number of views and we show clear improvement than previous per-scene optimization methods with 6 and 9 input views, despite without using any per-scene optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We also compare with previous methods on the LLFF dataset [18]. Since this is a forward-facing dataset, the baselines between different images are typically small. Previous methods usually evaluate on this dataset with 10 input views following IBRNet [32]'s setting. Surprisingly, our MuRF trained with 4 input views already outperforms all previous 10-view methods (Table 5), suggesting our model is also more data-efficient. Our performance gets further improve" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20046" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.274, + 0.678, + 0.288 + ], + "angle": 0, + "content": "Figure 4. Zero-shot generalization on Mip-NeRF 360 dataset." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.295, + 0.896, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.491, + 0.895, + 0.548 + ], + "angle": 0, + "content": "Figure 5. Our target view volume vs. reference (first image) view volume. Constructing the volume at the pre-defined reference view space might miss relevant information (e.g., red arrows regions) in other views since such information could be far away from the reference view's epipolar lines and thus hard to be sampled. In contrast, we construct the volume in the target view, which more effectively aggregates information from all input views and thus maximizes the information usage." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.56, + 0.468, + 0.626 + ], + "angle": 0, + "content": "
MethodDTUMip-NeRF 360 Dataset
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
AttnRend [8]11.350.5670.65114.000.4740.712
MuRF22.190.8940.21123.980.8000.293
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.627, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Table 6. Zero-shot generalization after trained on RealEstate10K. Our method outperforms AttnRend [8] by significantly large margins." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.726 + ], + "angle": 0, + "content": "ment with more input views. The visual comparison with previous state-of-the-art method GPNR [26] in Fig. 3 shows that our rendering has better structures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Generalization on DTU and Mip-NeRF 360 dataset. For feed-forward models, we are also interested in their zero-shot generalization abilities on unseen datasets, which is a practical problem setting. For this evaluation, we compare with AttnRend [8] using their released model trained on the same RealEstate10K dataset. The results are reported on the object-centric dataset DTU and scene-level Mip-NeRF 360 dataset with 2 input views. From Table 6, we observe our MuRF generalizes significantly better than AttnRend. The lack of geometric inductive biases makes AttnRend more data-hungry and prone to overfitting to the training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.895, + 0.649 + ], + "angle": 0, + "content": "To further explore the zero-shot performance limit on Mip-NeRF 360 dataset, we conduct additional fine-tuning with our RealEstate10K pre-trained model on a mixed dataset collected by IBRNet [32]. We achieve further improvement and the promising visual results shown in Fig. 4 indicate the general applicability of our method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.661, + 0.707, + 0.677 + ], + "angle": 0, + "content": "4.2. Ablation and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.894, + 0.756 + ], + "angle": 0, + "content": "For fast ablation experiments, we only train coarse models on both DTU and RealEstate10K datasets. Since the full RealEstate10K dataset is very large, which would take significant compute to train all the ablations, we use the subset provided by AttnRend [8] for training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Volume Orientation. One key difference of our method with previous approaches is that we construct the volume in the target view frustum, unlike the popular reference view-based volume in multi-view stereo related methods [3, 15]. Such a volume representation essentially aggregates information from multi-view input images, where the reference view volume would suffer from information loss since the regions outside its epipolar lines are less likely to be sampled. This issue would be more pronounced for large base" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20047" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.151, + 0.089, + 0.822, + 0.285 + ], + "angle": 0, + "content": "
ModuleMethodLarge Baseline (RealEstate10K)Small Baseline (DTU)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Volume Orientationtarget view21.290.8080.24626.890.9250.119
reference view20.200.7760.30926.500.9220.138
Volume Elementscosine & feature & color21.290.8080.24626.890.9250.119
w/o cosine21.060.8000.25526.680.9190.124
w/o feature20.910.7940.27526.790.9230.132
w/o color20.040.7890.27525.800.9160.133
Radiance Decoder(2+1)D CNN (spatial + depth)21.290.8080.24626.890.9250.119
3D CNN (spatial & depth)21.390.8090.24326.860.9240.121
2D CNN (spatial only)21.060.8010.25626.670.9210.124
1D CNN (depth only)20.780.7880.27226.370.9200.124
Ray Transformer (depth only)20.570.7800.28326.220.9120.129
MLP (point only)20.410.7780.29826.240.9180.126
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.288, + 0.893, + 0.318 + ], + "angle": 0, + "content": "Table 7. Ablations. Settings used in our final model are underlined. The 3D CNN model has more than \\(50\\%\\) parameters (16.6M vs. 10.4M) than our factorized \\((2 + 1)\\mathrm{D}\\) CNN model, and the performance is similar, thus we choose to use the \\((2 + 1)\\mathrm{D}\\) CNN." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.328, + 0.47, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.534 + ], + "angle": 0, + "content": "Figure 6. CNN vs. MLP vs. Ray Transformer. Our CNN-based decoder yields sharper structures thanks to its context modeling abilities in both spatial and depth dimensions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.645 + ], + "angle": 0, + "content": "lines, as illustrated in Fig. 5. In Table 7, we can observe a 1dB PSNR performance drop by replacing our target view volume with the reference view volume on the large baseline RealEstate10K dataset. Even on the small baseline DTU dataset, our target view volume is still better since it maximizes the information usage from input views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.47, + 0.865 + ], + "angle": 0, + "content": "Volume Elements. We construct the volume using information from the sampled colors, features, and features' cosine similarities, they are concatenated together as the volume elements. We ablate different inputs in Table 7 to understand their roles. It's interesting to see that for the large baseline setting, the features are more important than the cosine similarities, while cosine similarities become more important for small baselines. This could be expected since the feature matching information would be less reliable for large baselines due to insufficient scene overlap. It's also worth noting that for both small and large baselines, the sampled colors also contribute to the final performance. The image color provides valuable cues for predicting the 3D points' color, thus leading to improved performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Radiance Decoder. We compare different approaches to predict the radiance field from the target view frustum vol" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.895, + 0.537 + ], + "angle": 0, + "content": "ume in Table 7. It can be observed that the spatial context modeled by the 2D CNN is more important to the final performance than that on the depth dimension. Previous methods like IBRNet [32] and GNT [27] use Ray Transformer to only model the context information on the depth dimension. However, we observe in our experiments that the Ray Transformer not always produces better results than MLP (see Fig. 6), while our \\((2 + 1)\\mathrm{D}\\) CNN is consistently better. We also compare with the straightforward 3D CNN decoder and the performance is similar. However, 3D CNN has more than \\(50\\%\\) more parameters than our \\((2 + 1)\\mathrm{D}\\) CNN, and thus we choose to use the \\((2 + 1)\\mathrm{D}\\) CNN for better efficiency. The visual comparison in Fig. 6 illustrates that our CNN-based decoder produces clearly better structures." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.553, + 0.62, + 0.569 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.671 + ], + "angle": 0, + "content": "We present MuRF, a feed-forward method to sparse view synthesis from multiple different baseline settings. Key to our approach are the proposed target view frustum volume and CNN-based radiance field decoder. We achieve state-of-the-art performance on various evaluation settings, demonstrating the general applicability of our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.892, + 0.771 + ], + "angle": 0, + "content": "Limitation and Discussion. Currently, large-scale scene-level datasets are still scarce, we expect more diverse training data would improve our performance further. Besides, our model currently assumes known camera parameters and static scenes, extending our method to pose-free scenarios and dynamic scenes could be interesting future directions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Acknowledgements. We thank Shaofei Wang, Zehao Yu and Stefano Esposito for the insightful comments on the early draft of this work. We thank Tobias Fischer and Kashyap Chitta for the constructive discussions. We thank Takeru Miyato for the help with the RealEstate10K dataset. Andreas Geiger was supported by the ERC Starting Grant LEGO-3D (850533) and the DFG EXC number 2064/1 - project number 390727645." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20048" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.47, + 0.268 + ], + "angle": 0, + "content": "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.47, + 0.31 + ], + "angle": 0, + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[5] Yuedong Chen, Haofei Xu, Qianyi Wu, Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Explicit correspondence matching for generalizable neural radiance fields. In arXiv, 2023. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.422 + ], + "angle": 0, + "content": "[6] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In CVPR, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.47, + 0.506 + ], + "angle": 0, + "content": "[7] 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" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[8] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In CVPR, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.47, + 0.59 + ], + "angle": 0, + "content": "[9] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.47, + 0.633 + ], + "angle": 0, + "content": "[10] David Gallup, Jan-Michael Frahm, Philippos Mordohai, and Marc Pollefeys. Variable baseline/resolution stereo. In CVPR, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[11] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.47, + 0.717 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.76 + ], + "angle": 0, + "content": "[13] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[14] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[15] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In CVPR, 2022. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[16] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2, 4, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[17] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[19] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.292 + ], + "angle": 0, + "content": "[20] Masatoshi Okutomi and Takeo Kanade. A multiple-baseline stereo. TPAMI, 15(4):353-363, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.295, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[21] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.366, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.41, + 0.892, + 0.451 + ], + "angle": 0, + "content": "[23] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.453, + 0.892, + 0.535 + ], + "angle": 0, + "content": "[24] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.538, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[25] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.61, + 0.892, + 0.652 + ], + "angle": 0, + "content": "[26] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In ECCV. Springer, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.654, + 0.892, + 0.695 + ], + "angle": 0, + "content": "[27] Mukund Varma T, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that neRF needs? In ICLR, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.698, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[28] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.742, + 0.892, + 0.783 + ], + "angle": 0, + "content": "[29] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In CVPR, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.786, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[30] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NIPS, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.829, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[31] Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, Zhangyang Wang, et al. Is attention all nerf needs? arXiv preprint arXiv:2207.13298, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20049" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.468, + 0.175 + ], + "angle": 0, + "content": "[33] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In CVPR, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.468, + 0.219 + ], + "angle": 0, + "content": "[34] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.468, + 0.262 + ], + "angle": 0, + "content": "[35] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. PAMI, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.468, + 0.303 + ], + "angle": 0, + "content": "[36] Ruigang Yang and Marc Pollefeys. Multi-resolution real-time stereo on commodity graphics hardware. In CVPR, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.468, + 0.36 + ], + "angle": 0, + "content": "[37] Ruigang Yang, Greg Welch, and Gary Bishop. Real-time consensus-based scene reconstruction using commodity graphics hardware. In 10th Pacific Conference on Computer Graphics and Applications, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[38] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.468, + 0.447 + ], + "angle": 0, + "content": "[39] Jingyang Zhang, Yao Yao, Shiwei Li, Zixin Luo, and Tian Fang. Visibility-aware multi-view stereo network. arXiv preprint arXiv:2008.07928, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.468, + 0.502 + ], + "angle": 0, + "content": "[40] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view synthesis using multiplane images. ACM Trans. Graph., 37(4): 65, 2018. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "20050" + } + ] +] \ No newline at end of file diff --git a/2024/MuRF_ Multi-Baseline Radiance Fields/265b8c06-80fc-4e85-8d50-8ff0b6769d10_origin.pdf b/2024/MuRF_ Multi-Baseline Radiance Fields/265b8c06-80fc-4e85-8d50-8ff0b6769d10_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a39b40c25fb5d04f1526388770c551f00e07b2d0 --- /dev/null +++ b/2024/MuRF_ Multi-Baseline Radiance Fields/265b8c06-80fc-4e85-8d50-8ff0b6769d10_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:108e162c177cc79814528da233374c838b1660a328c315983d343ff4664f9a49 +size 4233812 diff --git a/2024/MuRF_ Multi-Baseline Radiance Fields/full.md b/2024/MuRF_ Multi-Baseline Radiance Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..20985d4325dabc31e12fca232b9f42bca0c9c0c0 --- /dev/null +++ b/2024/MuRF_ Multi-Baseline Radiance Fields/full.md @@ -0,0 +1,258 @@ +# MuRF: Multi-Baseline Radiance Fields + +Haofei $\mathrm{Xu}^{1,2}$ Anpei Chen $^{1,2}$ Yuedong Chen $^{3}$ Christos Sakaridis $^{1}$ Yulun Zhang $^{4}$ Marc Pollefeys $^{1,5}$ Andreas Geiger $^{2,\dagger}$ Fisher $\mathrm{Yu}^{1,\dagger}$ + +$^{1}$ ETH Zurich $^{2}$ University of Tübingen, Tübingen AI Center $^{3}$ Monash University $^{4}$ Shanghai Jiao Tong University $^{5}$ Microsoft $^{\dagger}$ Joint last author + +# Abstract + +We present Multi-Baseline Radiance Fields (MuRF), a general feed-forward approach to solving sparse view synthesis under multiple different baseline settings (small and large baselines, and different number of input views). To render a target novel view, we discretize the 3D space into planes parallel to the target image plane, and accordingly construct a target view frustum volume. Such a target volume representation is spatially aligned with the target view, which effectively aggregates relevant information from the input views for high-quality rendering. It also facilitates subsequent radiance field regression with a convolutional network thanks to its axis-aligned nature. The 3D context modeled by the convolutional network enables our method to synthesis sharper scene structures than prior works. Our MuRF achieves state-of-the-art performance across multiple different baseline settings and diverse scenarios ranging from simple objects (DTU) to complex indoor and outdoor scenes (RealEstate10K and LLFF). We also show promising zero-shot generalization abilities on the Mip-NeRF 360 dataset, demonstrating the general applicability of MuRF. + +# 1. Introduction + +Novel view synthesis from sparse input views is a critical and practical problem in computer vision and graphics. However, typical optimization-based Neural Radiance Fields (NeRFs) [1, 2, 4, 18] cannot cope well with sparse views, and additional regularizations [19, 29, 33] are usually required to better constrain the optimization process. + +In this paper, we aim at learning feed-forward NeRF models [3, 32, 38] that are able to perform feed-forward inference on new unseen data, without requiring per-scene optimization. Moreover, inductive biases could be acquired from data and better results can potentially be attained. + +To achieve this, existing sparse view methods can be broadly classified into methods that take small baseline images as input and methods that focus on large baseline settings. However, none of the existing methods works well across different baselines (see Fig. 1a and Fig. 1b). In par + +![](images/5b16c347522680150641c89b3ef6c8996a33d53409241d46a6ab3f594fa890dd.jpg) +(a) MuRF outperforms previous state-of-the-art small baseline method ENeRF. The larger baseline, the larger performance gap. + +![](images/0f74c0de3814f50ad9bc2ad0220474aa18c0732453508e8df039727a59f94f39.jpg) +(b) MuRF outperforms previous state-of-the-art large baseline method AttnRend. The smaller baseline, the larger performance gap. + +![](images/0d31ace908e061cc66813723fd604c73a8c4c2c0eeb837c007eb17ebec57785d.jpg) +Figure 1. MuRF supports multiple different baseline settings. Previous methods are specifically designed for either small (e.g., ENeRF [16]) or large (e.g., AttnRend1 [8]) baselines. However, no existing method performs well on both (see Table 1). + +ticular, small baseline methods [3, 5, 6, 15-17, 32] rely heavily on feature matching to extract relevant information from input views and suffer significantly when there is insufficient image overlap or occlusion occurs. In contrast, large baseline methods [8, 24] are mostly data-driven, and often discard matching cues. Instead, they resort to large-scale datasets and generic architectures (e.g., Transformers [7, 30]) to learn geometric inductive biases implicitly. However, due to their correspondence-unaware and data-driven nature, even state-of-the-art methods [8, 24] tend to produce blurry renderings and their generalization ability is still limited, as we show in Fig. 3 and Table 6. + +Our goal is to address sparse view synthesis under both + +small and large camera baselines, different number of input views, and diverse scenarios ranging from simple objects to complex scenes. To this end, we introduce Multi-Baseline Radiance Fields (MuRF), a method for sparse view synthesis from multiple different baseline settings (Fig. 1). + +More specifically, to render a target novel view, we discretize the 3D space with planes parallel to the target image plane, and accordingly construct a target view frustum volume that is spatially aligned with the target view to render. This is a crucial difference to previous volume-based methods like MVSNeRF [3] and GeoNeRF [15], where the volumes are constructed in a pre-defined reference input view by straightforwardly following the standard practice in multi-view stereo (MVS). However, a key difference between MVS and novel view synthesis (NVS) is that MVS estimates the depth of a reference input view, while NVS aims to render a novel view. Taking this difference into account is crucial, since the reference view volume will not be able to effectively aggregate information from input views when the overlap between the reference and target view is small, which results in failure for large baselines (Fig. 5). + +Our target view frustum volume is constructed by sampling multi-view input image colors and features, which provide appearance information to the prediction of the 3D point's color. We also compute the cosine similarities between sampled features to provide multi-view consistency cues to aid the prediction of volume density [5]. The sampled colors and features, and the computed cosine similarities are concatenated together to provide relevant information for both small and large baseline scenarios. + +Equipped with this axis-aligned target volume representation, we further propose to use a convolutional neural network (CNN) to reconstruct the radiance field from this volume. Thanks to the context modeling abilities of CNNs, our method yields more accurate scene structures than previous (MLP-based) point-wise [6, 8] and (Ray Transformer-based) ray-wise methods [5, 27, 32] (see Fig. 6). + +In practice, we perform $8 \times$ subsampling in the spatial dimension when constructing the target volume, which enables our method to efficiently handle high-resolution images. The full resolution radiance field is finally obtained with a lightweight $8 \times$ upsampler [25]. We also use a computationally more efficient 3D CNN alternative by factorizing the $3\mathrm{D}3 \times 3 \times 3$ convolution to a $2\mathrm{D}3 \times 3 \times 1$ convolution on the spatial dimension and a $1\mathrm{D}1 \times 1 \times 3$ convolution on the depth dimension, i.e., $(2 + 1)\mathrm{D}$ CNN, a popular strategy in video recognition works [9, 28]. + +We conduct extensive experiments to demonstrate the effectiveness of our proposed target view frustum volume and 3D context-aware radiance field decoder. More specifically, we outperform previous specialized models on both small (DTU [14], LLFF [18]) and large (RealEstate10K [40]) baselines, and achieve promising zero-short generalization + +abilities on the Mip-NeRF 360 [2] dataset, indicating the general applicability of our method. + +# 2. Related Work + +Multi-Baseline. The concept of "multi-baseline" dates back to multi-baseline stereo depth estimation [10, 20, 36, 37], where multiple stereo pairs with various baselines are used to surpass matching ambiguities and thus achieve improved precision. In this paper, we aim at designing a generally applicable approach to solving novel view synthesis from multiple different sparse view settings. When considering multiple images as input, our problem setup is somewhat similar to multi-baseline stereo, while with the radiance field as the output for volumetric rendering. Besides, we are also interested in novel view synthesis from very sparse (as few as 2) input views with different baselines. In this sense, our proposed multi-baseline radiance field not only draws connections to classic multi-baseline stereo, but also extends it to view synthesis from 2 input views with both small and large baselines. + +Sparse Input Novel View Synthesis. Applying NeRF to novel view synthesis with sparse input views has garnered considerable interest in recent years. Several works improve the optimization-based pipeline via the incorporation of additional training regularizations, e.g., depth [19], correspondence [29] and diffusion model [33]. An alternative area of research turns to the feed-forward designs by learning reliable representations from data, and accordingly removing the need of per-scene optimization. Among them, MVSNeRF [3], GeoNeRF [15] and NeuRay [17] follow the spirit of multi-view stereo (MVS) to construct cost volumes at pre-defined reference viewpoints. However, the reference volume will not be able to effectively aggregate the information for input views when the overlap between the reference and target novel view is small, which accordingly results in failure for large baselines. Besides, ENeRF [16] essentially relies on an MVS network to guide NeRF's sampling process with the estimated depth. However, the depth quality will be unreliable for large baselines and thus severely affects the subsequent rendering process. + +Another line of works such as SRF [6], IBRNet [32], GPNR [26], GNT [27] and MatchNeRF [5] mostly perform ray-based rendering, where each ray is rendered independently and no explicit 3D context between different rays is modeled. In contrast, our target view volume representation effectively encodes the scene geometry and naturally enables 3D context modeling with a convolutional network, yielding better scene structures. Despite various motivations and implementations, the above models are all designed and experimented only on small baseline settings, limiting their applications in real-world scenarios. Only a few existing approaches focus on the large baseline settings. Nevertheless, these works [8, 24] resort to large-scale + +![](images/b3439f4e5879f86ff2b7aa67e614d9600a2bff2123cf7c85447371c9202d55a1.jpg) +Figure 2. Overview. Given multiple input images, we first extract multi-view image features with a multi-view Transformer. To render a target image of resolution $H \times W$ , we construct a target view frustum volume by performing $8 \times$ subsampling in the spatial dimension while casting rays and sampling $D$ equidistant points on each ray. For each 3D point, we sample feature and color information from the extracted feature maps and input images, which consists of the elements of the target volume $z \in \mathbb{R}^{\frac{H}{8} \times \frac{W}{8} \times D \times C}$ . Here, $C$ denotes the channel dimension after aggregating sampled features and colors. To reconstruct the radiance field from the volume, we model the context information in the decoder with a (2+1)D CNN operating on low resolution and subsequently obtain the full-resolution radiance field with a lightweight $8 \times$ upsampler. The target image is finally rendered with volumetric rendering. + +datasets and generic architectures, without employing explicit 3D representations, which ultimately leads to blurry rendering and limited generalization. As a comparison, our MuRF is developed with both geometry-aware target view frustum volume representation and 3D context-aware convolutional network, which make it excel at both small and large baseline settings on different datasets. + +# 3. Approach + +Our MuRF (Multi-Baseline Radiance Fields) is an encoder-decoder architecture (Fig. 2), where the encoder maps the multi-view input images to features that are used to construct a volume at the target view's camera frustum, and the decoder regresses the radiance field from this volume representation. The target image is finally obtained using volume rendering [18]. The key components are introduced below. + +# 3.1. Multi-View Feature Encoder + +To aggregate essential information required as input for learning feed-forward NeRF models, we first extract features $\{F_k\}_{k=1}^K$ from $K$ input images $\{I_k\}_{k=1}^K$ . Our feature encoder consists of a weight-sharing per-image 2D CNN and a joint multi-view Transformer. The CNN consists of 6 residual blocks [12] and every 2 residual blocks include a stride-2 convolutional layer to achieve $2 \times$ downsampling. Accordingly, we obtain features at $1/2$ , $1/4$ and $1/8$ resolutions. The $1/8$ convolutional features are then fed into a joint multi-view Transformer to obtain features at $1/8$ resolution. Our Transformer is built on GMFlow [34, 35]'s 2-view Transformer architecture, where we extend the 2-view cross-attention to $K(K \geq 2)$ input views by performing cross-attention for all $K$ views simultaneously. This facilitates more efficient processing of additional input views without significantly increasing memory footprint compared to the pair-wise architecture [5]. The multi-scale + +features are sampled in next volume reconstruction step and the sampled features are concatenated together. + +# 3.2. Target View Frustum Volume + +To render an image for a target novel view, our key difference with previous methods [3, 15, 17] is that we construct our volume aligned with the target view frustum, instead of a pre-defined reference input view. Such a spatially-aligned target view frustum volume representation naturally enables effective aggregation of information from input images. This is in contrast to the reference volume construction approach which has a high risk missing the information outside the epipolar lines of the reference view. This phenomenon is more pronounced for large baselines where the scene overlap is small, as illustrated in Fig. 5. + +The volume elements are sampled from the input images and features to provide necessary cues to aid the prediction of color and density for volume rendering. More specifically, to render a target image of resolution $H \times W$ , we perform $8 \times$ subsampling in the spatial dimension while casting rays and uniformly sample $D$ points on each ray. The $8 \times$ subsampling enables our method to maintain an acceptable volume resolution for high-resolution images, and we finally obtain the full resolution volume with a lightweight upsampler [25]. Next, we introduce the sampling process. + +Color sampling within a window. Due to the $8 \times$ subsampling, we might risk losing some information if we only sample a single point compared to full resolution sampling. To remedy this, we instead sample a $9 \times 9$ window centered at the 2D projection of the 3D point and thus obtain a color vector of $\{\tilde{c}_k^{\mathrm{w}}\}_{k=1}^{K}$ for $K$ input views, where $\tilde{c}_k^{\mathrm{w}}$ denotes the concatenation of colors within the window. + +Feature sampling and matching. We also sample image features $\{\pmb{f}_i\}_{i=1}^K$ from $K$ feature maps. To obtain geometric + +cues, we quantify the multi-view consistency of these features by computing their pair-wise cosine similarities. We then use the cosine feature similarities to provide geometric cues for volume density prediction, based on the observation that the multi-view features of a surface point tend to exhibit high multi-view consistency [5]. All the pair-wise cosine similarities are aggregated with a learned weighted average, which can be expressed as + +$$ +\hat {\boldsymbol {s}} = w _ {i j} \sum_ {i < j} \cos \left(\boldsymbol {f} _ {i}, \boldsymbol {f} _ {j}\right), \quad i, j = 1, 2, \dots , K \tag {1} +$$ + +where $s_{ij}$ denotes the cosine similarity between pair $(\pmb{f}_i, \pmb{f}_j)$ , and $(i,j)$ iterates over all possible $K(K - 1)/2$ non-repeated pairs. Following previous work [39], $w_{ij}$ are normalized visibility weights learned from the entropy of the depth distribution on each ray. + +It's worth noting that the pair-wise cosine similarities $\cos (\pmb {f}_i,\pmb {f}_j)$ in Eq. (1) can be computed in parallel with a single matrix multiplication. More specifically, we first normalize the features to unit vectors and then collect them as a $K\times M$ matrix, where $M$ is the feature dimension. Finally all $K(K - 1) / 2$ pair-wise cosine similarities can be obtained by extracting the upper triangle of the multiplication of this matrix with its transpose. In practice, we use group-wise cosine similarities [5, 11] for further improved expressiveness, where the same computation still applies. + +Multi-view aggregation. Next, the sampled colors and features from $K$ input views are aggregated with a learned weighted average. More specifically, the aggregated color and features are + +$$ +\hat {\boldsymbol {c}} = w _ {i} \sum_ {i = 1} ^ {K} \tilde {\boldsymbol {c}} _ {i} ^ {\mathrm {w}}, \quad \hat {\boldsymbol {f}} = w _ {i} \sum_ {i = 1} ^ {K} \boldsymbol {f} _ {i}, \tag {2} +$$ + +where $w_{i}$ are normalized learned weights [16]. In particular, we use a small MLP network to predict $w_{i}$ from the concatenation of sampled features and view direction difference between the source and target views. + +The cosine similarity $\hat{s}$ in Eq. (1), colors $\hat{c}$ and features $\hat{f}$ in Eq. (2) are then concatenated and projected to a $C$ -dimensional vector with a linear layer. Accordingly we obtain a target volume $z \in \mathbb{R}^{\frac{H}{8} \times \frac{W}{8} \times D \times C}$ for all $\frac{H}{8} \times \frac{W}{8} \times D$ 3D points. This volume encodes appearance and geometry information from multi-view images and features, which is next fed as input to the decoder for radiance field prediction. + +# 3.3. Context-aware Radiance Field Decoder + +Given the target view frustum volume $\mathbf{z} \in \mathbb{R}^{\frac{H}{8} \times \frac{W}{8} \times D \times C}$ , our decoder learns to predict the 4-dimensional (color and density) radiance field $\mathbf{R} \in \mathbb{R}^{H \times W \times D \times 4}$ of all $H \times W \times D$ 3D points. In order to perform well on both small and large baseline input views, we observe that it's crucial to model the 3D context between different 3D points. The context information can help learn useful inductive biases from data + +and accordingly leads to better scene structures. + +To achieve this, we model the 3D context with a convolutional network. A straightforward approach would be using a 3D CNN. In this paper, we explore an alternative for better memory and parameter efficiency, while maintaining similar performance (Fig. 7). More specifically, we factorize the 3D $(3\times 3\times 3)$ convolution to a 2D $(3\times 3\times 1)$ convolution in the spatial dimension and a 1D $(1\times 1\times 3)$ convolution in the depth dimension, i.e., $(2 + 1)\mathrm{D}$ CNN, which is a popular strategy in video recognition works [9, 28]. + +The full decoder consists of two major components: a $(2 + 1)\mathrm{D}$ CNN which operates on the low resolution volume and a lightweight upsampler which achieves $8\times$ upsampling and thus outputs a full-resolution radiance filed. The low-resolution $(2 + 1)\mathrm{D}$ CNN architecture is composed of 12 stacked $(2 + 1)\mathrm{D}$ residual blocks [12]. The final color and density predictions are obtained using two linear layers, with output channels of 3 and 1, respectively. More architectural details are presented in the supplementary material. + +# 3.4. Hierarchical Volume Sampling + +Like other NeRF methods [1, 18, 32], our model also supports hierarchical volume sampling for further improved performance. When hierarchical sampling is used, the model presented before can be viewed as a coarse model, and the hierarchical stage as a fine model. The fine model has a very similar overall architecture as the coarse stage, with the ray sampling process as the key difference. More specifically, given the density prediction from the coarse model, we first compute the Probability Distribution Function (PDF) [18] on each ray by normalizing the color composition weights in the volume rendering equation. Next, we sample a new set of points on each ray according to this distribution. Thanks to the coarse geometry predicted by the coarse network, the fine model requires sampling fewer points since the influence of empty space or occluded regions can be removed. In our implementation, we uniformly sample 64 points on each ray in the coarse model, and only sample 16 points in the fine stage. Due to the smaller number of sampling points, we are able to directly construct a target view frustum volume at the full resolution. We also reduce the volume's channel dimension from 128 to 16, which makes it more efficient for subsequent regression. Thus the volume at the fine stage is more compact. We use a lightweight 3D U-Net [22, 23] to predict the color and density from this volume, and the final color prediction is similarly obtained using volume rendering [18]. More details are presented in the supplementary material. + +# 3.5. Training Loss + +We use random crops from the full image for training, the training loss is an addition of $\ell_1$ , SSIM and LPIPS losses between rendered and ground truth image color. + +# 4. Experiments + +Implementation details. We implement our method using PyTorch [21] and we adopt a two-stage training process, where we first train the coarse model only, and then train the fine model with the coarse model frozen. For all the experiments, we sample 64 points in the coarse model, and 16 points in the fine model. For experiments on the RealEstate10K [40] dataset, we follow AttnRend [8]'s setting to train and evaluate on the $256 \times 256$ resolution. On this resolution, our method can easily use $4 \times$ subsampling when constructing the volume, unlike the default $8 \times$ subsampling. We also reduce the number of channels of the volume from 128 to 64 when $4 \times$ subsampling is used. We don't include the additional hierarchical sampling stage for experiments on the RealEstate10K dataset since the single scale model already performs very competitively with $4 \times$ subsampling. Our code is available at https://github.com/autonomousvision/murf. + +Evaluation settings. Our evaluations include both small and large baselines, different number of views, and diverse scenarios, including object-centric, indoor and unbounded outdoor scenes, to test the model's general applicability. + +# 4.1. Main Results + +Both small & large baselines. To compare with existing methods on both small and large baselines, we choose previous state-of-the-art small baseline method ENeRF [16] and large baseline method AttnRend [8] as two representative approaches. The evaluations are conducted on DTU [14] and RealEstate10K [40] datasets. Since ENeRF only reported its results on DTU and AttnRend only on RealEstate10K, not both, we re-train ENeRF on RealEstate10K and AttnRend on DTU, which allows for more comprehensive comparisons. The results in Table 1 demonstrate that previous state-of-the-art methods are specialized to either small or large baselines, but they can not work well on both. In contrast, our MuRF performs consistently better on both small and large baselines. Next, we conduct comparisons in different specialized settings. + +3-view small baseline on DTU. In this setting, to render a target view, 3 nearest views are selected from all source views based on the distance to the target camera location. Thus this constitutes a small baseline setting. As shown in Table 2, we achieve more than 1dB PSNR improvement compared to previous best method ENeRF [16]. The visual comparisons are shown in Fig. 3, where our method produces significantly better scene structures. + +2-view large baseline on RealEstate10K. The RealEstate10K is a very large dataset, which consists of more than 66K training scenes and more than 7K testing scenes. We follow the evaluation setting of AttnRend [8] for comparisons. More specifically, the two input views are + +
MethodDTU (small baseline)RealEstate10K (large baseline)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
ENeRF [16]27.610.9560.09119.520.7390.341
AttnRend[8]18.570.7320.41921.380.8390.262
MuRF28.760.9610.07724.200.8650.170
+ +Table 1. Comparison on both small and large baselines. Previous state-of-the-art methods are specialized to either small (ENeRF) or large (AttnRend) baselines, but can not work well on both. + +
MethodPSNR↑SSIM↑LPIPS↓
PixelNeRF [38]19.310.7890.382
SRF [6]22.120.8450.292
IBRNet [32]26.040.9170.190
MVSNeRF [3]26.630.9310.168
GeoNeRF [15]26.760.8930.150
MatchNeRF [5]26.910.9340.159
ENeRF [16]27.610.9560.091
MuRF28.760.9610.077
+ +Table 2. DTU 3-view small baseline. + +
MethodPSNR↑SSIM↑LPIPS↓
PixelNeRF [38]13.910.4600.591
SRF [6]15.400.4860.604
GeoNeRF [15]16.650.5110.541
IBRNet [32]15.990.4840.532
GPNR [26]18.550.7480.459
AttnRend [8]21.380.8390.262
MuRF24.200.8650.170
+ +Table 3. RealEstate10K 2-view large baseline. + +
GTENeRFMuRF
GTAttnRenderMuRF
GTGPNRMuRF
+ +Figure 3. Visual comparisons with previous best methods on DTU, RealEstate10K and LLFF datasets. + +selected from a video with a distance of 128 frames, and the target view to synthesis is an intermediate frame. This is a challenging setting since the overlap between two input + +
MethodSetting3-view6-view9-view
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Mip-NeRF [1]8.680.5710.35316.650.7410.19823.580.8790.092
DietNeRF [13]11.850.6330.31420.630.7780.20123.830.8230.173
RegNeRF [19]per-scene optimization18.890.7450.19022.200.8410.11724.930.8840.089
DiffusioNeRF [33]16.200.6980.20720.340.8180.13925.180.8830.095
SPARF [29]21.010.8700.100------
SRF [6]15.320.6710.30417.540.7300.25018.350.7520.232
PixelNeRF [38]16.820.6950.27019.110.7450.23220.400.7680.220
MVSNeRF [3]feed-forward inference18.630.7690.19720.700.8230.15622.400.8530.135
ENeRF† [16]19.840.8560.17121.820.8940.13123.490.9190.106
MuRF21.310.8850.12723.740.9210.09525.280.9360.084
+ +views is usually small, useful priors should be acquired from the model and the data in order to obtain reliable synthesis results. In this large baseline setting, we achieve significant improvement ( $\sim$ 3dB PSNR) than previous best method AttnRend [8], as shown in Table 3. The visual comparison in Fig. 3 indicates that our method produces clearer rendering results than AttnRend [8]. We attribute such large improvements to our target view frustum volume and the convolutional radiance field decoder, which are two key missing components in AttnRend [8]. The target view frustum volume effectively encodes the geometric structure of the view synthesis task and the convolutional decoder enables context-modeling between neighbouring 3D points, both significantly contributing to our final performance. Their effects are more thoroughly analyzed in Sec. 4.2. + +Different number of input views on DTU and LLFF. We also evaluate on different number of input views to further understand the effect of different baselines. Firstly, we follow RegNeRF [19]'s setting of 3, 6 and 9 input views on DTU. Different from the aforementioned 3-view small baseline setting on DTU, the baseline of this setting is larger. More specifically, for each test scene, different numbers of views (3, 6 and 9) are sampled from a fixed set of 9 views. Each scene has 25 test views and they all share the same input views. This is different from the small baseline DTU setting before where the nearest views (instead of shared fixed views for all test views in each scene) are selected for each test view. Thus this setting is more close to the large baseline setting, especially when the number of input views is small. In this setting, our MuRF significantly outperforms previous representative methods like PixelNeRF [38] and MVSNeRF [3], as shown in Table 4. + +To better understand the performance of different methods on different baselines, we re-train ENeRF [16], the previous state-of-the-art method on DTU's small baseline setting, on this large baseline setting with their official code. Since the key component of ENeRF is to use depth-guided sampling to skip empty space in NeRF's volume rendering process, its performance heavily relies on the quality of the + +Table 4. DTU 3, 6 and 9 input views. ${}^{ \dagger }$ Enhanced ENeRF baseline by doubling the number of sampling points on each ray, otherwise we were not able to obtain meaningful results. We again outperform ENeRF by $1 \sim 2\mathrm{{dB}}$ PSNR even compared to this enhanced baseline. + +
Method#viewsPSNR↑SSIM↑LPIPS↓
PixelNeRF [38]1018.660.5880.463
MVSNeRF [3]1021.180.6910.301
IBRNet [32]1025.170.8130.200
NeuRay [17]1025.350.8180.198
GPNR [26]1025.720.8800.175
GNT [31]1025.530.8850.218
425.950.8970.149
MuRF626.040.9000.153
1026.490.9090.143
+ +Table 5. LLFF. Our 4-view model already outperforms previous 10-view methods. It improves further with more views. + +estimated depth. In this large baseline setting, our re-trained ENeRF is not able to produce reasonable results with their original hyper-parameters since the small overlap between input views makes the depth estimation quality lower and thus negatively affect NeRF's rendering process. Thus we train an enhanced ENeRF baseline by doubling the number of sampling points (from original 8 and 2 to 16 and 4 for 2-stage NeRFs) on each ray such that the model is more tolerant to the depth estimation error. The results are shown in Table 4. Our MuRF again outperforms ENeRF by $1 \sim 2\mathrm{dB}$ PSNR in different number of input views. + +Compared with per-scene optimization methods, our MuRF achieves similar performance with the state-of-the-art 3-view method SPARF [29] in the 3-view setting. It's worth noting that SPARF specifically focuses on 3 input views and they didn't report the performance on 6 and 9 views. In contrast, we purse thorough evaluation on different number of views and we show clear improvement than previous per-scene optimization methods with 6 and 9 input views, despite without using any per-scene optimization. + +We also compare with previous methods on the LLFF dataset [18]. Since this is a forward-facing dataset, the baselines between different images are typically small. Previous methods usually evaluate on this dataset with 10 input views following IBRNet [32]'s setting. Surprisingly, our MuRF trained with 4 input views already outperforms all previous 10-view methods (Table 5), suggesting our model is also more data-efficient. Our performance gets further improve + +![](images/d71efff9ad2a57d60f219caefe0497b39e898f15980b9ef5d076e48c5d5060d4.jpg) +Figure 4. Zero-shot generalization on Mip-NeRF 360 dataset. + +![](images/6e5d0051de422dbf8636620ce746498071839e7966ff4989658b5a34acb40d85.jpg) +Figure 5. Our target view volume vs. reference (first image) view volume. Constructing the volume at the pre-defined reference view space might miss relevant information (e.g., red arrows regions) in other views since such information could be far away from the reference view's epipolar lines and thus hard to be sampled. In contrast, we construct the volume in the target view, which more effectively aggregates information from all input views and thus maximizes the information usage. + +
MethodDTUMip-NeRF 360 Dataset
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
AttnRend [8]11.350.5670.65114.000.4740.712
MuRF22.190.8940.21123.980.8000.293
+ +Table 6. Zero-shot generalization after trained on RealEstate10K. Our method outperforms AttnRend [8] by significantly large margins. + +ment with more input views. The visual comparison with previous state-of-the-art method GPNR [26] in Fig. 3 shows that our rendering has better structures. + +Generalization on DTU and Mip-NeRF 360 dataset. For feed-forward models, we are also interested in their zero-shot generalization abilities on unseen datasets, which is a practical problem setting. For this evaluation, we compare with AttnRend [8] using their released model trained on the same RealEstate10K dataset. The results are reported on the object-centric dataset DTU and scene-level Mip-NeRF 360 dataset with 2 input views. From Table 6, we observe our MuRF generalizes significantly better than AttnRend. The lack of geometric inductive biases makes AttnRend more data-hungry and prone to overfitting to the training data. + +To further explore the zero-shot performance limit on Mip-NeRF 360 dataset, we conduct additional fine-tuning with our RealEstate10K pre-trained model on a mixed dataset collected by IBRNet [32]. We achieve further improvement and the promising visual results shown in Fig. 4 indicate the general applicability of our method. + +# 4.2. Ablation and Analysis + +For fast ablation experiments, we only train coarse models on both DTU and RealEstate10K datasets. Since the full RealEstate10K dataset is very large, which would take significant compute to train all the ablations, we use the subset provided by AttnRend [8] for training and evaluation. + +Volume Orientation. One key difference of our method with previous approaches is that we construct the volume in the target view frustum, unlike the popular reference view-based volume in multi-view stereo related methods [3, 15]. Such a volume representation essentially aggregates information from multi-view input images, where the reference view volume would suffer from information loss since the regions outside its epipolar lines are less likely to be sampled. This issue would be more pronounced for large base + +
ModuleMethodLarge Baseline (RealEstate10K)Small Baseline (DTU)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Volume Orientationtarget view21.290.8080.24626.890.9250.119
reference view20.200.7760.30926.500.9220.138
Volume Elementscosine & feature & color21.290.8080.24626.890.9250.119
w/o cosine21.060.8000.25526.680.9190.124
w/o feature20.910.7940.27526.790.9230.132
w/o color20.040.7890.27525.800.9160.133
Radiance Decoder(2+1)D CNN (spatial + depth)21.290.8080.24626.890.9250.119
3D CNN (spatial & depth)21.390.8090.24326.860.9240.121
2D CNN (spatial only)21.060.8010.25626.670.9210.124
1D CNN (depth only)20.780.7880.27226.370.9200.124
Ray Transformer (depth only)20.570.7800.28326.220.9120.129
MLP (point only)20.410.7780.29826.240.9180.126
+ +Table 7. Ablations. Settings used in our final model are underlined. The 3D CNN model has more than $50\%$ parameters (16.6M vs. 10.4M) than our factorized $(2 + 1)\mathrm{D}$ CNN model, and the performance is similar, thus we choose to use the $(2 + 1)\mathrm{D}$ CNN. + +![](images/01b0132d8d5ee1bdf46e49e4864c3d9c6d942634541c5b7ba974ca0c205bc82a.jpg) +Figure 6. CNN vs. MLP vs. Ray Transformer. Our CNN-based decoder yields sharper structures thanks to its context modeling abilities in both spatial and depth dimensions. + +lines, as illustrated in Fig. 5. In Table 7, we can observe a 1dB PSNR performance drop by replacing our target view volume with the reference view volume on the large baseline RealEstate10K dataset. Even on the small baseline DTU dataset, our target view volume is still better since it maximizes the information usage from input views. + +Volume Elements. We construct the volume using information from the sampled colors, features, and features' cosine similarities, they are concatenated together as the volume elements. We ablate different inputs in Table 7 to understand their roles. It's interesting to see that for the large baseline setting, the features are more important than the cosine similarities, while cosine similarities become more important for small baselines. This could be expected since the feature matching information would be less reliable for large baselines due to insufficient scene overlap. It's also worth noting that for both small and large baselines, the sampled colors also contribute to the final performance. The image color provides valuable cues for predicting the 3D points' color, thus leading to improved performance. + +Radiance Decoder. We compare different approaches to predict the radiance field from the target view frustum vol + +ume in Table 7. It can be observed that the spatial context modeled by the 2D CNN is more important to the final performance than that on the depth dimension. Previous methods like IBRNet [32] and GNT [27] use Ray Transformer to only model the context information on the depth dimension. However, we observe in our experiments that the Ray Transformer not always produces better results than MLP (see Fig. 6), while our $(2 + 1)\mathrm{D}$ CNN is consistently better. We also compare with the straightforward 3D CNN decoder and the performance is similar. However, 3D CNN has more than $50\%$ more parameters than our $(2 + 1)\mathrm{D}$ CNN, and thus we choose to use the $(2 + 1)\mathrm{D}$ CNN for better efficiency. The visual comparison in Fig. 6 illustrates that our CNN-based decoder produces clearly better structures. + +# 5. Conclusion + +We present MuRF, a feed-forward method to sparse view synthesis from multiple different baseline settings. Key to our approach are the proposed target view frustum volume and CNN-based radiance field decoder. We achieve state-of-the-art performance on various evaluation settings, demonstrating the general applicability of our method. + +Limitation and Discussion. Currently, large-scale scene-level datasets are still scarce, we expect more diverse training data would improve our performance further. Besides, our model currently assumes known camera parameters and static scenes, extending our method to pose-free scenarios and dynamic scenes could be interesting future directions. + +Acknowledgements. We thank Shaofei Wang, Zehao Yu and Stefano Esposito for the insightful comments on the early draft of this work. We thank Tobias Fischer and Kashyap Chitta for the constructive discussions. We thank Takeru Miyato for the help with the RealEstate10K dataset. Andreas Geiger was supported by the ERC Starting Grant LEGO-3D (850533) and the DFG EXC number 2064/1 - project number 390727645. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1, 4, 6 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1, 2 +[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 1, 2, 3, 5, 6, 7 +[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 1 +[5] Yuedong Chen, Haofei Xu, Qianyi Wu, Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Explicit correspondence matching for generalizable neural radiance fields. In arXiv, 2023. 1, 2, 3, 4, 5 +[6] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In CVPR, 2021. 1, 2, 5, 6 +[7] 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 +[8] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In CVPR, 2023. 1, 2, 5, 6, 7 +[9] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 2, 4 +[10] David Gallup, Jan-Michael Frahm, Philippos Mordohai, and Marc Pollefeys. Variable baseline/resolution stereo. In CVPR, 2008. 2 +[11] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In CVPR, 2019. 4 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 4 +[13] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In ICCV, 2021. 6 +[14] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2, 5 +[15] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In CVPR, 2022. 1, 2, 3, 5, 7 +[16] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2, 4, 5, 6 + +[17] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, 2022. 1, 2, 3, 6 +[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 4, 6 +[19] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 1, 2, 6 +[20] Masatoshi Okutomi and Takeo Kanade. A multiple-baseline stereo. TPAMI, 15(4):353-363, 1993. 2 +[21] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5 +[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4 +[23] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4 +[24] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In CVPR, 2022. 1, 2 +[25] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 2, 3 +[26] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In ECCV. Springer, 2022. 2, 5, 6, 7 +[27] Mukund Varma T, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that neRF needs? In ICLR, 2023. 2, 8 +[28] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 2, 4 +[29] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In CVPR, 2023. 1, 2, 6 +[30] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NIPS, 2017. 1 +[31] Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, Zhangyang Wang, et al. Is attention all nerf needs? arXiv preprint arXiv:2207.13298, 2022. 6 +[32] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo + +Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 1, 2, 4, 5, 6, 7, 8 +[33] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In CVPR, 2023. 1, 2, 6 +[34] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In CVPR, 2022. 3 +[35] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. PAMI, 2023. 3 +[36] Ruigang Yang and Marc Pollefeys. Multi-resolution real-time stereo on commodity graphics hardware. In CVPR, 2003. 2 +[37] Ruigang Yang, Greg Welch, and Gary Bishop. Real-time consensus-based scene reconstruction using commodity graphics hardware. In 10th Pacific Conference on Computer Graphics and Applications, 2002. 2 +[38] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 1, 5, 6 +[39] Jingyang Zhang, Yao Yao, Shiwei Li, Zixin Luo, and Tian Fang. Visibility-aware multi-view stereo network. arXiv preprint arXiv:2008.07928, 2020. 4 +[40] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view synthesis using multiplane images. ACM Trans. Graph., 37(4): 65, 2018. 2, 5 \ No newline at end of file diff --git a/2024/MuRF_ Multi-Baseline Radiance Fields/images.zip b/2024/MuRF_ Multi-Baseline Radiance Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2875ecf8b7951161629cea8da06e6ef7fa029e03 --- /dev/null +++ b/2024/MuRF_ Multi-Baseline Radiance Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1d35323afc26a2c40562f30b832e10a8fafeef14126d00d2a1fa61696b4ffe7 +size 729098 diff --git a/2024/MuRF_ Multi-Baseline Radiance Fields/layout.json b/2024/MuRF_ Multi-Baseline Radiance Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c0874e03a8525dfa22a91535d62308ac8ada8d5f --- /dev/null +++ b/2024/MuRF_ Multi-Baseline Radiance Fields/layout.json @@ -0,0 +1,7587 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 175, + 103, + 419, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 103, + 419, + 119 + ], + "spans": [ + { + "bbox": [ + 175, + 103, + 419, + 119 + ], + "type": "text", + "content": "MuRF: Multi-Baseline Radiance Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "spans": [ + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "text", + "content": "Haofei " + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{1,2}" + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "text", + "content": " Anpei Chen" + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "text", + "content": " Yuedong Chen" + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "text", + "content": " Christos Sakaridis" + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "text", + "content": " Yulun Zhang" + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "text", + "content": " Marc Pollefeys" + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "text", + "content": " Andreas Geiger" + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "text", + "content": " Fisher " + }, + { + "bbox": [ + 83, + 141, + 510, + 172 + ], + "type": "inline_equation", + "content": "\\mathrm{Yu}^{1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 174, + 496, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 174, + 496, + 203 + ], + "spans": [ + { + "bbox": [ + 99, + 174, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 174, + 496, + 203 + ], + "type": "text", + "content": "ETH Zurich " + }, + { + "bbox": [ + 99, + 174, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 174, + 496, + 203 + ], + "type": "text", + "content": "University of Tübingen, Tübingen AI Center " + }, + { + "bbox": [ + 99, + 174, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 99, + 174, + 496, + 203 + ], + "type": "text", + "content": "Monash University " + }, + { + "bbox": [ + 99, + 174, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 99, + 174, + 496, + 203 + ], + "type": "text", + "content": "Shanghai Jiao Tong University " + }, + { + "bbox": [ + 99, + 174, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 99, + 174, + 496, + 203 + ], + "type": "text", + "content": "Microsoft " + }, + { + "bbox": [ + 99, + 174, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 99, + 174, + 496, + 203 + ], + "type": "text", + "content": "Joint last author" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 255, + 290, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 290, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 290, + 495 + ], + "type": "text", + "content": "We present Multi-Baseline Radiance Fields (MuRF), a general feed-forward approach to solving sparse view synthesis under multiple different baseline settings (small and large baselines, and different number of input views). To render a target novel view, we discretize the 3D space into planes parallel to the target image plane, and accordingly construct a target view frustum volume. Such a target volume representation is spatially aligned with the target view, which effectively aggregates relevant information from the input views for high-quality rendering. It also facilitates subsequent radiance field regression with a convolutional network thanks to its axis-aligned nature. The 3D context modeled by the convolutional network enables our method to synthesis sharper scene structures than prior works. Our MuRF achieves state-of-the-art performance across multiple different baseline settings and diverse scenarios ranging from simple objects (DTU) to complex indoor and outdoor scenes (RealEstate10K and LLFF). We also show promising zero-shot generalization abilities on the Mip-NeRF 360 dataset, demonstrating the general applicability of MuRF." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 501, + 128, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 128, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 128, + 513 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 593 + ], + "type": "text", + "content": "Novel view synthesis from sparse input views is a critical and practical problem in computer vision and graphics. However, typical optimization-based Neural Radiance Fields (NeRFs) [1, 2, 4, 18] cannot cope well with sparse views, and additional regularizations [19, 29, 33] are usually required to better constrain the optimization process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 653 + ], + "type": "text", + "content": "In this paper, we aim at learning feed-forward NeRF models [3, 32, 38] that are able to perform feed-forward inference on new unseen data, without requiring per-scene optimization. Moreover, inductive biases could be acquired from data and better results can potentially be attained." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": "To achieve this, existing sparse view methods can be broadly classified into methods that take small baseline images as input and methods that focus on large baseline settings. However, none of the existing methods works well across different baselines (see Fig. 1a and Fig. 1b). In par" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 230, + 425, + 320 + ], + "blocks": [ + { + "bbox": [ + 307, + 230, + 425, + 320 + ], + "lines": [ + { + "bbox": [ + 307, + 230, + 425, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 425, + 320 + ], + "type": "image", + "image_path": "5b16c347522680150641c89b3ef6c8996a33d53409241d46a6ab3f594fa890dd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 323, + 425, + 357 + ], + "lines": [ + { + "bbox": [ + 306, + 323, + 425, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 425, + 357 + ], + "type": "text", + "content": "(a) MuRF outperforms previous state-of-the-art small baseline method ENeRF. The larger baseline, the larger performance gap." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 428, + 230, + 545, + 319 + ], + "blocks": [ + { + "bbox": [ + 428, + 230, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 428, + 230, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 428, + 230, + 545, + 319 + ], + "type": "image", + "image_path": "0f74c0de3814f50ad9bc2ad0220474aa18c0732453508e8df039727a59f94f39.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 323, + 545, + 356 + ], + "lines": [ + { + "bbox": [ + 427, + 323, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 427, + 323, + 545, + 356 + ], + "type": "text", + "content": "(b) MuRF outperforms previous state-of-the-art large baseline method AttnRend. The smaller baseline, the larger performance gap." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 361, + 539, + 471 + ], + "blocks": [ + { + "bbox": [ + 310, + 361, + 539, + 471 + ], + "lines": [ + { + "bbox": [ + 310, + 361, + 539, + 471 + ], + "spans": [ + { + "bbox": [ + 310, + 361, + 539, + 471 + ], + "type": "image", + "image_path": "0d31ace908e061cc66813723fd604c73a8c4c2c0eeb837c007eb17ebec57785d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 473, + 545, + 517 + ], + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 517 + ], + "type": "text", + "content": "Figure 1. MuRF supports multiple different baseline settings. Previous methods are specifically designed for either small (e.g., ENeRF [16]) or large (e.g., AttnRend1 [8]) baselines. However, no existing method performs well on both (see Table 1)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 525, + 546, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 546, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 546, + 670 + ], + "type": "text", + "content": "ticular, small baseline methods [3, 5, 6, 15-17, 32] rely heavily on feature matching to extract relevant information from input views and suffer significantly when there is insufficient image overlap or occlusion occurs. In contrast, large baseline methods [8, 24] are mostly data-driven, and often discard matching cues. Instead, they resort to large-scale datasets and generic architectures (e.g., Transformers [7, 30]) to learn geometric inductive biases implicitly. However, due to their correspondence-unaware and data-driven nature, even state-of-the-art methods [8, 24] tend to produce blurry renderings and their generalization ability is still limited, as we show in Fig. 3 and Table 6." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 670, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 545, + 682 + ], + "type": "text", + "content": "Our goal is to address sparse view synthesis under both" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1We name the previously unnamed method of Du et al. [8] as AttnRend, according its GitHub repository name." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20041" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "small and large camera baselines, different number of input views, and diverse scenarios ranging from simple objects to complex scenes. To this end, we introduce Multi-Baseline Radiance Fields (MuRF), a method for sparse view synthesis from multiple different baseline settings (Fig. 1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 326 + ], + "type": "text", + "content": "More specifically, to render a target novel view, we discretize the 3D space with planes parallel to the target image plane, and accordingly construct a target view frustum volume that is spatially aligned with the target view to render. This is a crucial difference to previous volume-based methods like MVSNeRF [3] and GeoNeRF [15], where the volumes are constructed in a pre-defined reference input view by straightforwardly following the standard practice in multi-view stereo (MVS). However, a key difference between MVS and novel view synthesis (NVS) is that MVS estimates the depth of a reference input view, while NVS aims to render a novel view. Taking this difference into account is crucial, since the reference view volume will not be able to effectively aggregate information from input views when the overlap between the reference and target view is small, which results in failure for large baselines (Fig. 5)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 326, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 434 + ], + "type": "text", + "content": "Our target view frustum volume is constructed by sampling multi-view input image colors and features, which provide appearance information to the prediction of the 3D point's color. We also compute the cosine similarities between sampled features to provide multi-view consistency cues to aid the prediction of volume density [5]. The sampled colors and features, and the computed cosine similarities are concatenated together to provide relevant information for both small and large baseline scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 435, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 519 + ], + "type": "text", + "content": "Equipped with this axis-aligned target volume representation, we further propose to use a convolutional neural network (CNN) to reconstruct the radiance field from this volume. Thanks to the context modeling abilities of CNNs, our method yields more accurate scene structures than previous (MLP-based) point-wise [6, 8] and (Ray Transformer-based) ray-wise methods [5, 27, 32] (see Fig. 6)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "text", + "content": "In practice, we perform " + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "text", + "content": " subsampling in the spatial dimension when constructing the target volume, which enables our method to efficiently handle high-resolution images. The full resolution radiance field is finally obtained with a lightweight " + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "text", + "content": " upsampler [25]. We also use a computationally more efficient 3D CNN alternative by factorizing the " + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "inline_equation", + "content": "3\\mathrm{D}3 \\times 3 \\times 3" + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "text", + "content": " convolution to a " + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "inline_equation", + "content": "2\\mathrm{D}3 \\times 3 \\times 1" + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "text", + "content": " convolution on the spatial dimension and a " + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "inline_equation", + "content": "1\\mathrm{D}1 \\times 1 \\times 3" + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "text", + "content": " convolution on the depth dimension, i.e., " + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "inline_equation", + "content": "(2 + 1)\\mathrm{D}" + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "text", + "content": " CNN, a popular strategy in video recognition works [9, 28]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "We conduct extensive experiments to demonstrate the effectiveness of our proposed target view frustum volume and 3D context-aware radiance field decoder. More specifically, we outperform previous specialized models on both small (DTU [14], LLFF [18]) and large (RealEstate10K [40]) baselines, and achieve promising zero-short generalization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "abilities on the Mip-NeRF 360 [2] dataset, indicating the general applicability of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 104, + 392, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 104, + 392, + 117 + ], + "spans": [ + { + "bbox": [ + 305, + 104, + 392, + 117 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 123, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 545, + 316 + ], + "type": "text", + "content": "Multi-Baseline. The concept of \"multi-baseline\" dates back to multi-baseline stereo depth estimation [10, 20, 36, 37], where multiple stereo pairs with various baselines are used to surpass matching ambiguities and thus achieve improved precision. In this paper, we aim at designing a generally applicable approach to solving novel view synthesis from multiple different sparse view settings. When considering multiple images as input, our problem setup is somewhat similar to multi-baseline stereo, while with the radiance field as the output for volumetric rendering. Besides, we are also interested in novel view synthesis from very sparse (as few as 2) input views with different baselines. In this sense, our proposed multi-baseline radiance field not only draws connections to classic multi-baseline stereo, but also extends it to view synthesis from 2 input views with both small and large baselines." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 317, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 557 + ], + "type": "text", + "content": "Sparse Input Novel View Synthesis. Applying NeRF to novel view synthesis with sparse input views has garnered considerable interest in recent years. Several works improve the optimization-based pipeline via the incorporation of additional training regularizations, e.g., depth [19], correspondence [29] and diffusion model [33]. An alternative area of research turns to the feed-forward designs by learning reliable representations from data, and accordingly removing the need of per-scene optimization. Among them, MVSNeRF [3], GeoNeRF [15] and NeuRay [17] follow the spirit of multi-view stereo (MVS) to construct cost volumes at pre-defined reference viewpoints. However, the reference volume will not be able to effectively aggregate the information for input views when the overlap between the reference and target novel view is small, which accordingly results in failure for large baselines. Besides, ENeRF [16] essentially relies on an MVS network to guide NeRF's sampling process with the estimated depth. However, the depth quality will be unreliable for large baselines and thus severely affects the subsequent rendering process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Another line of works such as SRF [6], IBRNet [32], GPNR [26], GNT [27] and MatchNeRF [5] mostly perform ray-based rendering, where each ray is rendered independently and no explicit 3D context between different rays is modeled. In contrast, our target view volume representation effectively encodes the scene geometry and naturally enables 3D context modeling with a convolutional network, yielding better scene structures. Despite various motivations and implementations, the above models are all designed and experimented only on small baseline settings, limiting their applications in real-world scenarios. Only a few existing approaches focus on the large baseline settings. Nevertheless, these works [8, 24] resort to large-scale" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20042" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 547, + 203 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 547, + 203 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 547, + 203 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 547, + 203 + ], + "type": "image", + "image_path": "b3439f4e5879f86ff2b7aa67e614d9600a2bff2123cf7c85447371c9202d55a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "type": "text", + "content": "Figure 2. Overview. Given multiple input images, we first extract multi-view image features with a multi-view Transformer. To render a target image of resolution " + }, + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "type": "text", + "content": ", we construct a target view frustum volume by performing " + }, + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "type": "text", + "content": " subsampling in the spatial dimension while casting rays and sampling " + }, + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "type": "text", + "content": " equidistant points on each ray. For each 3D point, we sample feature and color information from the extracted feature maps and input images, which consists of the elements of the target volume " + }, + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{\\frac{H}{8} \\times \\frac{W}{8} \\times D \\times C}" + }, + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "type": "text", + "content": " denotes the channel dimension after aggregating sampled features and colors. To reconstruct the radiance field from the volume, we model the context information in the decoder with a (2+1)D CNN operating on low resolution and subsequently obtain the full-resolution radiance field with a lightweight " + }, + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 46, + 206, + 547, + 285 + ], + "type": "text", + "content": " upsampler. The target image is finally rendered with volumetric rendering." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 293, + 289, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 376 + ], + "type": "text", + "content": "datasets and generic architectures, without employing explicit 3D representations, which ultimately leads to blurry rendering and limited generalization. As a comparison, our MuRF is developed with both geometry-aware target view frustum volume representation and 3D context-aware convolutional network, which make it excel at both small and large baseline settings on different datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 385, + 113, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 385, + 113, + 398 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 113, + 398 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 402, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 486 + ], + "type": "text", + "content": "Our MuRF (Multi-Baseline Radiance Fields) is an encoder-decoder architecture (Fig. 2), where the encoder maps the multi-view input images to features that are used to construct a volume at the target view's camera frustum, and the decoder regresses the radiance field from this volume representation. The target image is finally obtained using volume rendering [18]. The key components are introduced below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 493, + 204, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 493, + 204, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 493, + 204, + 504 + ], + "type": "text", + "content": "3.1. Multi-View Feature Encoder" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": "To aggregate essential information required as input for learning feed-forward NeRF models, we first extract features " + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\{F_k\\}_{k=1}^K" + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": " input images " + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\{I_k\\}_{k=1}^K" + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": ". Our feature encoder consists of a weight-sharing per-image 2D CNN and a joint multi-view Transformer. The CNN consists of 6 residual blocks [12] and every 2 residual blocks include a stride-2 convolutional layer to achieve " + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": " downsampling. Accordingly, we obtain features at " + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": " resolutions. The " + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": " convolutional features are then fed into a joint multi-view Transformer to obtain features at " + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": " resolution. Our Transformer is built on GMFlow [34, 35]'s 2-view Transformer architecture, where we extend the 2-view cross-attention to " + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "inline_equation", + "content": "K(K \\geq 2)" + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": " input views by performing cross-attention for all " + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": " views simultaneously. This facilitates more efficient processing of additional input views without significantly increasing memory footprint compared to the pair-wise architecture [5]. The multi-scale" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 293, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 317 + ], + "type": "text", + "content": "features are sampled in next volume reconstruction step and the sampled features are concatenated together." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 326, + 468, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 326, + 468, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 326, + 468, + 338 + ], + "type": "text", + "content": "3.2. Target View Frustum Volume" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 342, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 475 + ], + "type": "text", + "content": "To render an image for a target novel view, our key difference with previous methods [3, 15, 17] is that we construct our volume aligned with the target view frustum, instead of a pre-defined reference input view. Such a spatially-aligned target view frustum volume representation naturally enables effective aggregation of information from input images. This is in contrast to the reference volume construction approach which has a high risk missing the information outside the epipolar lines of the reference view. This phenomenon is more pronounced for large baselines where the scene overlap is small, as illustrated in Fig. 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 475, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 596 + ], + "type": "text", + "content": "The volume elements are sampled from the input images and features to provide necessary cues to aid the prediction of color and density for volume rendering. More specifically, to render a target image of resolution " + }, + { + "bbox": [ + 304, + 475, + 545, + 596 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 304, + 475, + 545, + 596 + ], + "type": "text", + "content": ", we perform " + }, + { + "bbox": [ + 304, + 475, + 545, + 596 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 304, + 475, + 545, + 596 + ], + "type": "text", + "content": " subsampling in the spatial dimension while casting rays and uniformly sample " + }, + { + "bbox": [ + 304, + 475, + 545, + 596 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 475, + 545, + 596 + ], + "type": "text", + "content": " points on each ray. The " + }, + { + "bbox": [ + 304, + 475, + 545, + 596 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 304, + 475, + 545, + 596 + ], + "type": "text", + "content": " subsampling enables our method to maintain an acceptable volume resolution for high-resolution images, and we finally obtain the full resolution volume with a lightweight upsampler [25]. Next, we introduce the sampling process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "text", + "content": "Color sampling within a window. Due to the " + }, + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "text", + "content": " subsampling, we might risk losing some information if we only sample a single point compared to full resolution sampling. To remedy this, we instead sample a " + }, + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "inline_equation", + "content": "9 \\times 9" + }, + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "text", + "content": " window centered at the 2D projection of the 3D point and thus obtain a color vector of " + }, + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\{\\tilde{c}_k^{\\mathrm{w}}\\}_{k=1}^{K}" + }, + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "text", + "content": " input views, where " + }, + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\tilde{c}_k^{\\mathrm{w}}" + }, + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "text", + "content": " denotes the concatenation of colors within the window." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "content": "Feature sampling and matching. We also sample image features " + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_i\\}_{i=1}^K" + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "content": " feature maps. To obtain geometric" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20043" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "cues, we quantify the multi-view consistency of these features by computing their pair-wise cosine similarities. We then use the cosine feature similarities to provide geometric cues for volume density prediction, based on the observation that the multi-view features of a surface point tend to exhibit high multi-view consistency [5]. All the pair-wise cosine similarities are aggregated with a learned weighted average, which can be expressed as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 172, + 287, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 172, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 74, + 172, + 287, + 198 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {s}} = w _ {i j} \\sum_ {i < j} \\cos \\left(\\boldsymbol {f} _ {i}, \\boldsymbol {f} _ {j}\\right), \\quad i, j = 1, 2, \\dots , K \\tag {1}", + "image_path": "fa13de92117828b85025436047a6a91d351c0739321642cc9bb649dac10cb29f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 201, + 287, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 201, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 287, + 261 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 201, + 287, + 261 + ], + "type": "inline_equation", + "content": "s_{ij}" + }, + { + "bbox": [ + 47, + 201, + 287, + 261 + ], + "type": "text", + "content": " denotes the cosine similarity between pair " + }, + { + "bbox": [ + 47, + 201, + 287, + 261 + ], + "type": "inline_equation", + "content": "(\\pmb{f}_i, \\pmb{f}_j)" + }, + { + "bbox": [ + 47, + 201, + 287, + 261 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 201, + 287, + 261 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 47, + 201, + 287, + 261 + ], + "type": "text", + "content": " iterates over all possible " + }, + { + "bbox": [ + 47, + 201, + 287, + 261 + ], + "type": "inline_equation", + "content": "K(K - 1)/2" + }, + { + "bbox": [ + 47, + 201, + 287, + 261 + ], + "type": "text", + "content": " non-repeated pairs. Following previous work [39], " + }, + { + "bbox": [ + 47, + 201, + 287, + 261 + ], + "type": "inline_equation", + "content": "w_{ij}" + }, + { + "bbox": [ + 47, + 201, + 287, + 261 + ], + "type": "text", + "content": " are normalized visibility weights learned from the entropy of the depth distribution on each ray." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 262, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 380 + ], + "type": "text", + "content": "It's worth noting that the pair-wise cosine similarities " + }, + { + "bbox": [ + 46, + 262, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\cos (\\pmb {f}_i,\\pmb {f}_j)" + }, + { + "bbox": [ + 46, + 262, + 287, + 380 + ], + "type": "text", + "content": " in Eq. (1) can be computed in parallel with a single matrix multiplication. More specifically, we first normalize the features to unit vectors and then collect them as a " + }, + { + "bbox": [ + 46, + 262, + 287, + 380 + ], + "type": "inline_equation", + "content": "K\\times M" + }, + { + "bbox": [ + 46, + 262, + 287, + 380 + ], + "type": "text", + "content": " matrix, where " + }, + { + "bbox": [ + 46, + 262, + 287, + 380 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 262, + 287, + 380 + ], + "type": "text", + "content": " is the feature dimension. Finally all " + }, + { + "bbox": [ + 46, + 262, + 287, + 380 + ], + "type": "inline_equation", + "content": "K(K - 1) / 2" + }, + { + "bbox": [ + 46, + 262, + 287, + 380 + ], + "type": "text", + "content": " pair-wise cosine similarities can be obtained by extracting the upper triangle of the multiplication of this matrix with its transpose. In practice, we use group-wise cosine similarities [5, 11] for further improved expressiveness, where the same computation still applies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 384, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 287, + 432 + ], + "type": "text", + "content": "Multi-view aggregation. Next, the sampled colors and features from " + }, + { + "bbox": [ + 47, + 384, + 287, + 432 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 384, + 287, + 432 + ], + "type": "text", + "content": " input views are aggregated with a learned weighted average. More specifically, the aggregated color and features are" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 98, + 434, + 287, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 434, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 98, + 434, + 287, + 467 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {c}} = w _ {i} \\sum_ {i = 1} ^ {K} \\tilde {\\boldsymbol {c}} _ {i} ^ {\\mathrm {w}}, \\quad \\hat {\\boldsymbol {f}} = w _ {i} \\sum_ {i = 1} ^ {K} \\boldsymbol {f} _ {i}, \\tag {2}", + "image_path": "ad57208bbac582b324b27876644171ba7103b7c1f181df8312e056294cf4386a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "text", + "content": " are normalized learned weights [16]. In particular, we use a small MLP network to predict " + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "text", + "content": " from the concatenation of sampled features and view direction difference between the source and target views." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "type": "text", + "content": "The cosine similarity " + }, + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\hat{s}" + }, + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "type": "text", + "content": " in Eq. (1), colors " + }, + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "type": "text", + "content": " and features " + }, + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "type": "text", + "content": " in Eq. (2) are then concatenated and projected to a " + }, + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "type": "text", + "content": "-dimensional vector with a linear layer. Accordingly we obtain a target volume " + }, + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{\\frac{H}{8} \\times \\frac{W}{8} \\times D \\times C}" + }, + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\frac{H}{8} \\times \\frac{W}{8} \\times D" + }, + { + "bbox": [ + 47, + 518, + 287, + 601 + ], + "type": "text", + "content": " 3D points. This volume encodes appearance and geometry information from multi-view images and features, which is next fed as input to the decoder for radiance field prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 610, + 253, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 253, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 253, + 623 + ], + "type": "text", + "content": "3.3. Context-aware Radiance Field Decoder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 627, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 287, + 713 + ], + "type": "text", + "content": "Given the target view frustum volume " + }, + { + "bbox": [ + 46, + 627, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^{\\frac{H}{8} \\times \\frac{W}{8} \\times D \\times C}" + }, + { + "bbox": [ + 46, + 627, + 287, + 713 + ], + "type": "text", + "content": ", our decoder learns to predict the 4-dimensional (color and density) radiance field " + }, + { + "bbox": [ + 46, + 627, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{H \\times W \\times D \\times 4}" + }, + { + "bbox": [ + 46, + 627, + 287, + 713 + ], + "type": "text", + "content": " of all " + }, + { + "bbox": [ + 46, + 627, + 287, + 713 + ], + "type": "inline_equation", + "content": "H \\times W \\times D" + }, + { + "bbox": [ + 46, + 627, + 287, + 713 + ], + "type": "text", + "content": " 3D points. In order to perform well on both small and large baseline input views, we observe that it's crucial to model the 3D context between different 3D points. The context information can help learn useful inductive biases from data" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 499, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 499, + 83 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 499, + 83 + ], + "type": "text", + "content": "and accordingly leads to better scene structures." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 84, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 192 + ], + "type": "text", + "content": "To achieve this, we model the 3D context with a convolutional network. A straightforward approach would be using a 3D CNN. In this paper, we explore an alternative for better memory and parameter efficiency, while maintaining similar performance (Fig. 7). More specifically, we factorize the 3D " + }, + { + "bbox": [ + 304, + 84, + 545, + 192 + ], + "type": "inline_equation", + "content": "(3\\times 3\\times 3)" + }, + { + "bbox": [ + 304, + 84, + 545, + 192 + ], + "type": "text", + "content": " convolution to a 2D " + }, + { + "bbox": [ + 304, + 84, + 545, + 192 + ], + "type": "inline_equation", + "content": "(3\\times 3\\times 1)" + }, + { + "bbox": [ + 304, + 84, + 545, + 192 + ], + "type": "text", + "content": " convolution in the spatial dimension and a 1D " + }, + { + "bbox": [ + 304, + 84, + 545, + 192 + ], + "type": "inline_equation", + "content": "(1\\times 1\\times 3)" + }, + { + "bbox": [ + 304, + 84, + 545, + 192 + ], + "type": "text", + "content": " convolution in the depth dimension, i.e., " + }, + { + "bbox": [ + 304, + 84, + 545, + 192 + ], + "type": "inline_equation", + "content": "(2 + 1)\\mathrm{D}" + }, + { + "bbox": [ + 304, + 84, + 545, + 192 + ], + "type": "text", + "content": " CNN, which is a popular strategy in video recognition works [9, 28]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 193, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 300 + ], + "type": "text", + "content": "The full decoder consists of two major components: a " + }, + { + "bbox": [ + 304, + 193, + 545, + 300 + ], + "type": "inline_equation", + "content": "(2 + 1)\\mathrm{D}" + }, + { + "bbox": [ + 304, + 193, + 545, + 300 + ], + "type": "text", + "content": " CNN which operates on the low resolution volume and a lightweight upsampler which achieves " + }, + { + "bbox": [ + 304, + 193, + 545, + 300 + ], + "type": "inline_equation", + "content": "8\\times" + }, + { + "bbox": [ + 304, + 193, + 545, + 300 + ], + "type": "text", + "content": " upsampling and thus outputs a full-resolution radiance filed. The low-resolution " + }, + { + "bbox": [ + 304, + 193, + 545, + 300 + ], + "type": "inline_equation", + "content": "(2 + 1)\\mathrm{D}" + }, + { + "bbox": [ + 304, + 193, + 545, + 300 + ], + "type": "text", + "content": " CNN architecture is composed of 12 stacked " + }, + { + "bbox": [ + 304, + 193, + 545, + 300 + ], + "type": "inline_equation", + "content": "(2 + 1)\\mathrm{D}" + }, + { + "bbox": [ + 304, + 193, + 545, + 300 + ], + "type": "text", + "content": " residual blocks [12]. The final color and density predictions are obtained using two linear layers, with output channels of 3 and 1, respectively. More architectural details are presented in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 309, + 473, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 309, + 473, + 322 + ], + "spans": [ + { + "bbox": [ + 305, + 309, + 473, + 322 + ], + "type": "text", + "content": "3.4. Hierarchical Volume Sampling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 328, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 652 + ], + "type": "text", + "content": "Like other NeRF methods [1, 18, 32], our model also supports hierarchical volume sampling for further improved performance. When hierarchical sampling is used, the model presented before can be viewed as a coarse model, and the hierarchical stage as a fine model. The fine model has a very similar overall architecture as the coarse stage, with the ray sampling process as the key difference. More specifically, given the density prediction from the coarse model, we first compute the Probability Distribution Function (PDF) [18] on each ray by normalizing the color composition weights in the volume rendering equation. Next, we sample a new set of points on each ray according to this distribution. Thanks to the coarse geometry predicted by the coarse network, the fine model requires sampling fewer points since the influence of empty space or occluded regions can be removed. In our implementation, we uniformly sample 64 points on each ray in the coarse model, and only sample 16 points in the fine stage. Due to the smaller number of sampling points, we are able to directly construct a target view frustum volume at the full resolution. We also reduce the volume's channel dimension from 128 to 16, which makes it more efficient for subsequent regression. Thus the volume at the fine stage is more compact. We use a lightweight 3D U-Net [22, 23] to predict the color and density from this volume, and the final color prediction is similarly obtained using volume rendering [18]. More details are presented in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 660, + 393, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 660, + 393, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 660, + 393, + 673 + ], + "type": "text", + "content": "3.5. Training Loss" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 679, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 679, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 679, + 545, + 715 + ], + "type": "text", + "content": "We use random crops from the full image for training, the training loss is an addition of " + }, + { + "bbox": [ + 304, + 679, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 304, + 679, + 545, + 715 + ], + "type": "text", + "content": ", SSIM and LPIPS losses between rendered and ground truth image color." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20044" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 289, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 289, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 289, + 293 + ], + "type": "text", + "content": "Implementation details. We implement our method using PyTorch [21] and we adopt a two-stage training process, where we first train the coarse model only, and then train the fine model with the coarse model frozen. For all the experiments, we sample 64 points in the coarse model, and 16 points in the fine model. For experiments on the RealEstate10K [40] dataset, we follow AttnRend [8]'s setting to train and evaluate on the " + }, + { + "bbox": [ + 46, + 90, + 289, + 293 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 90, + 289, + 293 + ], + "type": "text", + "content": " resolution. On this resolution, our method can easily use " + }, + { + "bbox": [ + 46, + 90, + 289, + 293 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 46, + 90, + 289, + 293 + ], + "type": "text", + "content": " subsampling when constructing the volume, unlike the default " + }, + { + "bbox": [ + 46, + 90, + 289, + 293 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 46, + 90, + 289, + 293 + ], + "type": "text", + "content": " subsampling. We also reduce the number of channels of the volume from 128 to 64 when " + }, + { + "bbox": [ + 46, + 90, + 289, + 293 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 46, + 90, + 289, + 293 + ], + "type": "text", + "content": " subsampling is used. We don't include the additional hierarchical sampling stage for experiments on the RealEstate10K dataset since the single scale model already performs very competitively with " + }, + { + "bbox": [ + 46, + 90, + 289, + 293 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 46, + 90, + 289, + 293 + ], + "type": "text", + "content": " subsampling. Our code is available at https://github.com/autonomousvision/murf." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 297, + 288, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 288, + 346 + ], + "type": "text", + "content": "Evaluation settings. Our evaluations include both small and large baselines, different number of views, and diverse scenarios, including object-centric, indoor and unbounded outdoor scenes, to test the model's general applicability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 354, + 132, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 132, + 366 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 132, + 366 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 371, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 552 + ], + "type": "text", + "content": "Both small & large baselines. To compare with existing methods on both small and large baselines, we choose previous state-of-the-art small baseline method ENeRF [16] and large baseline method AttnRend [8] as two representative approaches. The evaluations are conducted on DTU [14] and RealEstate10K [40] datasets. Since ENeRF only reported its results on DTU and AttnRend only on RealEstate10K, not both, we re-train ENeRF on RealEstate10K and AttnRend on DTU, which allows for more comprehensive comparisons. The results in Table 1 demonstrate that previous state-of-the-art methods are specialized to either small or large baselines, but they can not work well on both. In contrast, our MuRF performs consistently better on both small and large baselines. Next, we conduct comparisons in different specialized settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 554, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 650 + ], + "type": "text", + "content": "3-view small baseline on DTU. In this setting, to render a target view, 3 nearest views are selected from all source views based on the distance to the target camera location. Thus this constitutes a small baseline setting. As shown in Table 2, we achieve more than 1dB PSNR improvement compared to previous best method ENeRF [16]. The visual comparisons are shown in Fig. 3, where our method produces significantly better scene structures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "2-view large baseline on RealEstate10K. The RealEstate10K is a very large dataset, which consists of more than 66K training scenes and more than 7K testing scenes. We follow the evaluation setting of AttnRend [8] for comparisons. More specifically, the two input views are" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 74, + 544, + 133 + ], + "blocks": [ + { + "bbox": [ + 309, + 74, + 544, + 133 + ], + "lines": [ + { + "bbox": [ + 309, + 74, + 544, + 133 + ], + "spans": [ + { + "bbox": [ + 309, + 74, + 544, + 133 + ], + "type": "table", + "html": "
MethodDTU (small baseline)RealEstate10K (large baseline)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
ENeRF [16]27.610.9560.09119.520.7390.341
AttnRend[8]18.570.7320.41921.380.8390.262
MuRF28.760.9610.07724.200.8650.170
", + "image_path": "1b380ad9ff852756ebd0667abc707328a50c1e3bcb5ddec0d4d9c2711d98cacd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 332, + 177, + 520, + 288 + ], + "blocks": [ + { + "bbox": [ + 305, + 135, + 545, + 168 + ], + "lines": [ + { + "bbox": [ + 305, + 135, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 305, + 135, + 545, + 168 + ], + "type": "text", + "content": "Table 1. Comparison on both small and large baselines. Previous state-of-the-art methods are specialized to either small (ENeRF) or large (AttnRend) baselines, but can not work well on both." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 177, + 520, + 288 + ], + "lines": [ + { + "bbox": [ + 332, + 177, + 520, + 288 + ], + "spans": [ + { + "bbox": [ + 332, + 177, + 520, + 288 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓
PixelNeRF [38]19.310.7890.382
SRF [6]22.120.8450.292
IBRNet [32]26.040.9170.190
MVSNeRF [3]26.630.9310.168
GeoNeRF [15]26.760.8930.150
MatchNeRF [5]26.910.9340.159
ENeRF [16]27.610.9560.091
MuRF28.760.9610.077
", + "image_path": "f812cf1861ad1665082cead986f4c69f1a2d88db53fa75687b27f31c2d10f990.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 332, + 306, + 520, + 407 + ], + "blocks": [ + { + "bbox": [ + 356, + 290, + 494, + 300 + ], + "lines": [ + { + "bbox": [ + 356, + 290, + 494, + 300 + ], + "spans": [ + { + "bbox": [ + 356, + 290, + 494, + 300 + ], + "type": "text", + "content": "Table 2. DTU 3-view small baseline." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 306, + 520, + 407 + ], + "lines": [ + { + "bbox": [ + 332, + 306, + 520, + 407 + ], + "spans": [ + { + "bbox": [ + 332, + 306, + 520, + 407 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓
PixelNeRF [38]13.910.4600.591
SRF [6]15.400.4860.604
GeoNeRF [15]16.650.5110.541
IBRNet [32]15.990.4840.532
GPNR [26]18.550.7480.459
AttnRend [8]21.380.8390.262
MuRF24.200.8650.170
", + "image_path": "4144e479bb43dd97b5ae92ab25c784d8b9c8b49709d8d8a83829916b0b1b1aae.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 419, + 545, + 650 + ], + "blocks": [ + { + "bbox": [ + 338, + 408, + 512, + 418 + ], + "lines": [ + { + "bbox": [ + 338, + 408, + 512, + 418 + ], + "spans": [ + { + "bbox": [ + 338, + 408, + 512, + 418 + ], + "type": "text", + "content": "Table 3. RealEstate10K 2-view large baseline." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 419, + 545, + 650 + ], + "lines": [ + { + "bbox": [ + 307, + 419, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 307, + 419, + 545, + 650 + ], + "type": "table", + "html": "
GTENeRFMuRF
GTAttnRenderMuRF
GTGPNRMuRF
", + "image_path": "1f9e83f1d0a3b034393cdb1da4f26d457ba2fc521af8153cb42c01471a88586a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 651, + 545, + 672 + ], + "lines": [ + { + "bbox": [ + 305, + 651, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 305, + 651, + 545, + 672 + ], + "type": "text", + "content": "Figure 3. Visual comparisons with previous best methods on DTU, RealEstate10K and LLFF datasets." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "selected from a video with a distance of 128 frames, and the target view to synthesis is an intermediate frame. This is a challenging setting since the overlap between two input" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20045" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 70, + 488, + 201 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 488, + 201 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 488, + 201 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 488, + 201 + ], + "type": "table", + "html": "
MethodSetting3-view6-view9-view
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Mip-NeRF [1]8.680.5710.35316.650.7410.19823.580.8790.092
DietNeRF [13]11.850.6330.31420.630.7780.20123.830.8230.173
RegNeRF [19]per-scene optimization18.890.7450.19022.200.8410.11724.930.8840.089
DiffusioNeRF [33]16.200.6980.20720.340.8180.13925.180.8830.095
SPARF [29]21.010.8700.100------
SRF [6]15.320.6710.30417.540.7300.25018.350.7520.232
PixelNeRF [38]16.820.6950.27019.110.7450.23220.400.7680.220
MVSNeRF [3]feed-forward inference18.630.7690.19720.700.8230.15622.400.8530.135
ENeRF† [16]19.840.8560.17121.820.8940.13123.490.9190.106
MuRF21.310.8850.12723.740.9210.09525.280.9360.084
", + "image_path": "26e6407cf1d0206b6763cf0622714398e2fa1a7e54c53c715eaca2e47116f8b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 234, + 289, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 289, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 289, + 415 + ], + "type": "text", + "content": "views is usually small, useful priors should be acquired from the model and the data in order to obtain reliable synthesis results. In this large baseline setting, we achieve significant improvement (" + }, + { + "bbox": [ + 46, + 234, + 289, + 415 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 46, + 234, + 289, + 415 + ], + "type": "text", + "content": "3dB PSNR) than previous best method AttnRend [8], as shown in Table 3. The visual comparison in Fig. 3 indicates that our method produces clearer rendering results than AttnRend [8]. We attribute such large improvements to our target view frustum volume and the convolutional radiance field decoder, which are two key missing components in AttnRend [8]. The target view frustum volume effectively encodes the geometric structure of the view synthesis task and the convolutional decoder enables context-modeling between neighbouring 3D points, both significantly contributing to our final performance. Their effects are more thoroughly analyzed in Sec. 4.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 625 + ], + "type": "text", + "content": "Different number of input views on DTU and LLFF. We also evaluate on different number of input views to further understand the effect of different baselines. Firstly, we follow RegNeRF [19]'s setting of 3, 6 and 9 input views on DTU. Different from the aforementioned 3-view small baseline setting on DTU, the baseline of this setting is larger. More specifically, for each test scene, different numbers of views (3, 6 and 9) are sampled from a fixed set of 9 views. Each scene has 25 test views and they all share the same input views. This is different from the small baseline DTU setting before where the nearest views (instead of shared fixed views for all test views in each scene) are selected for each test view. Thus this setting is more close to the large baseline setting, especially when the number of input views is small. In this setting, our MuRF significantly outperforms previous representative methods like PixelNeRF [38] and MVSNeRF [3], as shown in Table 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "To better understand the performance of different methods on different baselines, we re-train ENeRF [16], the previous state-of-the-art method on DTU's small baseline setting, on this large baseline setting with their official code. Since the key component of ENeRF is to use depth-guided sampling to skip empty space in NeRF's volume rendering process, its performance heavily relies on the quality of the" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 322, + 232, + 531, + 345 + ], + "blocks": [ + { + "bbox": [ + 46, + 203, + 545, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 545, + 226 + ], + "type": "text", + "content": "Table 4. DTU 3, 6 and 9 input views. " + }, + { + "bbox": [ + 46, + 203, + 545, + 226 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 46, + 203, + 545, + 226 + ], + "type": "text", + "content": " Enhanced ENeRF baseline by doubling the number of sampling points on each ray, otherwise we were not able to obtain meaningful results. We again outperform ENeRF by " + }, + { + "bbox": [ + 46, + 203, + 545, + 226 + ], + "type": "inline_equation", + "content": "1 \\sim 2\\mathrm{{dB}}" + }, + { + "bbox": [ + 46, + 203, + 545, + 226 + ], + "type": "text", + "content": " PSNR even compared to this enhanced baseline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 232, + 531, + 345 + ], + "lines": [ + { + "bbox": [ + 322, + 232, + 531, + 345 + ], + "spans": [ + { + "bbox": [ + 322, + 232, + 531, + 345 + ], + "type": "table", + "html": "
Method#viewsPSNR↑SSIM↑LPIPS↓
PixelNeRF [38]1018.660.5880.463
MVSNeRF [3]1021.180.6910.301
IBRNet [32]1025.170.8130.200
NeuRay [17]1025.350.8180.198
GPNR [26]1025.720.8800.175
GNT [31]1025.530.8850.218
425.950.8970.149
MuRF626.040.9000.153
1026.490.9090.143
", + "image_path": "d8326132edc8756001db669a21eb32f7a5c333c64ac7bcceeee43f3f79fb0960.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 346, + 545, + 368 + ], + "lines": [ + { + "bbox": [ + 306, + 346, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 545, + 368 + ], + "type": "text", + "content": "Table 5. LLFF. Our 4-view model already outperforms previous 10-view methods. It improves further with more views." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": "estimated depth. In this large baseline setting, our re-trained ENeRF is not able to produce reasonable results with their original hyper-parameters since the small overlap between input views makes the depth estimation quality lower and thus negatively affect NeRF's rendering process. Thus we train an enhanced ENeRF baseline by doubling the number of sampling points (from original 8 and 2 to 16 and 4 for 2-stage NeRFs) on each ray such that the model is more tolerant to the depth estimation error. The results are shown in Table 4. Our MuRF again outperforms ENeRF by " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "1 \\sim 2\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": " PSNR in different number of input views." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 508, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 616 + ], + "type": "text", + "content": "Compared with per-scene optimization methods, our MuRF achieves similar performance with the state-of-the-art 3-view method SPARF [29] in the 3-view setting. It's worth noting that SPARF specifically focuses on 3 input views and they didn't report the performance on 6 and 9 views. In contrast, we purse thorough evaluation on different number of views and we show clear improvement than previous per-scene optimization methods with 6 and 9 input views, despite without using any per-scene optimization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We also compare with previous methods on the LLFF dataset [18]. Since this is a forward-facing dataset, the baselines between different images are typically small. Previous methods usually evaluate on this dataset with 10 input views following IBRNet [32]'s setting. Surprisingly, our MuRF trained with 4 input views already outperforms all previous 10-view methods (Table 5), suggesting our model is also more data-efficient. Our performance gets further improve" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20046" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 215 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 215 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 215 + ], + "type": "image", + "image_path": "d71efff9ad2a57d60f219caefe0497b39e898f15980b9ef5d076e48c5d5060d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 217, + 414, + 228 + ], + "lines": [ + { + "bbox": [ + 179, + 217, + 414, + 228 + ], + "spans": [ + { + "bbox": [ + 179, + 217, + 414, + 228 + ], + "type": "text", + "content": "Figure 4. Zero-shot generalization on Mip-NeRF 360 dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 233, + 548, + 387 + ], + "blocks": [ + { + "bbox": [ + 48, + 233, + 548, + 387 + ], + "lines": [ + { + "bbox": [ + 48, + 233, + 548, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 548, + 387 + ], + "type": "image", + "image_path": "6e5d0051de422dbf8636620ce746498071839e7966ff4989658b5a34acb40d85.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 388, + 547, + 434 + ], + "lines": [ + { + "bbox": [ + 46, + 388, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 547, + 434 + ], + "type": "text", + "content": "Figure 5. Our target view volume vs. reference (first image) view volume. Constructing the volume at the pre-defined reference view space might miss relevant information (e.g., red arrows regions) in other views since such information could be far away from the reference view's epipolar lines and thus hard to be sampled. In contrast, we construct the volume in the target view, which more effectively aggregates information from all input views and thus maximizes the information usage." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 443, + 286, + 495 + ], + "blocks": [ + { + "bbox": [ + 50, + 443, + 286, + 495 + ], + "lines": [ + { + "bbox": [ + 50, + 443, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 50, + 443, + 286, + 495 + ], + "type": "table", + "html": "
MethodDTUMip-NeRF 360 Dataset
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
AttnRend [8]11.350.5670.65114.000.4740.712
MuRF22.190.8940.21123.980.8000.293
", + "image_path": "5f171706def1c0d54c417257730618b138ad0755fccbb17aae9f70671b1eedb1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 496, + 287, + 531 + ], + "lines": [ + { + "bbox": [ + 47, + 496, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 287, + 531 + ], + "type": "text", + "content": "Table 6. Zero-shot generalization after trained on RealEstate10K. Our method outperforms AttnRend [8] by significantly large margins." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 539, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 574 + ], + "type": "text", + "content": "ment with more input views. The visual comparison with previous state-of-the-art method GPNR [26] in Fig. 3 shows that our rendering has better structures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Generalization on DTU and Mip-NeRF 360 dataset. For feed-forward models, we are also interested in their zero-shot generalization abilities on unseen datasets, which is a practical problem setting. For this evaluation, we compare with AttnRend [8] using their released model trained on the same RealEstate10K dataset. The results are reported on the object-centric dataset DTU and scene-level Mip-NeRF 360 dataset with 2 input views. From Table 6, we observe our MuRF generalizes significantly better than AttnRend. The lack of geometric inductive biases makes AttnRend more data-hungry and prone to overfitting to the training data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "text", + "content": "To further explore the zero-shot performance limit on Mip-NeRF 360 dataset, we conduct additional fine-tuning with our RealEstate10K pre-trained model on a mixed dataset collected by IBRNet [32]. We achieve further improvement and the promising visual results shown in Fig. 4 indicate the general applicability of our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 523, + 432, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 523, + 432, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 523, + 432, + 536 + ], + "type": "text", + "content": "4.2. Ablation and Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 537, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 547, + 598 + ], + "type": "text", + "content": "For fast ablation experiments, we only train coarse models on both DTU and RealEstate10K datasets. Since the full RealEstate10K dataset is very large, which would take significant compute to train all the ablations, we use the subset provided by AttnRend [8] for training and evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Volume Orientation. One key difference of our method with previous approaches is that we construct the volume in the target view frustum, unlike the popular reference view-based volume in multi-view stereo related methods [3, 15]. Such a volume representation essentially aggregates information from multi-view input images, where the reference view volume would suffer from information loss since the regions outside its epipolar lines are less likely to be sampled. This issue would be more pronounced for large base" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20047" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 92, + 70, + 503, + 225 + ], + "blocks": [ + { + "bbox": [ + 92, + 70, + 503, + 225 + ], + "lines": [ + { + "bbox": [ + 92, + 70, + 503, + 225 + ], + "spans": [ + { + "bbox": [ + 92, + 70, + 503, + 225 + ], + "type": "table", + "html": "
ModuleMethodLarge Baseline (RealEstate10K)Small Baseline (DTU)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Volume Orientationtarget view21.290.8080.24626.890.9250.119
reference view20.200.7760.30926.500.9220.138
Volume Elementscosine & feature & color21.290.8080.24626.890.9250.119
w/o cosine21.060.8000.25526.680.9190.124
w/o feature20.910.7940.27526.790.9230.132
w/o color20.040.7890.27525.800.9160.133
Radiance Decoder(2+1)D CNN (spatial + depth)21.290.8080.24626.890.9250.119
3D CNN (spatial & depth)21.390.8090.24326.860.9240.121
2D CNN (spatial only)21.060.8010.25626.670.9210.124
1D CNN (depth only)20.780.7880.27226.370.9200.124
Ray Transformer (depth only)20.570.7800.28326.220.9120.129
MLP (point only)20.410.7780.29826.240.9180.126
", + "image_path": "c8e778c65c0d205ee9648b89f76348283687a59aae3265894138c825e3dc8ced.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 546, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 546, + 251 + ], + "type": "text", + "content": "Table 7. Ablations. Settings used in our final model are underlined. The 3D CNN model has more than " + }, + { + "bbox": [ + 46, + 228, + 546, + 251 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 228, + 546, + 251 + ], + "type": "text", + "content": " parameters (16.6M vs. 10.4M) than our factorized " + }, + { + "bbox": [ + 46, + 228, + 546, + 251 + ], + "type": "inline_equation", + "content": "(2 + 1)\\mathrm{D}" + }, + { + "bbox": [ + 46, + 228, + 546, + 251 + ], + "type": "text", + "content": " CNN model, and the performance is similar, thus we choose to use the " + }, + { + "bbox": [ + 46, + 228, + 546, + 251 + ], + "type": "inline_equation", + "content": "(2 + 1)\\mathrm{D}" + }, + { + "bbox": [ + 46, + 228, + 546, + 251 + ], + "type": "text", + "content": " CNN." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 49, + 259, + 287, + 388 + ], + "blocks": [ + { + "bbox": [ + 49, + 259, + 287, + 388 + ], + "lines": [ + { + "bbox": [ + 49, + 259, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 49, + 259, + 287, + 388 + ], + "type": "image", + "image_path": "01b0132d8d5ee1bdf46e49e4864c3d9c6d942634541c5b7ba974ca0c205bc82a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 390, + 287, + 422 + ], + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 422 + ], + "type": "text", + "content": "Figure 6. CNN vs. MLP vs. Ray Transformer. Our CNN-based decoder yields sharper structures thanks to its context modeling abilities in both spatial and depth dimensions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 438, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 510 + ], + "type": "text", + "content": "lines, as illustrated in Fig. 5. In Table 7, we can observe a 1dB PSNR performance drop by replacing our target view volume with the reference view volume on the large baseline RealEstate10K dataset. Even on the small baseline DTU dataset, our target view volume is still better since it maximizes the information usage from input views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 516, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 685 + ], + "type": "text", + "content": "Volume Elements. We construct the volume using information from the sampled colors, features, and features' cosine similarities, they are concatenated together as the volume elements. We ablate different inputs in Table 7 to understand their roles. It's interesting to see that for the large baseline setting, the features are more important than the cosine similarities, while cosine similarities become more important for small baselines. This could be expected since the feature matching information would be less reliable for large baselines due to insufficient scene overlap. It's also worth noting that for both small and large baselines, the sampled colors also contribute to the final performance. The image color provides valuable cues for predicting the 3D points' color, thus leading to improved performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Radiance Decoder. We compare different approaches to predict the radiance field from the target view frustum vol" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 258, + 547, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 547, + 425 + ], + "type": "text", + "content": "ume in Table 7. It can be observed that the spatial context modeled by the 2D CNN is more important to the final performance than that on the depth dimension. Previous methods like IBRNet [32] and GNT [27] use Ray Transformer to only model the context information on the depth dimension. However, we observe in our experiments that the Ray Transformer not always produces better results than MLP (see Fig. 6), while our " + }, + { + "bbox": [ + 304, + 258, + 547, + 425 + ], + "type": "inline_equation", + "content": "(2 + 1)\\mathrm{D}" + }, + { + "bbox": [ + 304, + 258, + 547, + 425 + ], + "type": "text", + "content": " CNN is consistently better. We also compare with the straightforward 3D CNN decoder and the performance is similar. However, 3D CNN has more than " + }, + { + "bbox": [ + 304, + 258, + 547, + 425 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 258, + 547, + 425 + ], + "type": "text", + "content": " more parameters than our " + }, + { + "bbox": [ + 304, + 258, + 547, + 425 + ], + "type": "inline_equation", + "content": "(2 + 1)\\mathrm{D}" + }, + { + "bbox": [ + 304, + 258, + 547, + 425 + ], + "type": "text", + "content": " CNN, and thus we choose to use the " + }, + { + "bbox": [ + 304, + 258, + 547, + 425 + ], + "type": "inline_equation", + "content": "(2 + 1)\\mathrm{D}" + }, + { + "bbox": [ + 304, + 258, + 547, + 425 + ], + "type": "text", + "content": " CNN for better efficiency. The visual comparison in Fig. 6 illustrates that our CNN-based decoder produces clearly better structures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 437, + 379, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 379, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 379, + 450 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 458, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 531 + ], + "type": "text", + "content": "We present MuRF, a feed-forward method to sparse view synthesis from multiple different baseline settings. Key to our approach are the proposed target view frustum volume and CNN-based radiance field decoder. We achieve state-of-the-art performance on various evaluation settings, demonstrating the general applicability of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 537, + 545, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 545, + 610 + ], + "type": "text", + "content": "Limitation and Discussion. Currently, large-scale scene-level datasets are still scarce, we expect more diverse training data would improve our performance further. Besides, our model currently assumes known camera parameters and static scenes, extending our method to pose-free scenarios and dynamic scenes could be interesting future directions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "type": "text", + "content": "Acknowledgements. We thank Shaofei Wang, Zehao Yu and Stefano Esposito for the insightful comments on the early draft of this work. We thank Tobias Fischer and Kashyap Chitta for the constructive discussions. We thank Takeru Miyato for the help with the RealEstate10K dataset. Andreas Geiger was supported by the ERC Starting Grant LEGO-3D (850533) and the DFG EXC number 2064/1 - project number 390727645." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20048" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 287, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 287, + 212 + ], + "type": "text", + "content": "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 245 + ], + "type": "text", + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 246, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 287, + 290 + ], + "type": "text", + "content": "[5] Yuedong Chen, Haofei Xu, Qianyi Wu, Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Explicit correspondence matching for generalizable neural radiance fields. In arXiv, 2023. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "type": "text", + "content": "[6] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In CVPR, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 287, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 400 + ], + "type": "text", + "content": "[7] 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" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 402, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 287, + 435 + ], + "type": "text", + "content": "[8] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In CVPR, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 435, + 287, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 287, + 467 + ], + "type": "text", + "content": "[9] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 468, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 501 + ], + "type": "text", + "content": "[10] David Gallup, Jan-Michael Frahm, Philippos Mordohai, and Marc Pollefeys. Variable baseline/resolution stereo. In CVPR, 2008. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 502, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 534 + ], + "type": "text", + "content": "[11] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In CVPR, 2019. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "type": "text", + "content": "[13] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In ICCV, 2021. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 601, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 635 + ], + "type": "text", + "content": "[14] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "text", + "content": "[15] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In CVPR, 2022. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[16] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[17] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 164, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 208 + ], + "type": "text", + "content": "[19] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 209, + 545, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 231 + ], + "type": "text", + "content": "[20] Masatoshi Okutomi and Takeo Kanade. A multiple-baseline stereo. TPAMI, 15(4):353-363, 1993. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 233, + 545, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 233, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 307, + 233, + 545, + 287 + ], + "type": "text", + "content": "[21] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 289, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 289, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 545, + 323 + ], + "type": "text", + "content": "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 324, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 324, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 324, + 545, + 357 + ], + "type": "text", + "content": "[23] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 358, + 545, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 358, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 307, + 358, + 545, + 423 + ], + "type": "text", + "content": "[24] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "type": "text", + "content": "[25] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 483, + 545, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 483, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 307, + 483, + 545, + 516 + ], + "type": "text", + "content": "[26] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In ECCV. Springer, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 517, + 545, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 517, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 307, + 517, + 545, + 550 + ], + "type": "text", + "content": "[27] Mukund Varma T, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that neRF needs? In ICLR, 2023. 2, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 552, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 552, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 552, + 545, + 586 + ], + "type": "text", + "content": "[28] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 2, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 587, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 587, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 587, + 545, + 620 + ], + "type": "text", + "content": "[29] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In CVPR, 2023. 1, 2, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 622, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 622, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 622, + 545, + 655 + ], + "type": "text", + "content": "[30] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NIPS, 2017. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 656, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 656, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 656, + 545, + 689 + ], + "type": "text", + "content": "[31] Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, Zhangyang Wang, et al. Is attention all nerf needs? arXiv preprint arXiv:2207.13298, 2022. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "type": "text", + "content": "[32] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20049" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 397 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 286, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 286, + 138 + ], + "type": "text", + "content": "[33] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In CVPR, 2023. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 286, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 286, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 286, + 173 + ], + "type": "text", + "content": "[34] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In CVPR, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 286, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 286, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 286, + 207 + ], + "type": "text", + "content": "[35] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. PAMI, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 286, + 239 + ], + "type": "text", + "content": "[36] Ruigang Yang and Marc Pollefeys. Multi-resolution real-time stereo on commodity graphics hardware. In CVPR, 2003. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 286, + 285 + ], + "type": "text", + "content": "[37] Ruigang Yang, Greg Welch, and Gary Bishop. Real-time consensus-based scene reconstruction using commodity graphics hardware. In 10th Pacific Conference on Computer Graphics and Applications, 2002. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 288, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 286, + 319 + ], + "type": "text", + "content": "[38] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 1, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 321, + 286, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 286, + 354 + ], + "type": "text", + "content": "[39] Jingyang Zhang, Yao Yao, Shiwei Li, Zixin Luo, and Tian Fang. Visibility-aware multi-view stereo network. arXiv preprint arXiv:2008.07928, 2020. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 355, + 286, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 286, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 286, + 397 + ], + "type": "text", + "content": "[40] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view synthesis using multiplane images. ACM Trans. Graph., 37(4): 65, 2018. 2, 5" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20050" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/5a64aa1e-710a-4a50-b851-b34aaaa3cbbb_content_list.json b/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/5a64aa1e-710a-4a50-b851-b34aaaa3cbbb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b7be20d45900cebc4aadc116490e651c020c3fd6 --- /dev/null +++ b/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/5a64aa1e-710a-4a50-b851-b34aaaa3cbbb_content_list.json @@ -0,0 +1,1945 @@ +[ + { + "type": "text", + "text": "Mudslide: A Universal Nuclear Instance Segmentation Method", + "text_level": 1, + "bbox": [ + 166, + 130, + 803, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jun Wang \nSchool of Electronic and Computer Engineering, Peking University, Shenzhen, China \njunw@stu.pku.edu.cn", + "bbox": [ + 147, + 181, + 821, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nuclear instance segmentation has played a critical role in pathology image analysis. The main challenges arise from the difficulty in accurately segmenting densely overlapping instances and the high cost of precise mask-level annotations. Existing fully-supervised nuclear instance segmentation methods, such as boundary-based methods, struggle to capture differences between overlapping instances and thus fail in densely distributed blurry regions. They also face challenges transitioning to point supervision, where annotations are simple and effective. Inspired by natural mudslides, we propose a universal method called Mudslide that uses simple representations to characterize differences between different instances and can easily be extended from fully-supervised to point-supervised. Concretely, we introduce a collapse field and leverage it to construct a force map and initial boundary, enabling a distinctive representation for each instance. Each pixel is assigned a collapse force, with distinct directions between adjacent instances. Starting from the initial boundary, Mudslide executes a pixel-by-pixel collapse along various force directions. Pixels that collapse into the same region are considered as one instance, concurrently accounting for both inter-instance distinctions and intra-instance coherence. Experiments on public datasets show superior performance in both fully-supervised and point-supervised tasks.", + "bbox": [ + 75, + 301, + 473, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 723, + 209, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pathological slide analysis is widely regarded as the gold standard for cancer diagnosis, treatment, and prevention. Nuclear instance segmentation is a critical step in this process, as it enables the accurate identification and analysis of nuclei within pathology images. The morphology of nuclei, including their shape, appearance, and density, provides interpretable features that can be used as diagnostic and prognostic indicators of cancer [18].", + "bbox": [ + 75, + 748, + 468, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nuclear instance segmentation presents unique challenges compared to natural image instance segmentation", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e82e54116028c73b7f39769e95bdea5409bb493bc3187d6105305b306f6c5b33.jpg", + "image_caption": [ + "Figure 1. An analogy between mudslide and nuclear instance segmentation. (Top) Topographic map before and after the mudslide. The colored arrows indicate the collapse forces in different directions. (Bottom) Nuclear instance segmentation process in pathological images. The different layers on the left represent different feature representations of nuclear instances, and the right shows the result of instance segmentation." + ], + "image_footnote": [], + "bbox": [ + 504, + 268, + 890, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tasks. Firstly, the densely packed distribution of nuclei often results in clustered overlapping instances. Secondly, the blurry boundaries between touching or overlapping nuclei make it difficult to distinguish individual instances. Third, obtaining precise pixel-level annotations is time-consuming and expensive for professional pathologists.", + "bbox": [ + 496, + 551, + 893, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the first two challenges, many fully-supervised methods based on deep learning have made significant progress. Several algorithms based on constructing innovative network architectures [8, 20, 23] have been applied to this task. Some contour-based techniques have been investigated to enhance the accurate prediction of nucleus boundaries [3, 9, 29] which aids in splitting the touched and overlapped nuclei. However, these methods lack the capability to characterize the differences among neighboring nuclear instances, making it difficult to accurately segment densely stacked overlapping nuclei. Moreover, their applicability is hindered in scenarios where fine pixel-level annotations are insufficient, as obtaining precise boundary labels becomes a formidable task. Addressing the challenge posed by the scarcity of fine annotations, certain point annotation-based methods have emerged [2, 19, 21], leveraging the accessibility of coarse point labels. How", + "bbox": [ + 496, + 643, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11673", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ever, these approaches frequently prioritize the acquisition of pseudo-mask labels through design optimization methods, neglecting the differentiated representation of overlapping instances.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the mudslide in nature, we propose a simple universal method, called Mudslide. This approach excels in providing a differential representation of instances while remaining readily adaptable to the point-supervised domain.", + "bbox": [ + 76, + 152, + 468, + 214 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 1, for a continuous geographical region, mudslides induced by external forces lead to the creation of distinct partitions within the area. The unaffected section sets the partition boundary, while the collapsing segment forms a contiguous region. Analogously, this parallels the process of nucleus instance segmentation, where a method partitions the foreground with diverse instances into distinct instance regions. Furthermore, the mudslide collapse process concurrently considers the distinctions between regions and the coherence within each region. Variations in force directions acting on different regions induce noticeable changes in adjacent areas. Internal forces within a region, specifically the collapse of one soil mass exerting force on neighboring soil masses, facilitate the clustering and rationalization of soil masses in the same area. This aligns with the objectives of nuclear instance segmentation.", + "bbox": [ + 75, + 215, + 468, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, drawing inspiration from the characteristics of the muslide collapse process, we introduce Mudslide algorithm to induce the collapse of the foreground region of the nucleus and acquire instances. Specifically, we formulate a collapse field and leverage it to construct a force map to allocate forces directed to the center of the instance to individual pixels within the collapse field. This allocation ensures that the force direction in adjacent regions of distinct nuclei exhibits pronounced disparities, facilitating the distinctive representation of neighboring nuclei. Simultaneously, the forces within the same instance exhibit uniformity directed towards the center, enhancing a cohesive representation within instances. Subsequently, we devise a collapse algorithm commencing from the collapse boundary, proceeding pixel-by-pixel along the force direction. Ultimately, pixels collapsing into the same region are deemed components of the same instance. Due to the consideration of external differences between instances and the internal consistency within instances, Mudslide excels in accurately segmenting overlapping instances. The simplicity of collapse field design renders this approach easily extendable to the domain of point supervision.", + "bbox": [ + 75, + 458, + 468, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions can be summarized as follows:", + "bbox": [ + 94, + 792, + 421, + 806 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Mudslide, a novel nucleus instance segmentation algorithm inspired by natural phenomenon of mudslide, which excels in achieving precise segmentation of adjacent or overlapping nuclei.", + "- We propose a collapse field and generate a force map along with an initial boundary to provide differentiated" + ], + "bbox": [ + 76, + 810, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "representations for neighboring instances and enhance the internal coherence within instances.", + "bbox": [ + 509, + 90, + 888, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Mudslide is a universal method that can be easily extended from full supervision tasks to point supervision domains.", + "- We justify the superiority of our Mudslide on public nuclear datasets with state-of-the-art performance." + ], + "bbox": [ + 500, + 121, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 213, + 648, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Fully-supervised nuclei segmentation", + "text_level": 1, + "bbox": [ + 500, + 239, + 820, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fully-supervised nuclear instance segmentation has been widely studied and benefited from the application of deep learning techniques [3, 7, 22]. One of the most influential models in this field is the U-Net model proposed by Ronneberger et al. [23], which has become a benchmark model for medical image segmentation and has inspired many subsequent works. Qu et al. [20] proposed a full-resolution convolutional neural network (FullNet), a full-resolution convolutional neural network that avoids downsampling operations in the network structure to improve localization accuracy. Raza et al. [22] proposed Micro-Net, which utilizes multi-resolution and weighted loss functions to achieve robustness against large internal/external variances in nucleus size. He et al. [8] proposed Han-Net, a hybrid attention nested U-shaped network that extracts informative features from multiple layers to enhance segmentation performance.", + "bbox": [ + 496, + 263, + 890, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leveraging contour information for distinguishing interacting or overlapping nuclei, Chen et al. [3] introduced the integration of contour details into a multi-level Fully Convolutional Network (FCN), which resulted in the development of a profound contour-aware network designed specifically for nuclear instance segmentation. Following this, Zhou et al. [29] introduced the Contour-Aware Information Aggregation Network, which integrates spatial and textural features between nuclei and contours. Additionally, He et al. [9] proposed a Topology-Aware Loss to integrate meaningful topology constraints into the model optimization process for more precise boundary predictions.", + "bbox": [ + 496, + 506, + 890, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, certain studies [4, 6, 14, 16] have employed distance maps to address the segmentation challenges posed by contacting or overlapping nuclei. Naylor et al. [16] tackled the segmentation of touching nuclei by framing it as a regression task involving intra-nuclear distance maps. Graham et al. [6] introduced the Hover-Net, which utilizes vertical and horizontal distances between a nucleus pixel and its center of mass to delineate nuclei. Additionally, He et al. [7] proposed the Centripetal Directional Network (CDNet) for nuclear instance segmentation, integrating directional information into the network.", + "bbox": [ + 496, + 688, + 890, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The above models have achieved performance improvements in nuclear segmentation. However, the challenge of segmenting overlapping cell nuclei persists. To further en", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11674", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/49cb6aa336973ab8bf4dadd4a1ebba43e6efd75ee9dce3d03e17831f77026d5d.jpg", + "image_caption": [ + "Figure 2. The simple modeling of natural mudslide. (Left) Schematic diagram of the force on a soil mass on the slope. The green line represents the slope and the red dashed line represents the soil mass. (Right) A mountain where the mudslide occurred. The mudslide starts at the position pointed by the force indicated by the red arrow and stops at the horizontal slope." + ], + "image_footnote": [], + "bbox": [ + 84, + 90, + 467, + 231 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "hance the model's ability to distinguish instances, there is a need for a distinct and differential representation of nucleus instances.", + "bbox": [ + 75, + 354, + 470, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Pointly-supervised nuclei segmentation", + "text_level": 1, + "bbox": [ + 76, + 436, + 413, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pixel-wise annotation of nucleus instances incurs high costs. Point annotation, as one of the most efficient labeling manners, has garnered increasing attention in the field of nuclear segmentation. Current approaches [2, 19, 21] primarily transform point annotations into coarse pixel-level labels, including Voronoi labels [25], cluster labels [19], and pseudo edge maps [27]. These methods commonly integrate prior knowledge about the shape and texture of nuclei into these coarse pixel-level labels. However, these techniques face challenges related to incomplete coarse labels and imprecise boundary information. Several strategies have been proposed to address the issues arising from these coarse labels. These include multi-stage optimization for refining segmentation in a bootstrapping manner [21, 25], and the incorporation of additional constraints based on the local contrast between nuclei and their surrounding cytoplasm. Additionally, some unique training strategies have been employed to optimize pseudo-labels. [12] utilizes the co-training method, optimizing mask labels through the simultaneous training of two models with mutual supervision. [28] employs a collaborative knowledge-sharing approach, sharing knowledge between a principal model and a very lightweight collaborator model.", + "bbox": [ + 75, + 468, + 472, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite the progress in the field of making full use of point labels and generating accurate pseudo mask labels, however, the problem of nucleus overlapping instance segmentation in point-supervised prediction still needs further attention.", + "bbox": [ + 75, + 824, + 472, + 898 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/220afe0d0232dc0f9567b9df89bde1e6148d96ef7663b890e5ef5c8930ac5cbc.jpg", + "image_caption": [ + "Figure 3. The processing of Mudslide algorithm. (Left) The instance map, the four-channel collapse field, the eight-directional force map, and the initial boundary. Each color in the force map represents a direction. (Right) Schematic diagram of the mudslide collapse process performed on two adjacent nuclei." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 888, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 337, + 591, + 353 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Natural Mudslide", + "text_level": 1, + "bbox": [ + 500, + 362, + 671, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mudslide refers to the natural phenomenon where the soil or rock mass on a slope, influenced by external forces such as river erosion, rainfall, and seismic events, undergoes downhill movement collectively along the slope [5]. To simulate the mudslide process, we simplify it as a continuous sliding problem of soil mass on a slope [24]. This assumes an arc-shaped sliding surface adhering to the Coulomb criterion during soil movement, and the soil undergoes rigid sliding. This allows the assessment of the stability of the soil slope based on the magnitudes of the sliding moment and anti-sliding moment.", + "bbox": [ + 496, + 386, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, for the soil mass on the slope, the sliding moment is generated by the gravitational force $W$ , while the resisting moment is provided by the frictional force on the sliding surface. The sliding moment $M_{s}$ and resisting moment $M_{r}$ are", + "bbox": [ + 496, + 551, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {s} = W * d, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 632, + 890, + 647 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {r} = R * \\int_ {0} ^ {l} \\tau d l, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 647, + 890, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d$ represents the lever arm of gravity, $R$ and $l$ are respectively the radius and the length of the arc, and $\\tau$ represents the shear force. Introducing the safety factor $F_{os}$ [17]", + "bbox": [ + 498, + 684, + 890, + 726 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {o s} = \\frac {M _ {r}}{M _ {s}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 732, + 890, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "if $F_{os} > 1$ , the soil mass remains stable and will not collapse. If $F_{os} = 1$ , the soil mass is in a state of critical equilibrium, and if $F_{os} < 1$ , the soil mass undergoes collapse.", + "bbox": [ + 496, + 766, + 890, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assuming each soil mass on the slope is in a critical state, i.e., $F_{os} = 1$ , if an external force $F$ with moment $M_F$ is applied tangentially along the slope surface such that", + "bbox": [ + 496, + 823, + 890, + 868 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {o s} ^ {\\prime} = \\frac {M _ {r}}{M _ {s} ^ {\\prime}} = \\frac {M _ {r}}{M _ {s} + M _ {F}} < F _ {o s}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 871, + 890, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11675", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the corresponding soil mass undergoes collapse. This results in the application of a tangential force $F^{\\gamma}$ on the adjacent soil mass below, leading to a continued collapse of the lower soil mass, ultimately forming a mudslide. The collapse terminates when it reaches a horizontal slope surface. In the end, the mudslide induces the collapse of soil mass on the slope surface affected by external forces, while the unaffected soil mass remains intact. This delineates distinct regions on the slope.", + "bbox": [ + 76, + 90, + 472, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Mudslide Segmentation", + "text_level": 1, + "bbox": [ + 76, + 244, + 294, + 260 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Mudslide Processing", + "text_level": 1, + "bbox": [ + 76, + 272, + 274, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by the natural mudslide mentioned in Sec. 3.1, we propose a Mudslide algorithm for nuclear instance segmentation. In natural landslide phenomena, the main issues include the characteristics of the slope that collapses, the forces acting on the soil mass on the slope, the location where the collapse begins, and how the collapse occurs continuously. Corresponding to this, the Mudslide algorithm primarily involves three steps: constructing a collapse field, obtaining the force map and initial collapse boundary, and executing the collapse.", + "bbox": [ + 76, + 303, + 470, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Collapse field construction. In the first step, we initiate the process by constructing a collapse field, representing the initial slope surface of the mudslide. In image segmentation, we assume that each pixel represents a soil mass, located on a slope represented by the collapse field, and in a state of critical equilibrium. For simplicity, we utilize continuous normalized maps to characterize the collapse field, where pixel values close to 1 signify higher slope heights and values close to 0 indicate lower slope surfaces. Specifically, for ease of model learning and inspired by [6], we create a four-channel normalized map shown in Fig. 3. This map contains four distance maps generated in horizontal, vertical, and two diagonal directions, respectively, with each map indicating the distance from each pixel to the center of the nucleus in different directions. Pixel values are close to 0 when they are near the nucleus center and gradually increase outward until they reach 1 at the nucleus boundary. Thus, we construct a slope for the foreground nuclei in the image, where the slope is flat in the middle of each instance and steep around the edges, resembling the terrain that is prone to mudslide collapse.", + "bbox": [ + 76, + 458, + 472, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Force map and initial boundary obtaining. For the second step, we generate a force map and initial boundary from the collapse field. The force map assigns a force $F$ with different directions to each pixel, representing the directions in which pixels are expected to collapse during Mudslide. For two-dimensional images, we consider eight directions of force: up, down, left, right, upper-left, upper-right, lower-left, and lower-right (Fig. 3):", + "bbox": [ + 76, + 780, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {F} = \\left\\{\\binom {0} {1}, \\binom {0} {- 1}, \\binom {- 1} {0}, \\binom {1} {0}, \\right. \\\\ \\binom {- 1} {1}, \\binom {1} {1}, \\binom {- 1} {- 1}, \\binom {1} {- 1} \\}. \\end{array} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 87, + 890, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then the force map can be expressed as a set of forces:", + "bbox": [ + 503, + 166, + 867, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF M = \\{F (x, y) \\mid x, y \\in \\mathbb {Z}, F \\in \\mathbb {F} \\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 185, + 890, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{Z}$ indicates the foreground of nuclei and $x, y$ corresponds to the coordinate of each pixel. Here, we assign a force direction to each pixel that is approximately tangential to the slope. Since we assume that each pixel is in a state of critical equilibrium, we consider the force magnitude to be irrelevant to the collapse and therefore set the magnitude of each force to 1.", + "bbox": [ + 496, + 213, + 890, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain the starting position of mudslide collapse, we generate the initial boundary. We employ a straightforward edge detection filter, such as Sobel operator, on the collapse field to detect edges. The obtained edges are then utilized as the initial boundary for mudslide collapse. This approach is rational since, in our collapse field, there is a significant gradient between instance boundaries, allowing the Sobel operator to effectively identify them.", + "bbox": [ + 496, + 319, + 892, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mudslide collapse executing. After acquiring the force map and the initial boundary, we commence the mudslide collapse process. Originating from the initial boundary, each pixel $p$ applies an initial force (red arrows in 3) to the directed pixel $p'$ , which can be calculated by", + "bbox": [ + 496, + 439, + 890, + 515 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np ^ {\\prime} = p + F (p),\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 518, + 748, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np = (x, y), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 541, + 890, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np ^ {'} = (x ^ {'}, y ^ {'}).\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 559, + 733, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pixel $p'$ then undergoes collapse. Subsequently, the collapsed pixel exerts a sustained force (yellow arrows in 3) on the pointed pixel, initiating the collapse of more adjacent pixels. This iterative process results in a sustained mudslide collapse. Eventually, when all pixels directed to have undergone collapse, those pixels collapsing to the same region are identified as a single instance. Note that those pixels that are never directed by forces from other pixels will not collapse and the force assigned to them will be neglected (gray arrows in 3), corresponding to the soil mass that is not affected by the mudslide.", + "bbox": [ + 496, + 590, + 890, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Application to Nuclear Instance Segmentation", + "text_level": 1, + "bbox": [ + 498, + 771, + 874, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The framework of applying Mudslide to nuclear instance segmentation is illustrated in Fig. 4. Here, we unify the pipelines for both fully supervised and point-supervised scenarios, with the primary distinction lying in the labels used for training.", + "bbox": [ + 496, + 794, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we employ a dual-branch architecture, where one branch is dedicated to predicting a binary map", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11676", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9c569adf6916f15313c6c1194f95af122cf9dcc8702ae0bbe2364719bac39d29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 125, + 92, + 843, + 271 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0feec5d3f43da033cc3dfcd6818f06f5046869aca5f5184773ca9f33312dd437.jpg", + "image_caption": [ + "Figure 4. The overall framework of the proposed Mudslide. The network contains two basic prediction branches: the binary branch and the collapse field branch. During training, the difference between fully- and point-supervised is the type of ground truth. During testing, both branches' predictions are employed for the Mudslide algorithm." + ], + "image_footnote": [], + "bbox": [ + 125, + 273, + 843, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for the nuclear foreground, and the other branch focuses on predicting the collapse field. The collapse field consists of four channels, corresponding to horizontal, vertical, and two diagonal directions. The segmentation model adopted is the Trans-UNet architecture from [10]. In fully supervised training, the binary map labels are the ground truth representing the nuclear foreground, while the collapse field labels are constructed from both the foreground and instance ground truth. For point-supervised training, we leverage point labels and the original input image to generate a cluster map [19] and Voronoi map [25] as pseudolabels for foreground prediction, which is commonly observed in point-supervised tasks. Subsequently, we use the predicted foreground and point labels to generate collapse field labels.", + "bbox": [ + 75, + 508, + 468, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize cross-entropy loss as the binary classification loss and employ mean squared error loss and mean squared gradient error loss as the collapse field loss:", + "bbox": [ + 76, + 736, + 468, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {B} = \\lambda_ {B _ {C E}} \\mathcal {L} _ {C E}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 789, + 468, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C F} = \\lambda_ {C F _ {M S E}} \\mathcal {L} _ {M S E} + \\lambda_ {C F _ {M S G E}} \\mathcal {L} _ {M S G E}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 813, + 468, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the overall loss function is", + "bbox": [ + 81, + 835, + 295, + 849 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {B} + \\mathcal {L} _ {C F}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 858, + 468, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It's worth noting that in point-supervised training, the", + "bbox": [ + 84, + 885, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cross-entropy loss comprises two terms, corresponding to cluster and Voronoi:", + "bbox": [ + 498, + 508, + 890, + 539 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {B _ {\\text {p o i n t}}} = \\lambda_ {B _ {C E}} \\mathcal {L} _ {C E _ {\\text {c l u s t e r}}} + \\lambda_ {B _ {C E}} \\mathcal {L} _ {C E _ {\\text {v o r o n o i}}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 551, + 890, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the inference process, we first combine the foreground and collapse field predictions to construct the collapse field, ensuring subsequent mudslide acts on the nuclear foreground regions. Subsequently, as detailed in Sec. 3.1, we utilize the collapse field to obtain the initial boundary and force map, initiating the mudslide collapse process. Ultimately, this process yields the nuclear instance segmentation results.", + "bbox": [ + 496, + 580, + 890, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 715, + 624, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 739, + 705, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. I) CPM17 [26] contains 64 H&E stained histopathology images with 7,570 annotated nuclear boundaries. It is from the MICCAI 2017 Digital Pathology Challenge [26] and images on two different scales: $500 \\times 500$ and $600 \\times 600$ . This is a dataset for a single-class segmentation task. II) NuInsSeg [15] contains 665 image patches with more than 30,000 manually segmented nuclei from 31 human and mouse organs. It is one of the biggest fully manually annotated datasets of nuclei in Hematoxylin and", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11677", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eosin (H &E) stained histological images with the scale of $512 \\times 512$ . This is also a dataset for single-class segmentation task.", + "bbox": [ + 75, + 90, + 468, + 133 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use four instance-level evaluation metrics to measure the instance segmentation performance of the comparison models, which are: Detection Quality (DQ), Segmentation Quality (SQ), Panoptic Quality (PQ), and the Aggregated Jaccard Index (AJI).", + "bbox": [ + 75, + 136, + 467, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. The proposed Mudslide is trained and tested using the open-source software library Pytorch 1.8.0 on NVIDIA Tesla V100 with CUDA 10.1.", + "bbox": [ + 75, + 213, + 467, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order to build the dataset used for training, the CPM17 dataset is divided into multiple patches as it has an inconsistent original image size. We generate 512 training patches in the size of $256 \\times 256$ , thus allowing the network to be uniform and fully trained. In addition, we randomly divide NuInsSeg into training set and test set with the proportion of $80\\%$ and $20\\%$ , respectively. Moreover, for point-supervised tasks, we leverage the annotations of nucleus instances to generate a central point for each instance, serving as the point label. In the training phase, we apply a series of data augmentation methods, which include random flip, random rotation, random color, random crop, and Gaussian blur. The training epoch is set as 300, and the initial learning rate is set as $10^{-3}$ . The weights of each loss are $\\lambda_{BCE} = 1$ , $\\lambda_{CFMSE} = 2.5$ , $\\lambda_{CFMSGE} = 8$ .", + "bbox": [ + 75, + 258, + 467, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 496, + 228, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We performed ablation studies on the CPM17 and NuInsSeg datasets to validate the efficacy of the proposed Mudslide method. All the following experiments were conducted under the setting of fully supervised tasks.", + "bbox": [ + 75, + 521, + 467, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instance separation methods. Table 1 shows the comparison of segmentation performance obtained using different instance separation methods. We apply four different methods to the predictions of the same model: the traditional watershed method [1], boundary partition method, HVmap segmentation method from Hover-Net [6], and our proposed method.", + "bbox": [ + 75, + 582, + 467, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, the watershed method involves segmenting instances using the watershed algorithm on the binary probability map predicted by the model. The boundary separation method refers to predicting boundaries by adding an additional boundary prediction branch to the model and using the boundaries to directly separate different instances. The HVmap algorithm is an instance separation method in HoverNet that utilizes horizontal and vertical distance maps for segmentation.", + "bbox": [ + 75, + 688, + 467, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The experimental results indicate significant variations in model performance based on different instance separation methods. Traditional methods such as watershed and boundary separation exhibit subpar performance. The HVmap approach, utilizing horizontal and vertical maps,", + "bbox": [ + 75, + 824, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3c9744391d34e2065923ccfefb3582e297d552987a984d85f67e52f7a15a65c0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
Watershed0.7970.7300.5820.6940.6010.7350.4610.502
Boundary0.8370.7340.6140.7140.6400.7530.4910.532
HVMap0.8790.8100.7130.7370.7860.7540.5960.638
Ours0.8830.8080.7140.7400.7880.7560.5990.647
", + "bbox": [ + 506, + 90, + 885, + 188 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/efada8d7a07132abd73fb0b8055fd3980c6db39e5f82598c84d1c08fed605078.jpg", + "table_caption": [ + "Table 1. Comparison of different adjacent instance separation methods, including the watershed algorithm, boundary partition, HVmap and our method." + ], + "table_footnote": [], + "table_body": "
boundcollapseCPM17NuInsSeg
DQ ↑SQ ↑PQ ↑AJI ↑DQ ↑SQ ↑PQ ↑AJI ↑
✓0.7720.7100.5480.6100.5870.7030.4140.497
0.8770.8070.7070.7330.7750.7520.5820.630
✓✓0.8830.8080.7140.7400.7880.7560.5990.647
", + "bbox": [ + 500, + 244, + 898, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Ablation studies of components in Mudslide algorithm, where bound indicates initial boundary, and collapse indicates mudslide collapse processing.", + "bbox": [ + 498, + 335, + 890, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "enhances the differentiation between instances and achieves better results. Furthermore, our method, with more explicit instance differentiation representation and a more rational instance acquisition process, outperforms other methods. This underscores the superiority of our Mudslide approach.", + "bbox": [ + 498, + 383, + 890, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The ablation of components of Mudslide algorithm. We conducted ablation experiments on the key steps of the mudslide algorithm. Here, we considered two components: whether to use the initial boundary and whether to perform mudslide collapse. In Table 2, we explored three comparative experiments: a baseline approach with no additional components, relying solely on binary map partitioning based on connected regions; a scenario involving only the initial boundary extracted from the collapse field; and a comprehensive application of both the initial boundary and mudslide collapse, representing our full-fledged mudslide algorithm for instance segmentation. These experiments aim to scrutinize the individual contributions of each component to the overall segmentation performance.", + "bbox": [ + 498, + 460, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The experimental results indicate that incorporating an initial boundary leads to performance gains of $16\\%$ PQ and $12\\%$ AJI for the CPM dataset and $17\\%$ PQ and $14\\%$ AJI for NuInsSeg. Furthermore, the additional mudslide collapse results in a performance boost of $7\\%$ PQ and $7\\%$ AJI for the CPM dataset and $16\\%$ PQ and $17\\%$ AJI for NuInsSeg. These findings underscore the substantial benefits of each component within the mudslide method for enhancing instance segmentation outcomes.", + "bbox": [ + 498, + 672, + 890, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Force map construction methods. As a crucial manifestation of distinct instance differentiation, the construction of the force map is pivotal in the Mudslide algorithm. Therefore, we consider three methods for building the force map: direct prediction, collapse field segmentation, and collapse field extraction.", + "bbox": [ + 498, + 810, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11678", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/37a2ec5dab2c41eba39c37924a6fd0bd05236b0d0b0cf40d4bdda38cf8d6bad8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
predict0.8740.8070.7070.7350.7770.7510.5870.633
segment0.8640.8080.7010.7280.7720.7530.5820.629
extract0.8800.8070.7110.7390.7840.7550.5920.643
", + "bbox": [ + 78, + 92, + 470, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the first approach, we have the model predict the force map directly. Specifically, we augment the model with a branch that outputs nine channels, allowing it to learn the force map's eight directions and the background region. Subsequently, we obtain the force map by employing the argmax function to the predicted probability map. In the second approach, we perform segmentation on the collapse field. Here, we use a two-channel collapse field corresponding to the horizontal and vertical directions. By segmenting the collapse field based on angles, we categorize it into eight regions, representing the eight directions of the force map. In the third approach, we predict a four-channel collapse field for horizontal, vertical, and two diagonal directions. Then, we directly employ argmax to extract the direction of force that each pixel should possess.", + "bbox": [ + 75, + 218, + 470, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results presented in Table 3 indicate that the force map extracted through four-channel collapse fields performs the best. The direct prediction follows with slightly lower performance, while the force map obtained by splitting the two-channel collapse field performs the least favorably.", + "bbox": [ + 75, + 445, + 468, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The four-channel collapsing field extraction method outperformed direct prediction by $4\\%$ PQ and $4\\%$ AJI on CPM17, and outperformed the two-channel segmentation method by $6\\%$ PQ, and $11\\%$ AJI. On the NuInsSeg dataset, it outperformed direct prediction by $5\\%$ PQ, and $1\\%$ AJI, and outperformed the two-channel segmentation method by $10\\%$ PQ, and $14\\%$ AJI. These results suggest that the four-channel collapsing field can better capture and characterize the differences between cell nucleus instances, resulting in greater differences between adjacent instances and better instance segmentation results.", + "bbox": [ + 75, + 536, + 468, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Initial boundary generation methods. Regarding the issue of generating and selecting initial boundaries, we considered two methods: direct generation from the binary prediction map and generation from the collapsing field.", + "bbox": [ + 75, + 703, + 468, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, for the first method, we first obtained the binary probability map predicted by the model's binary classification branch and then selected a threshold of 0.5 to segment the predicted foreground and background regions of the nucleus. Then, we used the part of the pixels that contacted the background, i.e., the edge of the foreground, as the initial boundary. For the second method, we first obtained the four-channel collapsing field predicted by the model and then used some edge extraction methods, such as", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b55debf562f8ef3a2c33162710024a48a25ace8802db1d2135d7c9029f2978c5.jpg", + "table_caption": [ + "Table 3. Comparison of different force map construction methods." + ], + "table_footnote": [], + "table_body": "
MethodCPMNuInsSeg
DQ ↑SQ ↑PQ ↑AJI ↑DQ ↑SQ ↑PQ ↑AJI ↑
binary map collapse field0.8800.8070.7110.7390.7840.7550.5920.643
0.8830.8080.7140.7400.7880.7560.5990.647
", + "bbox": [ + 501, + 92, + 895, + 159 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison of different initial boundary generation methods.", + "bbox": [ + 498, + 167, + 890, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "using the Sobel operator to extract the regions with higher gradients in the collapsing field as the initial boundary.", + "bbox": [ + 498, + 203, + 890, + 233 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 4, the method of generating initial boundaries using the collapsing field outperforms the method of using only the binary prediction map. Specifically, on the CPM dataset, the collapsing field method improved performance by $3\\%$ PQ, and $1\\%$ AJI compared to the binary map method. On the NuInsSeg dataset, it improved performance by $7\\%$ PQ, and $4\\%$ AJI. As anticipated, the construction of the initial boundary is crucial in our Mudslide collapse algorithm, as it determines the algorithm's starting position. Obtaining foreground edges directly from a binary image makes it challenging to capture the boundaries between adjacent instances. Although the Mudslide algorithm can capture this difference to some extent, its performance is not as effective as constructing a better initial boundary. The initial boundary extracted through the collapse field takes into account the differentiated representations between cell nuclei, effectively separating nuclei to a certain degree and proving to be a superior boundary.", + "bbox": [ + 496, + 233, + 892, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Fully-Supervised Nuclei Segmentation", + "text_level": 1, + "bbox": [ + 498, + 513, + 830, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 5, the proposed Mudslide is compared with the state-of-the-art fully-supervised nuclear instance segmentation models on CPM17 and NuInsSeg. Our Mudslide achieves 0.714 PQ and 0.740 AJI on the CPM17 dataset, and 0.599 PQ and 0.647 AJI on the NuInsSeg dataset, respectively. The Mudslide achieves competitive results for fully-supervised nuclei segmentation tasks.", + "bbox": [ + 496, + 536, + 890, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further carry out a qualitative visualization analysis of our method and different existing methods. Compared to all the other methods in Fig. 5, our Mudslide achieves superior instance segmentation results for nuclear instance segmentation, especially in some regions where nuclei are densely distributed with a large amount of overlap, where our instance segmentation is more accurate. The results, both quantitative and qualitative, unequivocally establish the competitiveness of our Mudslide algorithm in accurately segmenting densely packed and overlapping nuclear instances.", + "bbox": [ + 496, + 643, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Pointly-Supervised Nuclei Segmentation", + "text_level": 1, + "bbox": [ + 498, + 816, + 844, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 6, we extend the proposed Mudslide from fully-supervised to point-supervised nuclear segmentation task and compare it with the state-of-the-art point-supervised nuclear instance segmentation models on CPM17 and", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11679", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aaaa8803377b0f6c6a24fd5d1dab9dc5b6f305fabb1f24b38a7f39e42a51b463.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 89, + 228, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/125633ffe895597bc4643391ed91a588962169c02a77af5e296c475763189d63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 166, + 227, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66186f44fb71f4cff435397bf4122e9138c00fb64c710f2132b49c8687cb8f6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 241, + 227, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/64ee2f5006a77cbef1b94e2ac689ac7aa3e76163828a5181e528f0be8fff67f7.jpg", + "image_caption": [ + "(a)", + "Image", + "Figure 5. Visualization of instance segmentation results with different methods. (a) Input image; (b) Ground Truth; (c) U-Net [23]; (d) DIST [16]; (e) FullNet [20]; (f) Hover-Net [6]; (g) Our Mudslide." + ], + "image_footnote": [], + "bbox": [ + 130, + 316, + 227, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5708ec99f9a7bd92be593b662715669722c556dc722f51b11b483b6c3f6b367a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 90, + 330, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f8e0f9de5635e1810e045b1aa2108fbd2d5074244e951531ad6531f859986121.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 166, + 330, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/471482f1465a64ceb3264a71576a892c45c19dbd6c0b6c5be3d0b76ae4dc7064.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 241, + 330, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/33b1f46c820e1b3400dbfe15d59d3af41002023a6930210126cee8fd86796439.jpg", + "image_caption": [ + "(b)", + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 233, + 316, + 330, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a86983016e9228e092eba7224f62db1c39d22ee49145b8b8961bb7f3253deb5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 90, + 431, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a0eacc8d16e1ea0e5159b7149451ee06cf4f02c391257ac23903cc7ba215618.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 166, + 431, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9ec58cd0715b9fb0f03ee8d06e8c3b7ea79c822bf532480fbddaff1e9d24fef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 241, + 431, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9635bf5f7691c1ecb44c4d55c2d3f88607f6a712d63e2203c6119a60c5c6a1d0.jpg", + "image_caption": [ + "(c)", + "U-Net" + ], + "image_footnote": [], + "bbox": [ + 334, + 316, + 431, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c546fdd12524abdc307ba770590f52c533c69a9fbce658069503f8ba0d6b355c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 90, + 532, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e8e0d7f7f354a03cce6e8f392facf7583d63da51f173be597e105406ccf43bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 166, + 532, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c696812f6984170a616234ed6f45cdf7ed30aef9789aa2a4c77fc9159cd06427.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 241, + 532, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a4cbb4925647bab34b1217499e374e85f8a6a9ad8e7606b866a42d0521cbeae.jpg", + "image_caption": [ + "(d)", + "DIST" + ], + "image_footnote": [], + "bbox": [ + 436, + 316, + 532, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/77c8e5a2c67baed7d469b070827aff673b6d098e11b7dd5c4665373a542f8434.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 90, + 633, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ca768617ffed6e199573c093cf47f7e46ba29519c54ed06146b2c0082859e67a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 166, + 633, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4198c1409f1634d95b6ef460ee8c7e2e896dfe9b96496764acf57e2fbac4b4db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 241, + 633, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c862aa2c699d6a5f9791afb001fdc49850fe37b802bc9d99202c77615337796.jpg", + "image_caption": [ + "(e)", + "FullNet" + ], + "image_footnote": [], + "bbox": [ + 537, + 316, + 633, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c235508c663315e854bbf20342a31c0e3d18200666c0cf848c8dcc4c398d3a1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 90, + 733, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2305446ffebd45dcf7ef74e5a225275d19c371c6643297da4930f31863cb7418.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 166, + 733, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c98a02cf5d83536d7093b2a28e8c8fce871f01fdbf608bde6b7418eb49b79aef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 241, + 733, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8fccbf4d87eb125159654ab79847fd67d18784c8398e5473bb68990989bf3748.jpg", + "image_caption": [ + "(f)", + "HoverNet" + ], + "image_footnote": [], + "bbox": [ + 638, + 316, + 733, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fec93d31d352012b7d08a188da620dc3d461ecf2147f978316311c66bb7e8e90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 90, + 836, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc91497ccdb1b9158db1921f1a40996505e3125c6ca6f478f79e278cb3228410.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 166, + 836, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2cce860532ab71e7280f8d24c59d12eaaedb06f6eead5cfb512f34a26783dbd9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 241, + 836, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/943ca03733899e071318603229e7bc7759770684b09b126c4bad9fc805c1370c.jpg", + "image_caption": [ + "(g)", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 738, + 316, + 836, + 391 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f67019a6e7cee528e43d8300720ef48dbc42199e1f437f1d971417e81998e3ac.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
UNet [23]0.7780.7340.5780.6430.6350.6880.4360.508
DCAN [3]0.7320.7400.5450.5610.5630.6440.3630.435
DIST [16]0.6630.7540.5040.6160.7030.7280.5120.542
FullNet [20]--0.6610.6680.7120.7600.5410.562
HoverNet [6]0.8540.8140.6970.7050.7430.7640.5680.611
CDNet [7]---0.733---0.640
CellViT [10]0.8790.8100.7130.7370.7860.7540.5960.638
Ours0.8830.8080.7140.7400.7880.7560.5990.647
", + "bbox": [ + 78, + 469, + 475, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "NuInsSeg. Our Mudslide achieves 0.529 PQ and 0.581 AJI on the CPM17 dataset, and 0.456 PQ and 0.558 AJI on the NuInsSeg dataset, respectively. In the task of nuclear instance segmentation with only point annotations, our method can also achieve competitive results.", + "bbox": [ + 75, + 674, + 468, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although methods relying solely on point annotations still have ample room for improvement compared to fully supervised approaches, our approach has already surpassed other methods in the realm of point annotations.", + "bbox": [ + 75, + 752, + 468, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 828, + 194, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes a novel universal approach for nuclear instance segmentation, called Mudslide, which is inspired by the natural phenomenon of mudslides. To enhance", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a7902bba82866f8621ae791c56289d6a61513e05598eb855c691d77365180992.jpg", + "table_caption": [ + "Table 5. Performance comparisons on fully-supervised nuclear segmentation." + ], + "table_footnote": [], + "table_body": "
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
PENet [27]0.4400.7220.3190.2900.4760.6300.3000.355
C2FNet [25]0.4290.6300.2710.4210.4050.5800.2350.400
S2Label [11]0.6540.6980.4590.5030.6130.6670.4090.496
SPN+IEN [13]0.7040.6890.4880.5430.6880.6600.4540.523
SCNet [12]0.7160.6850.4960.5540.6940.6550.4560.526
Ours0.7470.7030.5290.5810.7070.6890.4870.558
", + "bbox": [ + 500, + 469, + 906, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Performance comparisons on point supervised nuclear segmentation.", + "bbox": [ + 500, + 606, + 890, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the model's capability to differentiate between adjacent instances, we devise a collapsing field, and generate a force map along with an initial boundary, offering distinct representations for different instances and a unified representation for the same instance. We also propose Mudslide collapse algorithm that mimics the natural process by segmenting through the iterative collapse of individual pixels. Our collapse fields are simple and easily extendable to point-supervised tasks. The experimental results on public nuclei datasets demonstrate that Mudslide outperforms state-of-the-art methods.", + "bbox": [ + 496, + 643, + 890, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work was supported in part by the National Key R&D Program of China (No. 2022ZD0118201), Natural Science Foundation of China (No.61972217, 32071459, 62176249, 62006133, 62271465), and Shenzhen Medical Research Funds in China (No. B2302037).", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11680", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Serge Beucher. Use of watersheds in contour detection. In Proc. Int. Workshop on Image Processing, Sept. 1979, pages 17-21, 1979. 6", + "[2] Alireza Chamanzar and Yao Nie. Weakly supervised multitask learning for cell detection and segmentation. In 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI), pages 513-516. IEEE, 2020. 1, 3", + "[3] Hao Chen, Xiaojuan Qi, Lequan Yu, Qi Dou, Jing Qin, and Pheng-Ann Heng. Dcan: Deep contour-aware networks for object instance segmentation from histology images. Medical Image Analysis, 36:135-146, 2017. 1, 2, 8", + "[4] Shengcong Chen, Changxing Ding, Minfeng Liu, Jun Cheng, and Dacheng Tao. CPP-net: Context-aware polygon proposal network for nucleus segmentation. IEEE Transactions on Image Processing, 32:980-994, 2023. 2", + "[5] FC Dai, Chin Fei Lee, and Y Yip Ngai. Landslide risk assessment and management: an overview. Engineering Geology, 64(1):65-87, 2002. 3", + "[6] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:1-15, 2019. 2, 4, 6, 8", + "[7] Hongliang He, Zhongyi Huang, Yao Ding, Guoli Song, Lin Wang, Qian Ren, Pengxu Wei, Zhiqiang Gao, and Jie Chen. Cdnet: Centripetal direction network for nuclear instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4026-4035, 2021. 2, 8", + "[8] Hongliang He, Chi Zhang, Jie Chen, Ruizhe Geng, Luyang Chen, Yongsheng Liang, Yanchang Lu, Jihua Wu, and Yongjie Xu. A hybrid-attention nested unet for nuclear segmentation in histopathological images. Frontiers in Molecular Biosciences, 8:1–9, 2021. 1, 2", + "[9] Hongliang He, Jun Wang, Pengxu Wei, Fan Xu, Xiangyang Ji, Chang Liu, and Jie Chen. Toposeg: Topology-aware nuclear instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21307-21316, 2023. 1, 2", + "[10] Fabian Horst, Moritz Rempe, Lukas Heine, Constantin Seibold, Julius Keyl, Giulia Baldini, Selma Ugurel, Jens Siveke, Barbara Grünwald, Jan Egger, et al. Cellvit: Vision transformers for precise cell segmentation and classification. arXiv preprint arXiv:2306.15350, 2023. 5, 8", + "[11] Hyeonsoo Lee and Won-Ki Jeong. Scribble2label: Scribblesupervised cell segmentation via self-generating pseudolabels with consistency. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23, pages 14-23. Springer, 2020. 8", + "[12] Yi Lin, Zhiyong Qu, Hao Chen, Zhongke Gao, Yuexiang Li, Lili Xia, Kai Ma, Yefeng Zheng, and Kwang-Ting Cheng. Nuclei segmentation with point annotations from pathology images via self-supervised learning and co-training. Medical Image Analysis, 89:102933, 2023. 3, 8" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Weizhen Liu, Qian He, and Xuming He. Weakly supervised nuclei segmentation via instance learning. In 2022 IEEE 19th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2022. 8", + "[14] Xiaoming Liu, Zhengsheng Guo, Jun Cao, and Jinshan Tang. Mdc-net: a new convolutional neural network for nucleus segmentation in histopathology images with distance maps and contour information. Computers in Biology and Medicine, 135:1-12, 2021. 2", + "[15] Amirreza Mahbod, Christine Polak, Katharina Feldmann, Rumsha Khan, Katharina Gelles, Georg Dorffner, Ramona Woitek, Sepideh Hatamikia, and Isabella Ellinger. Nuinsseg: A fully annotated dataset for nuclei instance segmentation in h&E-stained histological images. arXiv preprint arXiv:2308.01760, 2023. 5", + "[16] Peter Naylor, Marick Laé, Fabien Reyal, and Thomas Walter. Segmentation of nuclei in histopathology images by deep regression of the distance map. IEEE Transactions on Medical Imaging, 38(2):448-459, 2018. 2, 8", + "[17] Hasmukh Pranshanker Oza and HP Oza. Soil mechanics and foundation engineering. Charotar Book Stall, 1969. 3", + "[18] H. Pinckaers, W. Bulten, J. van der Laak, and G. Litjens. Detection of prostate cancer in whole-slide images through end-to-end training with image-level labels. IEEE Transactions on Medical Imaging, 40(7):1817-1826, 2021. 1", + "[19] Hui Qu, Pengxiang Wu, Qiaoying Huang, Jingru Yi, Gregory M Riedlinger, Subhajyoti De, and Dimitris N Metaxas. Weakly supervised deep nuclei segmentation using points annotation in histopathology images. In International Conference on Medical Imaging with Deep Learning, pages 390-400. PMLR, 2019. 1, 3, 5", + "[20] Hui Qu, Zhennan Yan, Gregory M Riedlinger, Subhajyoti De, and Dimitris N Metaxas. Improving nuclei/gland instance segmentation in histopathology images by full resolution neural network and spatial constrained loss. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 378-386, 2019. 1, 2, 8", + "[21] Hui Qu, Pengxiang Wu, Qiaoying Huang, Jingru Yi, Zhennan Yan, Kang Li, Gregory M Riedlinger, Subhajyoti De, Shaoting Zhang, and Dimitris N Metaxas. Weakly supervised deep nuclei segmentation using partial points annotation in histopathology images. IEEE Transactions on Medical Imaging, 39(11):3655-3666, 2020. 1, 3", + "[22] Shan E Ahmed Raza, Linda Cheung, Muhammad Shaban, Simon Graham, David Epstein, Stella Pelengaris, Michael Khan, and Nasir M Rajpoot. Micro-net: A unified model for segmentation of various objects in microscopy images. Medical Image Analysis, 52:160-173, 2019. 2", + "[23] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 1, 2, 8", + "[24] Digvijay P Salunkhe, Rupa N Bartakke, Guruprasd Chvan, and Pooja R Kothavale. An overview on methods for slope stability analysis. International Journal of Engineering Research & Technology (IJERT), 6(03):2278-0181, 2017. 3" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11681", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Kuan Tian, Jun Zhang, Haocheng Shen, Kezhou Yan, Pei Dong, Jianhua Yao, Shannon Che, Pifu Luo, and Xiao Han. Weakly-supervised nucleus segmentation based on point annotations: A coarse-to-fine self-stimulated learning strategy. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part V 23, pages 299-308. Springer, 2020. 3, 5, 8", + "[26] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in Bioengineering and Biotechnology, 53: 1-15, 2019. 5", + "[27] Inwan Yoo, Donggeun Yoo, and Kyunghyun Paeng. Pseudoedgenet: Nuclei segmentation only with point annotations. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part I 22, pages 731-739. Springer, 2019. 3, 8", + "[28] Ji Yu. Point-supervised single-cell segmentation via collaborative knowledge sharing. arXiv preprint arXiv:2304.10671, 2023. 3", + "[29] Yanning Zhou, Omer Fahri Onder, Qi Dou, Efstratios Tsougenis, Hao Chen, and Pheng-Ann Heng. Cia-net: Robust nuclei instance segmentation with contour-aware information aggregation. In International Conference on Information Processing in Medical Imaging, pages 682–693, 2019. 1, 2" + ], + "bbox": [ + 78, + 90, + 470, + 512 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11682", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/5a64aa1e-710a-4a50-b851-b34aaaa3cbbb_model.json b/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/5a64aa1e-710a-4a50-b851-b34aaaa3cbbb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d13069442f4f1510820f644cd49bfa28a16e3cd5 --- /dev/null +++ b/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/5a64aa1e-710a-4a50-b851-b34aaaa3cbbb_model.json @@ -0,0 +1,2376 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.804, + 0.152 + ], + "angle": 0, + "content": "Mudslide: A Universal Nuclear Instance Segmentation Method" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.182, + 0.823, + 0.234 + ], + "angle": 0, + "content": "Jun Wang \nSchool of Electronic and Computer Engineering, Peking University, Shenzhen, China \njunw@stu.pku.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.681 + ], + "angle": 0, + "content": "Nuclear instance segmentation has played a critical role in pathology image analysis. The main challenges arise from the difficulty in accurately segmenting densely overlapping instances and the high cost of precise mask-level annotations. Existing fully-supervised nuclear instance segmentation methods, such as boundary-based methods, struggle to capture differences between overlapping instances and thus fail in densely distributed blurry regions. They also face challenges transitioning to point supervision, where annotations are simple and effective. Inspired by natural mudslides, we propose a universal method called Mudslide that uses simple representations to characterize differences between different instances and can easily be extended from fully-supervised to point-supervised. Concretely, we introduce a collapse field and leverage it to construct a force map and initial boundary, enabling a distinctive representation for each instance. Each pixel is assigned a collapse force, with distinct directions between adjacent instances. Starting from the initial boundary, Mudslide executes a pixel-by-pixel collapse along various force directions. Pixels that collapse into the same region are considered as one instance, concurrently accounting for both inter-instance distinctions and intra-instance coherence. Experiments on public datasets show superior performance in both fully-supervised and point-supervised tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.724, + 0.21, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Pathological slide analysis is widely regarded as the gold standard for cancer diagnosis, treatment, and prevention. Nuclear instance segmentation is a critical step in this process, as it enables the accurate identification and analysis of nuclei within pathology images. The morphology of nuclei, including their shape, appearance, and density, provides interpretable features that can be used as diagnostic and prognostic indicators of cancer [18]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Nuclear instance segmentation presents unique challenges compared to natural image instance segmentation" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.269, + 0.891, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.446, + 0.895, + 0.545 + ], + "angle": 0, + "content": "Figure 1. An analogy between mudslide and nuclear instance segmentation. (Top) Topographic map before and after the mudslide. The colored arrows indicate the collapse forces in different directions. (Bottom) Nuclear instance segmentation process in pathological images. The different layers on the left represent different feature representations of nuclear instances, and the right shows the result of instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.894, + 0.644 + ], + "angle": 0, + "content": "tasks. Firstly, the densely packed distribution of nuclei often results in clustered overlapping instances. Secondly, the blurry boundaries between touching or overlapping nuclei make it difficult to distinguish individual instances. Third, obtaining precise pixel-level annotations is time-consuming and expensive for professional pathologists." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To address the first two challenges, many fully-supervised methods based on deep learning have made significant progress. Several algorithms based on constructing innovative network architectures [8, 20, 23] have been applied to this task. Some contour-based techniques have been investigated to enhance the accurate prediction of nucleus boundaries [3, 9, 29] which aids in splitting the touched and overlapped nuclei. However, these methods lack the capability to characterize the differences among neighboring nuclear instances, making it difficult to accurately segment densely stacked overlapping nuclei. Moreover, their applicability is hindered in scenarios where fine pixel-level annotations are insufficient, as obtaining precise boundary labels becomes a formidable task. Addressing the challenge posed by the scarcity of fine annotations, certain point annotation-based methods have emerged [2, 19, 21], leveraging the accessibility of coarse point labels. How" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11673" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "ever, these approaches frequently prioritize the acquisition of pseudo-mask labels through design optimization methods, neglecting the differentiated representation of overlapping instances." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.154, + 0.47, + 0.215 + ], + "angle": 0, + "content": "Inspired by the mudslide in nature, we propose a simple universal method, called Mudslide. This approach excels in providing a differential representation of instances while remaining readily adaptable to the point-supervised domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.469, + 0.457 + ], + "angle": 0, + "content": "As shown in Fig. 1, for a continuous geographical region, mudslides induced by external forces lead to the creation of distinct partitions within the area. The unaffected section sets the partition boundary, while the collapsing segment forms a contiguous region. Analogously, this parallels the process of nucleus instance segmentation, where a method partitions the foreground with diverse instances into distinct instance regions. Furthermore, the mudslide collapse process concurrently considers the distinctions between regions and the coherence within each region. Variations in force directions acting on different regions induce noticeable changes in adjacent areas. Internal forces within a region, specifically the collapse of one soil mass exerting force on neighboring soil masses, facilitate the clustering and rationalization of soil masses in the same area. This aligns with the objectives of nuclear instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.791 + ], + "angle": 0, + "content": "Therefore, drawing inspiration from the characteristics of the muslide collapse process, we introduce Mudslide algorithm to induce the collapse of the foreground region of the nucleus and acquire instances. Specifically, we formulate a collapse field and leverage it to construct a force map to allocate forces directed to the center of the instance to individual pixels within the collapse field. This allocation ensures that the force direction in adjacent regions of distinct nuclei exhibits pronounced disparities, facilitating the distinctive representation of neighboring nuclei. Simultaneously, the forces within the same instance exhibit uniformity directed towards the center, enhancing a cohesive representation within instances. Subsequently, we devise a collapse algorithm commencing from the collapse boundary, proceeding pixel-by-pixel along the force direction. Ultimately, pixels collapsing into the same region are deemed components of the same instance. Due to the consideration of external differences between instances and the internal consistency within instances, Mudslide excels in accurately segmenting overlapping instances. The simplicity of collapse field design renders this approach easily extendable to the domain of point supervision." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.794, + 0.422, + 0.808 + ], + "angle": 0, + "content": "The contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.47, + 0.87 + ], + "angle": 0, + "content": "- We propose Mudslide, a novel nucleus instance segmentation algorithm inspired by natural phenomenon of mudslide, which excels in achieving precise segmentation of adjacent or overlapping nuclei." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- We propose a collapse field and generate a force map along with an initial boundary to provide differentiated" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.811, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "representations for neighboring instances and enhance the internal coherence within instances." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.891, + 0.166 + ], + "angle": 0, + "content": "- Mudslide is a universal method that can be easily extended from full supervision tasks to point supervision domains." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.167, + 0.891, + 0.197 + ], + "angle": 0, + "content": "- We justify the superiority of our Mudslide on public nuclear datasets with state-of-the-art performance." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.122, + 0.891, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.214, + 0.65, + 0.23 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.24, + 0.821, + 0.256 + ], + "angle": 0, + "content": "2.1. Fully-supervised nuclei segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.264, + 0.892, + 0.505 + ], + "angle": 0, + "content": "Fully-supervised nuclear instance segmentation has been widely studied and benefited from the application of deep learning techniques [3, 7, 22]. One of the most influential models in this field is the U-Net model proposed by Ronneberger et al. [23], which has become a benchmark model for medical image segmentation and has inspired many subsequent works. Qu et al. [20] proposed a full-resolution convolutional neural network (FullNet), a full-resolution convolutional neural network that avoids downsampling operations in the network structure to improve localization accuracy. Raza et al. [22] proposed Micro-Net, which utilizes multi-resolution and weighted loss functions to achieve robustness against large internal/external variances in nucleus size. He et al. [8] proposed Han-Net, a hybrid attention nested U-shaped network that extracts informative features from multiple layers to enhance segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.688 + ], + "angle": 0, + "content": "Leveraging contour information for distinguishing interacting or overlapping nuclei, Chen et al. [3] introduced the integration of contour details into a multi-level Fully Convolutional Network (FCN), which resulted in the development of a profound contour-aware network designed specifically for nuclear instance segmentation. Following this, Zhou et al. [29] introduced the Contour-Aware Information Aggregation Network, which integrates spatial and textural features between nuclei and contours. Additionally, He et al. [9] proposed a Topology-Aware Loss to integrate meaningful topology constraints into the model optimization process for more precise boundary predictions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Furthermore, certain studies [4, 6, 14, 16] have employed distance maps to address the segmentation challenges posed by contacting or overlapping nuclei. Naylor et al. [16] tackled the segmentation of touching nuclei by framing it as a regression task involving intra-nuclear distance maps. Graham et al. [6] introduced the Hover-Net, which utilizes vertical and horizontal distances between a nucleus pixel and its center of mass to delineate nuclei. Additionally, He et al. [7] proposed the Centripetal Directional Network (CDNet) for nuclear instance segmentation, integrating directional information into the network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The above models have achieved performance improvements in nuclear segmentation. However, the challenge of segmenting overlapping cell nuclei persists. To further en" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11674" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.092, + 0.468, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.244, + 0.473, + 0.329 + ], + "angle": 0, + "content": "Figure 2. The simple modeling of natural mudslide. (Left) Schematic diagram of the force on a soil mass on the slope. The green line represents the slope and the red dashed line represents the soil mass. (Right) A mountain where the mudslide occurred. The mudslide starts at the position pointed by the force indicated by the red arrow and stops at the horizontal slope." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.471, + 0.4 + ], + "angle": 0, + "content": "hance the model's ability to distinguish instances, there is a need for a distinct and differential representation of nucleus instances." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.437, + 0.414, + 0.453 + ], + "angle": 0, + "content": "2.2. Pointly-supervised nuclei segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.473, + 0.818 + ], + "angle": 0, + "content": "The pixel-wise annotation of nucleus instances incurs high costs. Point annotation, as one of the most efficient labeling manners, has garnered increasing attention in the field of nuclear segmentation. Current approaches [2, 19, 21] primarily transform point annotations into coarse pixel-level labels, including Voronoi labels [25], cluster labels [19], and pseudo edge maps [27]. These methods commonly integrate prior knowledge about the shape and texture of nuclei into these coarse pixel-level labels. However, these techniques face challenges related to incomplete coarse labels and imprecise boundary information. Several strategies have been proposed to address the issues arising from these coarse labels. These include multi-stage optimization for refining segmentation in a bootstrapping manner [21, 25], and the incorporation of additional constraints based on the local contrast between nuclei and their surrounding cytoplasm. Additionally, some unique training strategies have been employed to optimize pseudo-labels. [12] utilizes the co-training method, optimizing mask labels through the simultaneous training of two models with mutual supervision. [28] employs a collaborative knowledge-sharing approach, sharing knowledge between a principal model and a very lightweight collaborator model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.9 + ], + "angle": 0, + "content": "Despite the progress in the field of making full use of point labels and generating accurate pseudo mask labels, however, the problem of nucleus overlapping instance segmentation in point-supervised prediction still needs further attention." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.889, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.257, + 0.895, + 0.328 + ], + "angle": 0, + "content": "Figure 3. The processing of Mudslide algorithm. (Left) The instance map, the four-channel collapse field, the eight-directional force map, and the initial boundary. Each color in the force map represents a direction. (Right) Schematic diagram of the mudslide collapse process performed on two adjacent nuclei." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.338, + 0.593, + 0.354 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.363, + 0.673, + 0.379 + ], + "angle": 0, + "content": "3.1. Natural Mudslide" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Mudslide refers to the natural phenomenon where the soil or rock mass on a slope, influenced by external forces such as river erosion, rainfall, and seismic events, undergoes downhill movement collectively along the slope [5]. To simulate the mudslide process, we simplify it as a continuous sliding problem of soil mass on a slope [24]. This assumes an arc-shaped sliding surface adhering to the Coulomb criterion during soil movement, and the soil undergoes rigid sliding. This allows the assessment of the stability of the soil slope based on the magnitudes of the sliding moment and anti-sliding moment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.629 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, for the soil mass on the slope, the sliding moment is generated by the gravitational force \\( W \\), while the resisting moment is provided by the frictional force on the sliding surface. The sliding moment \\( M_{s} \\) and resisting moment \\( M_{r} \\) are" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.633, + 0.891, + 0.648 + ], + "angle": 0, + "content": "\\[\nM _ {s} = W * d, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.648, + 0.892, + 0.682 + ], + "angle": 0, + "content": "\\[\nM _ {r} = R * \\int_ {0} ^ {l} \\tau d l, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.685, + 0.892, + 0.727 + ], + "angle": 0, + "content": "where \\(d\\) represents the lever arm of gravity, \\(R\\) and \\(l\\) are respectively the radius and the length of the arc, and \\(\\tau\\) represents the shear force. Introducing the safety factor \\(F_{os}\\) [17]" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.733, + 0.892, + 0.765 + ], + "angle": 0, + "content": "\\[\nF _ {o s} = \\frac {M _ {r}}{M _ {s}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.767, + 0.892, + 0.822 + ], + "angle": 0, + "content": "if \\( F_{os} > 1 \\), the soil mass remains stable and will not collapse. If \\( F_{os} = 1 \\), the soil mass is in a state of critical equilibrium, and if \\( F_{os} < 1 \\), the soil mass undergoes collapse." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.824, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Assuming each soil mass on the slope is in a critical state, i.e., \\( F_{os} = 1 \\), if an external force \\( F \\) with moment \\( M_F \\) is applied tangentially along the slope surface such that" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.872, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\nF _ {o s} ^ {\\prime} = \\frac {M _ {r}}{M _ {s} ^ {\\prime}} = \\frac {M _ {r}}{M _ {s} + M _ {F}} < F _ {o s}, \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11675" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.224 + ], + "angle": 0, + "content": "the corresponding soil mass undergoes collapse. This results in the application of a tangential force \\( F^{\\gamma} \\) on the adjacent soil mass below, leading to a continued collapse of the lower soil mass, ultimately forming a mudslide. The collapse terminates when it reaches a horizontal slope surface. In the end, the mudslide induces the collapse of soil mass on the slope surface affected by external forces, while the unaffected soil mass remains intact. This delineates distinct regions on the slope." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.246, + 0.295, + 0.261 + ], + "angle": 0, + "content": "3.2. Mudslide Segmentation" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.273, + 0.276, + 0.287 + ], + "angle": 0, + "content": "3.2.1 Mudslide Processing" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.304, + 0.471, + 0.456 + ], + "angle": 0, + "content": "Inspired by the natural mudslide mentioned in Sec. 3.1, we propose a Mudslide algorithm for nuclear instance segmentation. In natural landslide phenomena, the main issues include the characteristics of the slope that collapses, the forces acting on the soil mass on the slope, the location where the collapse begins, and how the collapse occurs continuously. Corresponding to this, the Mudslide algorithm primarily involves three steps: constructing a collapse field, obtaining the force map and initial collapse boundary, and executing the collapse." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.459, + 0.473, + 0.777 + ], + "angle": 0, + "content": "Collapse field construction. In the first step, we initiate the process by constructing a collapse field, representing the initial slope surface of the mudslide. In image segmentation, we assume that each pixel represents a soil mass, located on a slope represented by the collapse field, and in a state of critical equilibrium. For simplicity, we utilize continuous normalized maps to characterize the collapse field, where pixel values close to 1 signify higher slope heights and values close to 0 indicate lower slope surfaces. Specifically, for ease of model learning and inspired by [6], we create a four-channel normalized map shown in Fig. 3. This map contains four distance maps generated in horizontal, vertical, and two diagonal directions, respectively, with each map indicating the distance from each pixel to the center of the nucleus in different directions. Pixel values are close to 0 when they are near the nucleus center and gradually increase outward until they reach 1 at the nucleus boundary. Thus, we construct a slope for the foreground nuclei in the image, where the slope is flat in the middle of each instance and steep around the edges, resembling the terrain that is prone to mudslide collapse." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.781, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Force map and initial boundary obtaining. For the second step, we generate a force map and initial boundary from the collapse field. The force map assigns a force \\( F \\) with different directions to each pixel, representing the directions in which pixels are expected to collapse during Mudslide. For two-dimensional images, we consider eight directions of force: up, down, left, right, upper-left, upper-right, lower-left, and lower-right (Fig. 3):" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.088, + 0.892, + 0.158 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {F} = \\left\\{\\binom {0} {1}, \\binom {0} {- 1}, \\binom {- 1} {0}, \\binom {1} {0}, \\right. \\\\ \\binom {- 1} {1}, \\binom {1} {1}, \\binom {- 1} {- 1}, \\binom {1} {- 1} \\}. \\end{array} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.167, + 0.868, + 0.182 + ], + "angle": 0, + "content": "Then the force map can be expressed as a set of forces:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.186, + 0.892, + 0.203 + ], + "angle": 0, + "content": "\\[\nF M = \\{F (x, y) \\mid x, y \\in \\mathbb {Z}, F \\in \\mathbb {F} \\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.214, + 0.892, + 0.318 + ], + "angle": 0, + "content": "where \\(\\mathbb{Z}\\) indicates the foreground of nuclei and \\(x, y\\) corresponds to the coordinate of each pixel. Here, we assign a force direction to each pixel that is approximately tangential to the slope. Since we assume that each pixel is in a state of critical equilibrium, we consider the force magnitude to be irrelevant to the collapse and therefore set the magnitude of each force to 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.893, + 0.44 + ], + "angle": 0, + "content": "To obtain the starting position of mudslide collapse, we generate the initial boundary. We employ a straightforward edge detection filter, such as Sobel operator, on the collapse field to detect edges. The obtained edges are then utilized as the initial boundary for mudslide collapse. This approach is rational since, in our collapse field, there is a significant gradient between instance boundaries, allowing the Sobel operator to effectively identify them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.516 + ], + "angle": 0, + "content": "Mudslide collapse executing. After acquiring the force map and the initial boundary, we commence the mudslide collapse process. Originating from the initial boundary, each pixel \\( p \\) applies an initial force (red arrows in 3) to the directed pixel \\( p' \\), which can be calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.52, + 0.749, + 0.539 + ], + "angle": 0, + "content": "\\[\np ^ {\\prime} = p + F (p),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.542, + 0.892, + 0.558 + ], + "angle": 0, + "content": "\\[\np = (x, y), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.56, + 0.735, + 0.58 + ], + "angle": 0, + "content": "\\[\np ^ {'} = (x ^ {'}, y ^ {'}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.753 + ], + "angle": 0, + "content": "The pixel \\( p' \\) then undergoes collapse. Subsequently, the collapsed pixel exerts a sustained force (yellow arrows in 3) on the pointed pixel, initiating the collapse of more adjacent pixels. This iterative process results in a sustained mudslide collapse. Eventually, when all pixels directed to have undergone collapse, those pixels collapsing to the same region are identified as a single instance. Note that those pixels that are never directed by forces from other pixels will not collapse and the force assigned to them will be neglected (gray arrows in 3), corresponding to the soil mass that is not affected by the mudslide." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.772, + 0.875, + 0.787 + ], + "angle": 0, + "content": "3.2.2 Application to Nuclear Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.87 + ], + "angle": 0, + "content": "The framework of applying Mudslide to nuclear instance segmentation is illustrated in Fig. 4. Here, we unify the pipelines for both fully supervised and point-supervised scenarios, with the primary distinction lying in the labels used for training." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Specifically, we employ a dual-branch architecture, where one branch is dedicated to predicting a binary map" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11676" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.093, + 0.844, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.274, + 0.844, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.459, + 0.892, + 0.503 + ], + "angle": 0, + "content": "Figure 4. The overall framework of the proposed Mudslide. The network contains two basic prediction branches: the binary branch and the collapse field branch. During training, the difference between fully- and point-supervised is the type of ground truth. During testing, both branches' predictions are employed for the Mudslide algorithm." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.47, + 0.735 + ], + "angle": 0, + "content": "for the nuclear foreground, and the other branch focuses on predicting the collapse field. The collapse field consists of four channels, corresponding to horizontal, vertical, and two diagonal directions. The segmentation model adopted is the Trans-UNet architecture from [10]. In fully supervised training, the binary map labels are the ground truth representing the nuclear foreground, while the collapse field labels are constructed from both the foreground and instance ground truth. For point-supervised training, we leverage point labels and the original input image to generate a cluster map [19] and Voronoi map [25] as pseudolabels for foreground prediction, which is commonly observed in point-supervised tasks. Subsequently, we use the predicted foreground and point labels to generate collapse field labels." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.737, + 0.47, + 0.782 + ], + "angle": 0, + "content": "We utilize cross-entropy loss as the binary classification loss and employ mean squared error loss and mean squared gradient error loss as the collapse field loss:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.79, + 0.47, + 0.807 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {B} = \\lambda_ {B _ {C E}} \\mathcal {L} _ {C E}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.814, + 0.469, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C F} = \\lambda_ {C F _ {M S E}} \\mathcal {L} _ {M S E} + \\lambda_ {C F _ {M S G E}} \\mathcal {L} _ {M S G E}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.836, + 0.296, + 0.85 + ], + "angle": 0, + "content": "Then the overall loss function is" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.859, + 0.469, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {B} + \\mathcal {L} _ {C F}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "It's worth noting that in point-supervised training, the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.51, + 0.892, + 0.54 + ], + "angle": 0, + "content": "cross-entropy loss comprises two terms, corresponding to cluster and Voronoi:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.553, + 0.892, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {B _ {\\text {p o i n t}}} = \\lambda_ {B _ {C E}} \\mathcal {L} _ {C E _ {\\text {c l u s t e r}}} + \\lambda_ {B _ {C E}} \\mathcal {L} _ {C E _ {\\text {v o r o n o i}}}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.703 + ], + "angle": 0, + "content": "During the inference process, we first combine the foreground and collapse field predictions to construct the collapse field, ensuring subsequent mudslide acts on the nuclear foreground regions. Subsequently, as detailed in Sec. 3.1, we utilize the collapse field to obtain the initial boundary and force map, initiating the mudslide collapse process. Ultimately, this process yields the nuclear instance segmentation results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.625, + 0.734 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.707, + 0.759 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets. I) CPM17 [26] contains 64 H&E stained histopathology images with 7,570 annotated nuclear boundaries. It is from the MICCAI 2017 Digital Pathology Challenge [26] and images on two different scales: \\(500 \\times 500\\) and \\(600 \\times 600\\). This is a dataset for a single-class segmentation task. II) NuInsSeg [15] contains 665 image patches with more than 30,000 manually segmented nuclei from 31 human and mouse organs. It is one of the biggest fully manually annotated datasets of nuclei in Hematoxylin and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11677" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.135 + ], + "angle": 0, + "content": "Eosin (H &E) stained histological images with the scale of \\(512 \\times 512\\). This is also a dataset for single-class segmentation task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.213 + ], + "angle": 0, + "content": "Evaluation Metrics. We use four instance-level evaluation metrics to measure the instance segmentation performance of the comparison models, which are: Detection Quality (DQ), Segmentation Quality (SQ), Panoptic Quality (PQ), and the Aggregated Jaccard Index (AJI)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.214, + 0.468, + 0.259 + ], + "angle": 0, + "content": "Implementation details. The proposed Mudslide is trained and tested using the open-source software library Pytorch 1.8.0 on NVIDIA Tesla V100 with CUDA 10.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.468, + 0.488 + ], + "angle": 0, + "content": "In order to build the dataset used for training, the CPM17 dataset is divided into multiple patches as it has an inconsistent original image size. We generate 512 training patches in the size of \\(256 \\times 256\\), thus allowing the network to be uniform and fully trained. In addition, we randomly divide NuInsSeg into training set and test set with the proportion of \\(80\\%\\) and \\(20\\%\\), respectively. Moreover, for point-supervised tasks, we leverage the annotations of nucleus instances to generate a central point for each instance, serving as the point label. In the training phase, we apply a series of data augmentation methods, which include random flip, random rotation, random color, random crop, and Gaussian blur. The training epoch is set as 300, and the initial learning rate is set as \\(10^{-3}\\). The weights of each loss are \\(\\lambda_{BCE} = 1\\), \\(\\lambda_{CFMSE} = 2.5\\), \\(\\lambda_{CFMSGE} = 8\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.497, + 0.229, + 0.513 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.468, + 0.581 + ], + "angle": 0, + "content": "We performed ablation studies on the CPM17 and NuInsSeg datasets to validate the efficacy of the proposed Mudslide method. All the following experiments were conducted under the setting of fully supervised tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.468, + 0.687 + ], + "angle": 0, + "content": "Instance separation methods. Table 1 shows the comparison of segmentation performance obtained using different instance separation methods. We apply four different methods to the predictions of the same model: the traditional watershed method [1], boundary partition method, HVmap segmentation method from Hover-Net [6], and our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.468, + 0.824 + ], + "angle": 0, + "content": "Specifically, the watershed method involves segmenting instances using the watershed algorithm on the binary probability map predicted by the model. The boundary separation method refers to predicting boundaries by adding an additional boundary prediction branch to the model and using the boundaries to directly separate different instances. The HVmap algorithm is an instance separation method in HoverNet that utilizes horizontal and vertical distance maps for segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The experimental results indicate significant variations in model performance based on different instance separation methods. Traditional methods such as watershed and boundary separation exhibit subpar performance. The HVmap approach, utilizing horizontal and vertical maps," + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.091, + 0.887, + 0.189 + ], + "angle": 0, + "content": "
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
Watershed0.7970.7300.5820.6940.6010.7350.4610.502
Boundary0.8370.7340.6140.7140.6400.7530.4910.532
HVMap0.8790.8100.7130.7370.7860.7540.5960.638
Ours0.8830.8080.7140.7400.7880.7560.5990.647
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.198, + 0.892, + 0.239 + ], + "angle": 0, + "content": "Table 1. Comparison of different adjacent instance separation methods, including the watershed algorithm, boundary partition, HVmap and our method." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.245, + 0.9, + 0.327 + ], + "angle": 0, + "content": "
boundcollapseCPM17NuInsSeg
DQ ↑SQ ↑PQ ↑AJI ↑DQ ↑SQ ↑PQ ↑AJI ↑
✓0.7720.7100.5480.6100.5870.7030.4140.497
0.8770.8070.7070.7330.7750.7520.5820.630
✓✓0.8830.8080.7140.7400.7880.7560.5990.647
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.336, + 0.892, + 0.378 + ], + "angle": 0, + "content": "Table 2. Ablation studies of components in Mudslide algorithm, where bound indicates initial boundary, and collapse indicates mudslide collapse processing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.385, + 0.892, + 0.461 + ], + "angle": 0, + "content": "enhances the differentiation between instances and achieves better results. Furthermore, our method, with more explicit instance differentiation representation and a more rational instance acquisition process, outperforms other methods. This underscores the superiority of our Mudslide approach." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.462, + 0.892, + 0.673 + ], + "angle": 0, + "content": "The ablation of components of Mudslide algorithm. We conducted ablation experiments on the key steps of the mudslide algorithm. Here, we considered two components: whether to use the initial boundary and whether to perform mudslide collapse. In Table 2, we explored three comparative experiments: a baseline approach with no additional components, relying solely on binary map partitioning based on connected regions; a scenario involving only the initial boundary extracted from the collapse field; and a comprehensive application of both the initial boundary and mudslide collapse, representing our full-fledged mudslide algorithm for instance segmentation. These experiments aim to scrutinize the individual contributions of each component to the overall segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.674, + 0.892, + 0.809 + ], + "angle": 0, + "content": "The experimental results indicate that incorporating an initial boundary leads to performance gains of \\(16\\%\\) PQ and \\(12\\%\\) AJI for the CPM dataset and \\(17\\%\\) PQ and \\(14\\%\\) AJI for NuInsSeg. Furthermore, the additional mudslide collapse results in a performance boost of \\(7\\%\\) PQ and \\(7\\%\\) AJI for the CPM dataset and \\(16\\%\\) PQ and \\(17\\%\\) AJI for NuInsSeg. These findings underscore the substantial benefits of each component within the mudslide method for enhancing instance segmentation outcomes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Force map construction methods. As a crucial manifestation of distinct instance differentiation, the construction of the force map is pivotal in the Mudslide algorithm. Therefore, we consider three methods for building the force map: direct prediction, collapse field segmentation, and collapse field extraction." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11678" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.173 + ], + "angle": 0, + "content": "
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
predict0.8740.8070.7070.7350.7770.7510.5870.633
segment0.8640.8080.7010.7280.7720.7530.5820.629
extract0.8800.8070.7110.7390.7840.7550.5920.643
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.184, + 0.47, + 0.211 + ], + "angle": 0, + "content": "Table 3. Comparison of different force map construction methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.219, + 0.471, + 0.446 + ], + "angle": 0, + "content": "In the first approach, we have the model predict the force map directly. Specifically, we augment the model with a branch that outputs nine channels, allowing it to learn the force map's eight directions and the background region. Subsequently, we obtain the force map by employing the argmax function to the predicted probability map. In the second approach, we perform segmentation on the collapse field. Here, we use a two-channel collapse field corresponding to the horizontal and vertical directions. By segmenting the collapse field based on angles, we categorize it into eight regions, representing the eight directions of the force map. In the third approach, we predict a four-channel collapse field for horizontal, vertical, and two diagonal directions. Then, we directly employ argmax to extract the direction of force that each pixel should possess." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.536 + ], + "angle": 0, + "content": "The results presented in Table 3 indicate that the force map extracted through four-channel collapse fields performs the best. The direct prediction follows with slightly lower performance, while the force map obtained by splitting the two-channel collapse field performs the least favorably." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.703 + ], + "angle": 0, + "content": "The four-channel collapsing field extraction method outperformed direct prediction by \\(4\\%\\) PQ and \\(4\\%\\) AJI on CPM17, and outperformed the two-channel segmentation method by \\(6\\%\\) PQ, and \\(11\\%\\) AJI. On the NuInsSeg dataset, it outperformed direct prediction by \\(5\\%\\) PQ, and \\(1\\%\\) AJI, and outperformed the two-channel segmentation method by \\(10\\%\\) PQ, and \\(14\\%\\) AJI. These results suggest that the four-channel collapsing field can better capture and characterize the differences between cell nucleus instances, resulting in greater differences between adjacent instances and better instance segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Initial boundary generation methods. Regarding the issue of generating and selecting initial boundaries, we considered two methods: direct generation from the binary prediction map and generation from the collapsing field." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Specifically, for the first method, we first obtained the binary probability map predicted by the model's binary classification branch and then selected a threshold of 0.5 to segment the predicted foreground and background regions of the nucleus. Then, we used the part of the pixels that contacted the background, i.e., the edge of the foreground, as the initial boundary. For the second method, we first obtained the four-channel collapsing field predicted by the model and then used some edge extraction methods, such as" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.093, + 0.897, + 0.16 + ], + "angle": 0, + "content": "
MethodCPMNuInsSeg
DQ ↑SQ ↑PQ ↑AJI ↑DQ ↑SQ ↑PQ ↑AJI ↑
binary map collapse field0.8800.8070.7110.7390.7840.7550.5920.643
0.8830.8080.7140.7400.7880.7560.5990.647
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.169, + 0.892, + 0.196 + ], + "angle": 0, + "content": "Table 4. Comparison of different initial boundary generation methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.204, + 0.892, + 0.234 + ], + "angle": 0, + "content": "using the Sobel operator to extract the regions with higher gradients in the collapsing field as the initial boundary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.893, + 0.506 + ], + "angle": 0, + "content": "As shown in Table 4, the method of generating initial boundaries using the collapsing field outperforms the method of using only the binary prediction map. Specifically, on the CPM dataset, the collapsing field method improved performance by \\(3\\%\\) PQ, and \\(1\\%\\) AJI compared to the binary map method. On the NuInsSeg dataset, it improved performance by \\(7\\%\\) PQ, and \\(4\\%\\) AJI. As anticipated, the construction of the initial boundary is crucial in our Mudslide collapse algorithm, as it determines the algorithm's starting position. Obtaining foreground edges directly from a binary image makes it challenging to capture the boundaries between adjacent instances. Although the Mudslide algorithm can capture this difference to some extent, its performance is not as effective as constructing a better initial boundary. The initial boundary extracted through the collapse field takes into account the differentiated representations between cell nuclei, effectively separating nuclei to a certain degree and proving to be a superior boundary." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.514, + 0.831, + 0.53 + ], + "angle": 0, + "content": "4.3. Fully-Supervised Nuclei Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.643 + ], + "angle": 0, + "content": "In Table 5, the proposed Mudslide is compared with the state-of-the-art fully-supervised nuclear instance segmentation models on CPM17 and NuInsSeg. Our Mudslide achieves 0.714 PQ and 0.740 AJI on the CPM17 dataset, and 0.599 PQ and 0.647 AJI on the NuInsSeg dataset, respectively. The Mudslide achieves competitive results for fully-supervised nuclei segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.809 + ], + "angle": 0, + "content": "We further carry out a qualitative visualization analysis of our method and different existing methods. Compared to all the other methods in Fig. 5, our Mudslide achieves superior instance segmentation results for nuclear instance segmentation, especially in some regions where nuclei are densely distributed with a large amount of overlap, where our instance segmentation is more accurate. The results, both quantitative and qualitative, unequivocally establish the competitiveness of our Mudslide algorithm in accurately segmenting densely packed and overlapping nuclear instances." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.818, + 0.845, + 0.834 + ], + "angle": 0, + "content": "4.4. Pointly-Supervised Nuclei Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Table 6, we extend the proposed Mudslide from fully-supervised to point-supervised nuclear segmentation task and compare it with the state-of-the-art point-supervised nuclear instance segmentation models on CPM17 and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11679" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.09, + 0.23, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.167, + 0.228, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.242, + 0.228, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.318, + 0.228, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.394, + 0.187, + 0.405 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.408, + 0.197, + 0.421 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.091, + 0.331, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.167, + 0.331, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.242, + 0.331, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.318, + 0.331, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.394, + 0.294, + 0.405 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.408, + 0.328, + 0.42 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.091, + 0.432, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.167, + 0.432, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.242, + 0.432, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.318, + 0.432, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.394, + 0.391, + 0.405 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.408, + 0.407, + 0.42 + ], + "angle": 0, + "content": "U-Net" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.091, + 0.533, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.167, + 0.533, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.242, + 0.533, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.318, + 0.533, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.478, + 0.394, + 0.496, + 0.405 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.408, + 0.506, + 0.42 + ], + "angle": 0, + "content": "DIST" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.091, + 0.635, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.167, + 0.635, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.242, + 0.635, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.318, + 0.635, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.394, + 0.599, + 0.405 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.408, + 0.619, + 0.42 + ], + "angle": 0, + "content": "FullNet" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.091, + 0.735, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.167, + 0.735, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.242, + 0.735, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.318, + 0.735, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.394, + 0.693, + 0.405 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.408, + 0.719, + 0.42 + ], + "angle": 0, + "content": "HoverNet" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.091, + 0.837, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.167, + 0.837, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.242, + 0.837, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.318, + 0.837, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.394, + 0.797, + 0.405 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.408, + 0.806, + 0.42 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.432, + 0.892, + 0.461 + ], + "angle": 0, + "content": "Figure 5. Visualization of instance segmentation results with different methods. (a) Input image; (b) Ground Truth; (c) U-Net [23]; (d) DIST [16]; (e) FullNet [20]; (f) Hover-Net [6]; (g) Our Mudslide." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.47, + 0.477, + 0.628 + ], + "angle": 0, + "content": "
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
UNet [23]0.7780.7340.5780.6430.6350.6880.4360.508
DCAN [3]0.7320.7400.5450.5610.5630.6440.3630.435
DIST [16]0.6630.7540.5040.6160.7030.7280.5120.542
FullNet [20]--0.6610.6680.7120.7600.5410.562
HoverNet [6]0.8540.8140.6970.7050.7430.7640.5680.611
CDNet [7]---0.733---0.640
CellViT [10]0.8790.8100.7130.7370.7860.7540.5960.638
Ours0.8830.8080.7140.7400.7880.7560.5990.647
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.637, + 0.47, + 0.666 + ], + "angle": 0, + "content": "Table 5. Performance comparisons on fully-supervised nuclear segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.752 + ], + "angle": 0, + "content": "NuInsSeg. Our Mudslide achieves 0.529 PQ and 0.581 AJI on the CPM17 dataset, and 0.456 PQ and 0.558 AJI on the NuInsSeg dataset, respectively. In the task of nuclear instance segmentation with only point annotations, our method can also achieve competitive results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.753, + 0.47, + 0.813 + ], + "angle": 0, + "content": "Although methods relying solely on point annotations still have ample room for improvement compared to fully supervised approaches, our approach has already surpassed other methods in the realm of point annotations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.829, + 0.196, + 0.845 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "This paper proposes a novel universal approach for nuclear instance segmentation, called Mudslide, which is inspired by the natural phenomenon of mudslides. To enhance" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.47, + 0.908, + 0.598 + ], + "angle": 0, + "content": "
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
PENet [27]0.4400.7220.3190.2900.4760.6300.3000.355
C2FNet [25]0.4290.6300.2710.4210.4050.5800.2350.400
S2Label [11]0.6540.6980.4590.5030.6130.6670.4090.496
SPN+IEN [13]0.7040.6890.4880.5430.6880.6600.4540.523
SCNet [12]0.7160.6850.4960.5540.6940.6550.4560.526
Ours0.7470.7030.5290.5810.7070.6890.4870.558
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.607, + 0.892, + 0.636 + ], + "angle": 0, + "content": "Table 6. Performance comparisons on point supervised nuclear segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.809 + ], + "angle": 0, + "content": "the model's capability to differentiate between adjacent instances, we devise a collapsing field, and generate a force map along with an initial boundary, offering distinct representations for different instances and a unified representation for the same instance. We also propose Mudslide collapse algorithm that mimics the natural process by segmenting through the iterative collapse of individual pixels. Our collapse fields are simple and easily extendable to point-supervised tasks. The experimental results on public nuclei datasets demonstrate that Mudslide outperforms state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments This work was supported in part by the National Key R&D Program of China (No. 2022ZD0118201), Natural Science Foundation of China (No.61972217, 32071459, 62176249, 62006133, 62271465), and Shenzhen Medical Research Funds in China (No. B2302037)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11680" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Serge Beucher. Use of watersheds in contour detection. In Proc. Int. Workshop on Image Processing, Sept. 1979, pages 17-21, 1979. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.214 + ], + "angle": 0, + "content": "[2] Alireza Chamanzar and Yao Nie. Weakly supervised multitask learning for cell detection and segmentation. In 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI), pages 513-516. IEEE, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.47, + 0.272 + ], + "angle": 0, + "content": "[3] Hao Chen, Xiaojuan Qi, Lequan Yu, Qi Dou, Jing Qin, and Pheng-Ann Heng. Dcan: Deep contour-aware networks for object instance segmentation from histology images. Medical Image Analysis, 36:135-146, 2017. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.47, + 0.33 + ], + "angle": 0, + "content": "[4] Shengcong Chen, Changxing Ding, Minfeng Liu, Jun Cheng, and Dacheng Tao. CPP-net: Context-aware polygon proposal network for nucleus segmentation. IEEE Transactions on Image Processing, 32:980-994, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.332, + 0.47, + 0.374 + ], + "angle": 0, + "content": "[5] FC Dai, Chin Fei Lee, and Y Yip Ngai. Landslide risk assessment and management: an overview. Engineering Geology, 64(1):65-87, 2002. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.376, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[6] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:1-15, 2019. 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.447, + 0.47, + 0.529 + ], + "angle": 0, + "content": "[7] Hongliang He, Zhongyi Huang, Yao Ding, Guoli Song, Lin Wang, Qian Ren, Pengxu Wei, Zhiqiang Gao, and Jie Chen. Cdnet: Centripetal direction network for nuclear instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4026-4035, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.532, + 0.47, + 0.601 + ], + "angle": 0, + "content": "[8] Hongliang He, Chi Zhang, Jie Chen, Ruizhe Geng, Luyang Chen, Yongsheng Liang, Yanchang Lu, Jihua Wu, and Yongjie Xu. A hybrid-attention nested unet for nuclear segmentation in histopathological images. Frontiers in Molecular Biosciences, 8:1–9, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.604, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[9] Hongliang He, Jun Wang, Pengxu Wei, Fan Xu, Xiangyang Ji, Chang Liu, and Jie Chen. Toposeg: Topology-aware nuclear instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21307-21316, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.676, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[10] Fabian Horst, Moritz Rempe, Lukas Heine, Constantin Seibold, Julius Keyl, Giulia Baldini, Selma Ugurel, Jens Siveke, Barbara Grünwald, Jan Egger, et al. Cellvit: Vision transformers for precise cell segmentation and classification. arXiv preprint arXiv:2306.15350, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.747, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[11] Hyeonsoo Lee and Won-Ki Jeong. Scribble2label: Scribblesupervised cell segmentation via self-generating pseudolabels with consistency. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23, pages 14-23. Springer, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Yi Lin, Zhiyong Qu, Hao Chen, Zhongke Gao, Yuexiang Li, Lili Xia, Kai Ma, Yefeng Zheng, and Kwang-Ting Cheng. Nuclei segmentation with point annotations from pathology images via self-supervised learning and co-training. Medical Image Analysis, 89:102933, 2023. 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[13] Weizhen Liu, Qian He, and Xuming He. Weakly supervised nuclei segmentation via instance learning. In 2022 IEEE 19th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[14] Xiaoming Liu, Zhengsheng Guo, Jun Cao, and Jinshan Tang. Mdc-net: a new convolutional neural network for nucleus segmentation in histopathology images with distance maps and contour information. Computers in Biology and Medicine, 135:1-12, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.894, + 0.3 + ], + "angle": 0, + "content": "[15] Amirreza Mahbod, Christine Polak, Katharina Feldmann, Rumsha Khan, Katharina Gelles, Georg Dorffner, Ramona Woitek, Sepideh Hatamikia, and Isabella Ellinger. Nuinsseg: A fully annotated dataset for nuclei instance segmentation in h&E-stained histological images. arXiv preprint arXiv:2308.01760, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.894, + 0.357 + ], + "angle": 0, + "content": "[16] Peter Naylor, Marick Laé, Fabien Reyal, and Thomas Walter. Segmentation of nuclei in histopathology images by deep regression of the distance map. IEEE Transactions on Medical Imaging, 38(2):448-459, 2018. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.894, + 0.385 + ], + "angle": 0, + "content": "[17] Hasmukh Pranshanker Oza and HP Oza. Soil mechanics and foundation engineering. Charotar Book Stall, 1969. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[18] H. Pinckaers, W. Bulten, J. van der Laak, and G. Litjens. Detection of prostate cancer in whole-slide images through end-to-end training with image-level labels. IEEE Transactions on Medical Imaging, 40(7):1817-1826, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[19] Hui Qu, Pengxiang Wu, Qiaoying Huang, Jingru Yi, Gregory M Riedlinger, Subhajyoti De, and Dimitris N Metaxas. Weakly supervised deep nuclei segmentation using points annotation in histopathology images. In International Conference on Medical Imaging with Deep Learning, pages 390-400. PMLR, 2019. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[20] Hui Qu, Zhennan Yan, Gregory M Riedlinger, Subhajyoti De, and Dimitris N Metaxas. Improving nuclei/gland instance segmentation in histopathology images by full resolution neural network and spatial constrained loss. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 378-386, 2019. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[21] Hui Qu, Pengxiang Wu, Qiaoying Huang, Jingru Yi, Zhennan Yan, Kang Li, Gregory M Riedlinger, Subhajyoti De, Shaoting Zhang, and Dimitris N Metaxas. Weakly supervised deep nuclei segmentation using partial points annotation in histopathology images. IEEE Transactions on Medical Imaging, 39(11):3655-3666, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[22] Shan E Ahmed Raza, Linda Cheung, Muhammad Shaban, Simon Graham, David Epstein, Stella Pelengaris, Michael Khan, and Nasir M Rajpoot. Micro-net: A unified model for segmentation of various objects in microscopy images. Medical Image Analysis, 52:160-173, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[23] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Digvijay P Salunkhe, Rupa N Bartakke, Guruprasd Chvan, and Pooja R Kothavale. An overview on methods for slope stability analysis. International Journal of Engineering Research & Technology (IJERT), 6(03):2278-0181, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11681" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[25] Kuan Tian, Jun Zhang, Haocheng Shen, Kezhou Yan, Pei Dong, Jianhua Yao, Shannon Che, Pifu Luo, and Xiao Han. Weakly-supervised nucleus segmentation based on point annotations: A coarse-to-fine self-stimulated learning strategy. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part V 23, pages 299-308. Springer, 2020. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.301 + ], + "angle": 0, + "content": "[26] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in Bioengineering and Biotechnology, 53: 1-15, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.471, + 0.386 + ], + "angle": 0, + "content": "[27] Inwan Yoo, Donggeun Yoo, and Kyunghyun Paeng. Pseudoedgenet: Nuclei segmentation only with point annotations. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part I 22, pages 731-739. Springer, 2019. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.387, + 0.47, + 0.428 + ], + "angle": 0, + "content": "[28] Ji Yu. Point-supervised single-cell segmentation via collaborative knowledge sharing. arXiv preprint arXiv:2304.10671, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.43, + 0.47, + 0.513 + ], + "angle": 0, + "content": "[29] Yanning Zhou, Omer Fahri Onder, Qi Dou, Efstratios Tsougenis, Hao Chen, and Pheng-Ann Heng. Cia-net: Robust nuclei instance segmentation with contour-aware information aggregation. In International Conference on Information Processing in Medical Imaging, pages 682–693, 2019. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11682" + } + ] +] \ No newline at end of file diff --git a/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/5a64aa1e-710a-4a50-b851-b34aaaa3cbbb_origin.pdf b/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/5a64aa1e-710a-4a50-b851-b34aaaa3cbbb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..af0f6bb8e7f505853d94b451c986b4c40d74e667 --- /dev/null +++ b/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/5a64aa1e-710a-4a50-b851-b34aaaa3cbbb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d246085527657a3c6ada9814d3819d6e419c9bbc5f9f051290e87763073ec33e +size 3398760 diff --git a/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/full.md b/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b4660dee75955108ee81755f4b953b6547d82e23 --- /dev/null +++ b/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/full.md @@ -0,0 +1,374 @@ +# Mudslide: A Universal Nuclear Instance Segmentation Method + +Jun Wang +School of Electronic and Computer Engineering, Peking University, Shenzhen, China +junw@stu.pku.edu.cn + +# Abstract + +Nuclear instance segmentation has played a critical role in pathology image analysis. The main challenges arise from the difficulty in accurately segmenting densely overlapping instances and the high cost of precise mask-level annotations. Existing fully-supervised nuclear instance segmentation methods, such as boundary-based methods, struggle to capture differences between overlapping instances and thus fail in densely distributed blurry regions. They also face challenges transitioning to point supervision, where annotations are simple and effective. Inspired by natural mudslides, we propose a universal method called Mudslide that uses simple representations to characterize differences between different instances and can easily be extended from fully-supervised to point-supervised. Concretely, we introduce a collapse field and leverage it to construct a force map and initial boundary, enabling a distinctive representation for each instance. Each pixel is assigned a collapse force, with distinct directions between adjacent instances. Starting from the initial boundary, Mudslide executes a pixel-by-pixel collapse along various force directions. Pixels that collapse into the same region are considered as one instance, concurrently accounting for both inter-instance distinctions and intra-instance coherence. Experiments on public datasets show superior performance in both fully-supervised and point-supervised tasks. + +# 1. Introduction + +Pathological slide analysis is widely regarded as the gold standard for cancer diagnosis, treatment, and prevention. Nuclear instance segmentation is a critical step in this process, as it enables the accurate identification and analysis of nuclei within pathology images. The morphology of nuclei, including their shape, appearance, and density, provides interpretable features that can be used as diagnostic and prognostic indicators of cancer [18]. + +Nuclear instance segmentation presents unique challenges compared to natural image instance segmentation + +![](images/e82e54116028c73b7f39769e95bdea5409bb493bc3187d6105305b306f6c5b33.jpg) +Figure 1. An analogy between mudslide and nuclear instance segmentation. (Top) Topographic map before and after the mudslide. The colored arrows indicate the collapse forces in different directions. (Bottom) Nuclear instance segmentation process in pathological images. The different layers on the left represent different feature representations of nuclear instances, and the right shows the result of instance segmentation. + +tasks. Firstly, the densely packed distribution of nuclei often results in clustered overlapping instances. Secondly, the blurry boundaries between touching or overlapping nuclei make it difficult to distinguish individual instances. Third, obtaining precise pixel-level annotations is time-consuming and expensive for professional pathologists. + +To address the first two challenges, many fully-supervised methods based on deep learning have made significant progress. Several algorithms based on constructing innovative network architectures [8, 20, 23] have been applied to this task. Some contour-based techniques have been investigated to enhance the accurate prediction of nucleus boundaries [3, 9, 29] which aids in splitting the touched and overlapped nuclei. However, these methods lack the capability to characterize the differences among neighboring nuclear instances, making it difficult to accurately segment densely stacked overlapping nuclei. Moreover, their applicability is hindered in scenarios where fine pixel-level annotations are insufficient, as obtaining precise boundary labels becomes a formidable task. Addressing the challenge posed by the scarcity of fine annotations, certain point annotation-based methods have emerged [2, 19, 21], leveraging the accessibility of coarse point labels. How + +ever, these approaches frequently prioritize the acquisition of pseudo-mask labels through design optimization methods, neglecting the differentiated representation of overlapping instances. + +Inspired by the mudslide in nature, we propose a simple universal method, called Mudslide. This approach excels in providing a differential representation of instances while remaining readily adaptable to the point-supervised domain. + +As shown in Fig. 1, for a continuous geographical region, mudslides induced by external forces lead to the creation of distinct partitions within the area. The unaffected section sets the partition boundary, while the collapsing segment forms a contiguous region. Analogously, this parallels the process of nucleus instance segmentation, where a method partitions the foreground with diverse instances into distinct instance regions. Furthermore, the mudslide collapse process concurrently considers the distinctions between regions and the coherence within each region. Variations in force directions acting on different regions induce noticeable changes in adjacent areas. Internal forces within a region, specifically the collapse of one soil mass exerting force on neighboring soil masses, facilitate the clustering and rationalization of soil masses in the same area. This aligns with the objectives of nuclear instance segmentation. + +Therefore, drawing inspiration from the characteristics of the muslide collapse process, we introduce Mudslide algorithm to induce the collapse of the foreground region of the nucleus and acquire instances. Specifically, we formulate a collapse field and leverage it to construct a force map to allocate forces directed to the center of the instance to individual pixels within the collapse field. This allocation ensures that the force direction in adjacent regions of distinct nuclei exhibits pronounced disparities, facilitating the distinctive representation of neighboring nuclei. Simultaneously, the forces within the same instance exhibit uniformity directed towards the center, enhancing a cohesive representation within instances. Subsequently, we devise a collapse algorithm commencing from the collapse boundary, proceeding pixel-by-pixel along the force direction. Ultimately, pixels collapsing into the same region are deemed components of the same instance. Due to the consideration of external differences between instances and the internal consistency within instances, Mudslide excels in accurately segmenting overlapping instances. The simplicity of collapse field design renders this approach easily extendable to the domain of point supervision. + +The contributions can be summarized as follows: + +- We propose Mudslide, a novel nucleus instance segmentation algorithm inspired by natural phenomenon of mudslide, which excels in achieving precise segmentation of adjacent or overlapping nuclei. +- We propose a collapse field and generate a force map along with an initial boundary to provide differentiated + +representations for neighboring instances and enhance the internal coherence within instances. + +- Mudslide is a universal method that can be easily extended from full supervision tasks to point supervision domains. +- We justify the superiority of our Mudslide on public nuclear datasets with state-of-the-art performance. + +# 2. Related Works + +# 2.1. Fully-supervised nuclei segmentation + +Fully-supervised nuclear instance segmentation has been widely studied and benefited from the application of deep learning techniques [3, 7, 22]. One of the most influential models in this field is the U-Net model proposed by Ronneberger et al. [23], which has become a benchmark model for medical image segmentation and has inspired many subsequent works. Qu et al. [20] proposed a full-resolution convolutional neural network (FullNet), a full-resolution convolutional neural network that avoids downsampling operations in the network structure to improve localization accuracy. Raza et al. [22] proposed Micro-Net, which utilizes multi-resolution and weighted loss functions to achieve robustness against large internal/external variances in nucleus size. He et al. [8] proposed Han-Net, a hybrid attention nested U-shaped network that extracts informative features from multiple layers to enhance segmentation performance. + +Leveraging contour information for distinguishing interacting or overlapping nuclei, Chen et al. [3] introduced the integration of contour details into a multi-level Fully Convolutional Network (FCN), which resulted in the development of a profound contour-aware network designed specifically for nuclear instance segmentation. Following this, Zhou et al. [29] introduced the Contour-Aware Information Aggregation Network, which integrates spatial and textural features between nuclei and contours. Additionally, He et al. [9] proposed a Topology-Aware Loss to integrate meaningful topology constraints into the model optimization process for more precise boundary predictions. + +Furthermore, certain studies [4, 6, 14, 16] have employed distance maps to address the segmentation challenges posed by contacting or overlapping nuclei. Naylor et al. [16] tackled the segmentation of touching nuclei by framing it as a regression task involving intra-nuclear distance maps. Graham et al. [6] introduced the Hover-Net, which utilizes vertical and horizontal distances between a nucleus pixel and its center of mass to delineate nuclei. Additionally, He et al. [7] proposed the Centripetal Directional Network (CDNet) for nuclear instance segmentation, integrating directional information into the network. + +The above models have achieved performance improvements in nuclear segmentation. However, the challenge of segmenting overlapping cell nuclei persists. To further en + +![](images/49cb6aa336973ab8bf4dadd4a1ebba43e6efd75ee9dce3d03e17831f77026d5d.jpg) +Figure 2. The simple modeling of natural mudslide. (Left) Schematic diagram of the force on a soil mass on the slope. The green line represents the slope and the red dashed line represents the soil mass. (Right) A mountain where the mudslide occurred. The mudslide starts at the position pointed by the force indicated by the red arrow and stops at the horizontal slope. + +hance the model's ability to distinguish instances, there is a need for a distinct and differential representation of nucleus instances. + +# 2.2. Pointly-supervised nuclei segmentation + +The pixel-wise annotation of nucleus instances incurs high costs. Point annotation, as one of the most efficient labeling manners, has garnered increasing attention in the field of nuclear segmentation. Current approaches [2, 19, 21] primarily transform point annotations into coarse pixel-level labels, including Voronoi labels [25], cluster labels [19], and pseudo edge maps [27]. These methods commonly integrate prior knowledge about the shape and texture of nuclei into these coarse pixel-level labels. However, these techniques face challenges related to incomplete coarse labels and imprecise boundary information. Several strategies have been proposed to address the issues arising from these coarse labels. These include multi-stage optimization for refining segmentation in a bootstrapping manner [21, 25], and the incorporation of additional constraints based on the local contrast between nuclei and their surrounding cytoplasm. Additionally, some unique training strategies have been employed to optimize pseudo-labels. [12] utilizes the co-training method, optimizing mask labels through the simultaneous training of two models with mutual supervision. [28] employs a collaborative knowledge-sharing approach, sharing knowledge between a principal model and a very lightweight collaborator model. + +Despite the progress in the field of making full use of point labels and generating accurate pseudo mask labels, however, the problem of nucleus overlapping instance segmentation in point-supervised prediction still needs further attention. + +![](images/220afe0d0232dc0f9567b9df89bde1e6148d96ef7663b890e5ef5c8930ac5cbc.jpg) +Figure 3. The processing of Mudslide algorithm. (Left) The instance map, the four-channel collapse field, the eight-directional force map, and the initial boundary. Each color in the force map represents a direction. (Right) Schematic diagram of the mudslide collapse process performed on two adjacent nuclei. + +# 3. Method + +# 3.1. Natural Mudslide + +Mudslide refers to the natural phenomenon where the soil or rock mass on a slope, influenced by external forces such as river erosion, rainfall, and seismic events, undergoes downhill movement collectively along the slope [5]. To simulate the mudslide process, we simplify it as a continuous sliding problem of soil mass on a slope [24]. This assumes an arc-shaped sliding surface adhering to the Coulomb criterion during soil movement, and the soil undergoes rigid sliding. This allows the assessment of the stability of the soil slope based on the magnitudes of the sliding moment and anti-sliding moment. + +As illustrated in Fig. 2, for the soil mass on the slope, the sliding moment is generated by the gravitational force $W$ , while the resisting moment is provided by the frictional force on the sliding surface. The sliding moment $M_{s}$ and resisting moment $M_{r}$ are + +$$ +M _ {s} = W * d, \tag {1} +$$ + +$$ +M _ {r} = R * \int_ {0} ^ {l} \tau d l, \tag {2} +$$ + +where $d$ represents the lever arm of gravity, $R$ and $l$ are respectively the radius and the length of the arc, and $\tau$ represents the shear force. Introducing the safety factor $F_{os}$ [17] + +$$ +F _ {o s} = \frac {M _ {r}}{M _ {s}}, \tag {3} +$$ + +if $F_{os} > 1$ , the soil mass remains stable and will not collapse. If $F_{os} = 1$ , the soil mass is in a state of critical equilibrium, and if $F_{os} < 1$ , the soil mass undergoes collapse. + +Assuming each soil mass on the slope is in a critical state, i.e., $F_{os} = 1$ , if an external force $F$ with moment $M_F$ is applied tangentially along the slope surface such that + +$$ +F _ {o s} ^ {\prime} = \frac {M _ {r}}{M _ {s} ^ {\prime}} = \frac {M _ {r}}{M _ {s} + M _ {F}} < F _ {o s}, \tag {4} +$$ + +the corresponding soil mass undergoes collapse. This results in the application of a tangential force $F^{\gamma}$ on the adjacent soil mass below, leading to a continued collapse of the lower soil mass, ultimately forming a mudslide. The collapse terminates when it reaches a horizontal slope surface. In the end, the mudslide induces the collapse of soil mass on the slope surface affected by external forces, while the unaffected soil mass remains intact. This delineates distinct regions on the slope. + +# 3.2. Mudslide Segmentation + +# 3.2.1 Mudslide Processing + +Inspired by the natural mudslide mentioned in Sec. 3.1, we propose a Mudslide algorithm for nuclear instance segmentation. In natural landslide phenomena, the main issues include the characteristics of the slope that collapses, the forces acting on the soil mass on the slope, the location where the collapse begins, and how the collapse occurs continuously. Corresponding to this, the Mudslide algorithm primarily involves three steps: constructing a collapse field, obtaining the force map and initial collapse boundary, and executing the collapse. + +Collapse field construction. In the first step, we initiate the process by constructing a collapse field, representing the initial slope surface of the mudslide. In image segmentation, we assume that each pixel represents a soil mass, located on a slope represented by the collapse field, and in a state of critical equilibrium. For simplicity, we utilize continuous normalized maps to characterize the collapse field, where pixel values close to 1 signify higher slope heights and values close to 0 indicate lower slope surfaces. Specifically, for ease of model learning and inspired by [6], we create a four-channel normalized map shown in Fig. 3. This map contains four distance maps generated in horizontal, vertical, and two diagonal directions, respectively, with each map indicating the distance from each pixel to the center of the nucleus in different directions. Pixel values are close to 0 when they are near the nucleus center and gradually increase outward until they reach 1 at the nucleus boundary. Thus, we construct a slope for the foreground nuclei in the image, where the slope is flat in the middle of each instance and steep around the edges, resembling the terrain that is prone to mudslide collapse. + +Force map and initial boundary obtaining. For the second step, we generate a force map and initial boundary from the collapse field. The force map assigns a force $F$ with different directions to each pixel, representing the directions in which pixels are expected to collapse during Mudslide. For two-dimensional images, we consider eight directions of force: up, down, left, right, upper-left, upper-right, lower-left, and lower-right (Fig. 3): + +$$ +\begin{array}{l} \mathbb {F} = \left\{\binom {0} {1}, \binom {0} {- 1}, \binom {- 1} {0}, \binom {1} {0}, \right. \\ \binom {- 1} {1}, \binom {1} {1}, \binom {- 1} {- 1}, \binom {1} {- 1} \}. \end{array} \tag {5} +$$ + +Then the force map can be expressed as a set of forces: + +$$ +F M = \{F (x, y) \mid x, y \in \mathbb {Z}, F \in \mathbb {F} \}, \tag {6} +$$ + +where $\mathbb{Z}$ indicates the foreground of nuclei and $x, y$ corresponds to the coordinate of each pixel. Here, we assign a force direction to each pixel that is approximately tangential to the slope. Since we assume that each pixel is in a state of critical equilibrium, we consider the force magnitude to be irrelevant to the collapse and therefore set the magnitude of each force to 1. + +To obtain the starting position of mudslide collapse, we generate the initial boundary. We employ a straightforward edge detection filter, such as Sobel operator, on the collapse field to detect edges. The obtained edges are then utilized as the initial boundary for mudslide collapse. This approach is rational since, in our collapse field, there is a significant gradient between instance boundaries, allowing the Sobel operator to effectively identify them. + +Mudslide collapse executing. After acquiring the force map and the initial boundary, we commence the mudslide collapse process. Originating from the initial boundary, each pixel $p$ applies an initial force (red arrows in 3) to the directed pixel $p'$ , which can be calculated by + +$$ +p ^ {\prime} = p + F (p), +$$ + +$$ +p = (x, y), \tag {7} +$$ + +$$ +p ^ {'} = (x ^ {'}, y ^ {'}). +$$ + +The pixel $p'$ then undergoes collapse. Subsequently, the collapsed pixel exerts a sustained force (yellow arrows in 3) on the pointed pixel, initiating the collapse of more adjacent pixels. This iterative process results in a sustained mudslide collapse. Eventually, when all pixels directed to have undergone collapse, those pixels collapsing to the same region are identified as a single instance. Note that those pixels that are never directed by forces from other pixels will not collapse and the force assigned to them will be neglected (gray arrows in 3), corresponding to the soil mass that is not affected by the mudslide. + +# 3.2.2 Application to Nuclear Instance Segmentation + +The framework of applying Mudslide to nuclear instance segmentation is illustrated in Fig. 4. Here, we unify the pipelines for both fully supervised and point-supervised scenarios, with the primary distinction lying in the labels used for training. + +Specifically, we employ a dual-branch architecture, where one branch is dedicated to predicting a binary map + +![](images/9c569adf6916f15313c6c1194f95af122cf9dcc8702ae0bbe2364719bac39d29.jpg) + +![](images/0feec5d3f43da033cc3dfcd6818f06f5046869aca5f5184773ca9f33312dd437.jpg) +Figure 4. The overall framework of the proposed Mudslide. The network contains two basic prediction branches: the binary branch and the collapse field branch. During training, the difference between fully- and point-supervised is the type of ground truth. During testing, both branches' predictions are employed for the Mudslide algorithm. + +for the nuclear foreground, and the other branch focuses on predicting the collapse field. The collapse field consists of four channels, corresponding to horizontal, vertical, and two diagonal directions. The segmentation model adopted is the Trans-UNet architecture from [10]. In fully supervised training, the binary map labels are the ground truth representing the nuclear foreground, while the collapse field labels are constructed from both the foreground and instance ground truth. For point-supervised training, we leverage point labels and the original input image to generate a cluster map [19] and Voronoi map [25] as pseudolabels for foreground prediction, which is commonly observed in point-supervised tasks. Subsequently, we use the predicted foreground and point labels to generate collapse field labels. + +We utilize cross-entropy loss as the binary classification loss and employ mean squared error loss and mean squared gradient error loss as the collapse field loss: + +$$ +\mathcal {L} _ {B} = \lambda_ {B _ {C E}} \mathcal {L} _ {C E}, \tag {8} +$$ + +$$ +\mathcal {L} _ {C F} = \lambda_ {C F _ {M S E}} \mathcal {L} _ {M S E} + \lambda_ {C F _ {M S G E}} \mathcal {L} _ {M S G E}. \tag {9} +$$ + +Then the overall loss function is + +$$ +\mathcal {L} = \mathcal {L} _ {B} + \mathcal {L} _ {C F}. \tag {10} +$$ + +It's worth noting that in point-supervised training, the + +cross-entropy loss comprises two terms, corresponding to cluster and Voronoi: + +$$ +\mathcal {L} _ {B _ {\text {p o i n t}}} = \lambda_ {B _ {C E}} \mathcal {L} _ {C E _ {\text {c l u s t e r}}} + \lambda_ {B _ {C E}} \mathcal {L} _ {C E _ {\text {v o r o n o i}}}. \tag {11} +$$ + +During the inference process, we first combine the foreground and collapse field predictions to construct the collapse field, ensuring subsequent mudslide acts on the nuclear foreground regions. Subsequently, as detailed in Sec. 3.1, we utilize the collapse field to obtain the initial boundary and force map, initiating the mudslide collapse process. Ultimately, this process yields the nuclear instance segmentation results. + +# 4. Experiment + +# 4.1. Experimental Settings + +Datasets. I) CPM17 [26] contains 64 H&E stained histopathology images with 7,570 annotated nuclear boundaries. It is from the MICCAI 2017 Digital Pathology Challenge [26] and images on two different scales: $500 \times 500$ and $600 \times 600$ . This is a dataset for a single-class segmentation task. II) NuInsSeg [15] contains 665 image patches with more than 30,000 manually segmented nuclei from 31 human and mouse organs. It is one of the biggest fully manually annotated datasets of nuclei in Hematoxylin and + +Eosin (H &E) stained histological images with the scale of $512 \times 512$ . This is also a dataset for single-class segmentation task. + +Evaluation Metrics. We use four instance-level evaluation metrics to measure the instance segmentation performance of the comparison models, which are: Detection Quality (DQ), Segmentation Quality (SQ), Panoptic Quality (PQ), and the Aggregated Jaccard Index (AJI). + +Implementation details. The proposed Mudslide is trained and tested using the open-source software library Pytorch 1.8.0 on NVIDIA Tesla V100 with CUDA 10.1. + +In order to build the dataset used for training, the CPM17 dataset is divided into multiple patches as it has an inconsistent original image size. We generate 512 training patches in the size of $256 \times 256$ , thus allowing the network to be uniform and fully trained. In addition, we randomly divide NuInsSeg into training set and test set with the proportion of $80\%$ and $20\%$ , respectively. Moreover, for point-supervised tasks, we leverage the annotations of nucleus instances to generate a central point for each instance, serving as the point label. In the training phase, we apply a series of data augmentation methods, which include random flip, random rotation, random color, random crop, and Gaussian blur. The training epoch is set as 300, and the initial learning rate is set as $10^{-3}$ . The weights of each loss are $\lambda_{BCE} = 1$ , $\lambda_{CFMSE} = 2.5$ , $\lambda_{CFMSGE} = 8$ . + +# 4.2. Ablation Study + +We performed ablation studies on the CPM17 and NuInsSeg datasets to validate the efficacy of the proposed Mudslide method. All the following experiments were conducted under the setting of fully supervised tasks. + +Instance separation methods. Table 1 shows the comparison of segmentation performance obtained using different instance separation methods. We apply four different methods to the predictions of the same model: the traditional watershed method [1], boundary partition method, HVmap segmentation method from Hover-Net [6], and our proposed method. + +Specifically, the watershed method involves segmenting instances using the watershed algorithm on the binary probability map predicted by the model. The boundary separation method refers to predicting boundaries by adding an additional boundary prediction branch to the model and using the boundaries to directly separate different instances. The HVmap algorithm is an instance separation method in HoverNet that utilizes horizontal and vertical distance maps for segmentation. + +The experimental results indicate significant variations in model performance based on different instance separation methods. Traditional methods such as watershed and boundary separation exhibit subpar performance. The HVmap approach, utilizing horizontal and vertical maps, + +
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
Watershed0.7970.7300.5820.6940.6010.7350.4610.502
Boundary0.8370.7340.6140.7140.6400.7530.4910.532
HVMap0.8790.8100.7130.7370.7860.7540.5960.638
Ours0.8830.8080.7140.7400.7880.7560.5990.647
+ +Table 1. Comparison of different adjacent instance separation methods, including the watershed algorithm, boundary partition, HVmap and our method. + +
boundcollapseCPM17NuInsSeg
DQ ↑SQ ↑PQ ↑AJI ↑DQ ↑SQ ↑PQ ↑AJI ↑
✓0.7720.7100.5480.6100.5870.7030.4140.497
0.8770.8070.7070.7330.7750.7520.5820.630
✓✓0.8830.8080.7140.7400.7880.7560.5990.647
+ +Table 2. Ablation studies of components in Mudslide algorithm, where bound indicates initial boundary, and collapse indicates mudslide collapse processing. + +enhances the differentiation between instances and achieves better results. Furthermore, our method, with more explicit instance differentiation representation and a more rational instance acquisition process, outperforms other methods. This underscores the superiority of our Mudslide approach. + +The ablation of components of Mudslide algorithm. We conducted ablation experiments on the key steps of the mudslide algorithm. Here, we considered two components: whether to use the initial boundary and whether to perform mudslide collapse. In Table 2, we explored three comparative experiments: a baseline approach with no additional components, relying solely on binary map partitioning based on connected regions; a scenario involving only the initial boundary extracted from the collapse field; and a comprehensive application of both the initial boundary and mudslide collapse, representing our full-fledged mudslide algorithm for instance segmentation. These experiments aim to scrutinize the individual contributions of each component to the overall segmentation performance. + +The experimental results indicate that incorporating an initial boundary leads to performance gains of $16\%$ PQ and $12\%$ AJI for the CPM dataset and $17\%$ PQ and $14\%$ AJI for NuInsSeg. Furthermore, the additional mudslide collapse results in a performance boost of $7\%$ PQ and $7\%$ AJI for the CPM dataset and $16\%$ PQ and $17\%$ AJI for NuInsSeg. These findings underscore the substantial benefits of each component within the mudslide method for enhancing instance segmentation outcomes. + +Force map construction methods. As a crucial manifestation of distinct instance differentiation, the construction of the force map is pivotal in the Mudslide algorithm. Therefore, we consider three methods for building the force map: direct prediction, collapse field segmentation, and collapse field extraction. + +
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
predict0.8740.8070.7070.7350.7770.7510.5870.633
segment0.8640.8080.7010.7280.7720.7530.5820.629
extract0.8800.8070.7110.7390.7840.7550.5920.643
+ +In the first approach, we have the model predict the force map directly. Specifically, we augment the model with a branch that outputs nine channels, allowing it to learn the force map's eight directions and the background region. Subsequently, we obtain the force map by employing the argmax function to the predicted probability map. In the second approach, we perform segmentation on the collapse field. Here, we use a two-channel collapse field corresponding to the horizontal and vertical directions. By segmenting the collapse field based on angles, we categorize it into eight regions, representing the eight directions of the force map. In the third approach, we predict a four-channel collapse field for horizontal, vertical, and two diagonal directions. Then, we directly employ argmax to extract the direction of force that each pixel should possess. + +The results presented in Table 3 indicate that the force map extracted through four-channel collapse fields performs the best. The direct prediction follows with slightly lower performance, while the force map obtained by splitting the two-channel collapse field performs the least favorably. + +The four-channel collapsing field extraction method outperformed direct prediction by $4\%$ PQ and $4\%$ AJI on CPM17, and outperformed the two-channel segmentation method by $6\%$ PQ, and $11\%$ AJI. On the NuInsSeg dataset, it outperformed direct prediction by $5\%$ PQ, and $1\%$ AJI, and outperformed the two-channel segmentation method by $10\%$ PQ, and $14\%$ AJI. These results suggest that the four-channel collapsing field can better capture and characterize the differences between cell nucleus instances, resulting in greater differences between adjacent instances and better instance segmentation results. + +Initial boundary generation methods. Regarding the issue of generating and selecting initial boundaries, we considered two methods: direct generation from the binary prediction map and generation from the collapsing field. + +Specifically, for the first method, we first obtained the binary probability map predicted by the model's binary classification branch and then selected a threshold of 0.5 to segment the predicted foreground and background regions of the nucleus. Then, we used the part of the pixels that contacted the background, i.e., the edge of the foreground, as the initial boundary. For the second method, we first obtained the four-channel collapsing field predicted by the model and then used some edge extraction methods, such as + +Table 3. Comparison of different force map construction methods. + +
MethodCPMNuInsSeg
DQ ↑SQ ↑PQ ↑AJI ↑DQ ↑SQ ↑PQ ↑AJI ↑
binary map collapse field0.8800.8070.7110.7390.7840.7550.5920.643
0.8830.8080.7140.7400.7880.7560.5990.647
+ +Table 4. Comparison of different initial boundary generation methods. + +using the Sobel operator to extract the regions with higher gradients in the collapsing field as the initial boundary. + +As shown in Table 4, the method of generating initial boundaries using the collapsing field outperforms the method of using only the binary prediction map. Specifically, on the CPM dataset, the collapsing field method improved performance by $3\%$ PQ, and $1\%$ AJI compared to the binary map method. On the NuInsSeg dataset, it improved performance by $7\%$ PQ, and $4\%$ AJI. As anticipated, the construction of the initial boundary is crucial in our Mudslide collapse algorithm, as it determines the algorithm's starting position. Obtaining foreground edges directly from a binary image makes it challenging to capture the boundaries between adjacent instances. Although the Mudslide algorithm can capture this difference to some extent, its performance is not as effective as constructing a better initial boundary. The initial boundary extracted through the collapse field takes into account the differentiated representations between cell nuclei, effectively separating nuclei to a certain degree and proving to be a superior boundary. + +# 4.3. Fully-Supervised Nuclei Segmentation + +In Table 5, the proposed Mudslide is compared with the state-of-the-art fully-supervised nuclear instance segmentation models on CPM17 and NuInsSeg. Our Mudslide achieves 0.714 PQ and 0.740 AJI on the CPM17 dataset, and 0.599 PQ and 0.647 AJI on the NuInsSeg dataset, respectively. The Mudslide achieves competitive results for fully-supervised nuclei segmentation tasks. + +We further carry out a qualitative visualization analysis of our method and different existing methods. Compared to all the other methods in Fig. 5, our Mudslide achieves superior instance segmentation results for nuclear instance segmentation, especially in some regions where nuclei are densely distributed with a large amount of overlap, where our instance segmentation is more accurate. The results, both quantitative and qualitative, unequivocally establish the competitiveness of our Mudslide algorithm in accurately segmenting densely packed and overlapping nuclear instances. + +# 4.4. Pointly-Supervised Nuclei Segmentation + +In Table 6, we extend the proposed Mudslide from fully-supervised to point-supervised nuclear segmentation task and compare it with the state-of-the-art point-supervised nuclear instance segmentation models on CPM17 and + +![](images/aaaa8803377b0f6c6a24fd5d1dab9dc5b6f305fabb1f24b38a7f39e42a51b463.jpg) + +![](images/125633ffe895597bc4643391ed91a588962169c02a77af5e296c475763189d63.jpg) + +![](images/66186f44fb71f4cff435397bf4122e9138c00fb64c710f2132b49c8687cb8f6e.jpg) + +![](images/64ee2f5006a77cbef1b94e2ac689ac7aa3e76163828a5181e528f0be8fff67f7.jpg) +(a) +Image +Figure 5. Visualization of instance segmentation results with different methods. (a) Input image; (b) Ground Truth; (c) U-Net [23]; (d) DIST [16]; (e) FullNet [20]; (f) Hover-Net [6]; (g) Our Mudslide. + +![](images/5708ec99f9a7bd92be593b662715669722c556dc722f51b11b483b6c3f6b367a.jpg) + +![](images/f8e0f9de5635e1810e045b1aa2108fbd2d5074244e951531ad6531f859986121.jpg) + +![](images/471482f1465a64ceb3264a71576a892c45c19dbd6c0b6c5be3d0b76ae4dc7064.jpg) + +![](images/33b1f46c820e1b3400dbfe15d59d3af41002023a6930210126cee8fd86796439.jpg) +(b) +Ground Truth + +![](images/a86983016e9228e092eba7224f62db1c39d22ee49145b8b8961bb7f3253deb5f.jpg) + +![](images/9a0eacc8d16e1ea0e5159b7149451ee06cf4f02c391257ac23903cc7ba215618.jpg) + +![](images/a9ec58cd0715b9fb0f03ee8d06e8c3b7ea79c822bf532480fbddaff1e9d24fef.jpg) + +![](images/9635bf5f7691c1ecb44c4d55c2d3f88607f6a712d63e2203c6119a60c5c6a1d0.jpg) +(c) +U-Net + +![](images/c546fdd12524abdc307ba770590f52c533c69a9fbce658069503f8ba0d6b355c.jpg) + +![](images/5e8e0d7f7f354a03cce6e8f392facf7583d63da51f173be597e105406ccf43bb.jpg) + +![](images/c696812f6984170a616234ed6f45cdf7ed30aef9789aa2a4c77fc9159cd06427.jpg) + +![](images/2a4cbb4925647bab34b1217499e374e85f8a6a9ad8e7606b866a42d0521cbeae.jpg) +(d) +DIST + +![](images/77c8e5a2c67baed7d469b070827aff673b6d098e11b7dd5c4665373a542f8434.jpg) + +![](images/ca768617ffed6e199573c093cf47f7e46ba29519c54ed06146b2c0082859e67a.jpg) + +![](images/4198c1409f1634d95b6ef460ee8c7e2e896dfe9b96496764acf57e2fbac4b4db.jpg) + +![](images/2c862aa2c699d6a5f9791afb001fdc49850fe37b802bc9d99202c77615337796.jpg) +(e) +FullNet + +![](images/c235508c663315e854bbf20342a31c0e3d18200666c0cf848c8dcc4c398d3a1a.jpg) + +![](images/2305446ffebd45dcf7ef74e5a225275d19c371c6643297da4930f31863cb7418.jpg) + +![](images/c98a02cf5d83536d7093b2a28e8c8fce871f01fdbf608bde6b7418eb49b79aef.jpg) + +![](images/8fccbf4d87eb125159654ab79847fd67d18784c8398e5473bb68990989bf3748.jpg) +(f) +HoverNet + +![](images/fec93d31d352012b7d08a188da620dc3d461ecf2147f978316311c66bb7e8e90.jpg) + +![](images/cc91497ccdb1b9158db1921f1a40996505e3125c6ca6f478f79e278cb3228410.jpg) + +![](images/2cce860532ab71e7280f8d24c59d12eaaedb06f6eead5cfb512f34a26783dbd9.jpg) + +![](images/943ca03733899e071318603229e7bc7759770684b09b126c4bad9fc805c1370c.jpg) +(g) +Ours + +
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
UNet [23]0.7780.7340.5780.6430.6350.6880.4360.508
DCAN [3]0.7320.7400.5450.5610.5630.6440.3630.435
DIST [16]0.6630.7540.5040.6160.7030.7280.5120.542
FullNet [20]--0.6610.6680.7120.7600.5410.562
HoverNet [6]0.8540.8140.6970.7050.7430.7640.5680.611
CDNet [7]---0.733---0.640
CellViT [10]0.8790.8100.7130.7370.7860.7540.5960.638
Ours0.8830.8080.7140.7400.7880.7560.5990.647
+ +NuInsSeg. Our Mudslide achieves 0.529 PQ and 0.581 AJI on the CPM17 dataset, and 0.456 PQ and 0.558 AJI on the NuInsSeg dataset, respectively. In the task of nuclear instance segmentation with only point annotations, our method can also achieve competitive results. + +Although methods relying solely on point annotations still have ample room for improvement compared to fully supervised approaches, our approach has already surpassed other methods in the realm of point annotations. + +# 5. Conclusion + +This paper proposes a novel universal approach for nuclear instance segmentation, called Mudslide, which is inspired by the natural phenomenon of mudslides. To enhance + +Table 5. Performance comparisons on fully-supervised nuclear segmentation. + +
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
PENet [27]0.4400.7220.3190.2900.4760.6300.3000.355
C2FNet [25]0.4290.6300.2710.4210.4050.5800.2350.400
S2Label [11]0.6540.6980.4590.5030.6130.6670.4090.496
SPN+IEN [13]0.7040.6890.4880.5430.6880.6600.4540.523
SCNet [12]0.7160.6850.4960.5540.6940.6550.4560.526
Ours0.7470.7030.5290.5810.7070.6890.4870.558
+ +Table 6. Performance comparisons on point supervised nuclear segmentation. + +the model's capability to differentiate between adjacent instances, we devise a collapsing field, and generate a force map along with an initial boundary, offering distinct representations for different instances and a unified representation for the same instance. We also propose Mudslide collapse algorithm that mimics the natural process by segmenting through the iterative collapse of individual pixels. Our collapse fields are simple and easily extendable to point-supervised tasks. The experimental results on public nuclei datasets demonstrate that Mudslide outperforms state-of-the-art methods. + +Acknowledgments This work was supported in part by the National Key R&D Program of China (No. 2022ZD0118201), Natural Science Foundation of China (No.61972217, 32071459, 62176249, 62006133, 62271465), and Shenzhen Medical Research Funds in China (No. B2302037). + +# References + +[1] Serge Beucher. Use of watersheds in contour detection. In Proc. Int. Workshop on Image Processing, Sept. 1979, pages 17-21, 1979. 6 +[2] Alireza Chamanzar and Yao Nie. Weakly supervised multitask learning for cell detection and segmentation. In 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI), pages 513-516. IEEE, 2020. 1, 3 +[3] Hao Chen, Xiaojuan Qi, Lequan Yu, Qi Dou, Jing Qin, and Pheng-Ann Heng. Dcan: Deep contour-aware networks for object instance segmentation from histology images. Medical Image Analysis, 36:135-146, 2017. 1, 2, 8 +[4] Shengcong Chen, Changxing Ding, Minfeng Liu, Jun Cheng, and Dacheng Tao. CPP-net: Context-aware polygon proposal network for nucleus segmentation. IEEE Transactions on Image Processing, 32:980-994, 2023. 2 +[5] FC Dai, Chin Fei Lee, and Y Yip Ngai. Landslide risk assessment and management: an overview. Engineering Geology, 64(1):65-87, 2002. 3 +[6] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:1-15, 2019. 2, 4, 6, 8 +[7] Hongliang He, Zhongyi Huang, Yao Ding, Guoli Song, Lin Wang, Qian Ren, Pengxu Wei, Zhiqiang Gao, and Jie Chen. Cdnet: Centripetal direction network for nuclear instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4026-4035, 2021. 2, 8 +[8] Hongliang He, Chi Zhang, Jie Chen, Ruizhe Geng, Luyang Chen, Yongsheng Liang, Yanchang Lu, Jihua Wu, and Yongjie Xu. A hybrid-attention nested unet for nuclear segmentation in histopathological images. Frontiers in Molecular Biosciences, 8:1–9, 2021. 1, 2 +[9] Hongliang He, Jun Wang, Pengxu Wei, Fan Xu, Xiangyang Ji, Chang Liu, and Jie Chen. Toposeg: Topology-aware nuclear instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21307-21316, 2023. 1, 2 +[10] Fabian Horst, Moritz Rempe, Lukas Heine, Constantin Seibold, Julius Keyl, Giulia Baldini, Selma Ugurel, Jens Siveke, Barbara Grünwald, Jan Egger, et al. Cellvit: Vision transformers for precise cell segmentation and classification. arXiv preprint arXiv:2306.15350, 2023. 5, 8 +[11] Hyeonsoo Lee and Won-Ki Jeong. Scribble2label: Scribblesupervised cell segmentation via self-generating pseudolabels with consistency. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23, pages 14-23. Springer, 2020. 8 +[12] Yi Lin, Zhiyong Qu, Hao Chen, Zhongke Gao, Yuexiang Li, Lili Xia, Kai Ma, Yefeng Zheng, and Kwang-Ting Cheng. Nuclei segmentation with point annotations from pathology images via self-supervised learning and co-training. Medical Image Analysis, 89:102933, 2023. 3, 8 + +[13] Weizhen Liu, Qian He, and Xuming He. Weakly supervised nuclei segmentation via instance learning. In 2022 IEEE 19th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2022. 8 +[14] Xiaoming Liu, Zhengsheng Guo, Jun Cao, and Jinshan Tang. Mdc-net: a new convolutional neural network for nucleus segmentation in histopathology images with distance maps and contour information. Computers in Biology and Medicine, 135:1-12, 2021. 2 +[15] Amirreza Mahbod, Christine Polak, Katharina Feldmann, Rumsha Khan, Katharina Gelles, Georg Dorffner, Ramona Woitek, Sepideh Hatamikia, and Isabella Ellinger. Nuinsseg: A fully annotated dataset for nuclei instance segmentation in h&E-stained histological images. arXiv preprint arXiv:2308.01760, 2023. 5 +[16] Peter Naylor, Marick Laé, Fabien Reyal, and Thomas Walter. Segmentation of nuclei in histopathology images by deep regression of the distance map. IEEE Transactions on Medical Imaging, 38(2):448-459, 2018. 2, 8 +[17] Hasmukh Pranshanker Oza and HP Oza. Soil mechanics and foundation engineering. Charotar Book Stall, 1969. 3 +[18] H. Pinckaers, W. Bulten, J. van der Laak, and G. Litjens. Detection of prostate cancer in whole-slide images through end-to-end training with image-level labels. IEEE Transactions on Medical Imaging, 40(7):1817-1826, 2021. 1 +[19] Hui Qu, Pengxiang Wu, Qiaoying Huang, Jingru Yi, Gregory M Riedlinger, Subhajyoti De, and Dimitris N Metaxas. Weakly supervised deep nuclei segmentation using points annotation in histopathology images. In International Conference on Medical Imaging with Deep Learning, pages 390-400. PMLR, 2019. 1, 3, 5 +[20] Hui Qu, Zhennan Yan, Gregory M Riedlinger, Subhajyoti De, and Dimitris N Metaxas. Improving nuclei/gland instance segmentation in histopathology images by full resolution neural network and spatial constrained loss. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 378-386, 2019. 1, 2, 8 +[21] Hui Qu, Pengxiang Wu, Qiaoying Huang, Jingru Yi, Zhennan Yan, Kang Li, Gregory M Riedlinger, Subhajyoti De, Shaoting Zhang, and Dimitris N Metaxas. Weakly supervised deep nuclei segmentation using partial points annotation in histopathology images. IEEE Transactions on Medical Imaging, 39(11):3655-3666, 2020. 1, 3 +[22] Shan E Ahmed Raza, Linda Cheung, Muhammad Shaban, Simon Graham, David Epstein, Stella Pelengaris, Michael Khan, and Nasir M Rajpoot. Micro-net: A unified model for segmentation of various objects in microscopy images. Medical Image Analysis, 52:160-173, 2019. 2 +[23] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 1, 2, 8 +[24] Digvijay P Salunkhe, Rupa N Bartakke, Guruprasd Chvan, and Pooja R Kothavale. An overview on methods for slope stability analysis. International Journal of Engineering Research & Technology (IJERT), 6(03):2278-0181, 2017. 3 + +[25] Kuan Tian, Jun Zhang, Haocheng Shen, Kezhou Yan, Pei Dong, Jianhua Yao, Shannon Che, Pifu Luo, and Xiao Han. Weakly-supervised nucleus segmentation based on point annotations: A coarse-to-fine self-stimulated learning strategy. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part V 23, pages 299-308. Springer, 2020. 3, 5, 8 +[26] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in Bioengineering and Biotechnology, 53: 1-15, 2019. 5 +[27] Inwan Yoo, Donggeun Yoo, and Kyunghyun Paeng. Pseudoedgenet: Nuclei segmentation only with point annotations. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part I 22, pages 731-739. Springer, 2019. 3, 8 +[28] Ji Yu. Point-supervised single-cell segmentation via collaborative knowledge sharing. arXiv preprint arXiv:2304.10671, 2023. 3 +[29] Yanning Zhou, Omer Fahri Onder, Qi Dou, Efstratios Tsougenis, Hao Chen, and Pheng-Ann Heng. Cia-net: Robust nuclei instance segmentation with contour-aware information aggregation. In International Conference on Information Processing in Medical Imaging, pages 682–693, 2019. 1, 2 \ No newline at end of file diff --git a/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/images.zip b/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b29d50198fb55ce350c59f745773863debbe4a15 --- /dev/null +++ b/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84bb651e57dcff334c783516661b6216418b072b8b7c2c6bbd4e80abe988fac4 +size 733568 diff --git a/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/layout.json b/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..98c69b65f9fcdaefc1abac753aaef85c9ad782cc --- /dev/null +++ b/2024/Mudslide_ A Universal Nuclear Instance Segmentation Method/layout.json @@ -0,0 +1,8590 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 492, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 492, + 120 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 492, + 120 + ], + "type": "text", + "content": "Mudslide: A Universal Nuclear Instance Segmentation Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 144, + 503, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 144, + 503, + 185 + ], + "spans": [ + { + "bbox": [ + 90, + 144, + 503, + 185 + ], + "type": "text", + "content": "Jun Wang \nSchool of Electronic and Computer Engineering, Peking University, Shenzhen, China \njunw@stu.pku.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 239, + 290, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 539 + ], + "type": "text", + "content": "Nuclear instance segmentation has played a critical role in pathology image analysis. The main challenges arise from the difficulty in accurately segmenting densely overlapping instances and the high cost of precise mask-level annotations. Existing fully-supervised nuclear instance segmentation methods, such as boundary-based methods, struggle to capture differences between overlapping instances and thus fail in densely distributed blurry regions. They also face challenges transitioning to point supervision, where annotations are simple and effective. Inspired by natural mudslides, we propose a universal method called Mudslide that uses simple representations to characterize differences between different instances and can easily be extended from fully-supervised to point-supervised. Concretely, we introduce a collapse field and leverage it to construct a force map and initial boundary, enabling a distinctive representation for each instance. Each pixel is assigned a collapse force, with distinct directions between adjacent instances. Starting from the initial boundary, Mudslide executes a pixel-by-pixel collapse along various force directions. Pixels that collapse into the same region are considered as one instance, concurrently accounting for both inter-instance distinctions and intra-instance coherence. Experiments on public datasets show superior performance in both fully-supervised and point-supervised tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 573, + 128, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 128, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 128, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "content": "Pathological slide analysis is widely regarded as the gold standard for cancer diagnosis, treatment, and prevention. Nuclear instance segmentation is a critical step in this process, as it enables the accurate identification and analysis of nuclei within pathology images. The morphology of nuclei, including their shape, appearance, and density, provides interpretable features that can be used as diagnostic and prognostic indicators of cancer [18]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Nuclear instance segmentation presents unique challenges compared to natural image instance segmentation" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 213, + 545, + 344 + ], + "blocks": [ + { + "bbox": [ + 309, + 213, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 309, + 213, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 309, + 213, + 545, + 344 + ], + "type": "image", + "image_path": "e82e54116028c73b7f39769e95bdea5409bb493bc3187d6105305b306f6c5b33.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 353, + 547, + 431 + ], + "lines": [ + { + "bbox": [ + 304, + 353, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 547, + 431 + ], + "type": "text", + "content": "Figure 1. An analogy between mudslide and nuclear instance segmentation. (Top) Topographic map before and after the mudslide. The colored arrows indicate the collapse forces in different directions. (Bottom) Nuclear instance segmentation process in pathological images. The different layers on the left represent different feature representations of nuclear instances, and the right shows the result of instance segmentation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "text", + "content": "tasks. Firstly, the densely packed distribution of nuclei often results in clustered overlapping instances. Secondly, the blurry boundaries between touching or overlapping nuclei make it difficult to distinguish individual instances. Third, obtaining precise pixel-level annotations is time-consuming and expensive for professional pathologists." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": "To address the first two challenges, many fully-supervised methods based on deep learning have made significant progress. Several algorithms based on constructing innovative network architectures [8, 20, 23] have been applied to this task. Some contour-based techniques have been investigated to enhance the accurate prediction of nucleus boundaries [3, 9, 29] which aids in splitting the touched and overlapped nuclei. However, these methods lack the capability to characterize the differences among neighboring nuclear instances, making it difficult to accurately segment densely stacked overlapping nuclei. Moreover, their applicability is hindered in scenarios where fine pixel-level annotations are insufficient, as obtaining precise boundary labels becomes a formidable task. Addressing the challenge posed by the scarcity of fine annotations, certain point annotation-based methods have emerged [2, 19, 21], leveraging the accessibility of coarse point labels. How" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11673" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "ever, these approaches frequently prioritize the acquisition of pseudo-mask labels through design optimization methods, neglecting the differentiated representation of overlapping instances." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 121, + 287, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 287, + 170 + ], + "type": "text", + "content": "Inspired by the mudslide in nature, we propose a simple universal method, called Mudslide. This approach excels in providing a differential representation of instances while remaining readily adaptable to the point-supervised domain." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 171, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 361 + ], + "type": "text", + "content": "As shown in Fig. 1, for a continuous geographical region, mudslides induced by external forces lead to the creation of distinct partitions within the area. The unaffected section sets the partition boundary, while the collapsing segment forms a contiguous region. Analogously, this parallels the process of nucleus instance segmentation, where a method partitions the foreground with diverse instances into distinct instance regions. Furthermore, the mudslide collapse process concurrently considers the distinctions between regions and the coherence within each region. Variations in force directions acting on different regions induce noticeable changes in adjacent areas. Internal forces within a region, specifically the collapse of one soil mass exerting force on neighboring soil masses, facilitate the clustering and rationalization of soil masses in the same area. This aligns with the objectives of nuclear instance segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 363, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 626 + ], + "type": "text", + "content": "Therefore, drawing inspiration from the characteristics of the muslide collapse process, we introduce Mudslide algorithm to induce the collapse of the foreground region of the nucleus and acquire instances. Specifically, we formulate a collapse field and leverage it to construct a force map to allocate forces directed to the center of the instance to individual pixels within the collapse field. This allocation ensures that the force direction in adjacent regions of distinct nuclei exhibits pronounced disparities, facilitating the distinctive representation of neighboring nuclei. Simultaneously, the forces within the same instance exhibit uniformity directed towards the center, enhancing a cohesive representation within instances. Subsequently, we devise a collapse algorithm commencing from the collapse boundary, proceeding pixel-by-pixel along the force direction. Ultimately, pixels collapsing into the same region are deemed components of the same instance. Due to the consideration of external differences between instances and the internal consistency within instances, Mudslide excels in accurately segmenting overlapping instances. The simplicity of collapse field design renders this approach easily extendable to the domain of point supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 628, + 258, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 628, + 258, + 639 + ], + "spans": [ + { + "bbox": [ + 58, + 628, + 258, + 639 + ], + "type": "text", + "content": "The contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 642, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 689 + ], + "type": "text", + "content": "- We propose Mudslide, a novel nucleus instance segmentation algorithm inspired by natural phenomenon of mudslide, which excels in achieving precise segmentation of adjacent or overlapping nuclei." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "- We propose a collapse field and generate a force map along with an initial boundary to provide differentiated" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 312, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 544, + 95 + ], + "type": "text", + "content": "representations for neighboring instances and enhance the internal coherence within instances." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 96, + 545, + 156 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 96, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 545, + 131 + ], + "type": "text", + "content": "- Mudslide is a universal method that can be easily extended from full supervision tasks to point supervision domains." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 132, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 132, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 132, + 545, + 156 + ], + "type": "text", + "content": "- We justify the superiority of our Mudslide on public nuclear datasets with state-of-the-art performance." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 169, + 397, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 169, + 397, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 169, + 397, + 182 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 190, + 502, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 190, + 502, + 202 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 502, + 202 + ], + "type": "text", + "content": "2.1. Fully-supervised nuclei segmentation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 209, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 399 + ], + "type": "text", + "content": "Fully-supervised nuclear instance segmentation has been widely studied and benefited from the application of deep learning techniques [3, 7, 22]. One of the most influential models in this field is the U-Net model proposed by Ronneberger et al. [23], which has become a benchmark model for medical image segmentation and has inspired many subsequent works. Qu et al. [20] proposed a full-resolution convolutional neural network (FullNet), a full-resolution convolutional neural network that avoids downsampling operations in the network structure to improve localization accuracy. Raza et al. [22] proposed Micro-Net, which utilizes multi-resolution and weighted loss functions to achieve robustness against large internal/external variances in nucleus size. He et al. [8] proposed Han-Net, a hybrid attention nested U-shaped network that extracts informative features from multiple layers to enhance segmentation performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 401, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 544 + ], + "type": "text", + "content": "Leveraging contour information for distinguishing interacting or overlapping nuclei, Chen et al. [3] introduced the integration of contour details into a multi-level Fully Convolutional Network (FCN), which resulted in the development of a profound contour-aware network designed specifically for nuclear instance segmentation. Following this, Zhou et al. [29] introduced the Contour-Aware Information Aggregation Network, which integrates spatial and textural features between nuclei and contours. Additionally, He et al. [9] proposed a Topology-Aware Loss to integrate meaningful topology constraints into the model optimization process for more precise boundary predictions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 545, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 675 + ], + "type": "text", + "content": "Furthermore, certain studies [4, 6, 14, 16] have employed distance maps to address the segmentation challenges posed by contacting or overlapping nuclei. Naylor et al. [16] tackled the segmentation of touching nuclei by framing it as a regression task involving intra-nuclear distance maps. Graham et al. [6] introduced the Hover-Net, which utilizes vertical and horizontal distances between a nucleus pixel and its center of mass to delineate nuclei. Additionally, He et al. [7] proposed the Centripetal Directional Network (CDNet) for nuclear instance segmentation, integrating directional information into the network." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "The above models have achieved performance improvements in nuclear segmentation. However, the challenge of segmenting overlapping cell nuclei persists. To further en" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11674" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 72, + 286, + 183 + ], + "blocks": [ + { + "bbox": [ + 52, + 72, + 286, + 183 + ], + "lines": [ + { + "bbox": [ + 52, + 72, + 286, + 183 + ], + "spans": [ + { + "bbox": [ + 52, + 72, + 286, + 183 + ], + "type": "image", + "image_path": "49cb6aa336973ab8bf4dadd4a1ebba43e6efd75ee9dce3d03e17831f77026d5d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 193, + 289, + 260 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 289, + 260 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 289, + 260 + ], + "type": "text", + "content": "Figure 2. The simple modeling of natural mudslide. (Left) Schematic diagram of the force on a soil mass on the slope. The green line represents the slope and the red dashed line represents the soil mass. (Right) A mountain where the mudslide occurred. The mudslide starts at the position pointed by the force indicated by the red arrow and stops at the horizontal slope." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 281, + 288, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 288, + 316 + ], + "type": "text", + "content": "hance the model's ability to distinguish instances, there is a need for a distinct and differential representation of nucleus instances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 346, + 253, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 253, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 253, + 358 + ], + "type": "text", + "content": "2.2. Pointly-supervised nuclei segmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 371, + 289, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 289, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 289, + 647 + ], + "type": "text", + "content": "The pixel-wise annotation of nucleus instances incurs high costs. Point annotation, as one of the most efficient labeling manners, has garnered increasing attention in the field of nuclear segmentation. Current approaches [2, 19, 21] primarily transform point annotations into coarse pixel-level labels, including Voronoi labels [25], cluster labels [19], and pseudo edge maps [27]. These methods commonly integrate prior knowledge about the shape and texture of nuclei into these coarse pixel-level labels. However, these techniques face challenges related to incomplete coarse labels and imprecise boundary information. Several strategies have been proposed to address the issues arising from these coarse labels. These include multi-stage optimization for refining segmentation in a bootstrapping manner [21, 25], and the incorporation of additional constraints based on the local contrast between nuclei and their surrounding cytoplasm. Additionally, some unique training strategies have been employed to optimize pseudo-labels. [12] utilizes the co-training method, optimizing mask labels through the simultaneous training of two models with mutual supervision. [28] employs a collaborative knowledge-sharing approach, sharing knowledge between a principal model and a very lightweight collaborator model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 712 + ], + "type": "text", + "content": "Despite the progress in the field of making full use of point labels and generating accurate pseudo mask labels, however, the problem of nucleus overlapping instance segmentation in point-supervised prediction still needs further attention." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 544, + 194 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 544, + 194 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 544, + 194 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 544, + 194 + ], + "type": "image", + "image_path": "220afe0d0232dc0f9567b9df89bde1e6148d96ef7663b890e5ef5c8930ac5cbc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 203, + 547, + 259 + ], + "lines": [ + { + "bbox": [ + 304, + 203, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 547, + 259 + ], + "type": "text", + "content": "Figure 3. The processing of Mudslide algorithm. (Left) The instance map, the four-channel collapse field, the eight-directional force map, and the initial boundary. Each color in the force map represents a direction. (Right) Schematic diagram of the mudslide collapse process performed on two adjacent nuclei." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 267, + 362, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 267, + 362, + 280 + ], + "spans": [ + { + "bbox": [ + 306, + 267, + 362, + 280 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 287, + 411, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 287, + 411, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 287, + 411, + 300 + ], + "type": "text", + "content": "3.1. Natural Mudslide" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 306, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 437 + ], + "type": "text", + "content": "Mudslide refers to the natural phenomenon where the soil or rock mass on a slope, influenced by external forces such as river erosion, rainfall, and seismic events, undergoes downhill movement collectively along the slope [5]. To simulate the mudslide process, we simplify it as a continuous sliding problem of soil mass on a slope [24]. This assumes an arc-shaped sliding surface adhering to the Coulomb criterion during soil movement, and the soil undergoes rigid sliding. This allows the assessment of the stability of the soil slope based on the magnitudes of the sliding moment and anti-sliding moment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": "As illustrated in Fig. 2, for the soil mass on the slope, the sliding moment is generated by the gravitational force " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": ", while the resisting moment is provided by the frictional force on the sliding surface. The sliding moment " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": " and resisting moment " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "M_{r}" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": " are" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 396, + 501, + 545, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 501, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 396, + 501, + 545, + 513 + ], + "type": "interline_equation", + "content": "M _ {s} = W * d, \\tag {1}", + "image_path": "5c41b6af3ee42ded88ecd7c07a4e9f886dc9eabcfcb75c3a5a0b947f30c6eacd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 387, + 513, + 545, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 513, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 387, + 513, + 545, + 540 + ], + "type": "interline_equation", + "content": "M _ {r} = R * \\int_ {0} ^ {l} \\tau d l, \\tag {2}", + "image_path": "4018d5374dc0cdebc02b59616eb27351eb45b31d9a19e785832327c287ed496d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 542, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 542, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 305, + 542, + 545, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 542, + 545, + 575 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 305, + 542, + 545, + 575 + ], + "type": "text", + "content": " represents the lever arm of gravity, " + }, + { + "bbox": [ + 305, + 542, + 545, + 575 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 542, + 545, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 542, + 545, + 575 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 542, + 545, + 575 + ], + "type": "text", + "content": " are respectively the radius and the length of the arc, and " + }, + { + "bbox": [ + 305, + 542, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 542, + 545, + 575 + ], + "type": "text", + "content": " represents the shear force. Introducing the safety factor " + }, + { + "bbox": [ + 305, + 542, + 545, + 575 + ], + "type": "inline_equation", + "content": "F_{os}" + }, + { + "bbox": [ + 305, + 542, + 545, + 575 + ], + "type": "text", + "content": " [17]" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 400, + 580, + 545, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 580, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 400, + 580, + 545, + 605 + ], + "type": "interline_equation", + "content": "F _ {o s} = \\frac {M _ {r}}{M _ {s}}, \\tag {3}", + "image_path": "bd1c1eb32b15600c2e64531206c4a21dd058078fedfd0817dd28435ac25b2fc0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 607, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 607, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 607, + 545, + 651 + ], + "type": "text", + "content": "if " + }, + { + "bbox": [ + 304, + 607, + 545, + 651 + ], + "type": "inline_equation", + "content": "F_{os} > 1" + }, + { + "bbox": [ + 304, + 607, + 545, + 651 + ], + "type": "text", + "content": ", the soil mass remains stable and will not collapse. If " + }, + { + "bbox": [ + 304, + 607, + 545, + 651 + ], + "type": "inline_equation", + "content": "F_{os} = 1" + }, + { + "bbox": [ + 304, + 607, + 545, + 651 + ], + "type": "text", + "content": ", the soil mass is in a state of critical equilibrium, and if " + }, + { + "bbox": [ + 304, + 607, + 545, + 651 + ], + "type": "inline_equation", + "content": "F_{os} < 1" + }, + { + "bbox": [ + 304, + 607, + 545, + 651 + ], + "type": "text", + "content": ", the soil mass undergoes collapse." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 652, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 652, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 652, + 545, + 688 + ], + "type": "text", + "content": "Assuming each soil mass on the slope is in a critical state, i.e., " + }, + { + "bbox": [ + 304, + 652, + 545, + 688 + ], + "type": "inline_equation", + "content": "F_{os} = 1" + }, + { + "bbox": [ + 304, + 652, + 545, + 688 + ], + "type": "text", + "content": ", if an external force " + }, + { + "bbox": [ + 304, + 652, + 545, + 688 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 652, + 545, + 688 + ], + "type": "text", + "content": " with moment " + }, + { + "bbox": [ + 304, + 652, + 545, + 688 + ], + "type": "inline_equation", + "content": "M_F" + }, + { + "bbox": [ + 304, + 652, + 545, + 688 + ], + "type": "text", + "content": " is applied tangentially along the slope surface such that" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 357, + 690, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 690, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 357, + 690, + 545, + 717 + ], + "type": "interline_equation", + "content": "F _ {o s} ^ {\\prime} = \\frac {M _ {r}}{M _ {s} ^ {\\prime}} = \\frac {M _ {r}}{M _ {s} + M _ {F}} < F _ {o s}, \\tag {4}", + "image_path": "f884ea92c978f25be0383590e556b103971d045cd8fdf36dd568daf356a2e4dc.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11675" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 177 + ], + "type": "text", + "content": "the corresponding soil mass undergoes collapse. This results in the application of a tangential force " + }, + { + "bbox": [ + 47, + 72, + 289, + 177 + ], + "type": "inline_equation", + "content": "F^{\\gamma}" + }, + { + "bbox": [ + 47, + 72, + 289, + 177 + ], + "type": "text", + "content": " on the adjacent soil mass below, leading to a continued collapse of the lower soil mass, ultimately forming a mudslide. The collapse terminates when it reaches a horizontal slope surface. In the end, the mudslide induces the collapse of soil mass on the slope surface affected by external forces, while the unaffected soil mass remains intact. This delineates distinct regions on the slope." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 194, + 180, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 180, + 206 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 180, + 206 + ], + "type": "text", + "content": "3.2. Mudslide Segmentation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 216, + 168, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 216, + 168, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 216, + 168, + 227 + ], + "type": "text", + "content": "3.2.1 Mudslide Processing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 240, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 288, + 361 + ], + "type": "text", + "content": "Inspired by the natural mudslide mentioned in Sec. 3.1, we propose a Mudslide algorithm for nuclear instance segmentation. In natural landslide phenomena, the main issues include the characteristics of the slope that collapses, the forces acting on the soil mass on the slope, the location where the collapse begins, and how the collapse occurs continuously. Corresponding to this, the Mudslide algorithm primarily involves three steps: constructing a collapse field, obtaining the force map and initial collapse boundary, and executing the collapse." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 363, + 289, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 289, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 289, + 615 + ], + "type": "text", + "content": "Collapse field construction. In the first step, we initiate the process by constructing a collapse field, representing the initial slope surface of the mudslide. In image segmentation, we assume that each pixel represents a soil mass, located on a slope represented by the collapse field, and in a state of critical equilibrium. For simplicity, we utilize continuous normalized maps to characterize the collapse field, where pixel values close to 1 signify higher slope heights and values close to 0 indicate lower slope surfaces. Specifically, for ease of model learning and inspired by [6], we create a four-channel normalized map shown in Fig. 3. This map contains four distance maps generated in horizontal, vertical, and two diagonal directions, respectively, with each map indicating the distance from each pixel to the center of the nucleus in different directions. Pixel values are close to 0 when they are near the nucleus center and gradually increase outward until they reach 1 at the nucleus boundary. Thus, we construct a slope for the foreground nuclei in the image, where the slope is flat in the middle of each instance and steep around the edges, resembling the terrain that is prone to mudslide collapse." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "content": "Force map and initial boundary obtaining. For the second step, we generate a force map and initial boundary from the collapse field. The force map assigns a force " + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "content": " with different directions to each pixel, representing the directions in which pixels are expected to collapse during Mudslide. For two-dimensional images, we consider eight directions of force: up, down, left, right, upper-left, upper-right, lower-left, and lower-right (Fig. 3):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 353, + 69, + 545, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 69, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 353, + 69, + 545, + 125 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {F} = \\left\\{\\binom {0} {1}, \\binom {0} {- 1}, \\binom {- 1} {0}, \\binom {1} {0}, \\right. \\\\ \\binom {- 1} {1}, \\binom {1} {1}, \\binom {- 1} {- 1}, \\binom {1} {- 1} \\}. \\end{array} \\tag {5}", + "image_path": "2192f25d582b5518766c3b83129531c3147249ca19005da2c8fed33da8828c9a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 132, + 531, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 132, + 531, + 144 + ], + "spans": [ + { + "bbox": [ + 308, + 132, + 531, + 144 + ], + "type": "text", + "content": "Then the force map can be expressed as a set of forces:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 350, + 147, + 545, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 147, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 350, + 147, + 545, + 160 + ], + "type": "interline_equation", + "content": "F M = \\{F (x, y) \\mid x, y \\in \\mathbb {Z}, F \\in \\mathbb {F} \\}, \\tag {6}", + "image_path": "b83349669b1030171b82453c3813971bc0368d3e3f27849a3ff622e5d0fc3dd4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 169, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 169, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 545, + 251 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 169, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}" + }, + { + "bbox": [ + 304, + 169, + 545, + 251 + ], + "type": "text", + "content": " indicates the foreground of nuclei and " + }, + { + "bbox": [ + 304, + 169, + 545, + 251 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 304, + 169, + 545, + 251 + ], + "type": "text", + "content": " corresponds to the coordinate of each pixel. Here, we assign a force direction to each pixel that is approximately tangential to the slope. Since we assume that each pixel is in a state of critical equilibrium, we consider the force magnitude to be irrelevant to the collapse and therefore set the magnitude of each force to 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 253, + 546, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 546, + 348 + ], + "type": "text", + "content": "To obtain the starting position of mudslide collapse, we generate the initial boundary. We employ a straightforward edge detection filter, such as Sobel operator, on the collapse field to detect edges. The obtained edges are then utilized as the initial boundary for mudslide collapse. This approach is rational since, in our collapse field, there is a significant gradient between instance boundaries, allowing the Sobel operator to effectively identify them." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 348, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 408 + ], + "type": "text", + "content": "Mudslide collapse executing. After acquiring the force map and the initial boundary, we commence the mudslide collapse process. Originating from the initial boundary, each pixel " + }, + { + "bbox": [ + 304, + 348, + 545, + 408 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 348, + 545, + 408 + ], + "type": "text", + "content": " applies an initial force (red arrows in 3) to the directed pixel " + }, + { + "bbox": [ + 304, + 348, + 545, + 408 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 304, + 348, + 545, + 408 + ], + "type": "text", + "content": ", which can be calculated by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 391, + 411, + 458, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 411, + 458, + 426 + ], + "spans": [ + { + "bbox": [ + 391, + 411, + 458, + 426 + ], + "type": "interline_equation", + "content": "p ^ {\\prime} = p + F (p),", + "image_path": "2968cb108179bbf0355e41b4a11ad4b307b4ca5cca58b9f8bde3d70c8a0cda2c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 393, + 429, + 545, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 429, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 393, + 429, + 545, + 441 + ], + "type": "interline_equation", + "content": "p = (x, y), \\tag {7}", + "image_path": "10ee4a652715875e7542b2070ccc96eb5357a021cac60251b696b468185b6879.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 393, + 443, + 449, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 443, + 449, + 459 + ], + "spans": [ + { + "bbox": [ + 393, + 443, + 449, + 459 + ], + "type": "interline_equation", + "content": "p ^ {'} = (x ^ {'}, y ^ {'}).", + "image_path": "ebe8d16ae6a2cbdee9b0e8d5e8393c8d4017c5297fc819dc83ac3197a61e5425.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 468, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 596 + ], + "type": "text", + "content": "The pixel " + }, + { + "bbox": [ + 304, + 468, + 545, + 596 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 304, + 468, + 545, + 596 + ], + "type": "text", + "content": " then undergoes collapse. Subsequently, the collapsed pixel exerts a sustained force (yellow arrows in 3) on the pointed pixel, initiating the collapse of more adjacent pixels. This iterative process results in a sustained mudslide collapse. Eventually, when all pixels directed to have undergone collapse, those pixels collapsing to the same region are identified as a single instance. Note that those pixels that are never directed by forces from other pixels will not collapse and the force assigned to them will be neglected (gray arrows in 3), corresponding to the soil mass that is not affected by the mudslide." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 611, + 535, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 611, + 535, + 623 + ], + "spans": [ + { + "bbox": [ + 305, + 611, + 535, + 623 + ], + "type": "text", + "content": "3.2.2 Application to Nuclear Instance Segmentation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": "The framework of applying Mudslide to nuclear instance segmentation is illustrated in Fig. 4. Here, we unify the pipelines for both fully supervised and point-supervised scenarios, with the primary distinction lying in the labels used for training." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Specifically, we employ a dual-branch architecture, where one branch is dedicated to predicting a binary map" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11676" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 73, + 516, + 215 + ], + "blocks": [ + { + "bbox": [ + 77, + 73, + 516, + 215 + ], + "lines": [ + { + "bbox": [ + 77, + 73, + 516, + 215 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 516, + 215 + ], + "type": "image", + "image_path": "9c569adf6916f15313c6c1194f95af122cf9dcc8702ae0bbe2364719bac39d29.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 77, + 217, + 516, + 352 + ], + "blocks": [ + { + "bbox": [ + 77, + 217, + 516, + 352 + ], + "lines": [ + { + "bbox": [ + 77, + 217, + 516, + 352 + ], + "spans": [ + { + "bbox": [ + 77, + 217, + 516, + 352 + ], + "type": "image", + "image_path": "0feec5d3f43da033cc3dfcd6818f06f5046869aca5f5184773ca9f33312dd437.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 363, + 545, + 398 + ], + "lines": [ + { + "bbox": [ + 46, + 363, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 545, + 398 + ], + "type": "text", + "content": "Figure 4. The overall framework of the proposed Mudslide. The network contains two basic prediction branches: the binary branch and the collapse field branch. During training, the difference between fully- and point-supervised is the type of ground truth. During testing, both branches' predictions are employed for the Mudslide algorithm." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 403, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 582 + ], + "type": "text", + "content": "for the nuclear foreground, and the other branch focuses on predicting the collapse field. The collapse field consists of four channels, corresponding to horizontal, vertical, and two diagonal directions. The segmentation model adopted is the Trans-UNet architecture from [10]. In fully supervised training, the binary map labels are the ground truth representing the nuclear foreground, while the collapse field labels are constructed from both the foreground and instance ground truth. For point-supervised training, we leverage point labels and the original input image to generate a cluster map [19] and Voronoi map [25] as pseudolabels for foreground prediction, which is commonly observed in point-supervised tasks. Subsequently, we use the predicted foreground and point labels to generate collapse field labels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 583, + 287, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 583, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 47, + 583, + 287, + 619 + ], + "type": "text", + "content": "We utilize cross-entropy loss as the binary classification loss and employ mean squared error loss and mean squared gradient error loss as the collapse field loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 625, + 287, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 625, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 129, + 625, + 287, + 639 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {B} = \\lambda_ {B _ {C E}} \\mathcal {L} _ {C E}, \\tag {8}", + "image_path": "fbfa73645d72344b33286a5f288cb71b573f29e2a2e4b058467034cb91e5b616.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 644, + 287, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 644, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 74, + 644, + 287, + 657 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C F} = \\lambda_ {C F _ {M S E}} \\mathcal {L} _ {M S E} + \\lambda_ {C F _ {M S G E}} \\mathcal {L} _ {M S G E}. \\tag {9}", + "image_path": "d405b2ee8f4224c8672c6c8660002172a2e06e73bf8cb2c9d325f9f071309cae.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 50, + 662, + 181, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 662, + 181, + 673 + ], + "spans": [ + { + "bbox": [ + 50, + 662, + 181, + 673 + ], + "type": "text", + "content": "Then the overall loss function is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 680, + 287, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 680, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 131, + 680, + 287, + 693 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {B} + \\mathcal {L} _ {C F}. \\tag {10}", + "image_path": "84d702fdf82a9838f2064d6e0e13c1d00e9e1b0d3e819d27f8ba3c1684bf6a34.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 52, + 701, + 287, + 714 + ], + "type": "text", + "content": "It's worth noting that in point-supervised training, the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 403, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 403, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 545, + 427 + ], + "type": "text", + "content": "cross-entropy loss comprises two terms, corresponding to cluster and Voronoi:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 437, + 545, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 437, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 318, + 437, + 545, + 452 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {B _ {\\text {p o i n t}}} = \\lambda_ {B _ {C E}} \\mathcal {L} _ {C E _ {\\text {c l u s t e r}}} + \\lambda_ {B _ {C E}} \\mathcal {L} _ {C E _ {\\text {v o r o n o i}}}. \\tag {11}", + "image_path": "57421006544e43ec312ea9bc9234f02d8bc2758285d804bd3ea89014ebf54b05.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "content": "During the inference process, we first combine the foreground and collapse field predictions to construct the collapse field, ensuring subsequent mudslide acts on the nuclear foreground regions. Subsequently, as detailed in Sec. 3.1, we utilize the collapse field to obtain the initial boundary and force map, initiating the mudslide collapse process. Ultimately, this process yields the nuclear instance segmentation results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 567, + 382, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 382, + 581 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 382, + 581 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 586, + 432, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 432, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 432, + 601 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Datasets. I) CPM17 [26] contains 64 H&E stained histopathology images with 7,570 annotated nuclear boundaries. It is from the MICCAI 2017 Digital Pathology Challenge [26] and images on two different scales: " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "500 \\times 500" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "600 \\times 600" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". This is a dataset for a single-class segmentation task. II) NuInsSeg [15] contains 665 image patches with more than 30,000 manually segmented nuclei from 31 human and mouse organs. It is one of the biggest fully manually annotated datasets of nuclei in Hematoxylin and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11677" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "type": "text", + "content": "Eosin (H &E) stained histological images with the scale of " + }, + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "type": "text", + "content": ". This is also a dataset for single-class segmentation task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 168 + ], + "type": "text", + "content": "Evaluation Metrics. We use four instance-level evaluation metrics to measure the instance segmentation performance of the comparison models, which are: Detection Quality (DQ), Segmentation Quality (SQ), Panoptic Quality (PQ), and the Aggregated Jaccard Index (AJI)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 169, + 286, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 169, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 286, + 205 + ], + "type": "text", + "content": "Implementation details. The proposed Mudslide is trained and tested using the open-source software library Pytorch 1.8.0 on NVIDIA Tesla V100 with CUDA 10.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "text", + "content": "In order to build the dataset used for training, the CPM17 dataset is divided into multiple patches as it has an inconsistent original image size. We generate 512 training patches in the size of " + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "text", + "content": ", thus allowing the network to be uniform and fully trained. In addition, we randomly divide NuInsSeg into training set and test set with the proportion of " + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "text", + "content": ", respectively. Moreover, for point-supervised tasks, we leverage the annotations of nucleus instances to generate a central point for each instance, serving as the point label. In the training phase, we apply a series of data augmentation methods, which include random flip, random rotation, random color, random crop, and Gaussian blur. The training epoch is set as 300, and the initial learning rate is set as " + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "text", + "content": ". The weights of each loss are " + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "inline_equation", + "content": "\\lambda_{BCE} = 1" + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "inline_equation", + "content": "\\lambda_{CFMSE} = 2.5" + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "inline_equation", + "content": "\\lambda_{CFMSGE} = 8" + }, + { + "bbox": [ + 46, + 205, + 286, + 386 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 393, + 140, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 140, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 140, + 406 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 413, + 286, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 286, + 460 + ], + "type": "text", + "content": "We performed ablation studies on the CPM17 and NuInsSeg datasets to validate the efficacy of the proposed Mudslide method. All the following experiments were conducted under the setting of fully supervised tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 461, + 286, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 286, + 544 + ], + "type": "text", + "content": "Instance separation methods. Table 1 shows the comparison of segmentation performance obtained using different instance separation methods. We apply four different methods to the predictions of the same model: the traditional watershed method [1], boundary partition method, HVmap segmentation method from Hover-Net [6], and our proposed method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 286, + 652 + ], + "type": "text", + "content": "Specifically, the watershed method involves segmenting instances using the watershed algorithm on the binary probability map predicted by the model. The boundary separation method refers to predicting boundaries by adding an additional boundary prediction branch to the model and using the boundaries to directly separate different instances. The HVmap algorithm is an instance separation method in HoverNet that utilizes horizontal and vertical distance maps for segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": "The experimental results indicate significant variations in model performance based on different instance separation methods. Traditional methods such as watershed and boundary separation exhibit subpar performance. The HVmap approach, utilizing horizontal and vertical maps," + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 72, + 542, + 149 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 542, + 149 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 542, + 149 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 542, + 149 + ], + "type": "table", + "html": "
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
Watershed0.7970.7300.5820.6940.6010.7350.4610.502
Boundary0.8370.7340.6140.7140.6400.7530.4910.532
HVMap0.8790.8100.7130.7370.7860.7540.5960.638
Ours0.8830.8080.7140.7400.7880.7560.5990.647
", + "image_path": "3c9744391d34e2065923ccfefb3582e297d552987a984d85f67e52f7a15a65c0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 194, + 550, + 258 + ], + "blocks": [ + { + "bbox": [ + 305, + 156, + 545, + 189 + ], + "lines": [ + { + "bbox": [ + 305, + 156, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 545, + 189 + ], + "type": "text", + "content": "Table 1. Comparison of different adjacent instance separation methods, including the watershed algorithm, boundary partition, HVmap and our method." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 194, + 550, + 258 + ], + "lines": [ + { + "bbox": [ + 306, + 194, + 550, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 194, + 550, + 258 + ], + "type": "table", + "html": "
boundcollapseCPM17NuInsSeg
DQ ↑SQ ↑PQ ↑AJI ↑DQ ↑SQ ↑PQ ↑AJI ↑
✓0.7720.7100.5480.6100.5870.7030.4140.497
0.8770.8070.7070.7330.7750.7520.5820.630
✓✓0.8830.8080.7140.7400.7880.7560.5990.647
", + "image_path": "efada8d7a07132abd73fb0b8055fd3980c6db39e5f82598c84d1c08fed605078.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 266, + 545, + 299 + ], + "lines": [ + { + "bbox": [ + 305, + 266, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 545, + 299 + ], + "type": "text", + "content": "Table 2. Ablation studies of components in Mudslide algorithm, where bound indicates initial boundary, and collapse indicates mudslide collapse processing." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 304, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 365 + ], + "type": "text", + "content": "enhances the differentiation between instances and achieves better results. Furthermore, our method, with more explicit instance differentiation representation and a more rational instance acquisition process, outperforms other methods. This underscores the superiority of our Mudslide approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 365, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 533 + ], + "type": "text", + "content": "The ablation of components of Mudslide algorithm. We conducted ablation experiments on the key steps of the mudslide algorithm. Here, we considered two components: whether to use the initial boundary and whether to perform mudslide collapse. In Table 2, we explored three comparative experiments: a baseline approach with no additional components, relying solely on binary map partitioning based on connected regions; a scenario involving only the initial boundary extracted from the collapse field; and a comprehensive application of both the initial boundary and mudslide collapse, representing our full-fledged mudslide algorithm for instance segmentation. These experiments aim to scrutinize the individual contributions of each component to the overall segmentation performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "text", + "content": "The experimental results indicate that incorporating an initial boundary leads to performance gains of " + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "text", + "content": " PQ and " + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "text", + "content": " AJI for the CPM dataset and " + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "text", + "content": " PQ and " + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "text", + "content": " AJI for NuInsSeg. Furthermore, the additional mudslide collapse results in a performance boost of " + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "text", + "content": " PQ and " + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "text", + "content": " AJI for the CPM dataset and " + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "text", + "content": " PQ and " + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 305, + 533, + 545, + 640 + ], + "type": "text", + "content": " AJI for NuInsSeg. These findings underscore the substantial benefits of each component within the mudslide method for enhancing instance segmentation outcomes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 642, + 545, + 712 + ], + "type": "text", + "content": "Force map construction methods. As a crucial manifestation of distinct instance differentiation, the construction of the force map is pivotal in the Mudslide algorithm. Therefore, we consider three methods for building the force map: direct prediction, collapse field segmentation, and collapse field extraction." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11678" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 73, + 288, + 137 + ], + "blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 137 + ], + "lines": [ + { + "bbox": [ + 48, + 73, + 288, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 288, + 137 + ], + "type": "table", + "html": "
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
predict0.8740.8070.7070.7350.7770.7510.5870.633
segment0.8640.8080.7010.7280.7720.7530.5820.629
extract0.8800.8070.7110.7390.7840.7550.5920.643
", + "image_path": "37a2ec5dab2c41eba39c37924a6fd0bd05236b0d0b0cf40d4bdda38cf8d6bad8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 173, + 288, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 173, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 173, + 288, + 353 + ], + "type": "text", + "content": "In the first approach, we have the model predict the force map directly. Specifically, we augment the model with a branch that outputs nine channels, allowing it to learn the force map's eight directions and the background region. Subsequently, we obtain the force map by employing the argmax function to the predicted probability map. In the second approach, we perform segmentation on the collapse field. Here, we use a two-channel collapse field corresponding to the horizontal and vertical directions. By segmenting the collapse field based on angles, we categorize it into eight regions, representing the eight directions of the force map. In the third approach, we predict a four-channel collapse field for horizontal, vertical, and two diagonal directions. Then, we directly employ argmax to extract the direction of force that each pixel should possess." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 353, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 424 + ], + "type": "text", + "content": "The results presented in Table 3 indicate that the force map extracted through four-channel collapse fields performs the best. The direct prediction follows with slightly lower performance, while the force map obtained by splitting the two-channel collapse field performs the least favorably." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "text", + "content": "The four-channel collapsing field extraction method outperformed direct prediction by " + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "text", + "content": " PQ and " + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "text", + "content": " AJI on CPM17, and outperformed the two-channel segmentation method by " + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "text", + "content": " PQ, and " + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "text", + "content": " AJI. On the NuInsSeg dataset, it outperformed direct prediction by " + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "text", + "content": " PQ, and " + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "text", + "content": " AJI, and outperformed the two-channel segmentation method by " + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "text", + "content": " PQ, and " + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 46, + 425, + 287, + 556 + ], + "type": "text", + "content": " AJI. These results suggest that the four-channel collapsing field can better capture and characterize the differences between cell nucleus instances, resulting in greater differences between adjacent instances and better instance segmentation results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "type": "text", + "content": "Initial boundary generation methods. Regarding the issue of generating and selecting initial boundaries, we considered two methods: direct generation from the binary prediction map and generation from the collapsing field." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Specifically, for the first method, we first obtained the binary probability map predicted by the model's binary classification branch and then selected a threshold of 0.5 to segment the predicted foreground and background regions of the nucleus. Then, we used the part of the pixels that contacted the background, i.e., the edge of the foreground, as the initial boundary. For the second method, we first obtained the four-channel collapsing field predicted by the model and then used some edge extraction methods, such as" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 73, + 548, + 126 + ], + "blocks": [ + { + "bbox": [ + 47, + 145, + 287, + 167 + ], + "lines": [ + { + "bbox": [ + 47, + 145, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 145, + 287, + 167 + ], + "type": "text", + "content": "Table 3. Comparison of different force map construction methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 73, + 548, + 126 + ], + "lines": [ + { + "bbox": [ + 307, + 73, + 548, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 548, + 126 + ], + "type": "table", + "html": "
MethodCPMNuInsSeg
DQ ↑SQ ↑PQ ↑AJI ↑DQ ↑SQ ↑PQ ↑AJI ↑
binary map collapse field0.8800.8070.7110.7390.7840.7550.5920.643
0.8830.8080.7140.7400.7880.7560.5990.647
", + "image_path": "b55debf562f8ef3a2c33162710024a48a25ace8802db1d2135d7c9029f2978c5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 133, + 545, + 155 + ], + "lines": [ + { + "bbox": [ + 305, + 133, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 545, + 155 + ], + "type": "text", + "content": "Table 4. Comparison of different initial boundary generation methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 161, + 545, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 161, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 161, + 545, + 185 + ], + "type": "text", + "content": "using the Sobel operator to extract the regions with higher gradients in the collapsing field as the initial boundary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "type": "text", + "content": "As shown in Table 4, the method of generating initial boundaries using the collapsing field outperforms the method of using only the binary prediction map. Specifically, on the CPM dataset, the collapsing field method improved performance by " + }, + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "type": "text", + "content": " PQ, and " + }, + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "type": "text", + "content": " AJI compared to the binary map method. On the NuInsSeg dataset, it improved performance by " + }, + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "type": "text", + "content": " PQ, and " + }, + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "type": "text", + "content": " AJI. As anticipated, the construction of the initial boundary is crucial in our Mudslide collapse algorithm, as it determines the algorithm's starting position. Obtaining foreground edges directly from a binary image makes it challenging to capture the boundaries between adjacent instances. Although the Mudslide algorithm can capture this difference to some extent, its performance is not as effective as constructing a better initial boundary. The initial boundary extracted through the collapse field takes into account the differentiated representations between cell nuclei, effectively separating nuclei to a certain degree and proving to be a superior boundary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 407, + 508, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 508, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 508, + 419 + ], + "type": "text", + "content": "4.3. Fully-Supervised Nuclei Segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 425, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 509 + ], + "type": "text", + "content": "In Table 5, the proposed Mudslide is compared with the state-of-the-art fully-supervised nuclear instance segmentation models on CPM17 and NuInsSeg. Our Mudslide achieves 0.714 PQ and 0.740 AJI on the CPM17 dataset, and 0.599 PQ and 0.647 AJI on the NuInsSeg dataset, respectively. The Mudslide achieves competitive results for fully-supervised nuclei segmentation tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "type": "text", + "content": "We further carry out a qualitative visualization analysis of our method and different existing methods. Compared to all the other methods in Fig. 5, our Mudslide achieves superior instance segmentation results for nuclear instance segmentation, especially in some regions where nuclei are densely distributed with a large amount of overlap, where our instance segmentation is more accurate. The results, both quantitative and qualitative, unequivocally establish the competitiveness of our Mudslide algorithm in accurately segmenting densely packed and overlapping nuclear instances." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 647, + 517, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 647, + 517, + 660 + ], + "spans": [ + { + "bbox": [ + 305, + 647, + 517, + 660 + ], + "type": "text", + "content": "4.4. Pointly-Supervised Nuclei Segmentation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "In Table 6, we extend the proposed Mudslide from fully-supervised to point-supervised nuclear segmentation task and compare it with the state-of-the-art point-supervised nuclear instance segmentation models on CPM17 and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11679" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 71, + 140, + 131 + ], + "blocks": [ + { + "bbox": [ + 80, + 71, + 140, + 131 + ], + "lines": [ + { + "bbox": [ + 80, + 71, + 140, + 131 + ], + "spans": [ + { + "bbox": [ + 80, + 71, + 140, + 131 + ], + "type": "image", + "image_path": "aaaa8803377b0f6c6a24fd5d1dab9dc5b6f305fabb1f24b38a7f39e42a51b463.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 80, + 132, + 139, + 190 + ], + "blocks": [ + { + "bbox": [ + 80, + 132, + 139, + 190 + ], + "lines": [ + { + "bbox": [ + 80, + 132, + 139, + 190 + ], + "spans": [ + { + "bbox": [ + 80, + 132, + 139, + 190 + ], + "type": "image", + "image_path": "125633ffe895597bc4643391ed91a588962169c02a77af5e296c475763189d63.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 80, + 191, + 139, + 251 + ], + "blocks": [ + { + "bbox": [ + 80, + 191, + 139, + 251 + ], + "lines": [ + { + "bbox": [ + 80, + 191, + 139, + 251 + ], + "spans": [ + { + "bbox": [ + 80, + 191, + 139, + 251 + ], + "type": "image", + "image_path": "66186f44fb71f4cff435397bf4122e9138c00fb64c710f2132b49c8687cb8f6e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 80, + 251, + 139, + 310 + ], + "blocks": [ + { + "bbox": [ + 80, + 251, + 139, + 310 + ], + "lines": [ + { + "bbox": [ + 80, + 251, + 139, + 310 + ], + "spans": [ + { + "bbox": [ + 80, + 251, + 139, + 310 + ], + "type": "image", + "image_path": "64ee2f5006a77cbef1b94e2ac689ac7aa3e76163828a5181e528f0be8fff67f7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 312, + 114, + 320 + ], + "lines": [ + { + "bbox": [ + 103, + 312, + 114, + 320 + ], + "spans": [ + { + "bbox": [ + 103, + 312, + 114, + 320 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 94, + 323, + 120, + 333 + ], + "lines": [ + { + "bbox": [ + 94, + 323, + 120, + 333 + ], + "spans": [ + { + "bbox": [ + 94, + 323, + 120, + 333 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 342, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 545, + 365 + ], + "type": "text", + "content": "Figure 5. Visualization of instance segmentation results with different methods. (a) Input image; (b) Ground Truth; (c) U-Net [23]; (d) DIST [16]; (e) FullNet [20]; (f) Hover-Net [6]; (g) Our Mudslide." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 143, + 72, + 202, + 131 + ], + "blocks": [ + { + "bbox": [ + 143, + 72, + 202, + 131 + ], + "lines": [ + { + "bbox": [ + 143, + 72, + 202, + 131 + ], + "spans": [ + { + "bbox": [ + 143, + 72, + 202, + 131 + ], + "type": "image", + "image_path": "5708ec99f9a7bd92be593b662715669722c556dc722f51b11b483b6c3f6b367a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 143, + 132, + 202, + 190 + ], + "blocks": [ + { + "bbox": [ + 143, + 132, + 202, + 190 + ], + "lines": [ + { + "bbox": [ + 143, + 132, + 202, + 190 + ], + "spans": [ + { + "bbox": [ + 143, + 132, + 202, + 190 + ], + "type": "image", + "image_path": "f8e0f9de5635e1810e045b1aa2108fbd2d5074244e951531ad6531f859986121.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 143, + 191, + 202, + 251 + ], + "blocks": [ + { + "bbox": [ + 143, + 191, + 202, + 251 + ], + "lines": [ + { + "bbox": [ + 143, + 191, + 202, + 251 + ], + "spans": [ + { + "bbox": [ + 143, + 191, + 202, + 251 + ], + "type": "image", + "image_path": "471482f1465a64ceb3264a71576a892c45c19dbd6c0b6c5be3d0b76ae4dc7064.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 143, + 251, + 202, + 310 + ], + "blocks": [ + { + "bbox": [ + 143, + 251, + 202, + 310 + ], + "lines": [ + { + "bbox": [ + 143, + 251, + 202, + 310 + ], + "spans": [ + { + "bbox": [ + 143, + 251, + 202, + 310 + ], + "type": "image", + "image_path": "33b1f46c820e1b3400dbfe15d59d3af41002023a6930210126cee8fd86796439.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 312, + 179, + 320 + ], + "lines": [ + { + "bbox": [ + 168, + 312, + 179, + 320 + ], + "spans": [ + { + "bbox": [ + 168, + 312, + 179, + 320 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 142, + 323, + 200, + 332 + ], + "lines": [ + { + "bbox": [ + 142, + 323, + 200, + 332 + ], + "spans": [ + { + "bbox": [ + 142, + 323, + 200, + 332 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 205, + 72, + 264, + 131 + ], + "blocks": [ + { + "bbox": [ + 205, + 72, + 264, + 131 + ], + "lines": [ + { + "bbox": [ + 205, + 72, + 264, + 131 + ], + "spans": [ + { + "bbox": [ + 205, + 72, + 264, + 131 + ], + "type": "image", + "image_path": "a86983016e9228e092eba7224f62db1c39d22ee49145b8b8961bb7f3253deb5f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 205, + 132, + 264, + 190 + ], + "blocks": [ + { + "bbox": [ + 205, + 132, + 264, + 190 + ], + "lines": [ + { + "bbox": [ + 205, + 132, + 264, + 190 + ], + "spans": [ + { + "bbox": [ + 205, + 132, + 264, + 190 + ], + "type": "image", + "image_path": "9a0eacc8d16e1ea0e5159b7149451ee06cf4f02c391257ac23903cc7ba215618.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 205, + 191, + 264, + 251 + ], + "blocks": [ + { + "bbox": [ + 205, + 191, + 264, + 251 + ], + "lines": [ + { + "bbox": [ + 205, + 191, + 264, + 251 + ], + "spans": [ + { + "bbox": [ + 205, + 191, + 264, + 251 + ], + "type": "image", + "image_path": "a9ec58cd0715b9fb0f03ee8d06e8c3b7ea79c822bf532480fbddaff1e9d24fef.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 205, + 251, + 264, + 310 + ], + "blocks": [ + { + "bbox": [ + 205, + 251, + 264, + 310 + ], + "lines": [ + { + "bbox": [ + 205, + 251, + 264, + 310 + ], + "spans": [ + { + "bbox": [ + 205, + 251, + 264, + 310 + ], + "type": "image", + "image_path": "9635bf5f7691c1ecb44c4d55c2d3f88607f6a712d63e2203c6119a60c5c6a1d0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 312, + 239, + 320 + ], + "lines": [ + { + "bbox": [ + 228, + 312, + 239, + 320 + ], + "spans": [ + { + "bbox": [ + 228, + 312, + 239, + 320 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 224, + 323, + 249, + 332 + ], + "lines": [ + { + "bbox": [ + 224, + 323, + 249, + 332 + ], + "spans": [ + { + "bbox": [ + 224, + 323, + 249, + 332 + ], + "type": "text", + "content": "U-Net" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 267, + 72, + 326, + 131 + ], + "blocks": [ + { + "bbox": [ + 267, + 72, + 326, + 131 + ], + "lines": [ + { + "bbox": [ + 267, + 72, + 326, + 131 + ], + "spans": [ + { + "bbox": [ + 267, + 72, + 326, + 131 + ], + "type": "image", + "image_path": "c546fdd12524abdc307ba770590f52c533c69a9fbce658069503f8ba0d6b355c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 267, + 132, + 326, + 190 + ], + "blocks": [ + { + "bbox": [ + 267, + 132, + 326, + 190 + ], + "lines": [ + { + "bbox": [ + 267, + 132, + 326, + 190 + ], + "spans": [ + { + "bbox": [ + 267, + 132, + 326, + 190 + ], + "type": "image", + "image_path": "5e8e0d7f7f354a03cce6e8f392facf7583d63da51f173be597e105406ccf43bb.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 267, + 191, + 326, + 251 + ], + "blocks": [ + { + "bbox": [ + 267, + 191, + 326, + 251 + ], + "lines": [ + { + "bbox": [ + 267, + 191, + 326, + 251 + ], + "spans": [ + { + "bbox": [ + 267, + 191, + 326, + 251 + ], + "type": "image", + "image_path": "c696812f6984170a616234ed6f45cdf7ed30aef9789aa2a4c77fc9159cd06427.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 267, + 251, + 326, + 310 + ], + "blocks": [ + { + "bbox": [ + 267, + 251, + 326, + 310 + ], + "lines": [ + { + "bbox": [ + 267, + 251, + 326, + 310 + ], + "spans": [ + { + "bbox": [ + 267, + 251, + 326, + 310 + ], + "type": "image", + "image_path": "2a4cbb4925647bab34b1217499e374e85f8a6a9ad8e7606b866a42d0521cbeae.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 292, + 312, + 303, + 320 + ], + "lines": [ + { + "bbox": [ + 292, + 312, + 303, + 320 + ], + "spans": [ + { + "bbox": [ + 292, + 312, + 303, + 320 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 287, + 323, + 309, + 332 + ], + "lines": [ + { + "bbox": [ + 287, + 323, + 309, + 332 + ], + "spans": [ + { + "bbox": [ + 287, + 323, + 309, + 332 + ], + "type": "text", + "content": "DIST" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 329, + 72, + 388, + 131 + ], + "blocks": [ + { + "bbox": [ + 329, + 72, + 388, + 131 + ], + "lines": [ + { + "bbox": [ + 329, + 72, + 388, + 131 + ], + "spans": [ + { + "bbox": [ + 329, + 72, + 388, + 131 + ], + "type": "image", + "image_path": "77c8e5a2c67baed7d469b070827aff673b6d098e11b7dd5c4665373a542f8434.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 329, + 132, + 388, + 190 + ], + "blocks": [ + { + "bbox": [ + 329, + 132, + 388, + 190 + ], + "lines": [ + { + "bbox": [ + 329, + 132, + 388, + 190 + ], + "spans": [ + { + "bbox": [ + 329, + 132, + 388, + 190 + ], + "type": "image", + "image_path": "ca768617ffed6e199573c093cf47f7e46ba29519c54ed06146b2c0082859e67a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 329, + 191, + 388, + 251 + ], + "blocks": [ + { + "bbox": [ + 329, + 191, + 388, + 251 + ], + "lines": [ + { + "bbox": [ + 329, + 191, + 388, + 251 + ], + "spans": [ + { + "bbox": [ + 329, + 191, + 388, + 251 + ], + "type": "image", + "image_path": "4198c1409f1634d95b6ef460ee8c7e2e896dfe9b96496764acf57e2fbac4b4db.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 329, + 251, + 388, + 310 + ], + "blocks": [ + { + "bbox": [ + 329, + 251, + 388, + 310 + ], + "lines": [ + { + "bbox": [ + 329, + 251, + 388, + 310 + ], + "spans": [ + { + "bbox": [ + 329, + 251, + 388, + 310 + ], + "type": "image", + "image_path": "2c862aa2c699d6a5f9791afb001fdc49850fe37b802bc9d99202c77615337796.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 312, + 366, + 320 + ], + "lines": [ + { + "bbox": [ + 356, + 312, + 366, + 320 + ], + "spans": [ + { + "bbox": [ + 356, + 312, + 366, + 320 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 348, + 323, + 378, + 332 + ], + "lines": [ + { + "bbox": [ + 348, + 323, + 378, + 332 + ], + "spans": [ + { + "bbox": [ + 348, + 323, + 378, + 332 + ], + "type": "text", + "content": "FullNet" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 391, + 72, + 449, + 131 + ], + "blocks": [ + { + "bbox": [ + 391, + 72, + 449, + 131 + ], + "lines": [ + { + "bbox": [ + 391, + 72, + 449, + 131 + ], + "spans": [ + { + "bbox": [ + 391, + 72, + 449, + 131 + ], + "type": "image", + "image_path": "c235508c663315e854bbf20342a31c0e3d18200666c0cf848c8dcc4c398d3a1a.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 391, + 132, + 449, + 190 + ], + "blocks": [ + { + "bbox": [ + 391, + 132, + 449, + 190 + ], + "lines": [ + { + "bbox": [ + 391, + 132, + 449, + 190 + ], + "spans": [ + { + "bbox": [ + 391, + 132, + 449, + 190 + ], + "type": "image", + "image_path": "2305446ffebd45dcf7ef74e5a225275d19c371c6643297da4930f31863cb7418.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 391, + 191, + 449, + 251 + ], + "blocks": [ + { + "bbox": [ + 391, + 191, + 449, + 251 + ], + "lines": [ + { + "bbox": [ + 391, + 191, + 449, + 251 + ], + "spans": [ + { + "bbox": [ + 391, + 191, + 449, + 251 + ], + "type": "image", + "image_path": "c98a02cf5d83536d7093b2a28e8c8fce871f01fdbf608bde6b7418eb49b79aef.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 391, + 251, + 449, + 310 + ], + "blocks": [ + { + "bbox": [ + 391, + 251, + 449, + 310 + ], + "lines": [ + { + "bbox": [ + 391, + 251, + 449, + 310 + ], + "spans": [ + { + "bbox": [ + 391, + 251, + 449, + 310 + ], + "type": "image", + "image_path": "8fccbf4d87eb125159654ab79847fd67d18784c8398e5473bb68990989bf3748.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 312, + 424, + 320 + ], + "lines": [ + { + "bbox": [ + 414, + 312, + 424, + 320 + ], + "spans": [ + { + "bbox": [ + 414, + 312, + 424, + 320 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 323, + 440, + 332 + ], + "lines": [ + { + "bbox": [ + 400, + 323, + 440, + 332 + ], + "spans": [ + { + "bbox": [ + 400, + 323, + 440, + 332 + ], + "type": "text", + "content": "HoverNet" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 452, + 72, + 512, + 131 + ], + "blocks": [ + { + "bbox": [ + 452, + 72, + 512, + 131 + ], + "lines": [ + { + "bbox": [ + 452, + 72, + 512, + 131 + ], + "spans": [ + { + "bbox": [ + 452, + 72, + 512, + 131 + ], + "type": "image", + "image_path": "fec93d31d352012b7d08a188da620dc3d461ecf2147f978316311c66bb7e8e90.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 452, + 132, + 512, + 190 + ], + "blocks": [ + { + "bbox": [ + 452, + 132, + 512, + 190 + ], + "lines": [ + { + "bbox": [ + 452, + 132, + 512, + 190 + ], + "spans": [ + { + "bbox": [ + 452, + 132, + 512, + 190 + ], + "type": "image", + "image_path": "cc91497ccdb1b9158db1921f1a40996505e3125c6ca6f478f79e278cb3228410.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 452, + 191, + 512, + 251 + ], + "blocks": [ + { + "bbox": [ + 452, + 191, + 512, + 251 + ], + "lines": [ + { + "bbox": [ + 452, + 191, + 512, + 251 + ], + "spans": [ + { + "bbox": [ + 452, + 191, + 512, + 251 + ], + "type": "image", + "image_path": "2cce860532ab71e7280f8d24c59d12eaaedb06f6eead5cfb512f34a26783dbd9.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 452, + 251, + 512, + 310 + ], + "blocks": [ + { + "bbox": [ + 452, + 251, + 512, + 310 + ], + "lines": [ + { + "bbox": [ + 452, + 251, + 512, + 310 + ], + "spans": [ + { + "bbox": [ + 452, + 251, + 512, + 310 + ], + "type": "image", + "image_path": "943ca03733899e071318603229e7bc7759770684b09b126c4bad9fc805c1370c.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 312, + 487, + 320 + ], + "lines": [ + { + "bbox": [ + 477, + 312, + 487, + 320 + ], + "spans": [ + { + "bbox": [ + 477, + 312, + 487, + 320 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 473, + 323, + 493, + 332 + ], + "lines": [ + { + "bbox": [ + 473, + 323, + 493, + 332 + ], + "spans": [ + { + "bbox": [ + 473, + 323, + 493, + 332 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "table", + "bbox": [ + 48, + 372, + 291, + 497 + ], + "blocks": [ + { + "bbox": [ + 48, + 372, + 291, + 497 + ], + "lines": [ + { + "bbox": [ + 48, + 372, + 291, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 291, + 497 + ], + "type": "table", + "html": "
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
UNet [23]0.7780.7340.5780.6430.6350.6880.4360.508
DCAN [3]0.7320.7400.5450.5610.5630.6440.3630.435
DIST [16]0.6630.7540.5040.6160.7030.7280.5120.542
FullNet [20]--0.6610.6680.7120.7600.5410.562
HoverNet [6]0.8540.8140.6970.7050.7430.7640.5680.611
CDNet [7]---0.733---0.640
CellViT [10]0.8790.8100.7130.7370.7860.7540.5960.638
Ours0.8830.8080.7140.7400.7880.7560.5990.647
", + "image_path": "f67019a6e7cee528e43d8300720ef48dbc42199e1f437f1d971417e81998e3ac.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "table_body" + } + ], + "index": 43 + }, + { + "bbox": [ + 46, + 534, + 287, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 595 + ], + "type": "text", + "content": "NuInsSeg. Our Mudslide achieves 0.529 PQ and 0.581 AJI on the CPM17 dataset, and 0.456 PQ and 0.558 AJI on the NuInsSeg dataset, respectively. In the task of nuclear instance segmentation with only point annotations, our method can also achieve competitive results." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 46, + 596, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 596, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 596, + 287, + 643 + ], + "type": "text", + "content": "Although methods relying solely on point annotations still have ample room for improvement compared to fully supervised approaches, our approach has already surpassed other methods in the realm of point annotations." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 47, + 656, + 119, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 119, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 119, + 669 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "This paper proposes a novel universal approach for nuclear instance segmentation, called Mudslide, which is inspired by the natural phenomenon of mudslides. To enhance" + } + ] + } + ], + "index": 48 + }, + { + "type": "table", + "bbox": [ + 306, + 372, + 555, + 473 + ], + "blocks": [ + { + "bbox": [ + 46, + 504, + 287, + 527 + ], + "lines": [ + { + "bbox": [ + 46, + 504, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 287, + 527 + ], + "type": "text", + "content": "Table 5. Performance comparisons on fully-supervised nuclear segmentation." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 372, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 306, + 372, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 555, + 473 + ], + "type": "table", + "html": "
MethodCPMNuInsSeg
DQ↑SQ↑PQ↑AJI↑DQ↑SQ↑PQ↑AJI↑
PENet [27]0.4400.7220.3190.2900.4760.6300.3000.355
C2FNet [25]0.4290.6300.2710.4210.4050.5800.2350.400
S2Label [11]0.6540.6980.4590.5030.6130.6670.4090.496
SPN+IEN [13]0.7040.6890.4880.5430.6880.6600.4540.523
SCNet [12]0.7160.6850.4960.5540.6940.6550.4560.526
Ours0.7470.7030.5290.5810.7070.6890.4870.558
", + "image_path": "a7902bba82866f8621ae791c56289d6a61513e05598eb855c691d77365180992.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "table_body" + } + ], + "index": 49 + }, + { + "bbox": [ + 306, + 480, + 545, + 503 + ], + "lines": [ + { + "bbox": [ + 306, + 480, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 480, + 545, + 503 + ], + "type": "text", + "content": "Table 6. Performance comparisons on point supervised nuclear segmentation." + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "type": "text", + "content": "the model's capability to differentiate between adjacent instances, we devise a collapsing field, and generate a force map along with an initial boundary, offering distinct representations for different instances and a unified representation for the same instance. We also propose Mudslide collapse algorithm that mimics the natural process by segmenting through the iterative collapse of individual pixels. Our collapse fields are simple and easily extendable to point-supervised tasks. The experimental results on public nuclei datasets demonstrate that Mudslide outperforms state-of-the-art methods." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgments This work was supported in part by the National Key R&D Program of China (No. 2022ZD0118201), Natural Science Foundation of China (No.61972217, 32071459, 62176249, 62006133, 62271465), and Shenzhen Medical Research Funds in China (No. B2302037)." + } + ] + } + ], + "index": 52 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11680" + } + ] + } + ], + "index": 53 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Serge Beucher. Use of watersheds in contour detection. In Proc. Int. Workshop on Image Processing, Sept. 1979, pages 17-21, 1979. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 169 + ], + "type": "text", + "content": "[2] Alireza Chamanzar and Yao Nie. Weakly supervised multitask learning for cell detection and segmentation. In 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI), pages 513-516. IEEE, 2020. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "type": "text", + "content": "[3] Hao Chen, Xiaojuan Qi, Lequan Yu, Qi Dou, Jing Qin, and Pheng-Ann Heng. Dcan: Deep contour-aware networks for object instance segmentation from histology images. Medical Image Analysis, 36:135-146, 2017. 1, 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 261 + ], + "type": "text", + "content": "[4] Shengcong Chen, Changxing Ding, Minfeng Liu, Jun Cheng, and Dacheng Tao. CPP-net: Context-aware polygon proposal network for nucleus segmentation. IEEE Transactions on Image Processing, 32:980-994, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 287, + 296 + ], + "type": "text", + "content": "[5] FC Dai, Chin Fei Lee, and Y Yip Ngai. Landslide risk assessment and management: an overview. Engineering Geology, 64(1):65-87, 2002. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 297, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 287, + 352 + ], + "type": "text", + "content": "[6] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:1-15, 2019. 2, 4, 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 354, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 287, + 418 + ], + "type": "text", + "content": "[7] Hongliang He, Zhongyi Huang, Yao Ding, Guoli Song, Lin Wang, Qian Ren, Pengxu Wei, Zhiqiang Gao, and Jie Chen. Cdnet: Centripetal direction network for nuclear instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4026-4035, 2021. 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 421, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 421, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 421, + 287, + 475 + ], + "type": "text", + "content": "[8] Hongliang He, Chi Zhang, Jie Chen, Ruizhe Geng, Luyang Chen, Yongsheng Liang, Yanchang Lu, Jihua Wu, and Yongjie Xu. A hybrid-attention nested unet for nuclear segmentation in histopathological images. Frontiers in Molecular Biosciences, 8:1–9, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 478, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 287, + 533 + ], + "type": "text", + "content": "[9] Hongliang He, Jun Wang, Pengxu Wei, Fan Xu, Xiangyang Ji, Chang Liu, and Jie Chen. Toposeg: Topology-aware nuclear instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21307-21316, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 589 + ], + "type": "text", + "content": "[10] Fabian Horst, Moritz Rempe, Lukas Heine, Constantin Seibold, Julius Keyl, Giulia Baldini, Selma Ugurel, Jens Siveke, Barbara Grünwald, Jan Egger, et al. Cellvit: Vision transformers for precise cell segmentation and classification. arXiv preprint arXiv:2306.15350, 2023. 5, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 657 + ], + "type": "text", + "content": "[11] Hyeonsoo Lee and Won-Ki Jeong. Scribble2label: Scribblesupervised cell segmentation via self-generating pseudolabels with consistency. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23, pages 14-23. Springer, 2020. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[12] Yi Lin, Zhiyong Qu, Hao Chen, Zhongke Gao, Yuexiang Li, Lili Xia, Kai Ma, Yefeng Zheng, and Kwang-Ting Cheng. Nuclei segmentation with point annotations from pathology images via self-supervised learning and co-training. Medical Image Analysis, 89:102933, 2023. 3, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[13] Weizhen Liu, Qian He, and Xuming He. Weakly supervised nuclei segmentation via instance learning. In 2022 IEEE 19th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2022. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "text", + "content": "[14] Xiaoming Liu, Zhengsheng Guo, Jun Cao, and Jinshan Tang. Mdc-net: a new convolutional neural network for nucleus segmentation in histopathology images with distance maps and contour information. Computers in Biology and Medicine, 135:1-12, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 173, + 547, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 547, + 237 + ], + "type": "text", + "content": "[15] Amirreza Mahbod, Christine Polak, Katharina Feldmann, Rumsha Khan, Katharina Gelles, Georg Dorffner, Ramona Woitek, Sepideh Hatamikia, and Isabella Ellinger. Nuinsseg: A fully annotated dataset for nuclei instance segmentation in h&E-stained histological images. arXiv preprint arXiv:2308.01760, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 239, + 547, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 547, + 282 + ], + "type": "text", + "content": "[16] Peter Naylor, Marick Laé, Fabien Reyal, and Thomas Walter. Segmentation of nuclei in histopathology images by deep regression of the distance map. IEEE Transactions on Medical Imaging, 38(2):448-459, 2018. 2, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 283, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 547, + 304 + ], + "type": "text", + "content": "[17] Hasmukh Pranshanker Oza and HP Oza. Soil mechanics and foundation engineering. Charotar Book Stall, 1969. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 305, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 545, + 350 + ], + "type": "text", + "content": "[18] H. Pinckaers, W. Bulten, J. van der Laak, and G. Litjens. Detection of prostate cancer in whole-slide images through end-to-end training with image-level labels. IEEE Transactions on Medical Imaging, 40(7):1817-1826, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 350, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 415 + ], + "type": "text", + "content": "[19] Hui Qu, Pengxiang Wu, Qiaoying Huang, Jingru Yi, Gregory M Riedlinger, Subhajyoti De, and Dimitris N Metaxas. Weakly supervised deep nuclei segmentation using points annotation in histopathology images. In International Conference on Medical Imaging with Deep Learning, pages 390-400. PMLR, 2019. 1, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 416, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 491 + ], + "type": "text", + "content": "[20] Hui Qu, Zhennan Yan, Gregory M Riedlinger, Subhajyoti De, and Dimitris N Metaxas. Improving nuclei/gland instance segmentation in histopathology images by full resolution neural network and spatial constrained loss. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 378-386, 2019. 1, 2, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 493, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 558 + ], + "type": "text", + "content": "[21] Hui Qu, Pengxiang Wu, Qiaoying Huang, Jingru Yi, Zhennan Yan, Kang Li, Gregory M Riedlinger, Subhajyoti De, Shaoting Zhang, and Dimitris N Metaxas. Weakly supervised deep nuclei segmentation using partial points annotation in histopathology images. IEEE Transactions on Medical Imaging, 39(11):3655-3666, 2020. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 613 + ], + "type": "text", + "content": "[22] Shan E Ahmed Raza, Linda Cheung, Muhammad Shaban, Simon Graham, David Epstein, Stella Pelengaris, Michael Khan, and Nasir M Rajpoot. Micro-net: A unified model for segmentation of various objects in microscopy images. Medical Image Analysis, 52:160-173, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "type": "text", + "content": "[23] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 1, 2, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[24] Digvijay P Salunkhe, Rupa N Bartakke, Guruprasd Chvan, and Pooja R Kothavale. An overview on methods for slope stability analysis. International Journal of Engineering Research & Technology (IJERT), 6(03):2278-0181, 2017. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11681" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 406 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 161 + ], + "type": "text", + "content": "[25] Kuan Tian, Jun Zhang, Haocheng Shen, Kezhou Yan, Pei Dong, Jianhua Yao, Shannon Che, Pifu Luo, and Xiao Han. Weakly-supervised nucleus segmentation based on point annotations: A coarse-to-fine self-stimulated learning strategy. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part V 23, pages 299-308. Springer, 2020. 3, 5, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 162, + 288, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 238 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 238 + ], + "type": "text", + "content": "[26] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in Bioengineering and Biotechnology, 53: 1-15, 2019. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 239, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 305 + ], + "type": "text", + "content": "[27] Inwan Yoo, Donggeun Yoo, and Kyunghyun Paeng. Pseudoedgenet: Nuclei segmentation only with point annotations. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part I 22, pages 731-739. Springer, 2019. 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 306, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 306, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 306, + 287, + 338 + ], + "type": "text", + "content": "[28] Ji Yu. Point-supervised single-cell segmentation via collaborative knowledge sharing. arXiv preprint arXiv:2304.10671, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 340, + 287, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 340, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 48, + 340, + 287, + 406 + ], + "type": "text", + "content": "[29] Yanning Zhou, Omer Fahri Onder, Qi Dou, Efstratios Tsougenis, Hao Chen, and Pheng-Ann Heng. Cia-net: Robust nuclei instance segmentation with contour-aware information aggregation. In International Conference on Information Processing in Medical Imaging, pages 682–693, 2019. 1, 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11682" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/477f7931-e1b2-442e-a45b-11cf7d648a49_content_list.json b/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/477f7931-e1b2-442e-a45b-11cf7d648a49_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..718c75084275df5a2efc2c67110728ece74fda8f --- /dev/null +++ b/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/477f7931-e1b2-442e-a45b-11cf7d648a49_content_list.json @@ -0,0 +1,1532 @@ +[ + { + "type": "text", + "text": "Multi-Attribute Interactions Matter for 3D Visual Grounding", + "text_level": 1, + "bbox": [ + 174, + 130, + 795, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Can Xu $^{1}$ , Yuehui Han $^{1}$ , Rui Xu $^{1}$ , Le Hui $^{2}$ , Jin Xie $^{3,4*}$ , Jian Yang $^{1}$ $^{1}$ PCA Lab, Nanjing University of Science and Technology, Nanjing, China \n $^{2}$ School of Electronics and Information, Northwestern Polytechnical University, Xi'an, China \n $^{3}$ State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China \n $^{4}$ School of Intelligence Science and Technology, Nanjing University, Suzhou, China \nvolcanoxl23@gmail.com; {hanyh,xu-ray,csjxie,csjyang}@njust.edu.cn; huile@nwpu.edu.cn", + "bbox": [ + 114, + 179, + 851, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D visual grounding aims to localize 3D objects described by free-form language sentences. Following the detection-then-matching paradigm, existing methods mainly focus on embedding object attributes in unimodal feature extraction and multimodal feature fusion, to enhance the discriminability of the proposal feature for accurate grounding. However, most of them ignore the explicit interaction of multiple attributes, causing a bias in unimodal representation and misalignment in multimodal fusion. In this paper, we propose a multi-attribute aware Transformer for 3D visual grounding, learning the multi-attribute interactions to refine the intra-modal and inter-modal grounding cues. Specifically, we first develop an attribute causal analysis module to quantify the causal effect of different attributes for the final prediction, which provides powerful supervision to correct the misleading attributes and adaptively capture other discriminative features. Then, we design an exchanging-based multimodal fusion module, which dynamically replaces tokens with low attribute attention between modalities before directly integrating low-dimensional global features. This ensures an attribute-level multimodal information fusion and helps align the language and vision details more efficiently for fine-grained multimodal features. Extensive experiments show that our method can achieve state-of-the-art performance on ScanRefer and Sr3D/Nr3D datasets. The code is publicly available at https://github.com/volcanoXC/MA2TransVG.", + "bbox": [ + 75, + 354, + 473, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 811, + 209, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Visual Grounding (VG) aims to locate the most relevant object in a given point cloud scene based on a language description. As a cornerstone of wide applications [6, 8, 14, 28, 43], such as vision-language navigation and au", + "bbox": [ + 76, + 839, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3af3f09543052496dd2e1c1bbc972bc409a2e8170c36d55b9b35066ea9f95be5.jpg", + "image_caption": [ + "Figure 1. Different from prior works, we capture the multi-attribute interactions to guide the unimodal representation and multimodal fusion for more accurate grounding." + ], + "image_footnote": [], + "bbox": [ + 504, + 320, + 890, + 601 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tonomous robots, 3D VG has attracted increasing attention from both academia and industry. However, limited by the complexity of free-form language description and the irregularity of sparse point clouds, achieving accurate 3D visual grounding remains an open issue.", + "bbox": [ + 496, + 686, + 890, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing 3D VG methods mainly follow a detection-then-matching pipeline. Specifically, a pool of proposals is first produced by the pre-trained general 3D object detectors [11, 29, 37]. Then, the linguistic description is associated with the visual feature of each proposal to generate multimodal object representation, which will be used to predict their referring confidence score for grounding. Since such a two-stage paradigm aims to collect enough semantic cues to distinguish the target object, the", + "bbox": [ + 496, + 763, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17253", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quality of unimodal feature representation and multimodal feature fusion always play a critical role, which decides the discriminativeness of each proposal and their matching difficulty with the target label. Most methods focus on enhancing the feature from various perspectives, such as sparse convolution [47], 2D image assistance [5, 27, 46, 52], coarse proposal refinement [13, 22, 50], or segmentation/captioning collaboration [3, 10, 18, 26]. As shown in Fig. 1(a), recent works [7, 18, 19, 35] enhance the feature by embedding object attributes, such as color, name and language-conditioned correlation (i.e., distance, view, direction), showing a promising future in advancing grounding performance.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Though prior works made significant progress, we argue few of them explore the multi-attribute interactions in either unimodal representation or multimodal fusion, which affects the grounding accuracy, especially for cases with multiple objects. As the CASE in Fig. 1, the model embeds the most salient attribute (object name of 'door') and learns other potential attributes (e.g., 'brown', 'wooden', 'left') depending on the self-attention mechanism. Since the loss supervises the final prediction rather than the intermediate attention, the model may be encouraged to focus on some main attributes instead of all attributes, which weakens the discrimination of intra-modal representation. Meanwhile, when fusing the low-dimensional global feature between modalities via cross-attention, attention to 'brown' and 'wooden' should be encouraged for more discriminative language expression of the object, while they may be suppressed in the proposal representation to maximize the proposal's distinctiveness. Unfortunately, this would bring a misalignment in multimodal features, causing the model to select the auxiliary object (right door) rather than the primary object (left door).", + "bbox": [ + 75, + 291, + 472, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate the bias in unimodal representation and misalignment in multimodal fusion, we propose a multi-attribute aware Transformer $(\\mathsf{MA}^2\\mathsf{TransVG})$ for 3D visual grounding, which enhances the unimodal representation with intrinsic attribute interaction and refines the cross-modal information fusion to the attribute level (See Fig. 1(b)). Concretely, we develop an attribute causal analysis module (ACAM), which quantifies the effect of each attribute (i.e., learned attentions) and counterfactual interference (i.e., wrong attention) for the final prediction (i.e., grounding score). By maximizing the causal effect, powerful supervision is provided to correct the biased attribute and adaptively assign more proper attention to all attributes. Guided by the learned interaction, we further propose an exchanging-based multimodal fusion module (EMFM) to replace tokens with low attribute attention between modalities, which helps align more vision and language details and generate fine-grained multimodal features for accurate grounding. Our contributions include:", + "bbox": [ + 75, + 613, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a multi-attribute aware Transformer for 3D object grounding, which explicitly models the multi-attribute interactions to prevent imbalanced and misaligned visual-textual representation.", + "- We propose the ACAM module to quantify the causal effect of multiple intrinsic attributes with the final prediction, which helps adaptively assign more proper attention to all attributes for a better intra-modal understanding.", + "- We propose the EMFM module to perform multimodal fusion based on multi-attribute exchanging, which helps align more attribute-level details between modalities for fine-grained multimodal features.", + "- We evaluate the proposed method on $\\mathrm{Nr3D/Sr3D}$ and ScanRefer benchmarks, which outperforms all state-of-the-art methods by a large margin." + ], + "bbox": [ + 500, + 90, + 893, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 330, + 638, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Visual Grounding. The interest in 3D VG has grown rapidly in recent years, and existing approaches include one-stage [21, 31] and two-stage methods. The two-stage methods [13, 16-18, 46, 47, 50] decouple the grounding into language-irrelevant object detection and cross-modal matching. Recent works focus on improving the intramodal representation and modeling the inter-modal relationship. For example, Graph-based approaches [1, 18, 47] infer spatial relations among proposals by connecting each proposal with its top-N nearest neighbors. Transformer-based approaches [7, 17, 19, 21, 31, 46] leverage the attention mechanism to concentrate the object attributes on proposals that are more essential to the referring. In these methods, only the most salient attribute is embedded and the fusion between modalities occurs in the global feature space, resulting in a biased intra-modal representation and inadequate inter-modal fusion. Instead, we capture the multi-attribute interactions and refine the cross-modal learning at the attribute level.", + "bbox": [ + 496, + 356, + 893, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Causal Inference in Vision. Causality analysis [32] has been successfully used in several areas, such as adversarial learning [24], reinforcement learning [23] and graph neural networks [42, 51]. In vision tasks [30, 40, 44], causality inference also works as an effect tool to discover hidden causal structures and confront data biases. For example, Rao et al. [38] learn more effective attention based on causal inference for fine-grained image categorization, person and vehicle re-identification. As the first attempt, we introduce the causality analysis into 3D VG to quantify the effect of each attribute, which helps model the complex interactions of various attributes to enhance the unimodal feature.", + "bbox": [ + 496, + 643, + 895, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformers in Vision-and-Language. Transformer shows a powerful ability to learn multimodal features in various tasks, such as image captioning [9, 10, 48], vision-and-language navigation [28, 43], and vision question-answering [15, 53]. For 3D VG, different from most", + "bbox": [ + 496, + 824, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17254", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2bf114a7b9284038875042191adacdd60fe419be31f3656ad11482cfe19d5db0.jpg", + "image_caption": [ + "Figure 2. The framework of proposed $\\mathsf{MA}^2\\mathsf{TransVG}$ . The embedding of pre-detected proposals, attributes, and language are extracted and then cross-attend in ACAM. By applying the causal effect analysis via $L_{attr}$ , ACAM models the multi-attribute interactions for final prediction to generate the Attribute-Vision/Text attention, which can enhance the proposal and text embedding. The attention further guides the attribute-level feature exchanging in EMFM, which helps fuse the multimodal feature for accurate grounding." + ], + "image_footnote": [], + "bbox": [ + 83, + 95, + 885, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "methods that directly fuse the global unimodal features by stacking cross-attention layers [7, 17, 19, 21, 31, 46], we refine the cross-modal fusion to the attribute level, which helps suppress the misalignment between modalities to capture a finer-grained matching relationship.", + "bbox": [ + 75, + 414, + 470, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 503, + 169, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 527, + 186, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 shows an overview of proposed $\\mathsf{MA}^2\\mathsf{TransVG}$ . Given a 3D point cloud with corresponding description, text features $T \\in \\mathbb{R}^{(m + 1)\\times d}$ , proposal features $O \\in \\mathbb{R}^{n\\times d}$ and attribute features $F \\in \\mathbb{R}^{n\\times d}$ are extracted (Sec. 3.2). The three streams then are fed into the attribute causal analysis module (ACAM) to capture the multi-attribute interactions (i.e., Attribute-Vision attention and Attribute-Text attention) to enhance initial text and proposal feature. The attention is supervised by an attribute loss $L_{attr}$ , which is designed by estimating the causal effect of attributes for the final prediction (Sec. 3.3). The enhanced proposal embedding $O^*$ and text embedding $T^*$ are further passed to the exchanging-based multimodal fusion module (EMFM), where their tokens with small attribute attention are replaced with the average of all the tokens in the other modality (Sec. 3.4). With the fused multimodal embedding $E \\in \\mathbb{R}^{n\\times d}$ , we predict the final grounding results.", + "bbox": [ + 75, + 550, + 468, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Input Modal Representation", + "text_level": 1, + "bbox": [ + 76, + 816, + 333, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text Encoding. Given a sentence $S$ with $m$ word tokens, we use a pre-trained BERT model [20] to produce the text features $(t_{cls}, t_1, \\ldots, t_m)$ . $t_i \\in \\mathbb{R}^d$ is the feature of each token with a dimensionality of $d$ and $t_{cls} \\in \\mathbb{R}^d$ is a special", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "token for text classification.", + "bbox": [ + 498, + 415, + 684, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object Encoding. Given a point cloud scene $\\pmb{P} \\in \\mathbb{R}^{K \\times 6}$ of $K$ points described by XYZ coordinates and RGB colors, we use the GroupFree [29] detector to obtain an object proposal list $\\pmb{o} = (o_1, o_2, \\dots, o_n)$ containing $n$ proposals and extract corresponding proposal feature $\\pmb{O} \\in \\mathbb{R}^{n \\times d}$ based on PointNet++ [36]. Since we mainly focus on the object referred in the text, each proposal feature $O_i \\in \\mathbb{R}^d$ will be added with the $t_{cls}$ before feeding into ACAM for attribute-vision causal effect analysis.", + "bbox": [ + 496, + 429, + 890, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attributes Encoding. Motivated by [3], for the proposal pool $o$ with $n$ objects, the multi-view RGB features $\\pmb{f}_{c} \\in \\mathbb{R}^{n \\times 128}$ that potentially contain the color and materials attribute is first extracted. Then, we use the 3-dimensional box center and 24-dimensional corner coordinates to describe the object size attribute $\\pmb{f}_{s} \\in \\mathbb{R}^{n \\times 27}$ . For the location attribute, we further consider the distance and angle of each pair of objects $(o_{i}, o_{j})$ . Connecting the object center point $(x_{c_{i}}, y_{c_{i}}, z_{c_{i}})$ and $(x_{c_{j}}, y_{c_{j}}, z_{c_{j}})$ , the distance is their Euclidean distance $d_{ij}$ . Similar to [7], the horizontal angle $\\theta_{h}$ and vertical angle $\\theta_{v}$ can be calculated as:", + "bbox": [ + 496, + 566, + 892, + 731 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c} \\theta_ {h} = \\arctan 2 \\left(\\left(y _ {c _ {i}} - y _ {c _ {j}}\\right) / \\left(x _ {c _ {i}} - x _ {c _ {j}}\\right) \\right. \\\\ \\theta_ {v} = \\arcsin \\left(\\left(z _ {c _ {i}} - z _ {c _ {j}}\\right) / d _ {i j}\\right) \\end{array} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 741, + 890, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The location attribute $f_{l}^{ij} \\in \\mathbb{R}^{5}$ between any pair of objects is defined, as:", + "bbox": [ + 496, + 785, + 890, + 815 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {l} ^ {i j} = \\left[ d _ {i j}, \\sin \\left(\\theta_ {h}\\right), \\cos \\left(\\theta_ {h}\\right), \\sin \\left(\\theta_ {v}\\right), \\cos \\left(\\theta_ {v}\\right) \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 825, + 890, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each proposal $o_i$ , the location attribute $\\pmb{f}_l^i \\in \\mathbb{R}^5$ can be calculated by averaging all $f_l^{ij}$ connected with $o_i$ , and the final attribute embedding can be generated by concatenating", + "bbox": [ + 496, + 853, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17255", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "all three attributes, as $\\pmb{f}^{i} = [\\pmb{f}_{l}^{i},\\pmb{f}_{c}^{i},\\pmb{f}_{s}^{i}]$ . We map $\\pmb{f}^{i}\\in \\mathbb{R}^{d}$ to the same dimension of text feature using a FC layer.", + "bbox": [ + 75, + 89, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Attribute Causal Analysis Module", + "text_level": 1, + "bbox": [ + 76, + 132, + 375, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The attribute causal analysis module (ACAM) models the multi-attribute interactions with vision/text as Attribute-Vision/Text attention based on a well-designed Attribute-aware Transformer. To capture more influential and diverse interactions, ACAM further guides the learning process with an extra loss $L_{attr}$ , which measures the causal effect of learned attribute-level attentions for the final prediction.", + "bbox": [ + 75, + 157, + 468, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attributes-Text Attention Learning. Given attribute embedding $\\pmb{F} \\in \\mathbb{R}^{n \\times d}$ , Attributes-Text attention $\\pmb{A}_{T}$ helps select related attribute features from the initial text feature $T \\in \\mathbb{R}^{(m + 1) \\times d}$ , thereby the text representation can be enhanced to better describe the target object. We define $\\pmb{A}_{T} \\in \\mathbb{R}^{n \\times (m + 1)}$ as the cross-attention between the Attribute embedding $\\pmb{F}$ and text embedding $\\pmb{T}$ , which can be computed with a standard Transformer [41].", + "bbox": [ + 75, + 263, + 468, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to enlarge the model capacity for diverse Attribute-Text correlations, we also introduce multi parallel attention layers $\\ell$ as different transformation heads, and the multi-head attentions will be fused with the standard self-attention [41]. As shown in Fig. 2, $A_{T}$ then guides to generate the enhanced text feature $T^{*} \\in \\mathbb{R}^{n \\times d}$ , which will be used to train the language classifier with the supervision of the language classification loss $L_{sent}$ , and also serves as input for subsequent EMFM for multi-modal fusion.", + "bbox": [ + 75, + 386, + 468, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attributes-Vision Attention Learning. Instead of directly calculating the cross-attention between proposal features $\\mathcal{O}$ and attribute features $\\mathbf{F}$ , we first guide the proposal features with language token $t_{cls}$ , as we prefer to care about specific categories of objects mentioned in the text [7]. Specifically, for the text-guided proposal feature, we adopt the similar Transformer structure in Attributes-Text attention learning to calculate its self-attention $A_V^s \\in \\mathbb{R}^{n \\times n}$ , and cross-attention $A_V^c \\in \\mathbb{R}^{n \\times n}$ with $\\mathbf{F}$ .", + "bbox": [ + 75, + 522, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to [7], we then integrate the self-attention of proposals $A_V^s$ with the cross attribute-proposal attention $A_V^c$ using a sigmoid softmax (sigsoftmax) fusion operation to calculate the Attributes-Vision attention $A_V$ , as:", + "bbox": [ + 75, + 659, + 468, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {v} = \\frac {\\sigma \\left(\\boldsymbol {A} _ {\\boldsymbol {V}} ^ {\\boldsymbol {c}} (i , j)\\right) \\exp \\left(\\boldsymbol {A} _ {\\boldsymbol {V}} ^ {\\boldsymbol {s}} (i , j)\\right)}{\\sum_ {j = 1} ^ {n} \\sigma \\left(\\boldsymbol {A} _ {\\boldsymbol {V}} ^ {\\boldsymbol {c}} (i , j)\\right) \\exp \\left(\\boldsymbol {A} _ {\\boldsymbol {V}} ^ {\\boldsymbol {s}} (i , j)\\right)} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 731, + 468, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma(\\cdot)$ is the sigmoid function. $A_V^s (i,j)$ and $A_V^c (i,j)$ are the element in $i$ -th row and $j$ -th column in attention matrix $A_V^s$ and $A_V^c$ , respectively. The same multi-head attention strategy is used in Attribute-Vision attention $A_V$ . In a similar manner, the enhanced proposal embedding $O^{*}$ can be obtained by performing a matrix multiplication operation between $O$ and $A_V$ , which will serve as the input of proposal classifier and EMFM module.", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/83d7e39ca8b8ee336fbb6f818734b170655925651d94eb642d55a8eac4a965ab.jpg", + "image_caption": [ + "Figure 3. Visualization of deriving causal effects through counterfactual causality. We subtract the counterfactual prediction from the original prediction to analyze the effects of learned attention and maximize them in the training process." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 887, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Strengthen Interaction With $L_{attri}$ . Though multiple attributes have been embedded to guide the language and vision representations via Attribute-Text/Vision attention, how these attributes interact with each other to affect predictions remains a black box. To address this issue, our core idea lies in measuring the causal effect of $A_V$ and $A_T$ for the final prediction. We design the attribute loss $L_{attri}$ based on the counterfactual analysis [33], providing extra supervision to encourage the network to explore more influential interactions.", + "bbox": [ + 496, + 281, + 890, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3, taking the Attribute-vision attention as an example, we conduct counterfactual intervention by imagining non-existent attention $\\bar{A}_V$ to replace the initial $A_V$ . We keep the vision representation $O$ unchanged, which will be multiplied with $\\bar{A}_V$ to calculate the counterfactual representation $\\bar{O}^*$ . $\\bar{O}^*$ are then fused with $T$ to obtain the counterfactual prediction $\\bar{Y} \\in \\mathbb{R}^{n \\times 1}$ , as:", + "bbox": [ + 496, + 433, + 890, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\boldsymbol {Y}} = d o \\left(\\boldsymbol {A} _ {V} = \\bar {\\boldsymbol {A}} _ {V}, \\boldsymbol {T}\\right) = G \\left(\\bar {\\boldsymbol {O}} ^ {*}, \\boldsymbol {T}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 544, + 890, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $G(\\cdot)$ is the grounding operation. In practice, the exact form of how counterfactual is achieved is not limited, and our starting point here is just to set the ball rolling. In our case, random attention, uniform attention, shuffle attention or reversed attention are used as the counterfactuals.", + "bbox": [ + 496, + 568, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following [33, 34, 40], we define the causal effect $\\mathbf{Y}_e$ of $\\mathbf{A}_V$ as the difference between the observed prediction $\\mathbf{Y}$ and the counterfactual prediction $\\bar{\\mathbf{Y}}$ , as:", + "bbox": [ + 496, + 643, + 890, + 689 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Y} _ {e} = \\mathbb {E} _ {\\bar {A} _ {V} \\sim \\gamma} (\\boldsymbol {Y} - \\bar {\\boldsymbol {Y}}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 695, + 890, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ means the distribution of counterfactual attention.", + "bbox": [ + 496, + 719, + 887, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the causal effect can be interpreted as how the attention improves the final prediction compared to wrong attentions, we use $\\mathbf{Y}_e$ to measure the learned attention, i.e., the multi-attribute interactions. Based on this observation, we further design an attribute loss $L_{attr}$ , as:", + "bbox": [ + 496, + 734, + 890, + 810 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {a t t r i}} = L _ {c e} \\left(\\boldsymbol {Y} _ {\\boldsymbol {e}}, \\boldsymbol {y}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 816, + 890, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{ce}$ is the cross-entropy loss and $\\pmb{y}$ is the GT label. Similarly, we can obtain the causal effect of Attribute-text attention and compute its attribute loss. By optimizing $L_{attr}$ from $A_{T}$ and $A_{V}$ , the model is expected to:", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17256", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(1) improve the prediction based on correct attentions as much as possible, which helps discover more discriminative attributes; (2) penalize the prediction based on wrong attentions, which forces the model adaptively capture more attributes to avoid biased unimodal embeddings.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Exchanging-based Multimodal Fusion Module", + "text_level": 1, + "bbox": [ + 76, + 176, + 468, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose the exchanging-based multimodal fusion module (EMFM) to communicate attribute-level details of objects between both input vectors for accurate grounding. Specifically, since text/visual encoders tend to focus on different attributes for better unimodal embeddings, directly fusing unimodal embeddings using cross-attention will hurt the distinguishability of multimodal features. To narrow the semantic misalignment between modalities, we select and exchange tokens with low attention score using the average of all tokens from another modal, which helps align attribute-level information between modalities sufficiently and eliminate semantic ambiguity in multimodal representations of proposals.", + "bbox": [ + 75, + 199, + 468, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Exchanging With Attributes Interactions. As shown in Fig. 3, we implement the EMFM by two Transformer encoders with shared parameters, where the shallow $\\mu$ layers are set as regular Transformer encoder layers, followed by $\\eta$ exchanging layers. Since $A_V$ and $A_T$ have modeled the multi-attribute interactions, they are taken as references to select tokens for exchanging. Meanwhile, $A_V$ and $A_T$ will be updated using the self-attention learned from unimodal embedding $T^*$ and $O^*$ . Considering the interdependence between the attribute attention and unimodal embedding, we concatenate them for the subsequent calculation, as:", + "bbox": [ + 75, + 396, + 468, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {T} ^ {e} (0) = \\operatorname {C o n c a t} \\left[ \\boldsymbol {A} _ {T}, \\boldsymbol {T} ^ {*} \\right] \\\\ \\boldsymbol {Q} ^ {e} (0) = \\operatorname {C o n c a t} \\left[ \\boldsymbol {A} _ {T}, \\boldsymbol {Q} ^ {*} \\right] \\end{array} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 574, + 468, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {O} ^ {e} (0) = \\operatorname {C o n c a t} \\left[ \\boldsymbol {A} _ {V}, \\boldsymbol {O} ^ {*} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 590, + 361, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T^{e}(0)$ and $O^{e}(0)$ are embedding for textual and visual modalities at layer 0, respectively.", + "bbox": [ + 75, + 617, + 468, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At layer $\\mu + 1$ with intermediate embedding $T^{e}(\\mu + 1)$ , inspired by [4, 25], we select tokens with a $\\theta$ -proportion of the smallest attention scores to $A_{T}$ and replace their embedding vectors with the average embedding of all the tokens in $O^{e}(\\mu + 1)$ . For instance, the selected $k$ -th token of $T^{e}(\\mu + 1)$ can be updated by:", + "bbox": [ + 75, + 647, + 468, + 731 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {T} ^ {\\boldsymbol {e}} (\\mu + 1) [ k,: ] = \\frac {1}{n} \\sum_ {j = 1} ^ {n} \\boldsymbol {O} ^ {\\boldsymbol {e}} (\\mu + 1) [ j,: ] + \\boldsymbol {T} ^ {\\boldsymbol {e}} (\\mu + 1) [ k,: ] \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 742, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The updated embeddings are used to calculate the multihead self-attention [41], which will guide the exchanging operation at the next layer, e.g., $A_T$ of the $(\\mu + 1)$ -th layer are updated by the self-attention of the previous layer $\\mu$ .", + "bbox": [ + 75, + 794, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal Fusion. We also apply the residual connection between adjacent exchanging layers to reduce the information loss caused by replacement. The exchanging process", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "continues until reaching the $\\eta$ layer to generate $O^e (\\eta)$ and $T^{e}(\\eta)$ , which will be fed into the feed-forward network (FFN) with a normalization layer and a fully connected layer to derive the final fusion embedding $\\pmb {E}\\in \\mathbb{R}^{n\\times n}$ .", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Grounding Head. We use a two-layer feed-forward neural network as the object grounding head. Given the multimodal representation $\\pmb{E}$ , the grounding head will predict the probability $\\pmb{p}_i \\in \\mathbb{R}^{n \\times 1}$ for each object proposal, where the proposal with the maximum probability is selected as the target in the inference stage.", + "bbox": [ + 496, + 151, + 890, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training Objectives", + "text_level": 1, + "bbox": [ + 498, + 250, + 687, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the previous works [1, 5, 17], the loss of $\\mathbf{MA}^2\\mathrm{TransVG}$ consists of the primary 3D object grounding loss $L_{vg}$ , language classification loss $L_{sent}$ , proposal classification loss $L_{obj}$ , and two attribute losses $L_{attri}$ for vision/text attention causal analysis, as:", + "bbox": [ + 496, + 273, + 890, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {v g} + L _ {s e n t} + L _ {o b j} + \\lambda L _ {a t t r i} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 359, + 890, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is the hyper-parameter to balance the loss.", + "bbox": [ + 500, + 385, + 834, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 412, + 622, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 438, + 785, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the proposed $\\mathsf{MA}^2\\mathsf{TransVG}$ on ScanRefer [5] and Sr3D/Nr3D [1] datasets. ScanRefer offers 51,583 descriptions of 11,046 objects from 800 ScanNet [12] scenes. A target object is labeled as 'unique' if it is the only object of its class in the scene; otherwise, it is labeled as 'multiple'. For ScanRefer, we use Acc@0.25 and Acc@0.5 as evaluation metrics, which measure the fraction of language queries whose predicted box overlaps the ground truth box with 3D IoU higher than 0.25/0.5. Compared to ScanRefer where objects need to be detected and then matched, Sr3D/Nr3D datasets provide object masks for each scene and only need to recognize the classes of the proposals to choose the target object. Nr3D consists of 41,503 descriptions from 707 scenes with human annotation, while Sr3D contains 83,572 simple machine-generated descriptions. Each scene in Sr3D/Nr3D can also be divided into 'easy' and 'hard' depending on whether there are more than two instances, and can be divided into 'view-dependent' and 'view-independent' according to whether the referring expression is dependent on the camera view. Following [1], the accuracy is used to verify the model.", + "bbox": [ + 496, + 460, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 786, + 715, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The text encoding module is initialized from the first three layers of BERT [20], and the object encoding module PointNet++ [36] samples 1024 points for all the objects. We generate an initial set of $n = 256$ proposals using [29]. We set the dimension of unimodal representation $d = 768$ and use a three-layer transformer with 12 heads for", + "bbox": [ + 496, + 809, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17257", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a4d504e5e19474e9bdc26fc4bf9292a9c3ea1f3b07525d4192eba0b10bc6b6f3.jpg", + "table_caption": [ + "Table 1. Comparison of 3D VG results with the state-of-the-art on ScanRefer." + ], + "table_footnote": [], + "table_body": "
MethodVenueUnique (19%)Multiple (81%)Overall
0.250.50.250.50.250.5
ScanRefer [5]ECCV2067.646.232.121.340.026.1
ReferIt3D [1]ECCV2053.837.521.012.826.416.9
TGNN [18]AAAI2168.656.829.823.237.429.7
InstanceRefer [47]ICCV2177.566.831.324.840.232.9
3DVG [50]ICCV2177.258.538.428.745.934.5
FFL-3DOG [13]ICCV2178.867.935.225.741.334.0
MVT [19]CVPR2277.766.531.925.340.833.3
3D-SPS [31]CVPR2281.664.839.529.647.736.4
ViL3DRef [7]NeurIPS2281.668.640.330.747.937.7
BUTD-DETR [22]ECCV22////52.239.8
EDA [45]CVPR2385.868.649.137.654.642.3
M3DRef-CLIP [49]ICCV23/77.2/36.8/44.7
3DJCG [3]CVPR2278.861.340.130.147.636.1
UniT3D [10]ICCV2382.873.136.431.142.339.1
3DRefTR-HR [26]ICCV2386.070.949.638.355.043.2
3DRefTR-SP [26]ICCV2386.171.050.138.755.543.5
Ours/86.374.153.841.457.945.7
", + "bbox": [ + 241, + 113, + 725, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the Attribute-Vision/Text attention learning. For a fair comparison, we keep the architecture parameters as same as previous works [7, 19, 46]. In ACAM, we use random attention to initialize the counterfactual attention. In EMFM, considering the balance of raw intra-modal knowledge and inter-modal communication, we empirically set $\\theta = 10\\%$ . The start layer $\\mu$ and end layer $\\eta$ are set to 1 and 4, respectively. The hyper-parameter of the whole loss in Eq. (9) is $\\lambda = 0.5$ . All models are trained on 4 NVIDIA RTX Titan GPUs, which are optimized using the AdamW algorithm with a batch size of 128 and a learning rate of 0.0005 with cosine decay scheduling.", + "bbox": [ + 75, + 402, + 470, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 76, + 589, + 444, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ScanRefer. In Tab. 1, we compare $\\mathsf{MA}^2\\mathsf{TransVG}$ with existing works on the ScanRef dataset and observe the following results: i) Compared to pure 3D VG works such as BUTD-DETR [22], 3D-SPS [31], ViL3DRef [7], and EDA [45], our method achieves state-of-the-art performance of $57.9\\%$ and $45.7\\%$ , by a substantial margin with an overall improvement over $3.3\\%$ and $3.4\\%$ . ii) To facilitate the 3D grounding, previous works [5, 31, 46, 50] learn better point cloud features with the supplementary of 2D images while most recent works [3, 10, 26] learn better language-visual alignment by connecting 3D dense captioning or segmentation tasks. However, we outperform second-place 3DRefTR-SP [26] with an overall improvement over $2\\%$ for both IoU = 0.25 and IoU = 0.5 setting. This superiority illustrates that mining the multi-attribute interactions help capture more efficient and discriminative multi-modal representation to distinguish objects. iii) Our $\\mathsf{MA}^2\\mathsf{TransVG}$ can reach a remarkable accuracy of $53.8\\%$ and $41.4\\%$ for 'Multiple' scenes, which verifies the effec", + "bbox": [ + 75, + 612, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tiveness of our attribute interaction for understanding finer-grained language-visual contextual dependency.", + "bbox": [ + 498, + 402, + 890, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Nr3D&Sr3D. Tab. 2 reports the grounding accuracy on Nr3D and Sr3D datasets. For a fair comparison, all the methods use the ground-truth object proposal with no ground-truth labels. In Nr3D, though the language descriptions are more complex to cause additional challenges for text understanding and cross-modal alignment, our method still outperforms the EDA [45], which is equipped with text component decoupling and dense matching between two modalities. Compared to EDA, our $\\mathsf{MA}^2\\mathsf{TransVG}$ not only achieves a similar decoupling function for both descriptions and point clouds via attribute representation, but also considers their correlationships for prediction during Attribute-Text/Vision causal attention learning, which contributes to the great improvements by $+13.1\\%$ in overall accuracy. Compared to other recent competitor ViL3DRef [7], 3DRefTRSR [26], and M3DRef-CLIP [49], we also achieve a consistent improvement of overall accuracy in all split settings by $+0.8\\%$ , $+2.6\\%$ and $+15.8\\%$ , respectively. In Sr3D, our method reaches an accuracy of $73.9\\%$ , which greatly surpasses the best competitor ViL3DRef [7] $1.1\\%$ and outperforms other state-of-the-art methods [26, 45] over $5\\%$ . Similar to ViL3DRef [7] that guided the grounding with languaged-conditioned spatial relation, we not only embed the spatial attributes but also explicitly consider other attributes in the unimodal representation and further perform exchanging-based fusion. All the results show that multi-attribute interactions can help capture more semantic clues for accurate grounding.", + "bbox": [ + 496, + 462, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17258", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a7e49266e07a30003872ae5e82b724fd0ab87b70537141ab8c14e48b10b733fc.jpg", + "table_caption": [ + "Table 2. Comparison of 3D VG results with the state-of-the-art on Sr3D/Nr3D." + ], + "table_footnote": [], + "table_body": "
MethodVenueNr3DSr3D
EasyHardView DepView IndepOverallEasyHardView DepView IndepOverall
ReferIt3D [1]ECCV2043.627.932.537.135.644.731.539.240.840.8
TGNN [18]AAAI2144.230.635.838.037.348.536.945.845.045.0
InstanceRefer [47]ICCV2146.031.834.541.938.851.140.545.448.148.0
3DVG [50]ICCV2148.534.834.843.740.854.244.944.651.751.4
FFL-3DOG [13]ICCV2148.235.037.144.741.7/////
TransRefer3D [17]ACMM2148.536.036.544.942.160.550.249.957.757.4
LanguageRefer [39]CoRL2151.036.641.745.043.958.949.349.256.356.0
SAT [46]ICCV2156.342.446.950.449.261.250.049.258.357.9
BUTD-DETR [22]ECCV22////43.3////52.1
LAR [2]NeurIPS2258.442.347.452.148.963.051.250.059.159.4
3D-SPS [31]CVPR2258.145.148.053.251.556.265.449.263.262.6
MVT [19]CVPR2261.349.154.354.355.466.958.858.464.764.5
M3DRef-CLIP [49]ICCV2355.643.442.352.949.4/////
EDA [45]CVPR23////52.1////68.1
3DRefTR-SR [26]ICCV23////52.6////68.5
ViL3DRef [7]NeurIPS2270.257.462.064.564.474.967.963.873.272.8
Ours/71.157.662.565.465.276.069.364.573.873.9
", + "bbox": [ + 117, + 114, + 846, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 429, + 240, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Each Component. We quantitatively investigate the contribution of each component in our method in Tab. 3. The baseline model simply extracts and fuses the object and language feature through a standard Transformer structure, where the modal relationships only implicitly depend on the self/cross-attention. Taking results on $\\mathrm{Nr3D}$ as an example, when enhancing the feature with $A_{T}$ or $A_{V}$ , we yield an accuracy of $55.2\\%$ and $51.9\\%$ , which can increase to $59.7\\%$ when combining two attentions. This indicates that embedding the multi-attribute within each single modal is the basis of accurate cross-modal understanding. Furthermore, we achieve a gain of $2.7\\%$ after introducing causal effect analysis, which helps adaptively capture more reliable multi-attribute interactions for feature enhancement. Exchanging tokens further improves the accuracy to $65.2\\%$ because multimodal knowledge exchanging boosts the quality of the learned embeddings for each modality. All the results demonstrate the effectiveness of each component in our proposed $\\mathrm{MA}^2\\mathrm{TransVG}$ .", + "bbox": [ + 75, + 457, + 470, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Multi-attribute. To verify the effectiveness of multi-attribute interactions, we conduct ablation on attribute quantity and multi-attribute interactions. In Tab. 3. We adopt commonly used attributes (location $f_{l}$ , size $f_{s}$ , color $f_{c}$ ) with simple embedding method and achieve SOTA performance. The accuracy is $44.9\\%$ on Nr3D when generating attribute embedding only using $f_{l}$ and $f_{s}$ , which increases to $46.1\\%$ ( $+1.2\\%$ ) and $60.2\\%$ ( $+15.3\\%$ ) after adding $f_{c}$ and interactions, respectively. Results show that our accuracy gain mainly comes from the multi-attribute", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/03829d7b4d246c196a13ecc798ace61586db465621937e1fdd19224114a2cbdb.jpg", + "table_caption": [ + "Table 3. Ablation studies of each component on $\\mathrm{{Nr}}3\\mathrm{D}$ and $\\mathrm{{Sr}}3\\mathrm{D}$ ." + ], + "table_footnote": [], + "table_body": "
flfsfcATAVLattriExgNr3DSr3D
✓✓44.955.6
✓✓✓✓✓60.269.2
✓✓✓✓✓59.368.4
✓✓✓✓✓✓61.871.2
✓✓✓46.157.8
✓✓✓✓55.264.6
✓✓✓✓51.962.2
✓✓✓✓✓59.768.5
✓✓✓✓✓✓62.472.0
✓✓✓✓✓✓62.071.3
✓✓✓✓✓✓✓65.273.9
", + "bbox": [ + 517, + 454, + 874, + 632 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5b0e79e2873a639f01ed7c33082ace8fa08778cb9d1c75cec7e94ce5546abd7d.jpg", + "table_caption": [ + "Table 4. Comparison with various counterfactual attentions." + ], + "table_footnote": [], + "table_body": "
Acc@0.25random \nattentionuniform \nattentionreversed \nattentionshuffle \nattention
ScanRefer57.857.456.957.5
Nr3D65.264.864.565.0
Sr3D73.973.373.173.5
", + "bbox": [ + 517, + 678, + 874, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "interactions rather than embedding more attitudes.", + "bbox": [ + 498, + 792, + 831, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of Counterfactual Attention. We implement different strategies to generate counterfactual attention including random attention, uniform attention, reversed attention, and shuffle attention, and report their result on $\\mathrm{Nr3D}$ in Tab. 4. We see all the counterfactual attention strategies achieve a similar performer gain in three datasets,", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17259", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb01b054e1463071777d5c89b7ad27b6f3bf07dfb43bab833cf0d11d90923c1e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 104, + 89, + 236, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bcd4b1b6912995469b95a18add85e50e816b24cf8883272caf1d6b9f83bf85a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 90, + 375, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e36ef93f4f35540859f5ef5d42ff26e73d643a3d3726d247340d06a65d4ed1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 382, + 90, + 516, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d3325333eb7e7e9a501c53172f5c701dd892e368cf4fdd2fb3d34fe5c070b173.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 90, + 653, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6a85df81ee1e3e12443d290b769e49b548bd2dcd94c7529d3e05a42aea879bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 90, + 882, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b13c49d1476c8c8c0eaf086ff0b8b9ea2c41a76eed082fc1bfc8a107bf14bd0.jpg", + "image_caption": [ + "(b)", + "wooden double bookcase filled with books. walking into the room, it is in the right hand most.", + "Figure 4. The visualization results. (a) and (b) present the correct/wrong prediction of $MA^2$ TransVG with/without attribute interaction." + ], + "image_footnote": [], + "bbox": [ + 104, + 191, + 238, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ffc6755253bfc4a2733baf192c333089b6fe3a0fbea990fed9240db91ad7a46e.jpg", + "image_caption": [ + "the trash can is in the corner next to the door, below the whiteboard. it is gray, tall, and narrow.", + "the black chair is located directly to the right of the left wall off the room. The chair is to the left of the burgundy chair and behind the desk." + ], + "image_footnote": [], + "bbox": [ + 243, + 191, + 375, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e098f3a4513655970d02a121b8f4839c5f3b11c27cdff957f2c4cccbddcaebf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 382, + 191, + 516, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd8f67419cd5daa5d188337ad67e40f127e7f6f92865dd1ceb4ad532d8257e98.jpg", + "image_caption": [ + "the pillow is on the bed. it is to the left of the lamp, resulting on the pillow that is behind it." + ], + "image_footnote": [], + "bbox": [ + 519, + 191, + 653, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0df27124929018779b595c1f7617e64973d3ffeb3175dafd6ce57e00500e192f.jpg", + "image_caption": [ + "this is a black leather loveseat. if you were sitting in it, the long, short bookshelf would be on the right." + ], + "image_footnote": [], + "bbox": [ + 658, + 191, + 882, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "while random attention can obtain the best accuracy. We think it is because random attention helps try more diverse combinations of attribute interactives and can provide a more effective signal to supervise the attention.", + "bbox": [ + 75, + 387, + 468, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of Exchanging Operation. We conduct a sensitivity analysis of hyper-parameters in our EMFM module, including: i) the exchange proportion $\\theta$ of tokens. As shown in Table. 5, the model performance increases and then drops with the increase of $\\theta$ , which achieves a peak value of $65.2\\%$ when $\\theta = 10\\%$ . The results show that too small or large $\\theta$ will adversely affect the inter-modal fusion, which may bring inadequate alignment or attenuate the intramodal knowledge. ii) the start layer $\\mu$ and end layer $\\eta$ for multimodal exchanging. Since the default number of layers in the regular Transformer is 6, we first fix the start layer $\\mu$ to 1 and investigate the value of $\\eta$ from 1, 2, 3, 4, 5, where the accuracy improves quickly from $62.4\\%$ to $65.2\\%$ . We then fix the end layer $\\eta$ to 5, where the accuracy decreases when $\\mu$ increases from 1 to 5. All the results verify the rationality of our hyper-parameters settings.", + "bbox": [ + 75, + 448, + 470, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualization and Limitation", + "text_level": 1, + "bbox": [ + 76, + 696, + 333, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show the visualization results in Fig. 4. The model may fail to recognize the spatial relationship among objects (the last three columns) or be significantly affected by other more salient objects (the first two columns) without considering the explicit multi-attribute interactions. Our $MA^2$ TransVG can better perceive discriminative clues such as the appearance or spatial correlation of objects for multimodal feature enhancement and accurate grounding (the first row). We also notice the accuracy still remains improvement room for 'View-dependent' descriptions compared to the 'View-independent' or 'Easy', especially in Sr3D (In Tab. 2). In the future, we will extend more at-", + "bbox": [ + 75, + 719, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dbaf90b94a46fc5c468d5e92406d8e6f80c8dd2cd7f1a181f900053d7f02ab7f.jpg", + "table_caption": [ + "Table 5. Hyper-parameter Sensitivity Analysis on Nr3D." + ], + "table_footnote": [], + "table_body": "
θ0%5%10%15%20%
64.464.765.263.961.4
η12345
μ = 162.463.964.865.264.2
μ12345
η = 564.263.362.962.662.4
", + "bbox": [ + 553, + 410, + 839, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tribute embedding strategies, e.g., better encoding absolute position or aggregating point clouds from various views.", + "bbox": [ + 498, + 556, + 890, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 601, + 617, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose an attribute-aware Transformer $(\\mathsf{MA}^2\\mathsf{TransVG})$ for 3D object grounding. As the first attempt to quantify the contribution of each attribute for final grounding, $\\mathsf{MA}^2\\mathsf{TransVG}$ first designs a newly attribute causal effect analysis module and an attribute loss, which help model the multi-attribute interactions of objects for a better visual/textual modal understanding. Based on the learned multi-attribute interactions, the exchanging-based multimodal fusion module further dynamically aligns more discriminative details between modalities for a fine-grained multimodal feature. The proposed model outperforms the SOTA methods on $\\mathrm{Nr3D / Sr3D}$ and ScanRefer datasets.", + "bbox": [ + 496, + 626, + 893, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 821, + 660, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Science Fund of China (Grant Nos. 62276144, 62306238) and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 498, + 847, + 892, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17260", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Ahmed Abdelreehem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 2, 5, 6, 7", + "[2] EslamMohamed Bakr, Yasmeen Alsaedy, and Mohamed Elhoseiny. Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. In NeurIPS, 2022. 7", + "[3] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In CVPR, pages 16464-16473, 2022. 2, 3, 6", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 5", + "[5] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221. Springer, 2020. 2, 5, 6", + "[6] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. NeurIPS, 34:5834-5847, 2021. 1", + "[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. NeurIPS, 35:20522-20535, 2022. 2, 3, 4, 6, 7", + "[8] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, pages 16537-16547, 2022. 1", + "[9] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In CVPR, pages 11124-11133, 2023. 2", + "[10] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In ICCV, pages 18109-18119, 2023. 2, 6", + "[11] Bowen Cheng, Lu Sheng, Shaoshuai Shi, Ming Yang, and Dong Xu. Back-tracing representative points for voting-based 3d object detection in point clouds. In CVPR, pages 8963-8972, 2021. 1", + "[12] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5", + "[13] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In CVPR, pages 3722-3731, 2021. 2, 6, 7", + "[14] Chen Gao, Xingyu Peng, Mi Yan, He Wang, Lirong Yang, Haibing Ren, Hongsheng Li, and Si Liu. Adaptive zone-aware hierarchical planner for vision-language navigation. In CVPR, pages 14911-14920, 2023. 1" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yi Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. In CVPR, pages 14773-14783, 2023. 2", + "[16] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In ICCV, pages 15372-15383, 2023. 2", + "[17] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-and-relation aware transformer for fine-grained 3d visual grounding. In ACM MM, pages 2344-2352, 2021. 2, 3, 5, 7", + "[18] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In AAAI, pages 1610-1618, 2021. 2, 6, 7", + "[19] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In CVPR, pages 15524-15533, 2022. 2, 3, 6, 7", + "[20] Kenton Lee Jacob Devlin, Ming-Wei Chang and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2, 2019. 3, 5", + "[21] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Looking outside the box to ground language in 3d scenes. arXiv preprint arXiv:2112.08879, 2021. 2, 3", + "[22] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, pages 417-433. Springer, 2022. 2, 6, 7", + "[23] Nathan Kallus and Angela Zhou. Confounding-robust policy improvement. NeurIPS, 31, 2018. 2", + "[24] Murat Kocaoglu, Christopher Snyder, Alexandros G Dimakis, and Sriram Vishwanath. Causalgan: Learning causal implicit generative models with adversarial training. In ICLR, 2018. 2", + "[25] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022. 5", + "[26] Haojia Lin, Yongdong Luo, Xiawu Zheng, Lijiang Li, Fei Chao, Taisong Jin, Donghao Luo, Chengjie Wang, Yan Wang, and Liujuan Cao. A unified framework for 3d point cloud visual grounding. arXiv preprint arXiv:2308.11887, 2023. 2, 6, 7", + "[27] Haolin Liu, Anran Lin, Xiaoguang Han, Lei Yang, Yizhou Yu, and Shuguang Cui. Refer-it-in-rgbd: A bottom-up approach for 3d visual grounding in rgb images. In CVPR, pages 6032-6041, 2021. 2", + "[28] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In ICCV, pages 10968-10980, 2023. 1, 2", + "[29] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, pages 2949-2958, 2021. 1, 3, 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17261", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] David Lopez-Paz, Robert Nishihara, Soumith Chintala, Bernhard Scholkopf, and Léon Bottou. Discovering causal signals in images. In CVPR, pages 6979-6987, 2017. 2", + "[31] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16454-16463, 2022. 2, 3, 6, 7", + "[32] Judea Pearl. Interpretation and identification of causal mediation. Psychological methods, 19(4):459, 2014. 2", + "[33] Judea Pearl. Direct and indirect effects. In Probabilistic and causal inference: the works of Judea Pearl, pages 373-392. 2022. 4", + "[34] Stephen Powell. The book of why: The new science of cause and effect. pearl, judea, and dana mackenzie. 2018. hachette uk. Journal of MultiDisciplinary Evaluation, 14(31):47-54, 2018. 4", + "[35] Mihir Prabhudesai, Hsiao-Yu Fish Tung, Syed Ashar Javed, Maximilian Sieb, Adam W Harley, and Katerina Fragkiadaki. Embodied language grounding with 3d visual feature representations. In CVPR, pages 2220-2229, 2020. 2", + "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3, 5", + "[37] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 1", + "[38] Yongming Rao, Guangyi Chen, Jiwen Lu, and Jie Zhou. Counterfactual attention learning for fine-grained visual categorization and re-identification. In ICCV, pages 1025–1034, 2021. 2", + "[39] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In Conference on Robot Learning, pages 1046-1056. PMLR, 2022. 7", + "[40] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In CVPR, pages 3716-3725, 2020. 2, 4", + "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurlPS, 2022. 4, 5", + "[42] Hongjun Wang, Jiyuan Chen, Lun Du, Qiang Fu, Shi Han, and Xuan Song. Causal-based supervision of attention in graph neural network: A better and simpler choice towards powerful attention. arXiv preprint arXiv:2305.13115, 2023. 2", + "[43] Hanqing Wang, Wei Liang, Luc Van Gool, and Wenguan Wang. Dreamwalker: Mental planning for continuous vision-language navigation. In ICCV, pages 10873-10883, 2023. 1, 2", + "[44] Tan Wang, Jianqiang Huang, Hanwang Zhang, and Qianru Sun. Visual commonsense r-cnn. In CVPR, pages 10760-10770, 2020. 2", + "[45] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In CVPR, pages 19231-19242, 2023. 6, 7" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 2, 3, 6, 7", + "[47] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In CVPR, pages 1791-1800, 2021. 2, 6, 7", + "[48] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Shuguang Cui, and Zhen Li. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In CVPR, pages 8563-8573, 2022. 2", + "[49] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, pages 15225-15236, 2023. 6, 7", + "[50] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2928-2937, 2021. 2, 6, 7", + "[51] Tong Zhao, Gang Liu, Daheng Wang, Wenhao Yu, and Meng Jiang. Learning from counterfactual links for link prediction. In Int. Conf. on Machine Learning, pages 26911-26926. PMLR, 2022. 2", + "[52] Dave Zhenyu Chen, Qirui Wu, Matthias Niener, and Angel X Chang. D3net: A speaker-listener architecture for semi-supervised dense captionin and visual grounding in rgb-d scans. In ECCV, 2022. 2", + "[53] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 2" + ], + "bbox": [ + 501, + 90, + 890, + 513 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17262", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/477f7931-e1b2-442e-a45b-11cf7d648a49_model.json b/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/477f7931-e1b2-442e-a45b-11cf7d648a49_model.json new file mode 100644 index 0000000000000000000000000000000000000000..86b89d8d63ba831eb33f6c6fe41fa591fbdff4c6 --- /dev/null +++ b/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/477f7931-e1b2-442e-a45b-11cf7d648a49_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.175, + 0.131, + 0.797, + 0.154 + ], + "angle": 0, + "content": "Multi-Attribute Interactions Matter for 3D Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.18, + 0.852, + 0.287 + ], + "angle": 0, + "content": "Can Xu\\(^{1}\\), Yuehui Han\\(^{1}\\), Rui Xu\\(^{1}\\), Le Hui\\(^{2}\\), Jin Xie\\(^{3,4*}\\), Jian Yang\\(^{1}\\) \n\\(^{1}\\)PCA Lab, Nanjing University of Science and Technology, Nanjing, China \n\\(^{2}\\)School of Electronics and Information, Northwestern Polytechnical University, Xi'an, China \n\\(^{3}\\)State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China \n\\(^{4}\\)School of Intelligence Science and Technology, Nanjing University, Suzhou, China \nvolcanoxl23@gmail.com; {hanyh,xu-ray,csjxie,csjyang}@njust.edu.cn; huile@nwpu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.474, + 0.78 + ], + "angle": 0, + "content": "3D visual grounding aims to localize 3D objects described by free-form language sentences. Following the detection-then-matching paradigm, existing methods mainly focus on embedding object attributes in unimodal feature extraction and multimodal feature fusion, to enhance the discriminability of the proposal feature for accurate grounding. However, most of them ignore the explicit interaction of multiple attributes, causing a bias in unimodal representation and misalignment in multimodal fusion. In this paper, we propose a multi-attribute aware Transformer for 3D visual grounding, learning the multi-attribute interactions to refine the intra-modal and inter-modal grounding cues. Specifically, we first develop an attribute causal analysis module to quantify the causal effect of different attributes for the final prediction, which provides powerful supervision to correct the misleading attributes and adaptively capture other discriminative features. Then, we design an exchanging-based multimodal fusion module, which dynamically replaces tokens with low attribute attention between modalities before directly integrating low-dimensional global features. This ensures an attribute-level multimodal information fusion and helps align the language and vision details more efficiently for fine-grained multimodal features. Extensive experiments show that our method can achieve state-of-the-art performance on ScanRefer and Sr3D/Nr3D datasets. The code is publicly available at https://github.com/volcanoXC/MA2TransVG." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.813, + 0.21, + 0.829 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "3D Visual Grounding (VG) aims to locate the most relevant object in a given point cloud scene based on a language description. As a cornerstone of wide applications [6, 8, 14, 28, 43], such as vision-language navigation and au" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.321, + 0.891, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.613, + 0.894, + 0.656 + ], + "angle": 0, + "content": "Figure 1. Different from prior works, we capture the multi-attribute interactions to guide the unimodal representation and multimodal fusion for more accurate grounding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.763 + ], + "angle": 0, + "content": "tonomous robots, 3D VG has attracted increasing attention from both academia and industry. However, limited by the complexity of free-form language description and the irregularity of sparse point clouds, achieving accurate 3D visual grounding remains an open issue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Existing 3D VG methods mainly follow a detection-then-matching pipeline. Specifically, a pool of proposals is first produced by the pre-trained general 3D object detectors [11, 29, 37]. Then, the linguistic description is associated with the visual feature of each proposal to generate multimodal object representation, which will be used to predict their referring confidence score for grounding. Since such a two-stage paradigm aims to collect enough semantic cues to distinguish the target object, the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17253" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.289 + ], + "angle": 0, + "content": "quality of unimodal feature representation and multimodal feature fusion always play a critical role, which decides the discriminativeness of each proposal and their matching difficulty with the target label. Most methods focus on enhancing the feature from various perspectives, such as sparse convolution [47], 2D image assistance [5, 27, 46, 52], coarse proposal refinement [13, 22, 50], or segmentation/captioning collaboration [3, 10, 18, 26]. As shown in Fig. 1(a), recent works [7, 18, 19, 35] enhance the feature by embedding object attributes, such as color, name and language-conditioned correlation (i.e., distance, view, direction), showing a promising future in advancing grounding performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.473, + 0.61 + ], + "angle": 0, + "content": "Though prior works made significant progress, we argue few of them explore the multi-attribute interactions in either unimodal representation or multimodal fusion, which affects the grounding accuracy, especially for cases with multiple objects. As the CASE in Fig. 1, the model embeds the most salient attribute (object name of 'door') and learns other potential attributes (e.g., 'brown', 'wooden', 'left') depending on the self-attention mechanism. Since the loss supervises the final prediction rather than the intermediate attention, the model may be encouraged to focus on some main attributes instead of all attributes, which weakens the discrimination of intra-modal representation. Meanwhile, when fusing the low-dimensional global feature between modalities via cross-attention, attention to 'brown' and 'wooden' should be encouraged for more discriminative language expression of the object, while they may be suppressed in the proposal representation to maximize the proposal's distinctiveness. Unfortunately, this would bring a misalignment in multimodal features, causing the model to select the auxiliary object (right door) rather than the primary object (left door)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.902 + ], + "angle": 0, + "content": "To alleviate the bias in unimodal representation and misalignment in multimodal fusion, we propose a multi-attribute aware Transformer \\((\\mathsf{MA}^2\\mathsf{TransVG})\\) for 3D visual grounding, which enhances the unimodal representation with intrinsic attribute interaction and refines the cross-modal information fusion to the attribute level (See Fig. 1(b)). Concretely, we develop an attribute causal analysis module (ACAM), which quantifies the effect of each attribute (i.e., learned attentions) and counterfactual interference (i.e., wrong attention) for the final prediction (i.e., grounding score). By maximizing the causal effect, powerful supervision is provided to correct the biased attribute and adaptively assign more proper attention to all attributes. Guided by the learned interaction, we further propose an exchanging-based multimodal fusion module (EMFM) to replace tokens with low attribute attention between modalities, which helps align more vision and language details and generate fine-grained multimodal features for accurate grounding. Our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "- We propose a multi-attribute aware Transformer for 3D object grounding, which explicitly models the multi-attribute interactions to prevent imbalanced and misaligned visual-textual representation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.152, + 0.895, + 0.213 + ], + "angle": 0, + "content": "- We propose the ACAM module to quantify the causal effect of multiple intrinsic attributes with the final prediction, which helps adaptively assign more proper attention to all attributes for a better intra-modal understanding." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.213, + 0.895, + 0.272 + ], + "angle": 0, + "content": "- We propose the EMFM module to perform multimodal fusion based on multi-attribute exchanging, which helps align more attribute-level details between modalities for fine-grained multimodal features." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.273, + 0.895, + 0.319 + ], + "angle": 0, + "content": "- We evaluate the proposed method on \\(\\mathrm{Nr3D/Sr3D}\\) and ScanRefer benchmarks, which outperforms all state-of-the-art methods by a large margin." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.895, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.332, + 0.64, + 0.348 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.895, + 0.644 + ], + "angle": 0, + "content": "3D Visual Grounding. The interest in 3D VG has grown rapidly in recent years, and existing approaches include one-stage [21, 31] and two-stage methods. The two-stage methods [13, 16-18, 46, 47, 50] decouple the grounding into language-irrelevant object detection and cross-modal matching. Recent works focus on improving the intramodal representation and modeling the inter-modal relationship. For example, Graph-based approaches [1, 18, 47] infer spatial relations among proposals by connecting each proposal with its top-N nearest neighbors. Transformer-based approaches [7, 17, 19, 21, 31, 46] leverage the attention mechanism to concentrate the object attributes on proposals that are more essential to the referring. In these methods, only the most salient attribute is embedded and the fusion between modalities occurs in the global feature space, resulting in a biased intra-modal representation and inadequate inter-modal fusion. Instead, we capture the multi-attribute interactions and refine the cross-modal learning at the attribute level." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.896, + 0.825 + ], + "angle": 0, + "content": "Causal Inference in Vision. Causality analysis [32] has been successfully used in several areas, such as adversarial learning [24], reinforcement learning [23] and graph neural networks [42, 51]. In vision tasks [30, 40, 44], causality inference also works as an effect tool to discover hidden causal structures and confront data biases. For example, Rao et al. [38] learn more effective attention based on causal inference for fine-grained image categorization, person and vehicle re-identification. As the first attempt, we introduce the causality analysis into 3D VG to quantify the effect of each attribute, which helps model the complex interactions of various attributes to enhance the unimodal feature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Transformers in Vision-and-Language. Transformer shows a powerful ability to learn multimodal features in various tasks, such as image captioning [9, 10, 48], vision-and-language navigation [28, 43], and vision question-answering [15, 53]. For 3D VG, different from most" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "17254" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.096, + 0.887, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.895, + 0.391 + ], + "angle": 0, + "content": "Figure 2. The framework of proposed \\(\\mathsf{MA}^2\\mathsf{TransVG}\\). The embedding of pre-detected proposals, attributes, and language are extracted and then cross-attend in ACAM. By applying the causal effect analysis via \\(L_{attr}\\), ACAM models the multi-attribute interactions for final prediction to generate the Attribute-Vision/Text attention, which can enhance the proposal and text embedding. The attention further guides the attribute-level feature exchanging in EMFM, which helps fuse the multimodal feature for accurate grounding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.471, + 0.492 + ], + "angle": 0, + "content": "methods that directly fuse the global unimodal features by stacking cross-attention layers [7, 17, 19, 21, 31, 46], we refine the cross-modal fusion to the attribute level, which helps suppress the misalignment between modalities to capture a finer-grained matching relationship." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.504, + 0.17, + 0.518 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.187, + 0.542 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Fig. 2 shows an overview of proposed \\(\\mathsf{MA}^2\\mathsf{TransVG}\\). Given a 3D point cloud with corresponding description, text features \\(T \\in \\mathbb{R}^{(m + 1)\\times d}\\), proposal features \\(O \\in \\mathbb{R}^{n\\times d}\\) and attribute features \\(F \\in \\mathbb{R}^{n\\times d}\\) are extracted (Sec. 3.2). The three streams then are fed into the attribute causal analysis module (ACAM) to capture the multi-attribute interactions (i.e., Attribute-Vision attention and Attribute-Text attention) to enhance initial text and proposal feature. The attention is supervised by an attribute loss \\(L_{attr}\\), which is designed by estimating the causal effect of attributes for the final prediction (Sec. 3.3). The enhanced proposal embedding \\(O^*\\) and text embedding \\(T^*\\) are further passed to the exchanging-based multimodal fusion module (EMFM), where their tokens with small attribute attention are replaced with the average of all the tokens in the other modality (Sec. 3.4). With the fused multimodal embedding \\(E \\in \\mathbb{R}^{n\\times d}\\), we predict the final grounding results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.334, + 0.833 + ], + "angle": 0, + "content": "3.2. Input Modal Representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Text Encoding. Given a sentence \\( S \\) with \\( m \\) word tokens, we use a pre-trained BERT model [20] to produce the text features \\( (t_{cls}, t_1, \\ldots, t_m) \\). \\( t_i \\in \\mathbb{R}^d \\) is the feature of each token with a dimensionality of \\( d \\) and \\( t_{cls} \\in \\mathbb{R}^d \\) is a special" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.416, + 0.686, + 0.429 + ], + "angle": 0, + "content": "token for text classification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.566 + ], + "angle": 0, + "content": "Object Encoding. Given a point cloud scene \\( \\pmb{P} \\in \\mathbb{R}^{K \\times 6} \\) of \\( K \\) points described by XYZ coordinates and RGB colors, we use the GroupFree [29] detector to obtain an object proposal list \\( \\pmb{o} = (o_1, o_2, \\dots, o_n) \\) containing \\( n \\) proposals and extract corresponding proposal feature \\( \\pmb{O} \\in \\mathbb{R}^{n \\times d} \\) based on PointNet++ [36]. Since we mainly focus on the object referred in the text, each proposal feature \\( O_i \\in \\mathbb{R}^d \\) will be added with the \\( t_{cls} \\) before feeding into ACAM for attribute-vision causal effect analysis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.893, + 0.732 + ], + "angle": 0, + "content": "Attributes Encoding. Motivated by [3], for the proposal pool \\(o\\) with \\(n\\) objects, the multi-view RGB features \\(\\pmb{f}_{c} \\in \\mathbb{R}^{n \\times 128}\\) that potentially contain the color and materials attribute is first extracted. Then, we use the 3-dimensional box center and 24-dimensional corner coordinates to describe the object size attribute \\(\\pmb{f}_{s} \\in \\mathbb{R}^{n \\times 27}\\). For the location attribute, we further consider the distance and angle of each pair of objects \\((o_{i}, o_{j})\\). Connecting the object center point \\((x_{c_{i}}, y_{c_{i}}, z_{c_{i}})\\) and \\((x_{c_{j}}, y_{c_{j}}, z_{c_{j}})\\), the distance is their Euclidean distance \\(d_{ij}\\). Similar to [7], the horizontal angle \\(\\theta_{h}\\) and vertical angle \\(\\theta_{v}\\) can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.742, + 0.891, + 0.774 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c} \\theta_ {h} = \\arctan 2 \\left(\\left(y _ {c _ {i}} - y _ {c _ {j}}\\right) / \\left(x _ {c _ {i}} - x _ {c _ {j}}\\right) \\right. \\\\ \\theta_ {v} = \\arcsin \\left(\\left(z _ {c _ {i}} - z _ {c _ {j}}\\right) / d _ {i j}\\right) \\end{array} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.786, + 0.891, + 0.816 + ], + "angle": 0, + "content": "The location attribute \\( f_{l}^{ij} \\in \\mathbb{R}^{5} \\) between any pair of objects is defined, as:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.826, + 0.891, + 0.845 + ], + "angle": 0, + "content": "\\[\nf _ {l} ^ {i j} = \\left[ d _ {i j}, \\sin \\left(\\theta_ {h}\\right), \\cos \\left(\\theta_ {h}\\right), \\sin \\left(\\theta_ {v}\\right), \\cos \\left(\\theta_ {v}\\right) \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.854, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For each proposal \\(o_i\\), the location attribute \\(\\pmb{f}_l^i \\in \\mathbb{R}^5\\) can be calculated by averaging all \\(f_l^{ij}\\) connected with \\(o_i\\), and the final attribute embedding can be generated by concatenating" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17255" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.47, + 0.122 + ], + "angle": 0, + "content": "all three attributes, as \\(\\pmb{f}^{i} = [\\pmb{f}_{l}^{i},\\pmb{f}_{c}^{i},\\pmb{f}_{s}^{i}]\\). We map \\(\\pmb{f}^{i}\\in \\mathbb{R}^{d}\\) to the same dimension of text feature using a FC layer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.133, + 0.376, + 0.15 + ], + "angle": 0, + "content": "3.3. Attribute Causal Analysis Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.158, + 0.469, + 0.263 + ], + "angle": 0, + "content": "The attribute causal analysis module (ACAM) models the multi-attribute interactions with vision/text as Attribute-Vision/Text attention based on a well-designed Attribute-aware Transformer. To capture more influential and diverse interactions, ACAM further guides the learning process with an extra loss \\( L_{attr} \\), which measures the causal effect of learned attribute-level attentions for the final prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.265, + 0.469, + 0.385 + ], + "angle": 0, + "content": "Attributes-Text Attention Learning. Given attribute embedding \\( \\pmb{F} \\in \\mathbb{R}^{n \\times d} \\), Attributes-Text attention \\( \\pmb{A}_{T} \\) helps select related attribute features from the initial text feature \\( T \\in \\mathbb{R}^{(m + 1) \\times d} \\), thereby the text representation can be enhanced to better describe the target object. We define \\( \\pmb{A}_{T} \\in \\mathbb{R}^{n \\times (m + 1)} \\) as the cross-attention between the Attribute embedding \\( \\pmb{F} \\) and text embedding \\( \\pmb{T} \\), which can be computed with a standard Transformer [41]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.469, + 0.522 + ], + "angle": 0, + "content": "In order to enlarge the model capacity for diverse Attribute-Text correlations, we also introduce multi parallel attention layers \\(\\ell\\) as different transformation heads, and the multi-head attentions will be fused with the standard self-attention [41]. As shown in Fig. 2, \\(A_{T}\\) then guides to generate the enhanced text feature \\(T^{*} \\in \\mathbb{R}^{n \\times d}\\), which will be used to train the language classifier with the supervision of the language classification loss \\(L_{sent}\\), and also serves as input for subsequent EMFM for multi-modal fusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.469, + 0.659 + ], + "angle": 0, + "content": "Attributes-Vision Attention Learning. Instead of directly calculating the cross-attention between proposal features \\( \\mathcal{O} \\) and attribute features \\( \\mathbf{F} \\), we first guide the proposal features with language token \\( t_{cls} \\), as we prefer to care about specific categories of objects mentioned in the text [7]. Specifically, for the text-guided proposal feature, we adopt the similar Transformer structure in Attributes-Text attention learning to calculate its self-attention \\( A_V^s \\in \\mathbb{R}^{n \\times n} \\), and cross-attention \\( A_V^c \\in \\mathbb{R}^{n \\times n} \\) with \\( \\mathbf{F} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.721 + ], + "angle": 0, + "content": "Similar to [7], we then integrate the self-attention of proposals \\( A_V^s \\) with the cross attribute-proposal attention \\( A_V^c \\) using a sigmoid softmax (sigsoftmax) fusion operation to calculate the Attributes-Vision attention \\( A_V \\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.732, + 0.469, + 0.768 + ], + "angle": 0, + "content": "\\[\nA _ {v} = \\frac {\\sigma \\left(\\boldsymbol {A} _ {\\boldsymbol {V}} ^ {\\boldsymbol {c}} (i , j)\\right) \\exp \\left(\\boldsymbol {A} _ {\\boldsymbol {V}} ^ {\\boldsymbol {s}} (i , j)\\right)}{\\sum_ {j = 1} ^ {n} \\sigma \\left(\\boldsymbol {A} _ {\\boldsymbol {V}} ^ {\\boldsymbol {c}} (i , j)\\right) \\exp \\left(\\boldsymbol {A} _ {\\boldsymbol {V}} ^ {\\boldsymbol {s}} (i , j)\\right)} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\sigma(\\cdot)\\) is the sigmoid function. \\(A_V^s (i,j)\\) and \\(A_V^c (i,j)\\) are the element in \\(i\\)-th row and \\(j\\)-th column in attention matrix \\(A_V^s\\) and \\(A_V^c\\), respectively. The same multi-head attention strategy is used in Attribute-Vision attention \\(A_V\\). In a similar manner, the enhanced proposal embedding \\(O^{*}\\) can be obtained by performing a matrix multiplication operation between \\(O\\) and \\(A_V\\), which will serve as the input of proposal classifier and EMFM module." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.091, + 0.888, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.204, + 0.892, + 0.261 + ], + "angle": 0, + "content": "Figure 3. Visualization of deriving causal effects through counterfactual causality. We subtract the counterfactual prediction from the original prediction to analyze the effects of learned attention and maximize them in the training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Strengthen Interaction With \\( L_{attri} \\). Though multiple attributes have been embedded to guide the language and vision representations via Attribute-Text/Vision attention, how these attributes interact with each other to affect predictions remains a black box. To address this issue, our core idea lies in measuring the causal effect of \\( A_V \\) and \\( A_T \\) for the final prediction. We design the attribute loss \\( L_{attri} \\) based on the counterfactual analysis [33], providing extra supervision to encourage the network to explore more influential interactions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.434, + 0.892, + 0.539 + ], + "angle": 0, + "content": "As shown in Fig. 3, taking the Attribute-vision attention as an example, we conduct counterfactual intervention by imagining non-existent attention \\(\\bar{A}_V\\) to replace the initial \\(A_V\\). We keep the vision representation \\(O\\) unchanged, which will be multiplied with \\(\\bar{A}_V\\) to calculate the counterfactual representation \\(\\bar{O}^*\\). \\(\\bar{O}^*\\) are then fused with \\(T\\) to obtain the counterfactual prediction \\(\\bar{Y} \\in \\mathbb{R}^{n \\times 1}\\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.545, + 0.891, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\bar {\\boldsymbol {Y}} = d o \\left(\\boldsymbol {A} _ {V} = \\bar {\\boldsymbol {A}} _ {V}, \\boldsymbol {T}\\right) = G \\left(\\bar {\\boldsymbol {O}} ^ {*}, \\boldsymbol {T}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.644 + ], + "angle": 0, + "content": "where \\( G(\\cdot) \\) is the grounding operation. In practice, the exact form of how counterfactual is achieved is not limited, and our starting point here is just to set the ball rolling. In our case, random attention, uniform attention, shuffle attention or reversed attention are used as the counterfactuals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.69 + ], + "angle": 0, + "content": "Following [33, 34, 40], we define the causal effect \\( \\mathbf{Y}_e \\) of \\( \\mathbf{A}_V \\) as the difference between the observed prediction \\( \\mathbf{Y} \\) and the counterfactual prediction \\( \\bar{\\mathbf{Y}} \\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.696, + 0.891, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Y} _ {e} = \\mathbb {E} _ {\\bar {A} _ {V} \\sim \\gamma} (\\boldsymbol {Y} - \\bar {\\boldsymbol {Y}}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.888, + 0.734 + ], + "angle": 0, + "content": "where \\(\\gamma\\) means the distribution of counterfactual attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.811 + ], + "angle": 0, + "content": "Since the causal effect can be interpreted as how the attention improves the final prediction compared to wrong attentions, we use \\( \\mathbf{Y}_e \\) to measure the learned attention, i.e., the multi-attribute interactions. Based on this observation, we further design an attribute loss \\( L_{attr} \\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.818, + 0.891, + 0.835 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {a t t r i}} = L _ {c e} \\left(\\boldsymbol {Y} _ {\\boldsymbol {e}}, \\boldsymbol {y}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(L_{ce}\\) is the cross-entropy loss and \\(\\pmb{y}\\) is the GT label. Similarly, we can obtain the causal effect of Attribute-text attention and compute its attribute loss. By optimizing \\(L_{attr}\\) from \\(A_{T}\\) and \\(A_{V}\\), the model is expected to:" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17256" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "(1) improve the prediction based on correct attentions as much as possible, which helps discover more discriminative attributes; (2) penalize the prediction based on wrong attentions, which forces the model adaptively capture more attributes to avoid biased unimodal embeddings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.177, + 0.469, + 0.193 + ], + "angle": 0, + "content": "3.4. Exchanging-based Multimodal Fusion Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.469, + 0.395 + ], + "angle": 0, + "content": "We propose the exchanging-based multimodal fusion module (EMFM) to communicate attribute-level details of objects between both input vectors for accurate grounding. Specifically, since text/visual encoders tend to focus on different attributes for better unimodal embeddings, directly fusing unimodal embeddings using cross-attention will hurt the distinguishability of multimodal features. To narrow the semantic misalignment between modalities, we select and exchange tokens with low attention score using the average of all tokens from another modal, which helps align attribute-level information between modalities sufficiently and eliminate semantic ambiguity in multimodal representations of proposals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.469, + 0.563 + ], + "angle": 0, + "content": "Exchanging With Attributes Interactions. As shown in Fig. 3, we implement the EMFM by two Transformer encoders with shared parameters, where the shallow \\(\\mu\\) layers are set as regular Transformer encoder layers, followed by \\(\\eta\\) exchanging layers. Since \\(A_V\\) and \\(A_T\\) have modeled the multi-attribute interactions, they are taken as references to select tokens for exchanging. Meanwhile, \\(A_V\\) and \\(A_T\\) will be updated using the self-attention learned from unimodal embedding \\(T^*\\) and \\(O^*\\). Considering the interdependence between the attribute attention and unimodal embedding, we concatenate them for the subsequent calculation, as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.575, + 0.469, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {T} ^ {e} (0) = \\operatorname {C o n c a t} \\left[ \\boldsymbol {A} _ {T}, \\boldsymbol {T} ^ {*} \\right] \\\\ \\boldsymbol {Q} ^ {e} (0) = \\operatorname {C o n c a t} \\left[ \\boldsymbol {A} _ {T}, \\boldsymbol {Q} ^ {*} \\right] \\end{array} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.591, + 0.362, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {O} ^ {e} (0) = \\operatorname {C o n c a t} \\left[ \\boldsymbol {A} _ {V}, \\boldsymbol {O} ^ {*} \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.469, + 0.648 + ], + "angle": 0, + "content": "where \\( T^{e}(0) \\) and \\( O^{e}(0) \\) are embedding for textual and visual modalities at layer 0, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.469, + 0.732 + ], + "angle": 0, + "content": "At layer \\(\\mu + 1\\) with intermediate embedding \\(T^{e}(\\mu + 1)\\), inspired by [4, 25], we select tokens with a \\(\\theta\\)-proportion of the smallest attention scores to \\(A_{T}\\) and replace their embedding vectors with the average embedding of all the tokens in \\(O^{e}(\\mu + 1)\\). For instance, the selected \\(k\\)-th token of \\(T^{e}(\\mu + 1)\\) can be updated by:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.743, + 0.469, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {T} ^ {\\boldsymbol {e}} (\\mu + 1) [ k,: ] = \\frac {1}{n} \\sum_ {j = 1} ^ {n} \\boldsymbol {O} ^ {\\boldsymbol {e}} (\\mu + 1) [ j,: ] + \\boldsymbol {T} ^ {\\boldsymbol {e}} (\\mu + 1) [ k,: ] \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.855 + ], + "angle": 0, + "content": "The updated embeddings are used to calculate the multihead self-attention [41], which will guide the exchanging operation at the next layer, e.g., \\( A_T \\) of the \\( (\\mu + 1) \\)-th layer are updated by the self-attention of the previous layer \\( \\mu \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Multimodal Fusion. We also apply the residual connection between adjacent exchanging layers to reduce the information loss caused by replacement. The exchanging process" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "continues until reaching the \\(\\eta\\) layer to generate \\(O^e (\\eta)\\) and \\(T^{e}(\\eta)\\), which will be fed into the feed-forward network (FFN) with a normalization layer and a fully connected layer to derive the final fusion embedding \\(\\pmb {E}\\in \\mathbb{R}^{n\\times n}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Grounding Head. We use a two-layer feed-forward neural network as the object grounding head. Given the multimodal representation \\(\\pmb{E}\\), the grounding head will predict the probability \\(\\pmb{p}_i \\in \\mathbb{R}^{n \\times 1}\\) for each object proposal, where the proposal with the maximum probability is selected as the target in the inference stage." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.251, + 0.688, + 0.267 + ], + "angle": 0, + "content": "3.5. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.892, + 0.35 + ], + "angle": 0, + "content": "Following the previous works [1, 5, 17], the loss of \\(\\mathbf{MA}^2\\mathrm{TransVG}\\) consists of the primary 3D object grounding loss \\(L_{vg}\\), language classification loss \\(L_{sent}\\), proposal classification loss \\(L_{obj}\\), and two attribute losses \\(L_{attri}\\) for vision/text attention causal analysis, as:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.36, + 0.892, + 0.377 + ], + "angle": 0, + "content": "\\[\nL = L _ {v g} + L _ {s e n t} + L _ {o b j} + \\lambda L _ {a t t r i} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.386, + 0.835, + 0.401 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the hyper-parameter to balance the loss." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.414, + 0.624, + 0.43 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.439, + 0.786, + 0.453 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.78 + ], + "angle": 0, + "content": "We evaluate the proposed \\(\\mathsf{MA}^2\\mathsf{TransVG}\\) on ScanRefer [5] and Sr3D/Nr3D [1] datasets. ScanRefer offers 51,583 descriptions of 11,046 objects from 800 ScanNet [12] scenes. A target object is labeled as 'unique' if it is the only object of its class in the scene; otherwise, it is labeled as 'multiple'. For ScanRefer, we use Acc@0.25 and Acc@0.5 as evaluation metrics, which measure the fraction of language queries whose predicted box overlaps the ground truth box with 3D IoU higher than 0.25/0.5. Compared to ScanRefer where objects need to be detected and then matched, Sr3D/Nr3D datasets provide object masks for each scene and only need to recognize the classes of the proposals to choose the target object. Nr3D consists of 41,503 descriptions from 707 scenes with human annotation, while Sr3D contains 83,572 simple machine-generated descriptions. Each scene in Sr3D/Nr3D can also be divided into 'easy' and 'hard' depending on whether there are more than two instances, and can be divided into 'view-dependent' and 'view-independent' according to whether the referring expression is dependent on the camera view. Following [1], the accuracy is used to verify the model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.716, + 0.803 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The text encoding module is initialized from the first three layers of BERT [20], and the object encoding module PointNet++ [36] samples 1024 points for all the objects. We generate an initial set of \\( n = 256 \\) proposals using [29]. We set the dimension of unimodal representation \\( d = 768 \\) and use a three-layer transformer with 12 heads for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17257" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.255, + 0.089, + 0.717, + 0.104 + ], + "angle": 0, + "content": "Table 1. Comparison of 3D VG results with the state-of-the-art on ScanRefer." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.114, + 0.727, + 0.378 + ], + "angle": 0, + "content": "
MethodVenueUnique (19%)Multiple (81%)Overall
0.250.50.250.50.250.5
ScanRefer [5]ECCV2067.646.232.121.340.026.1
ReferIt3D [1]ECCV2053.837.521.012.826.416.9
TGNN [18]AAAI2168.656.829.823.237.429.7
InstanceRefer [47]ICCV2177.566.831.324.840.232.9
3DVG [50]ICCV2177.258.538.428.745.934.5
FFL-3DOG [13]ICCV2178.867.935.225.741.334.0
MVT [19]CVPR2277.766.531.925.340.833.3
3D-SPS [31]CVPR2281.664.839.529.647.736.4
ViL3DRef [7]NeurIPS2281.668.640.330.747.937.7
BUTD-DETR [22]ECCV22////52.239.8
EDA [45]CVPR2385.868.649.137.654.642.3
M3DRef-CLIP [49]ICCV23/77.2/36.8/44.7
3DJCG [3]CVPR2278.861.340.130.147.636.1
UniT3D [10]ICCV2382.873.136.431.142.339.1
3DRefTR-HR [26]ICCV2386.070.949.638.355.043.2
3DRefTR-SP [26]ICCV2386.171.050.138.755.543.5
Ours/86.374.153.841.457.945.7
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.472, + 0.584 + ], + "angle": 0, + "content": "the Attribute-Vision/Text attention learning. For a fair comparison, we keep the architecture parameters as same as previous works [7, 19, 46]. In ACAM, we use random attention to initialize the counterfactual attention. In EMFM, considering the balance of raw intra-modal knowledge and inter-modal communication, we empirically set \\(\\theta = 10\\%\\). The start layer \\(\\mu\\) and end layer \\(\\eta\\) are set to 1 and 4, respectively. The hyper-parameter of the whole loss in Eq. (9) is \\(\\lambda = 0.5\\). All models are trained on 4 NVIDIA RTX Titan GPUs, which are optimized using the AdamW algorithm with a batch size of 128 and a learning rate of 0.0005 with cosine decay scheduling." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.59, + 0.446, + 0.607 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.471, + 0.901 + ], + "angle": 0, + "content": "ScanRefer. In Tab. 1, we compare \\(\\mathsf{MA}^2\\mathsf{TransVG}\\) with existing works on the ScanRef dataset and observe the following results: i) Compared to pure 3D VG works such as BUTD-DETR [22], 3D-SPS [31], ViL3DRef [7], and EDA [45], our method achieves state-of-the-art performance of \\(57.9\\%\\) and \\(45.7\\%\\), by a substantial margin with an overall improvement over \\(3.3\\%\\) and \\(3.4\\%\\). ii) To facilitate the 3D grounding, previous works [5, 31, 46, 50] learn better point cloud features with the supplementary of 2D images while most recent works [3, 10, 26] learn better language-visual alignment by connecting 3D dense captioning or segmentation tasks. However, we outperform second-place 3DRefTR-SP [26] with an overall improvement over \\(2\\%\\) for both IoU = 0.25 and IoU = 0.5 setting. This superiority illustrates that mining the multi-attribute interactions help capture more efficient and discriminative multi-modal representation to distinguish objects. iii) Our \\(\\mathsf{MA}^2\\mathsf{TransVG}\\) can reach a remarkable accuracy of \\(53.8\\%\\) and \\(41.4\\%\\) for 'Multiple' scenes, which verifies the effec" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.403, + 0.892, + 0.434 + ], + "angle": 0, + "content": "tiveness of our attribute interaction for understanding finer-grained language-visual contextual dependency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Nr3D&Sr3D. Tab. 2 reports the grounding accuracy on Nr3D and Sr3D datasets. For a fair comparison, all the methods use the ground-truth object proposal with no ground-truth labels. In Nr3D, though the language descriptions are more complex to cause additional challenges for text understanding and cross-modal alignment, our method still outperforms the EDA [45], which is equipped with text component decoupling and dense matching between two modalities. Compared to EDA, our \\(\\mathsf{MA}^2\\mathsf{TransVG}\\) not only achieves a similar decoupling function for both descriptions and point clouds via attribute representation, but also considers their correlationships for prediction during Attribute-Text/Vision causal attention learning, which contributes to the great improvements by \\(+13.1\\%\\) in overall accuracy. Compared to other recent competitor ViL3DRef [7], 3DRefTRSR [26], and M3DRef-CLIP [49], we also achieve a consistent improvement of overall accuracy in all split settings by \\(+0.8\\%\\), \\(+2.6\\%\\) and \\(+15.8\\%\\), respectively. In Sr3D, our method reaches an accuracy of \\(73.9\\%\\), which greatly surpasses the best competitor ViL3DRef [7] \\(1.1\\%\\) and outperforms other state-of-the-art methods [26, 45] over \\(5\\%\\). Similar to ViL3DRef [7] that guided the grounding with languaged-conditioned spatial relation, we not only embed the spatial attributes but also explicitly consider other attributes in the unimodal representation and further perform exchanging-based fusion. All the results show that multi-attribute interactions can help capture more semantic clues for accurate grounding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17258" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.251, + 0.089, + 0.719, + 0.104 + ], + "angle": 0, + "content": "Table 2. Comparison of 3D VG results with the state-of-the-art on Sr3D/Nr3D." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.115, + 0.847, + 0.407 + ], + "angle": 0, + "content": "
MethodVenueNr3DSr3D
EasyHardView DepView IndepOverallEasyHardView DepView IndepOverall
ReferIt3D [1]ECCV2043.627.932.537.135.644.731.539.240.840.8
TGNN [18]AAAI2144.230.635.838.037.348.536.945.845.045.0
InstanceRefer [47]ICCV2146.031.834.541.938.851.140.545.448.148.0
3DVG [50]ICCV2148.534.834.843.740.854.244.944.651.751.4
FFL-3DOG [13]ICCV2148.235.037.144.741.7/////
TransRefer3D [17]ACMM2148.536.036.544.942.160.550.249.957.757.4
LanguageRefer [39]CoRL2151.036.641.745.043.958.949.349.256.356.0
SAT [46]ICCV2156.342.446.950.449.261.250.049.258.357.9
BUTD-DETR [22]ECCV22////43.3////52.1
LAR [2]NeurIPS2258.442.347.452.148.963.051.250.059.159.4
3D-SPS [31]CVPR2258.145.148.053.251.556.265.449.263.262.6
MVT [19]CVPR2261.349.154.354.355.466.958.858.464.764.5
M3DRef-CLIP [49]ICCV2355.643.442.352.949.4/////
EDA [45]CVPR23////52.1////68.1
3DRefTR-SR [26]ICCV23////52.6////68.5
ViL3DRef [7]NeurIPS2270.257.462.064.564.474.967.963.873.272.8
Ours/71.157.662.565.465.276.069.364.573.873.9
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.43, + 0.241, + 0.445 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.471, + 0.746 + ], + "angle": 0, + "content": "Effectiveness of Each Component. We quantitatively investigate the contribution of each component in our method in Tab. 3. The baseline model simply extracts and fuses the object and language feature through a standard Transformer structure, where the modal relationships only implicitly depend on the self/cross-attention. Taking results on \\(\\mathrm{Nr3D}\\) as an example, when enhancing the feature with \\(A_{T}\\) or \\(A_{V}\\), we yield an accuracy of \\(55.2\\%\\) and \\(51.9\\%\\), which can increase to \\(59.7\\%\\) when combining two attentions. This indicates that embedding the multi-attribute within each single modal is the basis of accurate cross-modal understanding. Furthermore, we achieve a gain of \\(2.7\\%\\) after introducing causal effect analysis, which helps adaptively capture more reliable multi-attribute interactions for feature enhancement. Exchanging tokens further improves the accuracy to \\(65.2\\%\\) because multimodal knowledge exchanging boosts the quality of the learned embeddings for each modality. All the results demonstrate the effectiveness of each component in our proposed \\(\\mathrm{MA}^2\\mathrm{TransVG}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of Multi-attribute. To verify the effectiveness of multi-attribute interactions, we conduct ablation on attribute quantity and multi-attribute interactions. In Tab. 3. We adopt commonly used attributes (location \\( f_{l} \\), size \\( f_{s} \\), color \\( f_{c} \\)) with simple embedding method and achieve SOTA performance. The accuracy is \\( 44.9\\% \\) on Nr3D when generating attribute embedding only using \\( f_{l} \\) and \\( f_{s} \\), which increases to \\( 46.1\\% \\) (\\( +1.2\\% \\)) and \\( 60.2\\% \\) (\\( +15.3\\% \\)) after adding \\( f_{c} \\) and interactions, respectively. Results show that our accuracy gain mainly comes from the multi-attribute" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.429, + 0.889, + 0.443 + ], + "angle": 0, + "content": "Table 3. Ablation studies of each component on \\( \\mathrm{{Nr}}3\\mathrm{D} \\) and \\( \\mathrm{{Sr}}3\\mathrm{D} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.455, + 0.875, + 0.633 + ], + "angle": 0, + "content": "
flfsfcATAVLattriExgNr3DSr3D
✓✓44.955.6
✓✓✓✓✓60.269.2
✓✓✓✓✓59.368.4
✓✓✓✓✓✓61.871.2
✓✓✓46.157.8
✓✓✓✓55.264.6
✓✓✓✓51.962.2
✓✓✓✓✓59.768.5
✓✓✓✓✓✓62.472.0
✓✓✓✓✓✓62.071.3
✓✓✓✓✓✓✓65.273.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.654, + 0.872, + 0.668 + ], + "angle": 0, + "content": "Table 4. Comparison with various counterfactual attentions." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.679, + 0.875, + 0.762 + ], + "angle": 0, + "content": "
Acc@0.25random \nattentionuniform \nattentionreversed \nattentionshuffle \nattention
ScanRefer57.857.456.957.5
Nr3D65.264.864.565.0
Sr3D73.973.373.173.5
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.794, + 0.832, + 0.809 + ], + "angle": 0, + "content": "interactions rather than embedding more attitudes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Effects of Counterfactual Attention. We implement different strategies to generate counterfactual attention including random attention, uniform attention, reversed attention, and shuffle attention, and report their result on \\(\\mathrm{Nr3D}\\) in Tab. 4. We see all the counterfactual attention strategies achieve a similar performer gain in three datasets," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17259" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.135, + 0.104, + 0.145 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.09, + 0.238, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.091, + 0.377, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.091, + 0.517, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.091, + 0.654, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.091, + 0.883, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.245, + 0.105, + 0.256 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.193, + 0.239, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.3, + 0.227, + 0.335 + ], + "angle": 0, + "content": "wooden double bookcase filled with books. walking into the room, it is in the right hand most." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.193, + 0.377, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.3, + 0.366, + 0.335 + ], + "angle": 0, + "content": "the trash can is in the corner next to the door, below the whiteboard. it is gray, tall, and narrow." + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.193, + 0.517, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.299, + 0.526, + 0.335 + ], + "angle": 0, + "content": "the black chair is located directly to the right of the left wall off the room. The chair is to the left of the burgundy chair and behind the desk." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.193, + 0.655, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.299, + 0.645, + 0.335 + ], + "angle": 0, + "content": "the pillow is on the bed. it is to the left of the lamp, resulting on the pillow that is behind it." + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.193, + 0.883, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.3, + 0.866, + 0.327 + ], + "angle": 0, + "content": "this is a black leather loveseat. if you were sitting in it, the long, short bookshelf would be on the right." + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.346, + 0.882, + 0.362 + ], + "angle": 0, + "content": "Figure 4. The visualization results. (a) and (b) present the correct/wrong prediction of \\( MA^2 \\) TransVG with/without attribute interaction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.47, + 0.447 + ], + "angle": 0, + "content": "while random attention can obtain the best accuracy. We think it is because random attention helps try more diverse combinations of attribute interactives and can provide a more effective signal to supervise the attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.471, + 0.691 + ], + "angle": 0, + "content": "Effects of Exchanging Operation. We conduct a sensitivity analysis of hyper-parameters in our EMFM module, including: i) the exchange proportion \\(\\theta\\) of tokens. As shown in Table. 5, the model performance increases and then drops with the increase of \\(\\theta\\), which achieves a peak value of \\(65.2\\%\\) when \\(\\theta = 10\\%\\). The results show that too small or large \\(\\theta\\) will adversely affect the inter-modal fusion, which may bring inadequate alignment or attenuate the intramodal knowledge. ii) the start layer \\(\\mu\\) and end layer \\(\\eta\\) for multimodal exchanging. Since the default number of layers in the regular Transformer is 6, we first fix the start layer \\(\\mu\\) to 1 and investigate the value of \\(\\eta\\) from 1, 2, 3, 4, 5, where the accuracy improves quickly from \\(62.4\\%\\) to \\(65.2\\%\\). We then fix the end layer \\(\\eta\\) to 5, where the accuracy decreases when \\(\\mu\\) increases from 1 to 5. All the results verify the rationality of our hyper-parameters settings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.334, + 0.712 + ], + "angle": 0, + "content": "4.5. Visualization and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We show the visualization results in Fig. 4. The model may fail to recognize the spatial relationship among objects (the last three columns) or be significantly affected by other more salient objects (the first two columns) without considering the explicit multi-attribute interactions. Our \\( MA^2 \\) TransVG can better perceive discriminative clues such as the appearance or spatial correlation of objects for multimodal feature enhancement and accurate grounding (the first row). We also notice the accuracy still remains improvement room for 'View-dependent' descriptions compared to the 'View-independent' or 'Easy', especially in Sr3D (In Tab. 2). In the future, we will extend more at-" + }, + { + "type": "table_caption", + "bbox": [ + 0.528, + 0.386, + 0.864, + 0.401 + ], + "angle": 0, + "content": "Table 5. Hyper-parameter Sensitivity Analysis on Nr3D." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.411, + 0.84, + 0.535 + ], + "angle": 0, + "content": "
θ0%5%10%15%20%
64.464.765.263.961.4
η12345
μ = 162.463.964.865.264.2
μ12345
η = 564.263.362.962.662.4
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.558, + 0.892, + 0.59 + ], + "angle": 0, + "content": "tribute embedding strategies, e.g., better encoding absolute position or aggregating point clouds from various views." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.619, + 0.618 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.894, + 0.809 + ], + "angle": 0, + "content": "In this work, we propose an attribute-aware Transformer \\((\\mathsf{MA}^2\\mathsf{TransVG})\\) for 3D object grounding. As the first attempt to quantify the contribution of each attribute for final grounding, \\(\\mathsf{MA}^2\\mathsf{TransVG}\\) first designs a newly attribute causal effect analysis module and an attribute loss, which help model the multi-attribute interactions of objects for a better visual/textual modal understanding. Based on the learned multi-attribute interactions, the exchanging-based multimodal fusion module further dynamically aligns more discriminative details between modalities for a fine-grained multimodal feature. The proposed model outperforms the SOTA methods on \\(\\mathrm{Nr3D / Sr3D}\\) and ScanRefer datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.822, + 0.661, + 0.839 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.848, + 0.893, + 0.893 + ], + "angle": 0, + "content": "This work was supported by the National Science Fund of China (Grant Nos. 62276144, 62306238) and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17260" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Ahmed Abdelreehem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.225 + ], + "angle": 0, + "content": "[2] EslamMohamed Bakr, Yasmeen Alsaedy, and Mohamed Elhoseiny. Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. In NeurIPS, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.282 + ], + "angle": 0, + "content": "[3] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In CVPR, pages 16464-16473, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.468, + 0.338 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.468, + 0.382 + ], + "angle": 0, + "content": "[5] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221. Springer, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.468, + 0.436 + ], + "angle": 0, + "content": "[6] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. NeurIPS, 34:5834-5847, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.468, + 0.493 + ], + "angle": 0, + "content": "[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. NeurIPS, 35:20522-20535, 2022. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.468, + 0.55 + ], + "angle": 0, + "content": "[8] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, pages 16537-16547, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.468, + 0.592 + ], + "angle": 0, + "content": "[9] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In CVPR, pages 11124-11133, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.468, + 0.649 + ], + "angle": 0, + "content": "[10] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In ICCV, pages 18109-18119, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[11] Bowen Cheng, Lu Sheng, Shaoshuai Shi, Ming Yang, and Dong Xu. Back-tracing representative points for voting-based 3d object detection in point clouds. In CVPR, pages 8963-8972, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[12] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[13] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In CVPR, pages 3722-3731, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Chen Gao, Xingyu Peng, Mi Yan, He Wang, Lirong Yang, Haibing Ren, Hongsheng Li, and Si Liu. Adaptive zone-aware hierarchical planner for vision-language navigation. In CVPR, pages 14911-14920, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[15] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yi Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. In CVPR, pages 14773-14783, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[16] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In ICCV, pages 15372-15383, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.209, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[17] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-and-relation aware transformer for fine-grained 3d visual grounding. In ACM MM, pages 2344-2352, 2021. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.266, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[18] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In AAAI, pages 1610-1618, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.324, + 0.892, + 0.364 + ], + "angle": 0, + "content": "[19] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In CVPR, pages 15524-15533, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.368, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[20] Kenton Lee Jacob Devlin, Ming-Wei Chang and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.425, + 0.892, + 0.479 + ], + "angle": 0, + "content": "[21] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Looking outside the box to ground language in 3d scenes. arXiv preprint arXiv:2112.08879, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[22] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, pages 417-433. Springer, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.541, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[23] Nathan Kallus and Angela Zhou. Confounding-robust policy improvement. NeurIPS, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.571, + 0.892, + 0.625 + ], + "angle": 0, + "content": "[24] Murat Kocaoglu, Christopher Snyder, Alexandros G Dimakis, and Sriram Vishwanath. Causalgan: Learning causal implicit generative models with adversarial training. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.628, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[25] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.686, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[26] Haojia Lin, Yongdong Luo, Xiawu Zheng, Lijiang Li, Fei Chao, Taisong Jin, Donghao Luo, Chengjie Wang, Yan Wang, and Liujuan Cao. A unified framework for 3d point cloud visual grounding. arXiv preprint arXiv:2308.11887, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.758, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[27] Haolin Liu, Anran Lin, Xiaoguang Han, Lei Yang, Yizhou Yu, and Shuguang Cui. Refer-it-in-rgbd: A bottom-up approach for 3d visual grounding in rgb images. In CVPR, pages 6032-6041, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[28] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In ICCV, pages 10968-10980, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, pages 2949-2958, 2021. 1, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17261" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[30] David Lopez-Paz, Robert Nishihara, Soumith Chintala, Bernhard Scholkopf, and Léon Bottou. Discovering causal signals in images. In CVPR, pages 6979-6987, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.468, + 0.189 + ], + "angle": 0, + "content": "[31] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16454-16463, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.468, + 0.218 + ], + "angle": 0, + "content": "[32] Judea Pearl. Interpretation and identification of causal mediation. Psychological methods, 19(4):459, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.468, + 0.258 + ], + "angle": 0, + "content": "[33] Judea Pearl. Direct and indirect effects. In Probabilistic and causal inference: the works of Judea Pearl, pages 373-392. 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.468, + 0.314 + ], + "angle": 0, + "content": "[34] Stephen Powell. The book of why: The new science of cause and effect. pearl, judea, and dana mackenzie. 2018. hachette uk. Journal of MultiDisciplinary Evaluation, 14(31):47-54, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.468, + 0.383 + ], + "angle": 0, + "content": "[35] Mihir Prabhudesai, Hsiao-Yu Fish Tung, Syed Ashar Javed, Maximilian Sieb, Adam W Harley, and Katerina Fragkiadaki. Embodied language grounding with 3d visual feature representations. In CVPR, pages 2220-2229, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.468, + 0.427 + ], + "angle": 0, + "content": "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.468, + 0.468 + ], + "angle": 0, + "content": "[37] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.468, + 0.523 + ], + "angle": 0, + "content": "[38] Yongming Rao, Guangyi Chen, Jiwen Lu, and Jie Zhou. Counterfactual attention learning for fine-grained visual categorization and re-identification. In ICCV, pages 1025–1034, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.468, + 0.579 + ], + "angle": 0, + "content": "[39] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In Conference on Robot Learning, pages 1046-1056. PMLR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.468, + 0.622 + ], + "angle": 0, + "content": "[40] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In CVPR, pages 3716-3725, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.468, + 0.677 + ], + "angle": 0, + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurlPS, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.468, + 0.745 + ], + "angle": 0, + "content": "[42] Hongjun Wang, Jiyuan Chen, Lun Du, Qiang Fu, Shi Han, and Xuan Song. Causal-based supervision of attention in graph neural network: A better and simpler choice towards powerful attention. arXiv preprint arXiv:2305.13115, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.468, + 0.802 + ], + "angle": 0, + "content": "[43] Hanqing Wang, Wei Liang, Luc Van Gool, and Wenguan Wang. Dreamwalker: Mental planning for continuous vision-language navigation. In ICCV, pages 10873-10883, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[44] Tan Wang, Jianqiang Huang, Hanwang Zhang, and Qianru Sun. Visual commonsense r-cnn. In CVPR, pages 10760-10770, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[45] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In CVPR, pages 19231-19242, 2023. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[46] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[47] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In CVPR, pages 1791-1800, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[48] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Shuguang Cui, and Zhen Li. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In CVPR, pages 8563-8573, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[49] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, pages 15225-15236, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[50] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2928-2937, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.347, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[51] Tong Zhao, Gang Liu, Daheng Wang, Wenhao Yu, and Meng Jiang. Learning from counterfactual links for link prediction. In Int. Conf. on Machine Learning, pages 26911-26926. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[52] Dave Zhenyu Chen, Qirui Wu, Matthias Niener, and Angel X Chang. D3net: A speaker-listener architecture for semi-supervised dense captionin and visual grounding in rgb-d scans. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.46, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[53] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17262" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/477f7931-e1b2-442e-a45b-11cf7d648a49_origin.pdf b/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/477f7931-e1b2-442e-a45b-11cf7d648a49_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b9755d82eb73359c5805b1ab00415258aa20d123 --- /dev/null +++ b/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/477f7931-e1b2-442e-a45b-11cf7d648a49_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4965623b73799c24b46c4c582b9a9dce4a1564e06251c535b7c57a2e6ee7568d +size 2441206 diff --git a/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/full.md b/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5d17e367ab265b8d982ce1bcb279390010c20c7e --- /dev/null +++ b/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/full.md @@ -0,0 +1,324 @@ +# Multi-Attribute Interactions Matter for 3D Visual Grounding + +Can Xu $^{1}$ , Yuehui Han $^{1}$ , Rui Xu $^{1}$ , Le Hui $^{2}$ , Jin Xie $^{3,4*}$ , Jian Yang $^{1}$ $^{1}$ PCA Lab, Nanjing University of Science and Technology, Nanjing, China + $^{2}$ School of Electronics and Information, Northwestern Polytechnical University, Xi'an, China + $^{3}$ State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China + $^{4}$ School of Intelligence Science and Technology, Nanjing University, Suzhou, China +volcanoxl23@gmail.com; {hanyh,xu-ray,csjxie,csjyang}@njust.edu.cn; huile@nwpu.edu.cn + +# Abstract + +3D visual grounding aims to localize 3D objects described by free-form language sentences. Following the detection-then-matching paradigm, existing methods mainly focus on embedding object attributes in unimodal feature extraction and multimodal feature fusion, to enhance the discriminability of the proposal feature for accurate grounding. However, most of them ignore the explicit interaction of multiple attributes, causing a bias in unimodal representation and misalignment in multimodal fusion. In this paper, we propose a multi-attribute aware Transformer for 3D visual grounding, learning the multi-attribute interactions to refine the intra-modal and inter-modal grounding cues. Specifically, we first develop an attribute causal analysis module to quantify the causal effect of different attributes for the final prediction, which provides powerful supervision to correct the misleading attributes and adaptively capture other discriminative features. Then, we design an exchanging-based multimodal fusion module, which dynamically replaces tokens with low attribute attention between modalities before directly integrating low-dimensional global features. This ensures an attribute-level multimodal information fusion and helps align the language and vision details more efficiently for fine-grained multimodal features. Extensive experiments show that our method can achieve state-of-the-art performance on ScanRefer and Sr3D/Nr3D datasets. The code is publicly available at https://github.com/volcanoXC/MA2TransVG. + +# 1. Introduction + +3D Visual Grounding (VG) aims to locate the most relevant object in a given point cloud scene based on a language description. As a cornerstone of wide applications [6, 8, 14, 28, 43], such as vision-language navigation and au + +![](images/3af3f09543052496dd2e1c1bbc972bc409a2e8170c36d55b9b35066ea9f95be5.jpg) +Figure 1. Different from prior works, we capture the multi-attribute interactions to guide the unimodal representation and multimodal fusion for more accurate grounding. + +tonomous robots, 3D VG has attracted increasing attention from both academia and industry. However, limited by the complexity of free-form language description and the irregularity of sparse point clouds, achieving accurate 3D visual grounding remains an open issue. + +Existing 3D VG methods mainly follow a detection-then-matching pipeline. Specifically, a pool of proposals is first produced by the pre-trained general 3D object detectors [11, 29, 37]. Then, the linguistic description is associated with the visual feature of each proposal to generate multimodal object representation, which will be used to predict their referring confidence score for grounding. Since such a two-stage paradigm aims to collect enough semantic cues to distinguish the target object, the + +quality of unimodal feature representation and multimodal feature fusion always play a critical role, which decides the discriminativeness of each proposal and their matching difficulty with the target label. Most methods focus on enhancing the feature from various perspectives, such as sparse convolution [47], 2D image assistance [5, 27, 46, 52], coarse proposal refinement [13, 22, 50], or segmentation/captioning collaboration [3, 10, 18, 26]. As shown in Fig. 1(a), recent works [7, 18, 19, 35] enhance the feature by embedding object attributes, such as color, name and language-conditioned correlation (i.e., distance, view, direction), showing a promising future in advancing grounding performance. + +Though prior works made significant progress, we argue few of them explore the multi-attribute interactions in either unimodal representation or multimodal fusion, which affects the grounding accuracy, especially for cases with multiple objects. As the CASE in Fig. 1, the model embeds the most salient attribute (object name of 'door') and learns other potential attributes (e.g., 'brown', 'wooden', 'left') depending on the self-attention mechanism. Since the loss supervises the final prediction rather than the intermediate attention, the model may be encouraged to focus on some main attributes instead of all attributes, which weakens the discrimination of intra-modal representation. Meanwhile, when fusing the low-dimensional global feature between modalities via cross-attention, attention to 'brown' and 'wooden' should be encouraged for more discriminative language expression of the object, while they may be suppressed in the proposal representation to maximize the proposal's distinctiveness. Unfortunately, this would bring a misalignment in multimodal features, causing the model to select the auxiliary object (right door) rather than the primary object (left door). + +To alleviate the bias in unimodal representation and misalignment in multimodal fusion, we propose a multi-attribute aware Transformer $(\mathsf{MA}^2\mathsf{TransVG})$ for 3D visual grounding, which enhances the unimodal representation with intrinsic attribute interaction and refines the cross-modal information fusion to the attribute level (See Fig. 1(b)). Concretely, we develop an attribute causal analysis module (ACAM), which quantifies the effect of each attribute (i.e., learned attentions) and counterfactual interference (i.e., wrong attention) for the final prediction (i.e., grounding score). By maximizing the causal effect, powerful supervision is provided to correct the biased attribute and adaptively assign more proper attention to all attributes. Guided by the learned interaction, we further propose an exchanging-based multimodal fusion module (EMFM) to replace tokens with low attribute attention between modalities, which helps align more vision and language details and generate fine-grained multimodal features for accurate grounding. Our contributions include: + +- We propose a multi-attribute aware Transformer for 3D object grounding, which explicitly models the multi-attribute interactions to prevent imbalanced and misaligned visual-textual representation. +- We propose the ACAM module to quantify the causal effect of multiple intrinsic attributes with the final prediction, which helps adaptively assign more proper attention to all attributes for a better intra-modal understanding. +- We propose the EMFM module to perform multimodal fusion based on multi-attribute exchanging, which helps align more attribute-level details between modalities for fine-grained multimodal features. +- We evaluate the proposed method on $\mathrm{Nr3D/Sr3D}$ and ScanRefer benchmarks, which outperforms all state-of-the-art methods by a large margin. + +# 2. Related work + +3D Visual Grounding. The interest in 3D VG has grown rapidly in recent years, and existing approaches include one-stage [21, 31] and two-stage methods. The two-stage methods [13, 16-18, 46, 47, 50] decouple the grounding into language-irrelevant object detection and cross-modal matching. Recent works focus on improving the intramodal representation and modeling the inter-modal relationship. For example, Graph-based approaches [1, 18, 47] infer spatial relations among proposals by connecting each proposal with its top-N nearest neighbors. Transformer-based approaches [7, 17, 19, 21, 31, 46] leverage the attention mechanism to concentrate the object attributes on proposals that are more essential to the referring. In these methods, only the most salient attribute is embedded and the fusion between modalities occurs in the global feature space, resulting in a biased intra-modal representation and inadequate inter-modal fusion. Instead, we capture the multi-attribute interactions and refine the cross-modal learning at the attribute level. + +Causal Inference in Vision. Causality analysis [32] has been successfully used in several areas, such as adversarial learning [24], reinforcement learning [23] and graph neural networks [42, 51]. In vision tasks [30, 40, 44], causality inference also works as an effect tool to discover hidden causal structures and confront data biases. For example, Rao et al. [38] learn more effective attention based on causal inference for fine-grained image categorization, person and vehicle re-identification. As the first attempt, we introduce the causality analysis into 3D VG to quantify the effect of each attribute, which helps model the complex interactions of various attributes to enhance the unimodal feature. + +Transformers in Vision-and-Language. Transformer shows a powerful ability to learn multimodal features in various tasks, such as image captioning [9, 10, 48], vision-and-language navigation [28, 43], and vision question-answering [15, 53]. For 3D VG, different from most + +![](images/2bf114a7b9284038875042191adacdd60fe419be31f3656ad11482cfe19d5db0.jpg) +Figure 2. The framework of proposed $\mathsf{MA}^2\mathsf{TransVG}$ . The embedding of pre-detected proposals, attributes, and language are extracted and then cross-attend in ACAM. By applying the causal effect analysis via $L_{attr}$ , ACAM models the multi-attribute interactions for final prediction to generate the Attribute-Vision/Text attention, which can enhance the proposal and text embedding. The attention further guides the attribute-level feature exchanging in EMFM, which helps fuse the multimodal feature for accurate grounding. + +methods that directly fuse the global unimodal features by stacking cross-attention layers [7, 17, 19, 21, 31, 46], we refine the cross-modal fusion to the attribute level, which helps suppress the misalignment between modalities to capture a finer-grained matching relationship. + +# 3. Method + +# 3.1. Overview + +Fig. 2 shows an overview of proposed $\mathsf{MA}^2\mathsf{TransVG}$ . Given a 3D point cloud with corresponding description, text features $T \in \mathbb{R}^{(m + 1)\times d}$ , proposal features $O \in \mathbb{R}^{n\times d}$ and attribute features $F \in \mathbb{R}^{n\times d}$ are extracted (Sec. 3.2). The three streams then are fed into the attribute causal analysis module (ACAM) to capture the multi-attribute interactions (i.e., Attribute-Vision attention and Attribute-Text attention) to enhance initial text and proposal feature. The attention is supervised by an attribute loss $L_{attr}$ , which is designed by estimating the causal effect of attributes for the final prediction (Sec. 3.3). The enhanced proposal embedding $O^*$ and text embedding $T^*$ are further passed to the exchanging-based multimodal fusion module (EMFM), where their tokens with small attribute attention are replaced with the average of all the tokens in the other modality (Sec. 3.4). With the fused multimodal embedding $E \in \mathbb{R}^{n\times d}$ , we predict the final grounding results. + +# 3.2. Input Modal Representation + +Text Encoding. Given a sentence $S$ with $m$ word tokens, we use a pre-trained BERT model [20] to produce the text features $(t_{cls}, t_1, \ldots, t_m)$ . $t_i \in \mathbb{R}^d$ is the feature of each token with a dimensionality of $d$ and $t_{cls} \in \mathbb{R}^d$ is a special + +token for text classification. + +Object Encoding. Given a point cloud scene $\pmb{P} \in \mathbb{R}^{K \times 6}$ of $K$ points described by XYZ coordinates and RGB colors, we use the GroupFree [29] detector to obtain an object proposal list $\pmb{o} = (o_1, o_2, \dots, o_n)$ containing $n$ proposals and extract corresponding proposal feature $\pmb{O} \in \mathbb{R}^{n \times d}$ based on PointNet++ [36]. Since we mainly focus on the object referred in the text, each proposal feature $O_i \in \mathbb{R}^d$ will be added with the $t_{cls}$ before feeding into ACAM for attribute-vision causal effect analysis. + +Attributes Encoding. Motivated by [3], for the proposal pool $o$ with $n$ objects, the multi-view RGB features $\pmb{f}_{c} \in \mathbb{R}^{n \times 128}$ that potentially contain the color and materials attribute is first extracted. Then, we use the 3-dimensional box center and 24-dimensional corner coordinates to describe the object size attribute $\pmb{f}_{s} \in \mathbb{R}^{n \times 27}$ . For the location attribute, we further consider the distance and angle of each pair of objects $(o_{i}, o_{j})$ . Connecting the object center point $(x_{c_{i}}, y_{c_{i}}, z_{c_{i}})$ and $(x_{c_{j}}, y_{c_{j}}, z_{c_{j}})$ , the distance is their Euclidean distance $d_{ij}$ . Similar to [7], the horizontal angle $\theta_{h}$ and vertical angle $\theta_{v}$ can be calculated as: + +$$ +\begin{array}{c} \theta_ {h} = \arctan 2 \left(\left(y _ {c _ {i}} - y _ {c _ {j}}\right) / \left(x _ {c _ {i}} - x _ {c _ {j}}\right) \right. \\ \theta_ {v} = \arcsin \left(\left(z _ {c _ {i}} - z _ {c _ {j}}\right) / d _ {i j}\right) \end{array} \tag {1} +$$ + +The location attribute $f_{l}^{ij} \in \mathbb{R}^{5}$ between any pair of objects is defined, as: + +$$ +f _ {l} ^ {i j} = \left[ d _ {i j}, \sin \left(\theta_ {h}\right), \cos \left(\theta_ {h}\right), \sin \left(\theta_ {v}\right), \cos \left(\theta_ {v}\right) \right] \tag {2} +$$ + +For each proposal $o_i$ , the location attribute $\pmb{f}_l^i \in \mathbb{R}^5$ can be calculated by averaging all $f_l^{ij}$ connected with $o_i$ , and the final attribute embedding can be generated by concatenating + +all three attributes, as $\pmb{f}^{i} = [\pmb{f}_{l}^{i},\pmb{f}_{c}^{i},\pmb{f}_{s}^{i}]$ . We map $\pmb{f}^{i}\in \mathbb{R}^{d}$ to the same dimension of text feature using a FC layer. + +# 3.3. Attribute Causal Analysis Module + +The attribute causal analysis module (ACAM) models the multi-attribute interactions with vision/text as Attribute-Vision/Text attention based on a well-designed Attribute-aware Transformer. To capture more influential and diverse interactions, ACAM further guides the learning process with an extra loss $L_{attr}$ , which measures the causal effect of learned attribute-level attentions for the final prediction. + +Attributes-Text Attention Learning. Given attribute embedding $\pmb{F} \in \mathbb{R}^{n \times d}$ , Attributes-Text attention $\pmb{A}_{T}$ helps select related attribute features from the initial text feature $T \in \mathbb{R}^{(m + 1) \times d}$ , thereby the text representation can be enhanced to better describe the target object. We define $\pmb{A}_{T} \in \mathbb{R}^{n \times (m + 1)}$ as the cross-attention between the Attribute embedding $\pmb{F}$ and text embedding $\pmb{T}$ , which can be computed with a standard Transformer [41]. + +In order to enlarge the model capacity for diverse Attribute-Text correlations, we also introduce multi parallel attention layers $\ell$ as different transformation heads, and the multi-head attentions will be fused with the standard self-attention [41]. As shown in Fig. 2, $A_{T}$ then guides to generate the enhanced text feature $T^{*} \in \mathbb{R}^{n \times d}$ , which will be used to train the language classifier with the supervision of the language classification loss $L_{sent}$ , and also serves as input for subsequent EMFM for multi-modal fusion. + +Attributes-Vision Attention Learning. Instead of directly calculating the cross-attention between proposal features $\mathcal{O}$ and attribute features $\mathbf{F}$ , we first guide the proposal features with language token $t_{cls}$ , as we prefer to care about specific categories of objects mentioned in the text [7]. Specifically, for the text-guided proposal feature, we adopt the similar Transformer structure in Attributes-Text attention learning to calculate its self-attention $A_V^s \in \mathbb{R}^{n \times n}$ , and cross-attention $A_V^c \in \mathbb{R}^{n \times n}$ with $\mathbf{F}$ . + +Similar to [7], we then integrate the self-attention of proposals $A_V^s$ with the cross attribute-proposal attention $A_V^c$ using a sigmoid softmax (sigsoftmax) fusion operation to calculate the Attributes-Vision attention $A_V$ , as: + +$$ +A _ {v} = \frac {\sigma \left(\boldsymbol {A} _ {\boldsymbol {V}} ^ {\boldsymbol {c}} (i , j)\right) \exp \left(\boldsymbol {A} _ {\boldsymbol {V}} ^ {\boldsymbol {s}} (i , j)\right)}{\sum_ {j = 1} ^ {n} \sigma \left(\boldsymbol {A} _ {\boldsymbol {V}} ^ {\boldsymbol {c}} (i , j)\right) \exp \left(\boldsymbol {A} _ {\boldsymbol {V}} ^ {\boldsymbol {s}} (i , j)\right)} \tag {3} +$$ + +where $\sigma(\cdot)$ is the sigmoid function. $A_V^s (i,j)$ and $A_V^c (i,j)$ are the element in $i$ -th row and $j$ -th column in attention matrix $A_V^s$ and $A_V^c$ , respectively. The same multi-head attention strategy is used in Attribute-Vision attention $A_V$ . In a similar manner, the enhanced proposal embedding $O^{*}$ can be obtained by performing a matrix multiplication operation between $O$ and $A_V$ , which will serve as the input of proposal classifier and EMFM module. + +![](images/83d7e39ca8b8ee336fbb6f818734b170655925651d94eb642d55a8eac4a965ab.jpg) +Figure 3. Visualization of deriving causal effects through counterfactual causality. We subtract the counterfactual prediction from the original prediction to analyze the effects of learned attention and maximize them in the training process. + +Strengthen Interaction With $L_{attri}$ . Though multiple attributes have been embedded to guide the language and vision representations via Attribute-Text/Vision attention, how these attributes interact with each other to affect predictions remains a black box. To address this issue, our core idea lies in measuring the causal effect of $A_V$ and $A_T$ for the final prediction. We design the attribute loss $L_{attri}$ based on the counterfactual analysis [33], providing extra supervision to encourage the network to explore more influential interactions. + +As shown in Fig. 3, taking the Attribute-vision attention as an example, we conduct counterfactual intervention by imagining non-existent attention $\bar{A}_V$ to replace the initial $A_V$ . We keep the vision representation $O$ unchanged, which will be multiplied with $\bar{A}_V$ to calculate the counterfactual representation $\bar{O}^*$ . $\bar{O}^*$ are then fused with $T$ to obtain the counterfactual prediction $\bar{Y} \in \mathbb{R}^{n \times 1}$ , as: + +$$ +\bar {\boldsymbol {Y}} = d o \left(\boldsymbol {A} _ {V} = \bar {\boldsymbol {A}} _ {V}, \boldsymbol {T}\right) = G \left(\bar {\boldsymbol {O}} ^ {*}, \boldsymbol {T}\right) \tag {4} +$$ + +where $G(\cdot)$ is the grounding operation. In practice, the exact form of how counterfactual is achieved is not limited, and our starting point here is just to set the ball rolling. In our case, random attention, uniform attention, shuffle attention or reversed attention are used as the counterfactuals. + +Following [33, 34, 40], we define the causal effect $\mathbf{Y}_e$ of $\mathbf{A}_V$ as the difference between the observed prediction $\mathbf{Y}$ and the counterfactual prediction $\bar{\mathbf{Y}}$ , as: + +$$ +\boldsymbol {Y} _ {e} = \mathbb {E} _ {\bar {A} _ {V} \sim \gamma} (\boldsymbol {Y} - \bar {\boldsymbol {Y}}) \tag {5} +$$ + +where $\gamma$ means the distribution of counterfactual attention. + +Since the causal effect can be interpreted as how the attention improves the final prediction compared to wrong attentions, we use $\mathbf{Y}_e$ to measure the learned attention, i.e., the multi-attribute interactions. Based on this observation, we further design an attribute loss $L_{attr}$ , as: + +$$ +L _ {\text {a t t r i}} = L _ {c e} \left(\boldsymbol {Y} _ {\boldsymbol {e}}, \boldsymbol {y}\right) \tag {6} +$$ + +where $L_{ce}$ is the cross-entropy loss and $\pmb{y}$ is the GT label. Similarly, we can obtain the causal effect of Attribute-text attention and compute its attribute loss. By optimizing $L_{attr}$ from $A_{T}$ and $A_{V}$ , the model is expected to: + +(1) improve the prediction based on correct attentions as much as possible, which helps discover more discriminative attributes; (2) penalize the prediction based on wrong attentions, which forces the model adaptively capture more attributes to avoid biased unimodal embeddings. + +# 3.4. Exchanging-based Multimodal Fusion Module + +We propose the exchanging-based multimodal fusion module (EMFM) to communicate attribute-level details of objects between both input vectors for accurate grounding. Specifically, since text/visual encoders tend to focus on different attributes for better unimodal embeddings, directly fusing unimodal embeddings using cross-attention will hurt the distinguishability of multimodal features. To narrow the semantic misalignment between modalities, we select and exchange tokens with low attention score using the average of all tokens from another modal, which helps align attribute-level information between modalities sufficiently and eliminate semantic ambiguity in multimodal representations of proposals. + +Exchanging With Attributes Interactions. As shown in Fig. 3, we implement the EMFM by two Transformer encoders with shared parameters, where the shallow $\mu$ layers are set as regular Transformer encoder layers, followed by $\eta$ exchanging layers. Since $A_V$ and $A_T$ have modeled the multi-attribute interactions, they are taken as references to select tokens for exchanging. Meanwhile, $A_V$ and $A_T$ will be updated using the self-attention learned from unimodal embedding $T^*$ and $O^*$ . Considering the interdependence between the attribute attention and unimodal embedding, we concatenate them for the subsequent calculation, as: + +$$ +\begin{array}{l} \boldsymbol {T} ^ {e} (0) = \operatorname {C o n c a t} \left[ \boldsymbol {A} _ {T}, \boldsymbol {T} ^ {*} \right] \\ \boldsymbol {Q} ^ {e} (0) = \operatorname {C o n c a t} \left[ \boldsymbol {A} _ {T}, \boldsymbol {Q} ^ {*} \right] \end{array} \tag {7} +$$ + +$$ +\boldsymbol {O} ^ {e} (0) = \operatorname {C o n c a t} \left[ \boldsymbol {A} _ {V}, \boldsymbol {O} ^ {*} \right] +$$ + +where $T^{e}(0)$ and $O^{e}(0)$ are embedding for textual and visual modalities at layer 0, respectively. + +At layer $\mu + 1$ with intermediate embedding $T^{e}(\mu + 1)$ , inspired by [4, 25], we select tokens with a $\theta$ -proportion of the smallest attention scores to $A_{T}$ and replace their embedding vectors with the average embedding of all the tokens in $O^{e}(\mu + 1)$ . For instance, the selected $k$ -th token of $T^{e}(\mu + 1)$ can be updated by: + +$$ +\boldsymbol {T} ^ {\boldsymbol {e}} (\mu + 1) [ k,: ] = \frac {1}{n} \sum_ {j = 1} ^ {n} \boldsymbol {O} ^ {\boldsymbol {e}} (\mu + 1) [ j,: ] + \boldsymbol {T} ^ {\boldsymbol {e}} (\mu + 1) [ k,: ] \tag {8} +$$ + +The updated embeddings are used to calculate the multihead self-attention [41], which will guide the exchanging operation at the next layer, e.g., $A_T$ of the $(\mu + 1)$ -th layer are updated by the self-attention of the previous layer $\mu$ . + +Multimodal Fusion. We also apply the residual connection between adjacent exchanging layers to reduce the information loss caused by replacement. The exchanging process + +continues until reaching the $\eta$ layer to generate $O^e (\eta)$ and $T^{e}(\eta)$ , which will be fed into the feed-forward network (FFN) with a normalization layer and a fully connected layer to derive the final fusion embedding $\pmb {E}\in \mathbb{R}^{n\times n}$ . + +Grounding Head. We use a two-layer feed-forward neural network as the object grounding head. Given the multimodal representation $\pmb{E}$ , the grounding head will predict the probability $\pmb{p}_i \in \mathbb{R}^{n \times 1}$ for each object proposal, where the proposal with the maximum probability is selected as the target in the inference stage. + +# 3.5. Training Objectives + +Following the previous works [1, 5, 17], the loss of $\mathbf{MA}^2\mathrm{TransVG}$ consists of the primary 3D object grounding loss $L_{vg}$ , language classification loss $L_{sent}$ , proposal classification loss $L_{obj}$ , and two attribute losses $L_{attri}$ for vision/text attention causal analysis, as: + +$$ +L = L _ {v g} + L _ {s e n t} + L _ {o b j} + \lambda L _ {a t t r i} \tag {9} +$$ + +where $\lambda$ is the hyper-parameter to balance the loss. + +# 4. Experiment + +# 4.1. Datasets and Evaluation Metrics + +We evaluate the proposed $\mathsf{MA}^2\mathsf{TransVG}$ on ScanRefer [5] and Sr3D/Nr3D [1] datasets. ScanRefer offers 51,583 descriptions of 11,046 objects from 800 ScanNet [12] scenes. A target object is labeled as 'unique' if it is the only object of its class in the scene; otherwise, it is labeled as 'multiple'. For ScanRefer, we use Acc@0.25 and Acc@0.5 as evaluation metrics, which measure the fraction of language queries whose predicted box overlaps the ground truth box with 3D IoU higher than 0.25/0.5. Compared to ScanRefer where objects need to be detected and then matched, Sr3D/Nr3D datasets provide object masks for each scene and only need to recognize the classes of the proposals to choose the target object. Nr3D consists of 41,503 descriptions from 707 scenes with human annotation, while Sr3D contains 83,572 simple machine-generated descriptions. Each scene in Sr3D/Nr3D can also be divided into 'easy' and 'hard' depending on whether there are more than two instances, and can be divided into 'view-dependent' and 'view-independent' according to whether the referring expression is dependent on the camera view. Following [1], the accuracy is used to verify the model. + +# 4.2. Implementation Details + +The text encoding module is initialized from the first three layers of BERT [20], and the object encoding module PointNet++ [36] samples 1024 points for all the objects. We generate an initial set of $n = 256$ proposals using [29]. We set the dimension of unimodal representation $d = 768$ and use a three-layer transformer with 12 heads for + +Table 1. Comparison of 3D VG results with the state-of-the-art on ScanRefer. + +
MethodVenueUnique (19%)Multiple (81%)Overall
0.250.50.250.50.250.5
ScanRefer [5]ECCV2067.646.232.121.340.026.1
ReferIt3D [1]ECCV2053.837.521.012.826.416.9
TGNN [18]AAAI2168.656.829.823.237.429.7
InstanceRefer [47]ICCV2177.566.831.324.840.232.9
3DVG [50]ICCV2177.258.538.428.745.934.5
FFL-3DOG [13]ICCV2178.867.935.225.741.334.0
MVT [19]CVPR2277.766.531.925.340.833.3
3D-SPS [31]CVPR2281.664.839.529.647.736.4
ViL3DRef [7]NeurIPS2281.668.640.330.747.937.7
BUTD-DETR [22]ECCV22////52.239.8
EDA [45]CVPR2385.868.649.137.654.642.3
M3DRef-CLIP [49]ICCV23/77.2/36.8/44.7
3DJCG [3]CVPR2278.861.340.130.147.636.1
UniT3D [10]ICCV2382.873.136.431.142.339.1
3DRefTR-HR [26]ICCV2386.070.949.638.355.043.2
3DRefTR-SP [26]ICCV2386.171.050.138.755.543.5
Ours/86.374.153.841.457.945.7
+ +the Attribute-Vision/Text attention learning. For a fair comparison, we keep the architecture parameters as same as previous works [7, 19, 46]. In ACAM, we use random attention to initialize the counterfactual attention. In EMFM, considering the balance of raw intra-modal knowledge and inter-modal communication, we empirically set $\theta = 10\%$ . The start layer $\mu$ and end layer $\eta$ are set to 1 and 4, respectively. The hyper-parameter of the whole loss in Eq. (9) is $\lambda = 0.5$ . All models are trained on 4 NVIDIA RTX Titan GPUs, which are optimized using the AdamW algorithm with a batch size of 128 and a learning rate of 0.0005 with cosine decay scheduling. + +# 4.3. Comparison with State-of-the-Art Methods + +ScanRefer. In Tab. 1, we compare $\mathsf{MA}^2\mathsf{TransVG}$ with existing works on the ScanRef dataset and observe the following results: i) Compared to pure 3D VG works such as BUTD-DETR [22], 3D-SPS [31], ViL3DRef [7], and EDA [45], our method achieves state-of-the-art performance of $57.9\%$ and $45.7\%$ , by a substantial margin with an overall improvement over $3.3\%$ and $3.4\%$ . ii) To facilitate the 3D grounding, previous works [5, 31, 46, 50] learn better point cloud features with the supplementary of 2D images while most recent works [3, 10, 26] learn better language-visual alignment by connecting 3D dense captioning or segmentation tasks. However, we outperform second-place 3DRefTR-SP [26] with an overall improvement over $2\%$ for both IoU = 0.25 and IoU = 0.5 setting. This superiority illustrates that mining the multi-attribute interactions help capture more efficient and discriminative multi-modal representation to distinguish objects. iii) Our $\mathsf{MA}^2\mathsf{TransVG}$ can reach a remarkable accuracy of $53.8\%$ and $41.4\%$ for 'Multiple' scenes, which verifies the effec + +tiveness of our attribute interaction for understanding finer-grained language-visual contextual dependency. + +Nr3D&Sr3D. Tab. 2 reports the grounding accuracy on Nr3D and Sr3D datasets. For a fair comparison, all the methods use the ground-truth object proposal with no ground-truth labels. In Nr3D, though the language descriptions are more complex to cause additional challenges for text understanding and cross-modal alignment, our method still outperforms the EDA [45], which is equipped with text component decoupling and dense matching between two modalities. Compared to EDA, our $\mathsf{MA}^2\mathsf{TransVG}$ not only achieves a similar decoupling function for both descriptions and point clouds via attribute representation, but also considers their correlationships for prediction during Attribute-Text/Vision causal attention learning, which contributes to the great improvements by $+13.1\%$ in overall accuracy. Compared to other recent competitor ViL3DRef [7], 3DRefTRSR [26], and M3DRef-CLIP [49], we also achieve a consistent improvement of overall accuracy in all split settings by $+0.8\%$ , $+2.6\%$ and $+15.8\%$ , respectively. In Sr3D, our method reaches an accuracy of $73.9\%$ , which greatly surpasses the best competitor ViL3DRef [7] $1.1\%$ and outperforms other state-of-the-art methods [26, 45] over $5\%$ . Similar to ViL3DRef [7] that guided the grounding with languaged-conditioned spatial relation, we not only embed the spatial attributes but also explicitly consider other attributes in the unimodal representation and further perform exchanging-based fusion. All the results show that multi-attribute interactions can help capture more semantic clues for accurate grounding. + +Table 2. Comparison of 3D VG results with the state-of-the-art on Sr3D/Nr3D. + +
MethodVenueNr3DSr3D
EasyHardView DepView IndepOverallEasyHardView DepView IndepOverall
ReferIt3D [1]ECCV2043.627.932.537.135.644.731.539.240.840.8
TGNN [18]AAAI2144.230.635.838.037.348.536.945.845.045.0
InstanceRefer [47]ICCV2146.031.834.541.938.851.140.545.448.148.0
3DVG [50]ICCV2148.534.834.843.740.854.244.944.651.751.4
FFL-3DOG [13]ICCV2148.235.037.144.741.7/////
TransRefer3D [17]ACMM2148.536.036.544.942.160.550.249.957.757.4
LanguageRefer [39]CoRL2151.036.641.745.043.958.949.349.256.356.0
SAT [46]ICCV2156.342.446.950.449.261.250.049.258.357.9
BUTD-DETR [22]ECCV22////43.3////52.1
LAR [2]NeurIPS2258.442.347.452.148.963.051.250.059.159.4
3D-SPS [31]CVPR2258.145.148.053.251.556.265.449.263.262.6
MVT [19]CVPR2261.349.154.354.355.466.958.858.464.764.5
M3DRef-CLIP [49]ICCV2355.643.442.352.949.4/////
EDA [45]CVPR23////52.1////68.1
3DRefTR-SR [26]ICCV23////52.6////68.5
ViL3DRef [7]NeurIPS2270.257.462.064.564.474.967.963.873.272.8
Ours/71.157.662.565.465.276.069.364.573.873.9
+ +# 4.4. Ablation Studies + +Effectiveness of Each Component. We quantitatively investigate the contribution of each component in our method in Tab. 3. The baseline model simply extracts and fuses the object and language feature through a standard Transformer structure, where the modal relationships only implicitly depend on the self/cross-attention. Taking results on $\mathrm{Nr3D}$ as an example, when enhancing the feature with $A_{T}$ or $A_{V}$ , we yield an accuracy of $55.2\%$ and $51.9\%$ , which can increase to $59.7\%$ when combining two attentions. This indicates that embedding the multi-attribute within each single modal is the basis of accurate cross-modal understanding. Furthermore, we achieve a gain of $2.7\%$ after introducing causal effect analysis, which helps adaptively capture more reliable multi-attribute interactions for feature enhancement. Exchanging tokens further improves the accuracy to $65.2\%$ because multimodal knowledge exchanging boosts the quality of the learned embeddings for each modality. All the results demonstrate the effectiveness of each component in our proposed $\mathrm{MA}^2\mathrm{TransVG}$ . + +Effectiveness of Multi-attribute. To verify the effectiveness of multi-attribute interactions, we conduct ablation on attribute quantity and multi-attribute interactions. In Tab. 3. We adopt commonly used attributes (location $f_{l}$ , size $f_{s}$ , color $f_{c}$ ) with simple embedding method and achieve SOTA performance. The accuracy is $44.9\%$ on Nr3D when generating attribute embedding only using $f_{l}$ and $f_{s}$ , which increases to $46.1\%$ ( $+1.2\%$ ) and $60.2\%$ ( $+15.3\%$ ) after adding $f_{c}$ and interactions, respectively. Results show that our accuracy gain mainly comes from the multi-attribute + +Table 3. Ablation studies of each component on $\mathrm{{Nr}}3\mathrm{D}$ and $\mathrm{{Sr}}3\mathrm{D}$ . + +
flfsfcATAVLattriExgNr3DSr3D
✓✓44.955.6
✓✓✓✓✓60.269.2
✓✓✓✓✓59.368.4
✓✓✓✓✓✓61.871.2
✓✓✓46.157.8
✓✓✓✓55.264.6
✓✓✓✓51.962.2
✓✓✓✓✓59.768.5
✓✓✓✓✓✓62.472.0
✓✓✓✓✓✓62.071.3
✓✓✓✓✓✓✓65.273.9
+ +Table 4. Comparison with various counterfactual attentions. + +
Acc@0.25random +attentionuniform +attentionreversed +attentionshuffle +attention
ScanRefer57.857.456.957.5
Nr3D65.264.864.565.0
Sr3D73.973.373.173.5
+ +interactions rather than embedding more attitudes. + +Effects of Counterfactual Attention. We implement different strategies to generate counterfactual attention including random attention, uniform attention, reversed attention, and shuffle attention, and report their result on $\mathrm{Nr3D}$ in Tab. 4. We see all the counterfactual attention strategies achieve a similar performer gain in three datasets, + +![](images/eb01b054e1463071777d5c89b7ad27b6f3bf07dfb43bab833cf0d11d90923c1e.jpg) +(a) + +![](images/bcd4b1b6912995469b95a18add85e50e816b24cf8883272caf1d6b9f83bf85a5.jpg) + +![](images/6e36ef93f4f35540859f5ef5d42ff26e73d643a3d3726d247340d06a65d4ed1e.jpg) + +![](images/d3325333eb7e7e9a501c53172f5c701dd892e368cf4fdd2fb3d34fe5c070b173.jpg) + +![](images/d6a85df81ee1e3e12443d290b769e49b548bd2dcd94c7529d3e05a42aea879bf.jpg) + +![](images/5b13c49d1476c8c8c0eaf086ff0b8b9ea2c41a76eed082fc1bfc8a107bf14bd0.jpg) +(b) +wooden double bookcase filled with books. walking into the room, it is in the right hand most. +Figure 4. The visualization results. (a) and (b) present the correct/wrong prediction of $MA^2$ TransVG with/without attribute interaction. + +![](images/ffc6755253bfc4a2733baf192c333089b6fe3a0fbea990fed9240db91ad7a46e.jpg) +the trash can is in the corner next to the door, below the whiteboard. it is gray, tall, and narrow. +the black chair is located directly to the right of the left wall off the room. The chair is to the left of the burgundy chair and behind the desk. + +![](images/4e098f3a4513655970d02a121b8f4839c5f3b11c27cdff957f2c4cccbddcaebf.jpg) + +![](images/bd8f67419cd5daa5d188337ad67e40f127e7f6f92865dd1ceb4ad532d8257e98.jpg) +the pillow is on the bed. it is to the left of the lamp, resulting on the pillow that is behind it. + +![](images/0df27124929018779b595c1f7617e64973d3ffeb3175dafd6ce57e00500e192f.jpg) +this is a black leather loveseat. if you were sitting in it, the long, short bookshelf would be on the right. + +while random attention can obtain the best accuracy. We think it is because random attention helps try more diverse combinations of attribute interactives and can provide a more effective signal to supervise the attention. + +Effects of Exchanging Operation. We conduct a sensitivity analysis of hyper-parameters in our EMFM module, including: i) the exchange proportion $\theta$ of tokens. As shown in Table. 5, the model performance increases and then drops with the increase of $\theta$ , which achieves a peak value of $65.2\%$ when $\theta = 10\%$ . The results show that too small or large $\theta$ will adversely affect the inter-modal fusion, which may bring inadequate alignment or attenuate the intramodal knowledge. ii) the start layer $\mu$ and end layer $\eta$ for multimodal exchanging. Since the default number of layers in the regular Transformer is 6, we first fix the start layer $\mu$ to 1 and investigate the value of $\eta$ from 1, 2, 3, 4, 5, where the accuracy improves quickly from $62.4\%$ to $65.2\%$ . We then fix the end layer $\eta$ to 5, where the accuracy decreases when $\mu$ increases from 1 to 5. All the results verify the rationality of our hyper-parameters settings. + +# 4.5. Visualization and Limitation + +We show the visualization results in Fig. 4. The model may fail to recognize the spatial relationship among objects (the last three columns) or be significantly affected by other more salient objects (the first two columns) without considering the explicit multi-attribute interactions. Our $MA^2$ TransVG can better perceive discriminative clues such as the appearance or spatial correlation of objects for multimodal feature enhancement and accurate grounding (the first row). We also notice the accuracy still remains improvement room for 'View-dependent' descriptions compared to the 'View-independent' or 'Easy', especially in Sr3D (In Tab. 2). In the future, we will extend more at- + +Table 5. Hyper-parameter Sensitivity Analysis on Nr3D. + +
θ0%5%10%15%20%
64.464.765.263.961.4
η12345
μ = 162.463.964.865.264.2
μ12345
η = 564.263.362.962.662.4
+ +tribute embedding strategies, e.g., better encoding absolute position or aggregating point clouds from various views. + +# 5. Conclusion + +In this work, we propose an attribute-aware Transformer $(\mathsf{MA}^2\mathsf{TransVG})$ for 3D object grounding. As the first attempt to quantify the contribution of each attribute for final grounding, $\mathsf{MA}^2\mathsf{TransVG}$ first designs a newly attribute causal effect analysis module and an attribute loss, which help model the multi-attribute interactions of objects for a better visual/textual modal understanding. Based on the learned multi-attribute interactions, the exchanging-based multimodal fusion module further dynamically aligns more discriminative details between modalities for a fine-grained multimodal feature. The proposed model outperforms the SOTA methods on $\mathrm{Nr3D / Sr3D}$ and ScanRefer datasets. + +# Acknowledgement + +This work was supported by the National Science Fund of China (Grant Nos. 62276144, 62306238) and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Panos Achlioptas, Ahmed Abdelreehem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 2, 5, 6, 7 +[2] EslamMohamed Bakr, Yasmeen Alsaedy, and Mohamed Elhoseiny. Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. In NeurIPS, 2022. 7 +[3] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In CVPR, pages 16464-16473, 2022. 2, 3, 6 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 5 +[5] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221. Springer, 2020. 2, 5, 6 +[6] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. NeurIPS, 34:5834-5847, 2021. 1 +[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. NeurIPS, 35:20522-20535, 2022. 2, 3, 4, 6, 7 +[8] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, pages 16537-16547, 2022. 1 +[9] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In CVPR, pages 11124-11133, 2023. 2 +[10] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In ICCV, pages 18109-18119, 2023. 2, 6 +[11] Bowen Cheng, Lu Sheng, Shaoshuai Shi, Ming Yang, and Dong Xu. Back-tracing representative points for voting-based 3d object detection in point clouds. In CVPR, pages 8963-8972, 2021. 1 +[12] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5 +[13] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In CVPR, pages 3722-3731, 2021. 2, 6, 7 +[14] Chen Gao, Xingyu Peng, Mi Yan, He Wang, Lirong Yang, Haibing Ren, Hongsheng Li, and Si Liu. Adaptive zone-aware hierarchical planner for vision-language navigation. In CVPR, pages 14911-14920, 2023. 1 + +[15] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yi Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. In CVPR, pages 14773-14783, 2023. 2 +[16] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In ICCV, pages 15372-15383, 2023. 2 +[17] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-and-relation aware transformer for fine-grained 3d visual grounding. In ACM MM, pages 2344-2352, 2021. 2, 3, 5, 7 +[18] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In AAAI, pages 1610-1618, 2021. 2, 6, 7 +[19] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In CVPR, pages 15524-15533, 2022. 2, 3, 6, 7 +[20] Kenton Lee Jacob Devlin, Ming-Wei Chang and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2, 2019. 3, 5 +[21] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Looking outside the box to ground language in 3d scenes. arXiv preprint arXiv:2112.08879, 2021. 2, 3 +[22] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, pages 417-433. Springer, 2022. 2, 6, 7 +[23] Nathan Kallus and Angela Zhou. Confounding-robust policy improvement. NeurIPS, 31, 2018. 2 +[24] Murat Kocaoglu, Christopher Snyder, Alexandros G Dimakis, and Sriram Vishwanath. Causalgan: Learning causal implicit generative models with adversarial training. In ICLR, 2018. 2 +[25] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022. 5 +[26] Haojia Lin, Yongdong Luo, Xiawu Zheng, Lijiang Li, Fei Chao, Taisong Jin, Donghao Luo, Chengjie Wang, Yan Wang, and Liujuan Cao. A unified framework for 3d point cloud visual grounding. arXiv preprint arXiv:2308.11887, 2023. 2, 6, 7 +[27] Haolin Liu, Anran Lin, Xiaoguang Han, Lei Yang, Yizhou Yu, and Shuguang Cui. Refer-it-in-rgbd: A bottom-up approach for 3d visual grounding in rgb images. In CVPR, pages 6032-6041, 2021. 2 +[28] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In ICCV, pages 10968-10980, 2023. 1, 2 +[29] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, pages 2949-2958, 2021. 1, 3, 5 + +[30] David Lopez-Paz, Robert Nishihara, Soumith Chintala, Bernhard Scholkopf, and Léon Bottou. Discovering causal signals in images. In CVPR, pages 6979-6987, 2017. 2 +[31] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16454-16463, 2022. 2, 3, 6, 7 +[32] Judea Pearl. Interpretation and identification of causal mediation. Psychological methods, 19(4):459, 2014. 2 +[33] Judea Pearl. Direct and indirect effects. In Probabilistic and causal inference: the works of Judea Pearl, pages 373-392. 2022. 4 +[34] Stephen Powell. The book of why: The new science of cause and effect. pearl, judea, and dana mackenzie. 2018. hachette uk. Journal of MultiDisciplinary Evaluation, 14(31):47-54, 2018. 4 +[35] Mihir Prabhudesai, Hsiao-Yu Fish Tung, Syed Ashar Javed, Maximilian Sieb, Adam W Harley, and Katerina Fragkiadaki. Embodied language grounding with 3d visual feature representations. In CVPR, pages 2220-2229, 2020. 2 +[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3, 5 +[37] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 1 +[38] Yongming Rao, Guangyi Chen, Jiwen Lu, and Jie Zhou. Counterfactual attention learning for fine-grained visual categorization and re-identification. In ICCV, pages 1025–1034, 2021. 2 +[39] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In Conference on Robot Learning, pages 1046-1056. PMLR, 2022. 7 +[40] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In CVPR, pages 3716-3725, 2020. 2, 4 +[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurlPS, 2022. 4, 5 +[42] Hongjun Wang, Jiyuan Chen, Lun Du, Qiang Fu, Shi Han, and Xuan Song. Causal-based supervision of attention in graph neural network: A better and simpler choice towards powerful attention. arXiv preprint arXiv:2305.13115, 2023. 2 +[43] Hanqing Wang, Wei Liang, Luc Van Gool, and Wenguan Wang. Dreamwalker: Mental planning for continuous vision-language navigation. In ICCV, pages 10873-10883, 2023. 1, 2 +[44] Tan Wang, Jianqiang Huang, Hanwang Zhang, and Qianru Sun. Visual commonsense r-cnn. In CVPR, pages 10760-10770, 2020. 2 +[45] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In CVPR, pages 19231-19242, 2023. 6, 7 + +[46] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 2, 3, 6, 7 +[47] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In CVPR, pages 1791-1800, 2021. 2, 6, 7 +[48] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Shuguang Cui, and Zhen Li. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In CVPR, pages 8563-8573, 2022. 2 +[49] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, pages 15225-15236, 2023. 6, 7 +[50] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2928-2937, 2021. 2, 6, 7 +[51] Tong Zhao, Gang Liu, Daheng Wang, Wenhao Yu, and Meng Jiang. Learning from counterfactual links for link prediction. In Int. Conf. on Machine Learning, pages 26911-26926. PMLR, 2022. 2 +[52] Dave Zhenyu Chen, Qirui Wu, Matthias Niener, and Angel X Chang. D3net: A speaker-listener architecture for semi-supervised dense captionin and visual grounding in rgb-d scans. In ECCV, 2022. 2 +[53] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 2 \ No newline at end of file diff --git a/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/images.zip b/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..82d7c492dc35d7369af437f3a29e4738a62a61ad --- /dev/null +++ b/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b0894516dc0b14c99f3cfc98fb8d0ccca86e0d8528c00bf252d20d363c2f968 +size 603475 diff --git a/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/layout.json b/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0553037799a011c74058794ab7db41967813b4e8 --- /dev/null +++ b/2024/Multi-Attribute Interactions Matter for 3D Visual Grounding/layout.json @@ -0,0 +1,10843 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 103, + 487, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 103, + 487, + 121 + ], + "spans": [ + { + "bbox": [ + 107, + 103, + 487, + 121 + ], + "type": "text", + "content": "Multi-Attribute Interactions Matter for 3D Visual Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "spans": [ + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "text", + "content": "Can Xu" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "text", + "content": ", Yuehui Han" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "text", + "content": ", Rui Xu" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "text", + "content": ", Le Hui" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "text", + "content": ", Jin Xie" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "inline_equation", + "content": "^{3,4*}" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "text", + "content": ", Jian Yang" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "text", + "content": "PCA Lab, Nanjing University of Science and Technology, Nanjing, China \n" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "text", + "content": "School of Electronics and Information, Northwestern Polytechnical University, Xi'an, China \n" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "text", + "content": "State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China \n" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 70, + 142, + 521, + 227 + ], + "type": "text", + "content": "School of Intelligence Science and Technology, Nanjing University, Suzhou, China \nvolcanoxl23@gmail.com; {hanyh,xu-ray,csjxie,csjyang}@njust.edu.cn; huile@nwpu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 281, + 290, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 290, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 290, + 617 + ], + "type": "text", + "content": "3D visual grounding aims to localize 3D objects described by free-form language sentences. Following the detection-then-matching paradigm, existing methods mainly focus on embedding object attributes in unimodal feature extraction and multimodal feature fusion, to enhance the discriminability of the proposal feature for accurate grounding. However, most of them ignore the explicit interaction of multiple attributes, causing a bias in unimodal representation and misalignment in multimodal fusion. In this paper, we propose a multi-attribute aware Transformer for 3D visual grounding, learning the multi-attribute interactions to refine the intra-modal and inter-modal grounding cues. Specifically, we first develop an attribute causal analysis module to quantify the causal effect of different attributes for the final prediction, which provides powerful supervision to correct the misleading attributes and adaptively capture other discriminative features. Then, we design an exchanging-based multimodal fusion module, which dynamically replaces tokens with low attribute attention between modalities before directly integrating low-dimensional global features. This ensures an attribute-level multimodal information fusion and helps align the language and vision details more efficiently for fine-grained multimodal features. Extensive experiments show that our method can achieve state-of-the-art performance on ScanRefer and Sr3D/Nr3D datasets. The code is publicly available at https://github.com/volcanoXC/MA2TransVG." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 643, + 128, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 643, + 128, + 656 + ], + "spans": [ + { + "bbox": [ + 47, + 643, + 128, + 656 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "text", + "content": "3D Visual Grounding (VG) aims to locate the most relevant object in a given point cloud scene based on a language description. As a cornerstone of wide applications [6, 8, 14, 28, 43], such as vision-language navigation and au" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 254, + 545, + 476 + ], + "blocks": [ + { + "bbox": [ + 309, + 254, + 545, + 476 + ], + "lines": [ + { + "bbox": [ + 309, + 254, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 309, + 254, + 545, + 476 + ], + "type": "image", + "image_path": "3af3f09543052496dd2e1c1bbc972bc409a2e8170c36d55b9b35066ea9f95be5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 485, + 547, + 519 + ], + "lines": [ + { + "bbox": [ + 305, + 485, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 305, + 485, + 547, + 519 + ], + "type": "text", + "content": "Figure 1. Different from prior works, we capture the multi-attribute interactions to guide the unimodal representation and multimodal fusion for more accurate grounding." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": "tonomous robots, 3D VG has attracted increasing attention from both academia and industry. However, limited by the complexity of free-form language description and the irregularity of sparse point clouds, achieving accurate 3D visual grounding remains an open issue." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "content": "Existing 3D VG methods mainly follow a detection-then-matching pipeline. Specifically, a pool of proposals is first produced by the pre-trained general 3D object detectors [11, 29, 37]. Then, the linguistic description is associated with the visual feature of each proposal to generate multimodal object representation, which will be used to predict their referring confidence score for grounding. Since such a two-stage paradigm aims to collect enough semantic cues to distinguish the target object, the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17253" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "quality of unimodal feature representation and multimodal feature fusion always play a critical role, which decides the discriminativeness of each proposal and their matching difficulty with the target label. Most methods focus on enhancing the feature from various perspectives, such as sparse convolution [47], 2D image assistance [5, 27, 46, 52], coarse proposal refinement [13, 22, 50], or segmentation/captioning collaboration [3, 10, 18, 26]. As shown in Fig. 1(a), recent works [7, 18, 19, 35] enhance the feature by embedding object attributes, such as color, name and language-conditioned correlation (i.e., distance, view, direction), showing a promising future in advancing grounding performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 231, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 289, + 483 + ], + "type": "text", + "content": "Though prior works made significant progress, we argue few of them explore the multi-attribute interactions in either unimodal representation or multimodal fusion, which affects the grounding accuracy, especially for cases with multiple objects. As the CASE in Fig. 1, the model embeds the most salient attribute (object name of 'door') and learns other potential attributes (e.g., 'brown', 'wooden', 'left') depending on the self-attention mechanism. Since the loss supervises the final prediction rather than the intermediate attention, the model may be encouraged to focus on some main attributes instead of all attributes, which weakens the discrimination of intra-modal representation. Meanwhile, when fusing the low-dimensional global feature between modalities via cross-attention, attention to 'brown' and 'wooden' should be encouraged for more discriminative language expression of the object, while they may be suppressed in the proposal representation to maximize the proposal's distinctiveness. Unfortunately, this would bring a misalignment in multimodal features, causing the model to select the auxiliary object (right door) rather than the primary object (left door)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": "To alleviate the bias in unimodal representation and misalignment in multimodal fusion, we propose a multi-attribute aware Transformer " + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "inline_equation", + "content": "(\\mathsf{MA}^2\\mathsf{TransVG})" + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": " for 3D visual grounding, which enhances the unimodal representation with intrinsic attribute interaction and refines the cross-modal information fusion to the attribute level (See Fig. 1(b)). Concretely, we develop an attribute causal analysis module (ACAM), which quantifies the effect of each attribute (i.e., learned attentions) and counterfactual interference (i.e., wrong attention) for the final prediction (i.e., grounding score). By maximizing the causal effect, powerful supervision is provided to correct the biased attribute and adaptively assign more proper attention to all attributes. Guided by the learned interaction, we further propose an exchanging-based multimodal fusion module (EMFM) to replace tokens with low attribute attention between modalities, which helps align more vision and language details and generate fine-grained multimodal features for accurate grounding. Our contributions include:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 72, + 547, + 252 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 120 + ], + "type": "text", + "content": "- We propose a multi-attribute aware Transformer for 3D object grounding, which explicitly models the multi-attribute interactions to prevent imbalanced and misaligned visual-textual representation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 120, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 120, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 120, + 547, + 168 + ], + "type": "text", + "content": "- We propose the ACAM module to quantify the causal effect of multiple intrinsic attributes with the final prediction, which helps adaptively assign more proper attention to all attributes for a better intra-modal understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 168, + 547, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 168, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 547, + 215 + ], + "type": "text", + "content": "- We propose the EMFM module to perform multimodal fusion based on multi-attribute exchanging, which helps align more attribute-level details between modalities for fine-grained multimodal features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 216, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 216, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 306, + 216, + 547, + 252 + ], + "type": "text", + "content": "- We evaluate the proposed method on " + }, + { + "bbox": [ + 306, + 216, + 547, + 252 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D/Sr3D}" + }, + { + "bbox": [ + 306, + 216, + 547, + 252 + ], + "type": "text", + "content": " and ScanRefer benchmarks, which outperforms all state-of-the-art methods by a large margin." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 262, + 391, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 391, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 391, + 275 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 282, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 547, + 510 + ], + "type": "text", + "content": "3D Visual Grounding. The interest in 3D VG has grown rapidly in recent years, and existing approaches include one-stage [21, 31] and two-stage methods. The two-stage methods [13, 16-18, 46, 47, 50] decouple the grounding into language-irrelevant object detection and cross-modal matching. Recent works focus on improving the intramodal representation and modeling the inter-modal relationship. For example, Graph-based approaches [1, 18, 47] infer spatial relations among proposals by connecting each proposal with its top-N nearest neighbors. Transformer-based approaches [7, 17, 19, 21, 31, 46] leverage the attention mechanism to concentrate the object attributes on proposals that are more essential to the referring. In these methods, only the most salient attribute is embedded and the fusion between modalities occurs in the global feature space, resulting in a biased intra-modal representation and inadequate inter-modal fusion. Instead, we capture the multi-attribute interactions and refine the cross-modal learning at the attribute level." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 548, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 548, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 548, + 653 + ], + "type": "text", + "content": "Causal Inference in Vision. Causality analysis [32] has been successfully used in several areas, such as adversarial learning [24], reinforcement learning [23] and graph neural networks [42, 51]. In vision tasks [30, 40, 44], causality inference also works as an effect tool to discover hidden causal structures and confront data biases. For example, Rao et al. [38] learn more effective attention based on causal inference for fine-grained image categorization, person and vehicle re-identification. As the first attempt, we introduce the causality analysis into 3D VG to quantify the effect of each attribute, which helps model the complex interactions of various attributes to enhance the unimodal feature." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "type": "text", + "content": "Transformers in Vision-and-Language. Transformer shows a powerful ability to learn multimodal features in various tasks, such as image captioning [9, 10, 48], vision-and-language navigation [28, 43], and vision question-answering [15, 53]. For 3D VG, different from most" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17254" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 76, + 542, + 252 + ], + "blocks": [ + { + "bbox": [ + 51, + 76, + 542, + 252 + ], + "lines": [ + { + "bbox": [ + 51, + 76, + 542, + 252 + ], + "spans": [ + { + "bbox": [ + 51, + 76, + 542, + 252 + ], + "type": "image", + "image_path": "2bf114a7b9284038875042191adacdd60fe419be31f3656ad11482cfe19d5db0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 547, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 547, + 309 + ], + "type": "text", + "content": "Figure 2. The framework of proposed " + }, + { + "bbox": [ + 46, + 262, + 547, + 309 + ], + "type": "inline_equation", + "content": "\\mathsf{MA}^2\\mathsf{TransVG}" + }, + { + "bbox": [ + 46, + 262, + 547, + 309 + ], + "type": "text", + "content": ". The embedding of pre-detected proposals, attributes, and language are extracted and then cross-attend in ACAM. By applying the causal effect analysis via " + }, + { + "bbox": [ + 46, + 262, + 547, + 309 + ], + "type": "inline_equation", + "content": "L_{attr}" + }, + { + "bbox": [ + 46, + 262, + 547, + 309 + ], + "type": "text", + "content": ", ACAM models the multi-attribute interactions for final prediction to generate the Attribute-Vision/Text attention, which can enhance the proposal and text embedding. The attention further guides the attribute-level feature exchanging in EMFM, which helps fuse the multimodal feature for accurate grounding." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 328, + 288, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 288, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 288, + 389 + ], + "type": "text", + "content": "methods that directly fuse the global unimodal features by stacking cross-attention layers [7, 17, 19, 21, 31, 46], we refine the cross-modal fusion to the attribute level, which helps suppress the misalignment between modalities to capture a finer-grained matching relationship." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 399, + 104, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 104, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 104, + 410 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 418, + 114, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 114, + 429 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 114, + 429 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "text", + "content": "Fig. 2 shows an overview of proposed " + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathsf{MA}^2\\mathsf{TransVG}" + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "text", + "content": ". Given a 3D point cloud with corresponding description, text features " + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^{(m + 1)\\times d}" + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "text", + "content": ", proposal features " + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "inline_equation", + "content": "O \\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "text", + "content": " and attribute features " + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "text", + "content": " are extracted (Sec. 3.2). The three streams then are fed into the attribute causal analysis module (ACAM) to capture the multi-attribute interactions (i.e., Attribute-Vision attention and Attribute-Text attention) to enhance initial text and proposal feature. The attention is supervised by an attribute loss " + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "inline_equation", + "content": "L_{attr}" + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "text", + "content": ", which is designed by estimating the causal effect of attributes for the final prediction (Sec. 3.3). The enhanced proposal embedding " + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "inline_equation", + "content": "O^*" + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "text", + "content": " and text embedding " + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "inline_equation", + "content": "T^*" + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "text", + "content": " are further passed to the exchanging-based multimodal fusion module (EMFM), where their tokens with small attribute attention are replaced with the average of all the tokens in the other modality (Sec. 3.4). With the fused multimodal embedding " + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "inline_equation", + "content": "E \\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 46, + 436, + 287, + 640 + ], + "type": "text", + "content": ", we predict the final grounding results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 204, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 204, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 204, + 659 + ], + "type": "text", + "content": "3.2. Input Modal Representation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Text Encoding. Given a sentence " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " word tokens, we use a pre-trained BERT model [20] to produce the text features " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "(t_{cls}, t_1, \\ldots, t_m)" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "t_i \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is the feature of each token with a dimensionality of " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "t_{cls} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is a special" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 329, + 419, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 329, + 419, + 339 + ], + "spans": [ + { + "bbox": [ + 305, + 329, + 419, + 339 + ], + "type": "text", + "content": "token for text classification." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "text", + "content": "Object Encoding. Given a point cloud scene " + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\pmb{P} \\in \\mathbb{R}^{K \\times 6}" + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "text", + "content": " points described by XYZ coordinates and RGB colors, we use the GroupFree [29] detector to obtain an object proposal list " + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\pmb{o} = (o_1, o_2, \\dots, o_n)" + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "text", + "content": " proposals and extract corresponding proposal feature " + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\pmb{O} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "text", + "content": " based on PointNet++ [36]. Since we mainly focus on the object referred in the text, each proposal feature " + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "inline_equation", + "content": "O_i \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "text", + "content": " will be added with the " + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "inline_equation", + "content": "t_{cls}" + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "text", + "content": " before feeding into ACAM for attribute-vision causal effect analysis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "text", + "content": "Attributes Encoding. Motivated by [3], for the proposal pool " + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "text", + "content": " objects, the multi-view RGB features " + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{c} \\in \\mathbb{R}^{n \\times 128}" + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "text", + "content": " that potentially contain the color and materials attribute is first extracted. Then, we use the 3-dimensional box center and 24-dimensional corner coordinates to describe the object size attribute " + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{s} \\in \\mathbb{R}^{n \\times 27}" + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "text", + "content": ". For the location attribute, we further consider the distance and angle of each pair of objects " + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "inline_equation", + "content": "(o_{i}, o_{j})" + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "text", + "content": ". Connecting the object center point " + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "inline_equation", + "content": "(x_{c_{i}}, y_{c_{i}}, z_{c_{i}})" + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "inline_equation", + "content": "(x_{c_{j}}, y_{c_{j}}, z_{c_{j}})" + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "text", + "content": ", the distance is their Euclidean distance " + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "inline_equation", + "content": "d_{ij}" + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "text", + "content": ". Similar to [7], the horizontal angle " + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "inline_equation", + "content": "\\theta_{h}" + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "text", + "content": " and vertical angle " + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "inline_equation", + "content": "\\theta_{v}" + }, + { + "bbox": [ + 304, + 449, + 546, + 579 + ], + "type": "text", + "content": " can be calculated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 344, + 587, + 545, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 587, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 344, + 587, + 545, + 613 + ], + "type": "interline_equation", + "content": "\\begin{array}{c} \\theta_ {h} = \\arctan 2 \\left(\\left(y _ {c _ {i}} - y _ {c _ {j}}\\right) / \\left(x _ {c _ {i}} - x _ {c _ {j}}\\right) \\right. \\\\ \\theta_ {v} = \\arcsin \\left(\\left(z _ {c _ {i}} - z _ {c _ {j}}\\right) / d _ {i j}\\right) \\end{array} \\tag {1}", + "image_path": "910446fc5d1954ea12604b13f4a5a866e0a8826ecbb6184eff9dfb42d45e96ec.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "type": "text", + "content": "The location attribute " + }, + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "type": "inline_equation", + "content": "f_{l}^{ij} \\in \\mathbb{R}^{5}" + }, + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "type": "text", + "content": " between any pair of objects is defined, as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 332, + 654, + 545, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 654, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 332, + 654, + 545, + 669 + ], + "type": "interline_equation", + "content": "f _ {l} ^ {i j} = \\left[ d _ {i j}, \\sin \\left(\\theta_ {h}\\right), \\cos \\left(\\theta_ {h}\\right), \\sin \\left(\\theta_ {v}\\right), \\cos \\left(\\theta_ {v}\\right) \\right] \\tag {2}", + "image_path": "cbe11ad1d9b9bfdd7246f354d4d29197bf0e64a6f351f981094904577a30885e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 676, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 676, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 676, + 545, + 713 + ], + "type": "text", + "content": "For each proposal " + }, + { + "bbox": [ + 304, + 676, + 545, + 713 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 304, + 676, + 545, + 713 + ], + "type": "text", + "content": ", the location attribute " + }, + { + "bbox": [ + 304, + 676, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{f}_l^i \\in \\mathbb{R}^5" + }, + { + "bbox": [ + 304, + 676, + 545, + 713 + ], + "type": "text", + "content": " can be calculated by averaging all " + }, + { + "bbox": [ + 304, + 676, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_l^{ij}" + }, + { + "bbox": [ + 304, + 676, + 545, + 713 + ], + "type": "text", + "content": " connected with " + }, + { + "bbox": [ + 304, + 676, + 545, + 713 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 304, + 676, + 545, + 713 + ], + "type": "text", + "content": ", and the final attribute embedding can be generated by concatenating" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17255" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 287, + 96 + ], + "type": "text", + "content": "all three attributes, as " + }, + { + "bbox": [ + 46, + 71, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{i} = [\\pmb{f}_{l}^{i},\\pmb{f}_{c}^{i},\\pmb{f}_{s}^{i}]" + }, + { + "bbox": [ + 46, + 71, + 287, + 96 + ], + "type": "text", + "content": ". We map " + }, + { + "bbox": [ + 46, + 71, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{i}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 46, + 71, + 287, + 96 + ], + "type": "text", + "content": " to the same dimension of text feature using a FC layer." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 105, + 230, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 230, + 118 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 230, + 118 + ], + "type": "text", + "content": "3.3. Attribute Causal Analysis Module" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 125, + 287, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 125, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 125, + 287, + 208 + ], + "type": "text", + "content": "The attribute causal analysis module (ACAM) models the multi-attribute interactions with vision/text as Attribute-Vision/Text attention based on a well-designed Attribute-aware Transformer. To capture more influential and diverse interactions, ACAM further guides the learning process with an extra loss " + }, + { + "bbox": [ + 46, + 125, + 287, + 208 + ], + "type": "inline_equation", + "content": "L_{attr}" + }, + { + "bbox": [ + 46, + 125, + 287, + 208 + ], + "type": "text", + "content": ", which measures the causal effect of learned attribute-level attentions for the final prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "type": "text", + "content": "Attributes-Text Attention Learning. Given attribute embedding " + }, + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{F} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "type": "text", + "content": ", Attributes-Text attention " + }, + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{A}_{T}" + }, + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "type": "text", + "content": " helps select related attribute features from the initial text feature " + }, + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^{(m + 1) \\times d}" + }, + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "type": "text", + "content": ", thereby the text representation can be enhanced to better describe the target object. We define " + }, + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{A}_{T} \\in \\mathbb{R}^{n \\times (m + 1)}" + }, + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "type": "text", + "content": " as the cross-attention between the Attribute embedding " + }, + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "type": "text", + "content": " and text embedding " + }, + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 46, + 209, + 287, + 304 + ], + "type": "text", + "content": ", which can be computed with a standard Transformer [41]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 306, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 413 + ], + "type": "text", + "content": "In order to enlarge the model capacity for diverse Attribute-Text correlations, we also introduce multi parallel attention layers " + }, + { + "bbox": [ + 46, + 306, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 306, + 287, + 413 + ], + "type": "text", + "content": " as different transformation heads, and the multi-head attentions will be fused with the standard self-attention [41]. As shown in Fig. 2, " + }, + { + "bbox": [ + 46, + 306, + 287, + 413 + ], + "type": "inline_equation", + "content": "A_{T}" + }, + { + "bbox": [ + 46, + 306, + 287, + 413 + ], + "type": "text", + "content": " then guides to generate the enhanced text feature " + }, + { + "bbox": [ + 46, + 306, + 287, + 413 + ], + "type": "inline_equation", + "content": "T^{*} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 46, + 306, + 287, + 413 + ], + "type": "text", + "content": ", which will be used to train the language classifier with the supervision of the language classification loss " + }, + { + "bbox": [ + 46, + 306, + 287, + 413 + ], + "type": "inline_equation", + "content": "L_{sent}" + }, + { + "bbox": [ + 46, + 306, + 287, + 413 + ], + "type": "text", + "content": ", and also serves as input for subsequent EMFM for multi-modal fusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "text", + "content": "Attributes-Vision Attention Learning. Instead of directly calculating the cross-attention between proposal features " + }, + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "text", + "content": " and attribute features " + }, + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "text", + "content": ", we first guide the proposal features with language token " + }, + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "inline_equation", + "content": "t_{cls}" + }, + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "text", + "content": ", as we prefer to care about specific categories of objects mentioned in the text [7]. Specifically, for the text-guided proposal feature, we adopt the similar Transformer structure in Attributes-Text attention learning to calculate its self-attention " + }, + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "inline_equation", + "content": "A_V^s \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "text", + "content": ", and cross-attention " + }, + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "inline_equation", + "content": "A_V^c \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 414, + 287, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 571 + ], + "type": "text", + "content": "Similar to [7], we then integrate the self-attention of proposals " + }, + { + "bbox": [ + 46, + 522, + 287, + 571 + ], + "type": "inline_equation", + "content": "A_V^s" + }, + { + "bbox": [ + 46, + 522, + 287, + 571 + ], + "type": "text", + "content": " with the cross attribute-proposal attention " + }, + { + "bbox": [ + 46, + 522, + 287, + 571 + ], + "type": "inline_equation", + "content": "A_V^c" + }, + { + "bbox": [ + 46, + 522, + 287, + 571 + ], + "type": "text", + "content": " using a sigmoid softmax (sigsoftmax) fusion operation to calculate the Attributes-Vision attention " + }, + { + "bbox": [ + 46, + 522, + 287, + 571 + ], + "type": "inline_equation", + "content": "A_V" + }, + { + "bbox": [ + 46, + 522, + 287, + 571 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 579, + 287, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 579, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 82, + 579, + 287, + 608 + ], + "type": "interline_equation", + "content": "A _ {v} = \\frac {\\sigma \\left(\\boldsymbol {A} _ {\\boldsymbol {V}} ^ {\\boldsymbol {c}} (i , j)\\right) \\exp \\left(\\boldsymbol {A} _ {\\boldsymbol {V}} ^ {\\boldsymbol {s}} (i , j)\\right)}{\\sum_ {j = 1} ^ {n} \\sigma \\left(\\boldsymbol {A} _ {\\boldsymbol {V}} ^ {\\boldsymbol {c}} (i , j)\\right) \\exp \\left(\\boldsymbol {A} _ {\\boldsymbol {V}} ^ {\\boldsymbol {s}} (i , j)\\right)} \\tag {3}", + "image_path": "ff5f339f68dc65c726610525f67613144604836ba1d449c47b168f705fb5d202.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " is the sigmoid function. " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "A_V^s (i,j)" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "A_V^c (i,j)" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " are the element in " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "-th column in attention matrix " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "A_V^s" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "A_V^c" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ", respectively. The same multi-head attention strategy is used in Attribute-Vision attention " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "A_V" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ". In a similar manner, the enhanced proposal embedding " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "O^{*}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " can be obtained by performing a matrix multiplication operation between " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "A_V" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ", which will serve as the input of proposal classifier and EMFM module." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 543, + 153 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 543, + 153 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 543, + 153 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 543, + 153 + ], + "type": "image", + "image_path": "83d7e39ca8b8ee336fbb6f818734b170655925651d94eb642d55a8eac4a965ab.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 161, + 545, + 206 + ], + "lines": [ + { + "bbox": [ + 304, + 161, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 304, + 161, + 545, + 206 + ], + "type": "text", + "content": "Figure 3. Visualization of deriving causal effects through counterfactual causality. We subtract the counterfactual prediction from the original prediction to analyze the effects of learned attention and maximize them in the training process." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 223, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 342 + ], + "type": "text", + "content": "Strengthen Interaction With " + }, + { + "bbox": [ + 304, + 223, + 545, + 342 + ], + "type": "inline_equation", + "content": "L_{attri}" + }, + { + "bbox": [ + 304, + 223, + 545, + 342 + ], + "type": "text", + "content": ". Though multiple attributes have been embedded to guide the language and vision representations via Attribute-Text/Vision attention, how these attributes interact with each other to affect predictions remains a black box. To address this issue, our core idea lies in measuring the causal effect of " + }, + { + "bbox": [ + 304, + 223, + 545, + 342 + ], + "type": "inline_equation", + "content": "A_V" + }, + { + "bbox": [ + 304, + 223, + 545, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 223, + 545, + 342 + ], + "type": "inline_equation", + "content": "A_T" + }, + { + "bbox": [ + 304, + 223, + 545, + 342 + ], + "type": "text", + "content": " for the final prediction. We design the attribute loss " + }, + { + "bbox": [ + 304, + 223, + 545, + 342 + ], + "type": "inline_equation", + "content": "L_{attri}" + }, + { + "bbox": [ + 304, + 223, + 545, + 342 + ], + "type": "text", + "content": " based on the counterfactual analysis [33], providing extra supervision to encourage the network to explore more influential interactions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "text", + "content": "As shown in Fig. 3, taking the Attribute-vision attention as an example, we conduct counterfactual intervention by imagining non-existent attention " + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\bar{A}_V" + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "text", + "content": " to replace the initial " + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "inline_equation", + "content": "A_V" + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "text", + "content": ". We keep the vision representation " + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "text", + "content": " unchanged, which will be multiplied with " + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\bar{A}_V" + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "text", + "content": " to calculate the counterfactual representation " + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\bar{O}^*" + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\bar{O}^*" + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "text", + "content": " are then fused with " + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "text", + "content": " to obtain the counterfactual prediction " + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\bar{Y} \\in \\mathbb{R}^{n \\times 1}" + }, + { + "bbox": [ + 304, + 343, + 545, + 426 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 431, + 545, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 431, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 347, + 431, + 545, + 445 + ], + "type": "interline_equation", + "content": "\\bar {\\boldsymbol {Y}} = d o \\left(\\boldsymbol {A} _ {V} = \\bar {\\boldsymbol {A}} _ {V}, \\boldsymbol {T}\\right) = G \\left(\\bar {\\boldsymbol {O}} ^ {*}, \\boldsymbol {T}\\right) \\tag {4}", + "image_path": "2e8bb40bad2768da02a1e61884c6225e6bf1c393f1c8435f3e26714948a5eba7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 450, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 450, + 545, + 510 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 304, + 450, + 545, + 510 + ], + "type": "text", + "content": " is the grounding operation. In practice, the exact form of how counterfactual is achieved is not limited, and our starting point here is just to set the ball rolling. In our case, random attention, uniform attention, shuffle attention or reversed attention are used as the counterfactuals." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 510, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 546 + ], + "type": "text", + "content": "Following [33, 34, 40], we define the causal effect " + }, + { + "bbox": [ + 304, + 510, + 545, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_e" + }, + { + "bbox": [ + 304, + 510, + 545, + 546 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 510, + 545, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_V" + }, + { + "bbox": [ + 304, + 510, + 545, + 546 + ], + "type": "text", + "content": " as the difference between the observed prediction " + }, + { + "bbox": [ + 304, + 510, + 545, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 510, + 545, + 546 + ], + "type": "text", + "content": " and the counterfactual prediction " + }, + { + "bbox": [ + 304, + 510, + 545, + 546 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{Y}}" + }, + { + "bbox": [ + 304, + 510, + 545, + 546 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 377, + 551, + 545, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 551, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 377, + 551, + 545, + 566 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Y} _ {e} = \\mathbb {E} _ {\\bar {A} _ {V} \\sim \\gamma} (\\boldsymbol {Y} - \\bar {\\boldsymbol {Y}}) \\tag {5}", + "image_path": "2d4c32cdd97c9200ff47f0a47f06dda54eb73cdb5fdfd0519b1b6f0c60101cd9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 543, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 543, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 543, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 570, + 543, + 581 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 570, + 543, + 581 + ], + "type": "text", + "content": " means the distribution of counterfactual attention." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 642 + ], + "type": "text", + "content": "Since the causal effect can be interpreted as how the attention improves the final prediction compared to wrong attentions, we use " + }, + { + "bbox": [ + 304, + 582, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_e" + }, + { + "bbox": [ + 304, + 582, + 545, + 642 + ], + "type": "text", + "content": " to measure the learned attention, i.e., the multi-attribute interactions. Based on this observation, we further design an attribute loss " + }, + { + "bbox": [ + 304, + 582, + 545, + 642 + ], + "type": "inline_equation", + "content": "L_{attr}" + }, + { + "bbox": [ + 304, + 582, + 545, + 642 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 382, + 647, + 545, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 647, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 382, + 647, + 545, + 661 + ], + "type": "interline_equation", + "content": "L _ {\\text {a t t r i}} = L _ {c e} \\left(\\boldsymbol {Y} _ {\\boldsymbol {e}}, \\boldsymbol {y}\\right) \\tag {6}", + "image_path": "029dd9ed23fe88e666fe1219da014efc10c3db1f3791bb0e46f4fe0fd8032c55.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "L_{ce}" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " is the cross-entropy loss and " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " is the GT label. Similarly, we can obtain the causal effect of Attribute-text attention and compute its attribute loss. By optimizing " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "L_{attr}" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "A_{T}" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "A_{V}" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": ", the model is expected to:" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17256" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "(1) improve the prediction based on correct attentions as much as possible, which helps discover more discriminative attributes; (2) penalize the prediction based on wrong attentions, which forces the model adaptively capture more attributes to avoid biased unimodal embeddings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 140, + 287, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 152 + ], + "type": "text", + "content": "3.4. Exchanging-based Multimodal Fusion Module" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 158, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 287, + 312 + ], + "type": "text", + "content": "We propose the exchanging-based multimodal fusion module (EMFM) to communicate attribute-level details of objects between both input vectors for accurate grounding. Specifically, since text/visual encoders tend to focus on different attributes for better unimodal embeddings, directly fusing unimodal embeddings using cross-attention will hurt the distinguishability of multimodal features. To narrow the semantic misalignment between modalities, we select and exchange tokens with low attention score using the average of all tokens from another modal, which helps align attribute-level information between modalities sufficiently and eliminate semantic ambiguity in multimodal representations of proposals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "text", + "content": "Exchanging With Attributes Interactions. As shown in Fig. 3, we implement the EMFM by two Transformer encoders with shared parameters, where the shallow " + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "text", + "content": " layers are set as regular Transformer encoder layers, followed by " + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "text", + "content": " exchanging layers. Since " + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "inline_equation", + "content": "A_V" + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "inline_equation", + "content": "A_T" + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "text", + "content": " have modeled the multi-attribute interactions, they are taken as references to select tokens for exchanging. Meanwhile, " + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "inline_equation", + "content": "A_V" + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "inline_equation", + "content": "A_T" + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "text", + "content": " will be updated using the self-attention learned from unimodal embedding " + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "inline_equation", + "content": "T^*" + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "inline_equation", + "content": "O^*" + }, + { + "bbox": [ + 46, + 314, + 287, + 445 + ], + "type": "text", + "content": ". Considering the interdependence between the attribute attention and unimodal embedding, we concatenate them for the subsequent calculation, as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 455, + 287, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 455, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 111, + 455, + 287, + 472 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {T} ^ {e} (0) = \\operatorname {C o n c a t} \\left[ \\boldsymbol {A} _ {T}, \\boldsymbol {T} ^ {*} \\right] \\\\ \\boldsymbol {Q} ^ {e} (0) = \\operatorname {C o n c a t} \\left[ \\boldsymbol {A} _ {T}, \\boldsymbol {Q} ^ {*} \\right] \\end{array} \\tag {7}", + "image_path": "66a36781de1514b8a1b256a45becba3dea297b134e8701725b0ff44f7723f303.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 468, + 221, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 468, + 221, + 479 + ], + "spans": [ + { + "bbox": [ + 113, + 468, + 221, + 479 + ], + "type": "interline_equation", + "content": "\\boldsymbol {O} ^ {e} (0) = \\operatorname {C o n c a t} \\left[ \\boldsymbol {A} _ {V}, \\boldsymbol {O} ^ {*} \\right]", + "image_path": "6cc6100805631a8345d79f314faabfe12dda61a22969323b0db13a1610342c2b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 489, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 287, + 513 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 489, + 287, + 513 + ], + "type": "inline_equation", + "content": "T^{e}(0)" + }, + { + "bbox": [ + 46, + 489, + 287, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 489, + 287, + 513 + ], + "type": "inline_equation", + "content": "O^{e}(0)" + }, + { + "bbox": [ + 46, + 489, + 287, + 513 + ], + "type": "text", + "content": " are embedding for textual and visual modalities at layer 0, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "text", + "content": "At layer " + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\mu + 1" + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "text", + "content": " with intermediate embedding " + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "inline_equation", + "content": "T^{e}(\\mu + 1)" + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "text", + "content": ", inspired by [4, 25], we select tokens with a " + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "text", + "content": "-proportion of the smallest attention scores to " + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "inline_equation", + "content": "A_{T}" + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "text", + "content": " and replace their embedding vectors with the average embedding of all the tokens in " + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "inline_equation", + "content": "O^{e}(\\mu + 1)" + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "text", + "content": ". For instance, the selected " + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "text", + "content": "-th token of " + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "inline_equation", + "content": "T^{e}(\\mu + 1)" + }, + { + "bbox": [ + 46, + 513, + 287, + 579 + ], + "type": "text", + "content": " can be updated by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 588, + 287, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 588, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 52, + 588, + 287, + 616 + ], + "type": "interline_equation", + "content": "\\boldsymbol {T} ^ {\\boldsymbol {e}} (\\mu + 1) [ k,: ] = \\frac {1}{n} \\sum_ {j = 1} ^ {n} \\boldsymbol {O} ^ {\\boldsymbol {e}} (\\mu + 1) [ j,: ] + \\boldsymbol {T} ^ {\\boldsymbol {e}} (\\mu + 1) [ k,: ] \\tag {8}", + "image_path": "84258b0e56470c3b9ad742a369dddbc00cbae3ddd3ed82b03d55296843a77cdc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": "The updated embeddings are used to calculate the multihead self-attention [41], which will guide the exchanging operation at the next layer, e.g., " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "A_T" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "(\\mu + 1)" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": "-th layer are updated by the self-attention of the previous layer " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Multimodal Fusion. We also apply the residual connection between adjacent exchanging layers to reduce the information loss caused by replacement. The exchanging process" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "continues until reaching the " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " layer to generate " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "O^e (\\eta)" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "T^{e}(\\eta)" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ", which will be fed into the feed-forward network (FFN) with a normalization layer and a fully connected layer to derive the final fusion embedding " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\pmb {E}\\in \\mathbb{R}^{n\\times n}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 120, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 545, + 192 + ], + "type": "text", + "content": "Grounding Head. We use a two-layer feed-forward neural network as the object grounding head. Given the multimodal representation " + }, + { + "bbox": [ + 304, + 120, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 304, + 120, + 545, + 192 + ], + "type": "text", + "content": ", the grounding head will predict the probability " + }, + { + "bbox": [ + 304, + 120, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i \\in \\mathbb{R}^{n \\times 1}" + }, + { + "bbox": [ + 304, + 120, + 545, + 192 + ], + "type": "text", + "content": " for each object proposal, where the proposal with the maximum probability is selected as the target in the inference stage." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 198, + 421, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 198, + 421, + 211 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 421, + 211 + ], + "type": "text", + "content": "3.5. Training Objectives" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 217, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 277 + ], + "type": "text", + "content": "Following the previous works [1, 5, 17], the loss of " + }, + { + "bbox": [ + 304, + 217, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{MA}^2\\mathrm{TransVG}" + }, + { + "bbox": [ + 304, + 217, + 545, + 277 + ], + "type": "text", + "content": " consists of the primary 3D object grounding loss " + }, + { + "bbox": [ + 304, + 217, + 545, + 277 + ], + "type": "inline_equation", + "content": "L_{vg}" + }, + { + "bbox": [ + 304, + 217, + 545, + 277 + ], + "type": "text", + "content": ", language classification loss " + }, + { + "bbox": [ + 304, + 217, + 545, + 277 + ], + "type": "inline_equation", + "content": "L_{sent}" + }, + { + "bbox": [ + 304, + 217, + 545, + 277 + ], + "type": "text", + "content": ", proposal classification loss " + }, + { + "bbox": [ + 304, + 217, + 545, + 277 + ], + "type": "inline_equation", + "content": "L_{obj}" + }, + { + "bbox": [ + 304, + 217, + 545, + 277 + ], + "type": "text", + "content": ", and two attribute losses " + }, + { + "bbox": [ + 304, + 217, + 545, + 277 + ], + "type": "inline_equation", + "content": "L_{attri}" + }, + { + "bbox": [ + 304, + 217, + 545, + 277 + ], + "type": "text", + "content": " for vision/text attention causal analysis, as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 352, + 285, + 545, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 285, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 352, + 285, + 545, + 298 + ], + "type": "interline_equation", + "content": "L = L _ {v g} + L _ {s e n t} + L _ {o b j} + \\lambda L _ {a t t r i} \\tag {9}", + "image_path": "1ff6c362581a911caf8cc6e2855a3b6a769d8963a1f378a74df9ded09b4a6069.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 305, + 511, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 305, + 511, + 317 + ], + "spans": [ + { + "bbox": [ + 306, + 305, + 511, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 305, + 511, + 317 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 305, + 511, + 317 + ], + "type": "text", + "content": " is the hyper-parameter to balance the loss." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 327, + 381, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 327, + 381, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 381, + 340 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 347, + 481, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 347, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 481, + 358 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 365, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 617 + ], + "type": "text", + "content": "We evaluate the proposed " + }, + { + "bbox": [ + 304, + 365, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathsf{MA}^2\\mathsf{TransVG}" + }, + { + "bbox": [ + 304, + 365, + 545, + 617 + ], + "type": "text", + "content": " on ScanRefer [5] and Sr3D/Nr3D [1] datasets. ScanRefer offers 51,583 descriptions of 11,046 objects from 800 ScanNet [12] scenes. A target object is labeled as 'unique' if it is the only object of its class in the scene; otherwise, it is labeled as 'multiple'. For ScanRefer, we use Acc@0.25 and Acc@0.5 as evaluation metrics, which measure the fraction of language queries whose predicted box overlaps the ground truth box with 3D IoU higher than 0.25/0.5. Compared to ScanRefer where objects need to be detected and then matched, Sr3D/Nr3D datasets provide object masks for each scene and only need to recognize the classes of the proposals to choose the target object. Nr3D consists of 41,503 descriptions from 707 scenes with human annotation, while Sr3D contains 83,572 simple machine-generated descriptions. Each scene in Sr3D/Nr3D can also be divided into 'easy' and 'hard' depending on whether there are more than two instances, and can be divided into 'view-dependent' and 'view-independent' according to whether the referring expression is dependent on the camera view. Following [1], the accuracy is used to verify the model." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 623, + 438, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 438, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 438, + 635 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "The text encoding module is initialized from the first three layers of BERT [20], and the object encoding module PointNet++ [36] samples 1024 points for all the objects. We generate an initial set of " + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "inline_equation", + "content": "n = 256" + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": " proposals using [29]. We set the dimension of unimodal representation " + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "inline_equation", + "content": "d = 768" + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": " and use a three-layer transformer with 12 heads for" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17257" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 90, + 444, + 299 + ], + "blocks": [ + { + "bbox": [ + 156, + 70, + 438, + 82 + ], + "lines": [ + { + "bbox": [ + 156, + 70, + 438, + 82 + ], + "spans": [ + { + "bbox": [ + 156, + 70, + 438, + 82 + ], + "type": "text", + "content": "Table 1. Comparison of 3D VG results with the state-of-the-art on ScanRefer." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 90, + 444, + 299 + ], + "lines": [ + { + "bbox": [ + 148, + 90, + 444, + 299 + ], + "spans": [ + { + "bbox": [ + 148, + 90, + 444, + 299 + ], + "type": "table", + "html": "
MethodVenueUnique (19%)Multiple (81%)Overall
0.250.50.250.50.250.5
ScanRefer [5]ECCV2067.646.232.121.340.026.1
ReferIt3D [1]ECCV2053.837.521.012.826.416.9
TGNN [18]AAAI2168.656.829.823.237.429.7
InstanceRefer [47]ICCV2177.566.831.324.840.232.9
3DVG [50]ICCV2177.258.538.428.745.934.5
FFL-3DOG [13]ICCV2178.867.935.225.741.334.0
MVT [19]CVPR2277.766.531.925.340.833.3
3D-SPS [31]CVPR2281.664.839.529.647.736.4
ViL3DRef [7]NeurIPS2281.668.640.330.747.937.7
BUTD-DETR [22]ECCV22////52.239.8
EDA [45]CVPR2385.868.649.137.654.642.3
M3DRef-CLIP [49]ICCV23/77.2/36.8/44.7
3DJCG [3]CVPR2278.861.340.130.147.636.1
UniT3D [10]ICCV2382.873.136.431.142.339.1
3DRefTR-HR [26]ICCV2386.070.949.638.355.043.2
3DRefTR-SP [26]ICCV2386.171.050.138.755.543.5
Ours/86.374.153.841.457.945.7
", + "image_path": "a4d504e5e19474e9bdc26fc4bf9292a9c3ea1f3b07525d4192eba0b10bc6b6f3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 319, + 288, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 288, + 462 + ], + "type": "text", + "content": "the Attribute-Vision/Text attention learning. For a fair comparison, we keep the architecture parameters as same as previous works [7, 19, 46]. In ACAM, we use random attention to initialize the counterfactual attention. In EMFM, considering the balance of raw intra-modal knowledge and inter-modal communication, we empirically set " + }, + { + "bbox": [ + 46, + 319, + 288, + 462 + ], + "type": "inline_equation", + "content": "\\theta = 10\\%" + }, + { + "bbox": [ + 46, + 319, + 288, + 462 + ], + "type": "text", + "content": ". The start layer " + }, + { + "bbox": [ + 46, + 319, + 288, + 462 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 319, + 288, + 462 + ], + "type": "text", + "content": " and end layer " + }, + { + "bbox": [ + 46, + 319, + 288, + 462 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 319, + 288, + 462 + ], + "type": "text", + "content": " are set to 1 and 4, respectively. The hyper-parameter of the whole loss in Eq. (9) is " + }, + { + "bbox": [ + 46, + 319, + 288, + 462 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 46, + 319, + 288, + 462 + ], + "type": "text", + "content": ". All models are trained on 4 NVIDIA RTX Titan GPUs, which are optimized using the AdamW algorithm with a batch size of 128 and a learning rate of 0.0005 with cosine decay scheduling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 467, + 272, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 272, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 272, + 480 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "text", + "content": "ScanRefer. In Tab. 1, we compare " + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{MA}^2\\mathsf{TransVG}" + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "text", + "content": " with existing works on the ScanRef dataset and observe the following results: i) Compared to pure 3D VG works such as BUTD-DETR [22], 3D-SPS [31], ViL3DRef [7], and EDA [45], our method achieves state-of-the-art performance of " + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "inline_equation", + "content": "57.9\\%" + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "inline_equation", + "content": "45.7\\%" + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "text", + "content": ", by a substantial margin with an overall improvement over " + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "text", + "content": ". ii) To facilitate the 3D grounding, previous works [5, 31, 46, 50] learn better point cloud features with the supplementary of 2D images while most recent works [3, 10, 26] learn better language-visual alignment by connecting 3D dense captioning or segmentation tasks. However, we outperform second-place 3DRefTR-SP [26] with an overall improvement over " + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "text", + "content": " for both IoU = 0.25 and IoU = 0.5 setting. This superiority illustrates that mining the multi-attribute interactions help capture more efficient and discriminative multi-modal representation to distinguish objects. iii) Our " + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{MA}^2\\mathsf{TransVG}" + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "text", + "content": " can reach a remarkable accuracy of " + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "inline_equation", + "content": "53.8\\%" + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "inline_equation", + "content": "41.4\\%" + }, + { + "bbox": [ + 46, + 485, + 288, + 713 + ], + "type": "text", + "content": " for 'Multiple' scenes, which verifies the effec" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 319, + 545, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 343 + ], + "type": "text", + "content": "tiveness of our attribute interaction for understanding finer-grained language-visual contextual dependency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "content": "Nr3D&Sr3D. Tab. 2 reports the grounding accuracy on Nr3D and Sr3D datasets. For a fair comparison, all the methods use the ground-truth object proposal with no ground-truth labels. In Nr3D, though the language descriptions are more complex to cause additional challenges for text understanding and cross-modal alignment, our method still outperforms the EDA [45], which is equipped with text component decoupling and dense matching between two modalities. Compared to EDA, our " + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathsf{MA}^2\\mathsf{TransVG}" + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "content": " not only achieves a similar decoupling function for both descriptions and point clouds via attribute representation, but also considers their correlationships for prediction during Attribute-Text/Vision causal attention learning, which contributes to the great improvements by " + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "inline_equation", + "content": "+13.1\\%" + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "content": " in overall accuracy. Compared to other recent competitor ViL3DRef [7], 3DRefTRSR [26], and M3DRef-CLIP [49], we also achieve a consistent improvement of overall accuracy in all split settings by " + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "inline_equation", + "content": "+0.8\\%" + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "inline_equation", + "content": "+2.6\\%" + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "inline_equation", + "content": "+15.8\\%" + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "content": ", respectively. In Sr3D, our method reaches an accuracy of " + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "inline_equation", + "content": "73.9\\%" + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "content": ", which greatly surpasses the best competitor ViL3DRef [7] " + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "content": " and outperforms other state-of-the-art methods [26, 45] over " + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "content": ". Similar to ViL3DRef [7] that guided the grounding with languaged-conditioned spatial relation, we not only embed the spatial attributes but also explicitly consider other attributes in the unimodal representation and further perform exchanging-based fusion. All the results show that multi-attribute interactions can help capture more semantic clues for accurate grounding." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17258" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 91, + 518, + 322 + ], + "blocks": [ + { + "bbox": [ + 153, + 70, + 440, + 82 + ], + "lines": [ + { + "bbox": [ + 153, + 70, + 440, + 82 + ], + "spans": [ + { + "bbox": [ + 153, + 70, + 440, + 82 + ], + "type": "text", + "content": "Table 2. Comparison of 3D VG results with the state-of-the-art on Sr3D/Nr3D." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 91, + 518, + 322 + ], + "lines": [ + { + "bbox": [ + 72, + 91, + 518, + 322 + ], + "spans": [ + { + "bbox": [ + 72, + 91, + 518, + 322 + ], + "type": "table", + "html": "
MethodVenueNr3DSr3D
EasyHardView DepView IndepOverallEasyHardView DepView IndepOverall
ReferIt3D [1]ECCV2043.627.932.537.135.644.731.539.240.840.8
TGNN [18]AAAI2144.230.635.838.037.348.536.945.845.045.0
InstanceRefer [47]ICCV2146.031.834.541.938.851.140.545.448.148.0
3DVG [50]ICCV2148.534.834.843.740.854.244.944.651.751.4
FFL-3DOG [13]ICCV2148.235.037.144.741.7/////
TransRefer3D [17]ACMM2148.536.036.544.942.160.550.249.957.757.4
LanguageRefer [39]CoRL2151.036.641.745.043.958.949.349.256.356.0
SAT [46]ICCV2156.342.446.950.449.261.250.049.258.357.9
BUTD-DETR [22]ECCV22////43.3////52.1
LAR [2]NeurIPS2258.442.347.452.148.963.051.250.059.159.4
3D-SPS [31]CVPR2258.145.148.053.251.556.265.449.263.262.6
MVT [19]CVPR2261.349.154.354.355.466.958.858.464.764.5
M3DRef-CLIP [49]ICCV2355.643.442.352.949.4/////
EDA [45]CVPR23////52.1////68.1
3DRefTR-SR [26]ICCV23////52.6////68.5
ViL3DRef [7]NeurIPS2270.257.462.064.564.474.967.963.873.272.8
Ours/71.157.662.565.465.276.069.364.573.873.9
", + "image_path": "a7e49266e07a30003872ae5e82b724fd0ab87b70537141ab8c14e48b10b733fc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 340, + 147, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 340, + 147, + 352 + ], + "spans": [ + { + "bbox": [ + 47, + 340, + 147, + 352 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "text", + "content": "Effectiveness of Each Component. We quantitatively investigate the contribution of each component in our method in Tab. 3. The baseline model simply extracts and fuses the object and language feature through a standard Transformer structure, where the modal relationships only implicitly depend on the self/cross-attention. Taking results on " + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D}" + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "text", + "content": " as an example, when enhancing the feature with " + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "inline_equation", + "content": "A_{T}" + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "inline_equation", + "content": "A_{V}" + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "text", + "content": ", we yield an accuracy of " + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "inline_equation", + "content": "55.2\\%" + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "inline_equation", + "content": "51.9\\%" + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "text", + "content": ", which can increase to " + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "inline_equation", + "content": "59.7\\%" + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "text", + "content": " when combining two attentions. This indicates that embedding the multi-attribute within each single modal is the basis of accurate cross-modal understanding. Furthermore, we achieve a gain of " + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "text", + "content": " after introducing causal effect analysis, which helps adaptively capture more reliable multi-attribute interactions for feature enhancement. Exchanging tokens further improves the accuracy to " + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "inline_equation", + "content": "65.2\\%" + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "text", + "content": " because multimodal knowledge exchanging boosts the quality of the learned embeddings for each modality. All the results demonstrate the effectiveness of each component in our proposed " + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{MA}^2\\mathrm{TransVG}" + }, + { + "bbox": [ + 46, + 362, + 288, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Effectiveness of Multi-attribute. To verify the effectiveness of multi-attribute interactions, we conduct ablation on attribute quantity and multi-attribute interactions. In Tab. 3. We adopt commonly used attributes (location " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{l}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", size " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", color " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ") with simple embedding method and achieve SOTA performance. The accuracy is " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "44.9\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " on Nr3D when generating attribute embedding only using " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{l}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", which increases to " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "46.1\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "+1.2\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "60.2\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "+15.3\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ") after adding " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " and interactions, respectively. Results show that our accuracy gain mainly comes from the multi-attribute" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 360, + 535, + 501 + ], + "blocks": [ + { + "bbox": [ + 307, + 339, + 544, + 350 + ], + "lines": [ + { + "bbox": [ + 307, + 339, + 544, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 339, + 544, + 350 + ], + "type": "text", + "content": "Table 3. Ablation studies of each component on " + }, + { + "bbox": [ + 307, + 339, + 544, + 350 + ], + "type": "inline_equation", + "content": "\\mathrm{{Nr}}3\\mathrm{D}" + }, + { + "bbox": [ + 307, + 339, + 544, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 307, + 339, + 544, + 350 + ], + "type": "inline_equation", + "content": "\\mathrm{{Sr}}3\\mathrm{D}" + }, + { + "bbox": [ + 307, + 339, + 544, + 350 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 360, + 535, + 501 + ], + "lines": [ + { + "bbox": [ + 317, + 360, + 535, + 501 + ], + "spans": [ + { + "bbox": [ + 317, + 360, + 535, + 501 + ], + "type": "table", + "html": "
flfsfcATAVLattriExgNr3DSr3D
✓✓44.955.6
✓✓✓✓✓60.269.2
✓✓✓✓✓59.368.4
✓✓✓✓✓✓61.871.2
✓✓✓46.157.8
✓✓✓✓55.264.6
✓✓✓✓51.962.2
✓✓✓✓✓59.768.5
✓✓✓✓✓✓62.472.0
✓✓✓✓✓✓62.071.3
✓✓✓✓✓✓✓65.273.9
", + "image_path": "03829d7b4d246c196a13ecc798ace61586db465621937e1fdd19224114a2cbdb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 537, + 535, + 603 + ], + "blocks": [ + { + "bbox": [ + 317, + 517, + 533, + 529 + ], + "lines": [ + { + "bbox": [ + 317, + 517, + 533, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 517, + 533, + 529 + ], + "type": "text", + "content": "Table 4. Comparison with various counterfactual attentions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 537, + 535, + 603 + ], + "lines": [ + { + "bbox": [ + 317, + 537, + 535, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 535, + 603 + ], + "type": "table", + "html": "
Acc@0.25random \nattentionuniform \nattentionreversed \nattentionshuffle \nattention
ScanRefer57.857.456.957.5
Nr3D65.264.864.565.0
Sr3D73.973.373.173.5
", + "image_path": "5b0e79e2873a639f01ed7c33082ace8fa08778cb9d1c75cec7e94ce5546abd7d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 628, + 509, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 628, + 509, + 640 + ], + "spans": [ + { + "bbox": [ + 305, + 628, + 509, + 640 + ], + "type": "text", + "content": "interactions rather than embedding more attitudes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "Effects of Counterfactual Attention. We implement different strategies to generate counterfactual attention including random attention, uniform attention, reversed attention, and shuffle attention, and report their result on " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " in Tab. 4. We see all the counterfactual attention strategies achieve a similar performer gain in three datasets," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17259" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 71, + 145, + 151 + ], + "blocks": [ + { + "bbox": [ + 53, + 106, + 63, + 114 + ], + "lines": [ + { + "bbox": [ + 53, + 106, + 63, + 114 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 63, + 114 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 71, + 145, + 151 + ], + "lines": [ + { + "bbox": [ + 64, + 71, + 145, + 151 + ], + "spans": [ + { + "bbox": [ + 64, + 71, + 145, + 151 + ], + "type": "image", + "image_path": "eb01b054e1463071777d5c89b7ad27b6f3bf07dfb43bab833cf0d11d90923c1e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 149, + 72, + 230, + 151 + ], + "blocks": [ + { + "bbox": [ + 149, + 72, + 230, + 151 + ], + "lines": [ + { + "bbox": [ + 149, + 72, + 230, + 151 + ], + "spans": [ + { + "bbox": [ + 149, + 72, + 230, + 151 + ], + "type": "image", + "image_path": "bcd4b1b6912995469b95a18add85e50e816b24cf8883272caf1d6b9f83bf85a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 234, + 72, + 316, + 151 + ], + "blocks": [ + { + "bbox": [ + 234, + 72, + 316, + 151 + ], + "lines": [ + { + "bbox": [ + 234, + 72, + 316, + 151 + ], + "spans": [ + { + "bbox": [ + 234, + 72, + 316, + 151 + ], + "type": "image", + "image_path": "6e36ef93f4f35540859f5ef5d42ff26e73d643a3d3726d247340d06a65d4ed1e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 400, + 151 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 400, + 151 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 400, + 151 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 400, + 151 + ], + "type": "image", + "image_path": "d3325333eb7e7e9a501c53172f5c701dd892e368cf4fdd2fb3d34fe5c070b173.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 403, + 72, + 540, + 151 + ], + "blocks": [ + { + "bbox": [ + 403, + 72, + 540, + 151 + ], + "lines": [ + { + "bbox": [ + 403, + 72, + 540, + 151 + ], + "spans": [ + { + "bbox": [ + 403, + 72, + 540, + 151 + ], + "type": "image", + "image_path": "d6a85df81ee1e3e12443d290b769e49b548bd2dcd94c7529d3e05a42aea879bf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 64, + 152, + 146, + 233 + ], + "blocks": [ + { + "bbox": [ + 54, + 194, + 64, + 202 + ], + "lines": [ + { + "bbox": [ + 54, + 194, + 64, + 202 + ], + "spans": [ + { + "bbox": [ + 54, + 194, + 64, + 202 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 152, + 146, + 233 + ], + "lines": [ + { + "bbox": [ + 64, + 152, + 146, + 233 + ], + "spans": [ + { + "bbox": [ + 64, + 152, + 146, + 233 + ], + "type": "image", + "image_path": "5b13c49d1476c8c8c0eaf086ff0b8b9ea2c41a76eed082fc1bfc8a107bf14bd0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 237, + 138, + 265 + ], + "lines": [ + { + "bbox": [ + 70, + 237, + 138, + 265 + ], + "spans": [ + { + "bbox": [ + 70, + 237, + 138, + 265 + ], + "type": "text", + "content": "wooden double bookcase filled with books. walking into the room, it is in the right hand most." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 274, + 539, + 286 + ], + "lines": [ + { + "bbox": [ + 51, + 274, + 539, + 286 + ], + "spans": [ + { + "bbox": [ + 51, + 274, + 539, + 286 + ], + "type": "text", + "content": "Figure 4. The visualization results. (a) and (b) present the correct/wrong prediction of " + }, + { + "bbox": [ + 51, + 274, + 539, + 286 + ], + "type": "inline_equation", + "content": "MA^2" + }, + { + "bbox": [ + 51, + 274, + 539, + 286 + ], + "type": "text", + "content": " TransVG with/without attribute interaction." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 149, + 152, + 230, + 233 + ], + "blocks": [ + { + "bbox": [ + 149, + 152, + 230, + 233 + ], + "lines": [ + { + "bbox": [ + 149, + 152, + 230, + 233 + ], + "spans": [ + { + "bbox": [ + 149, + 152, + 230, + 233 + ], + "type": "image", + "image_path": "ffc6755253bfc4a2733baf192c333089b6fe3a0fbea990fed9240db91ad7a46e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 237, + 223, + 265 + ], + "lines": [ + { + "bbox": [ + 156, + 237, + 223, + 265 + ], + "spans": [ + { + "bbox": [ + 156, + 237, + 223, + 265 + ], + "type": "text", + "content": "the trash can is in the corner next to the door, below the whiteboard. it is gray, tall, and narrow." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 230, + 236, + 321, + 265 + ], + "lines": [ + { + "bbox": [ + 230, + 236, + 321, + 265 + ], + "spans": [ + { + "bbox": [ + 230, + 236, + 321, + 265 + ], + "type": "text", + "content": "the black chair is located directly to the right of the left wall off the room. The chair is to the left of the burgundy chair and behind the desk." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 234, + 152, + 316, + 233 + ], + "blocks": [ + { + "bbox": [ + 234, + 152, + 316, + 233 + ], + "lines": [ + { + "bbox": [ + 234, + 152, + 316, + 233 + ], + "spans": [ + { + "bbox": [ + 234, + 152, + 316, + 233 + ], + "type": "image", + "image_path": "4e098f3a4513655970d02a121b8f4839c5f3b11c27cdff957f2c4cccbddcaebf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 152, + 400, + 233 + ], + "blocks": [ + { + "bbox": [ + 318, + 152, + 400, + 233 + ], + "lines": [ + { + "bbox": [ + 318, + 152, + 400, + 233 + ], + "spans": [ + { + "bbox": [ + 318, + 152, + 400, + 233 + ], + "type": "image", + "image_path": "bd8f67419cd5daa5d188337ad67e40f127e7f6f92865dd1ceb4ad532d8257e98.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 236, + 394, + 265 + ], + "lines": [ + { + "bbox": [ + 324, + 236, + 394, + 265 + ], + "spans": [ + { + "bbox": [ + 324, + 236, + 394, + 265 + ], + "type": "text", + "content": "the pillow is on the bed. it is to the left of the lamp, resulting on the pillow that is behind it." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 403, + 152, + 540, + 233 + ], + "blocks": [ + { + "bbox": [ + 403, + 152, + 540, + 233 + ], + "lines": [ + { + "bbox": [ + 403, + 152, + 540, + 233 + ], + "spans": [ + { + "bbox": [ + 403, + 152, + 540, + 233 + ], + "type": "image", + "image_path": "0df27124929018779b595c1f7617e64973d3ffeb3175dafd6ce57e00500e192f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 237, + 529, + 258 + ], + "lines": [ + { + "bbox": [ + 414, + 237, + 529, + 258 + ], + "spans": [ + { + "bbox": [ + 414, + 237, + 529, + 258 + ], + "type": "text", + "content": "this is a black leather loveseat. if you were sitting in it, the long, short bookshelf would be on the right." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 307, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 287, + 354 + ], + "type": "text", + "content": "while random attention can obtain the best accuracy. We think it is because random attention helps try more diverse combinations of attribute interactives and can provide a more effective signal to supervise the attention." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "content": "Effects of Exchanging Operation. We conduct a sensitivity analysis of hyper-parameters in our EMFM module, including: i) the exchange proportion " + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "content": " of tokens. As shown in Table. 5, the model performance increases and then drops with the increase of " + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "content": ", which achieves a peak value of " + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "inline_equation", + "content": "65.2\\%" + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "inline_equation", + "content": "\\theta = 10\\%" + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "content": ". The results show that too small or large " + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "content": " will adversely affect the inter-modal fusion, which may bring inadequate alignment or attenuate the intramodal knowledge. ii) the start layer " + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "content": " and end layer " + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "content": " for multimodal exchanging. Since the default number of layers in the regular Transformer is 6, we first fix the start layer " + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "content": " to 1 and investigate the value of " + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "content": " from 1, 2, 3, 4, 5, where the accuracy improves quickly from " + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "inline_equation", + "content": "65.2\\%" + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "content": ". We then fix the end layer " + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "content": " to 5, where the accuracy decreases when " + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 355, + 288, + 547 + ], + "type": "text", + "content": " increases from 1 to 5. All the results verify the rationality of our hyper-parameters settings." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 552, + 204, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 204, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 204, + 563 + ], + "type": "text", + "content": "4.5. Visualization and Limitation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": "We show the visualization results in Fig. 4. The model may fail to recognize the spatial relationship among objects (the last three columns) or be significantly affected by other more salient objects (the first two columns) without considering the explicit multi-attribute interactions. Our " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "MA^2" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " TransVG can better perceive discriminative clues such as the appearance or spatial correlation of objects for multimodal feature enhancement and accurate grounding (the first row). We also notice the accuracy still remains improvement room for 'View-dependent' descriptions compared to the 'View-independent' or 'Easy', especially in Sr3D (In Tab. 2). In the future, we will extend more at-" + } + ] + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 339, + 325, + 514, + 423 + ], + "blocks": [ + { + "bbox": [ + 323, + 305, + 528, + 317 + ], + "lines": [ + { + "bbox": [ + 323, + 305, + 528, + 317 + ], + "spans": [ + { + "bbox": [ + 323, + 305, + 528, + 317 + ], + "type": "text", + "content": "Table 5. Hyper-parameter Sensitivity Analysis on Nr3D." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 325, + 514, + 423 + ], + "lines": [ + { + "bbox": [ + 339, + 325, + 514, + 423 + ], + "spans": [ + { + "bbox": [ + 339, + 325, + 514, + 423 + ], + "type": "table", + "html": "
θ0%5%10%15%20%
64.464.765.263.961.4
η12345
μ = 162.463.964.865.264.2
μ12345
η = 564.263.362.962.662.4
", + "image_path": "dbaf90b94a46fc5c468d5e92406d8e6f80c8dd2cd7f1a181f900053d7f02ab7f.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 441, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 467 + ], + "type": "text", + "content": "tribute embedding strategies, e.g., better encoding absolute position or aggregating point clouds from various views." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 476, + 378, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 378, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 378, + 489 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "type": "text", + "content": "In this work, we propose an attribute-aware Transformer " + }, + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "type": "inline_equation", + "content": "(\\mathsf{MA}^2\\mathsf{TransVG})" + }, + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "type": "text", + "content": " for 3D object grounding. As the first attempt to quantify the contribution of each attribute for final grounding, " + }, + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "type": "inline_equation", + "content": "\\mathsf{MA}^2\\mathsf{TransVG}" + }, + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "type": "text", + "content": " first designs a newly attribute causal effect analysis module and an attribute loss, which help model the multi-attribute interactions of objects for a better visual/textual modal understanding. Based on the learned multi-attribute interactions, the exchanging-based multimodal fusion module further dynamically aligns more discriminative details between modalities for a fine-grained multimodal feature. The proposed model outperforms the SOTA methods on " + }, + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D / Sr3D}" + }, + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "type": "text", + "content": " and ScanRefer datasets." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 651, + 404, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 404, + 664 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 404, + 664 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 671, + 546, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 671, + 546, + 707 + ], + "spans": [ + { + "bbox": [ + 305, + 671, + 546, + 707 + ], + "type": "text", + "content": "This work was supported by the National Science Fund of China (Grant Nos. 62276144, 62306238) and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17260" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Ahmed Abdelreehem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "type": "text", + "content": "[2] EslamMohamed Bakr, Yasmeen Alsaedy, and Mohamed Elhoseiny. Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. In NeurIPS, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "type": "text", + "content": "[3] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In CVPR, pages 16464-16473, 2022. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 286, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 286, + 267 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 286, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 286, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 286, + 302 + ], + "type": "text", + "content": "[5] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221. Springer, 2020. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 303, + 286, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 286, + 345 + ], + "type": "text", + "content": "[6] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. NeurIPS, 34:5834-5847, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 347, + 286, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 286, + 390 + ], + "type": "text", + "content": "[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. NeurIPS, 35:20522-20535, 2022. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 286, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 286, + 435 + ], + "type": "text", + "content": "[8] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, pages 16537-16547, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 286, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 286, + 468 + ], + "type": "text", + "content": "[9] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In CVPR, pages 11124-11133, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 470, + 286, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 286, + 514 + ], + "type": "text", + "content": "[10] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In ICCV, pages 18109-18119, 2023. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "type": "text", + "content": "[11] Bowen Cheng, Lu Sheng, Shaoshuai Shi, Ming Yang, and Dong Xu. Back-tracing representative points for voting-based 3d object detection in point clouds. In CVPR, pages 8963-8972, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "type": "text", + "content": "[12] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 668 + ], + "type": "text", + "content": "[13] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In CVPR, pages 3722-3731, 2021. 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "text", + "content": "[14] Chen Gao, Xingyu Peng, Mi Yan, He Wang, Lirong Yang, Haibing Ren, Hongsheng Li, and Si Liu. Adaptive zone-aware hierarchical planner for vision-language navigation. In CVPR, pages 14911-14920, 2023. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[15] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yi Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. In CVPR, pages 14773-14783, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "type": "text", + "content": "[16] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In ICCV, pages 15372-15383, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 165, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 165, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 308, + 165, + 545, + 208 + ], + "type": "text", + "content": "[17] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-and-relation aware transformer for fine-grained 3d visual grounding. In ACM MM, pages 2344-2352, 2021. 2, 3, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 210, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 210, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 308, + 210, + 545, + 253 + ], + "type": "text", + "content": "[18] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In AAAI, pages 1610-1618, 2021. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 256, + 545, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 256, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 308, + 256, + 545, + 288 + ], + "type": "text", + "content": "[19] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In CVPR, pages 15524-15533, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 291, + 545, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 291, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 308, + 291, + 545, + 334 + ], + "type": "text", + "content": "[20] Kenton Lee Jacob Devlin, Ming-Wei Chang and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2, 2019. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 336, + 545, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 336, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 308, + 336, + 545, + 379 + ], + "type": "text", + "content": "[21] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Looking outside the box to ground language in 3d scenes. arXiv preprint arXiv:2112.08879, 2021. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "type": "text", + "content": "[22] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, pages 417-433. Springer, 2022. 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 428, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 428, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 545, + 449 + ], + "type": "text", + "content": "[23] Nathan Kallus and Angela Zhou. Confounding-robust policy improvement. NeurIPS, 31, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 452, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 452, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 308, + 452, + 545, + 495 + ], + "type": "text", + "content": "[24] Murat Kocaoglu, Christopher Snyder, Alexandros G Dimakis, and Sriram Vishwanath. Causalgan: Learning causal implicit generative models with adversarial training. In ICLR, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 497, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 497, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 308, + 497, + 545, + 540 + ], + "type": "text", + "content": "[25] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 543, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 543, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 308, + 543, + 545, + 597 + ], + "type": "text", + "content": "[26] Haojia Lin, Yongdong Luo, Xiawu Zheng, Lijiang Li, Fei Chao, Taisong Jin, Donghao Luo, Chengjie Wang, Yan Wang, and Liujuan Cao. A unified framework for 3d point cloud visual grounding. arXiv preprint arXiv:2308.11887, 2023. 2, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 600, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 600, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 600, + 545, + 643 + ], + "type": "text", + "content": "[27] Haolin Liu, Anran Lin, Xiaoguang Han, Lei Yang, Yizhou Yu, and Shuguang Cui. Refer-it-in-rgbd: A bottom-up approach for 3d visual grounding in rgb images. In CVPR, pages 6032-6041, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "type": "text", + "content": "[28] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In ICCV, pages 10968-10980, 2023. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[29] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, pages 2949-2958, 2021. 1, 3, 5" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17261" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[30] David Lopez-Paz, Robert Nishihara, Soumith Chintala, Bernhard Scholkopf, and Léon Bottou. Discovering causal signals in images. In CVPR, pages 6979-6987, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 286, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 286, + 149 + ], + "type": "text", + "content": "[31] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16454-16463, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 286, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 286, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 286, + 172 + ], + "type": "text", + "content": "[32] Judea Pearl. Interpretation and identification of causal mediation. Psychological methods, 19(4):459, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 172, + 286, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 286, + 204 + ], + "type": "text", + "content": "[33] Judea Pearl. Direct and indirect effects. In Probabilistic and causal inference: the works of Judea Pearl, pages 373-392. 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 205, + 286, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 286, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 286, + 248 + ], + "type": "text", + "content": "[34] Stephen Powell. The book of why: The new science of cause and effect. pearl, judea, and dana mackenzie. 2018. hachette uk. Journal of MultiDisciplinary Evaluation, 14(31):47-54, 2018. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 250, + 286, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 286, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 286, + 303 + ], + "type": "text", + "content": "[35] Mihir Prabhudesai, Hsiao-Yu Fish Tung, Syed Ashar Javed, Maximilian Sieb, Adam W Harley, and Katerina Fragkiadaki. Embodied language grounding with 3d visual feature representations. In CVPR, pages 2220-2229, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 304, + 286, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 286, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 286, + 338 + ], + "type": "text", + "content": "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 338, + 286, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 286, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 286, + 370 + ], + "type": "text", + "content": "[37] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 371, + 286, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 286, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 286, + 414 + ], + "type": "text", + "content": "[38] Yongming Rao, Guangyi Chen, Jiwen Lu, and Jie Zhou. Counterfactual attention learning for fine-grained visual categorization and re-identification. In ICCV, pages 1025–1034, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 415, + 286, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 286, + 458 + ], + "type": "text", + "content": "[39] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In Conference on Robot Learning, pages 1046-1056. PMLR, 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 459, + 286, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 286, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 286, + 492 + ], + "type": "text", + "content": "[40] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In CVPR, pages 3716-3725, 2020. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 492, + 286, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 286, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 286, + 536 + ], + "type": "text", + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurlPS, 2022. 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 537, + 286, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 286, + 590 + ], + "type": "text", + "content": "[42] Hongjun Wang, Jiyuan Chen, Lun Du, Qiang Fu, Shi Han, and Xuan Song. Causal-based supervision of attention in graph neural network: A better and simpler choice towards powerful attention. arXiv preprint arXiv:2305.13115, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "type": "text", + "content": "[43] Hanqing Wang, Wei Liang, Luc Van Gool, and Wenguan Wang. Dreamwalker: Mental planning for continuous vision-language navigation. In ICCV, pages 10873-10883, 2023. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 286, + 667 + ], + "type": "text", + "content": "[44] Tan Wang, Jianqiang Huang, Hanwang Zhang, and Qianru Sun. Visual commonsense r-cnn. In CVPR, pages 10760-10770, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "type": "text", + "content": "[45] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In CVPR, pages 19231-19242, 2023. 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 407 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[46] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[47] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In CVPR, pages 1791-1800, 2021. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 162, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 206 + ], + "type": "text", + "content": "[48] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Shuguang Cui, and Zhen Li. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In CVPR, pages 8563-8573, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "type": "text", + "content": "[49] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, pages 15225-15236, 2023. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[50] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2928-2937, 2021. 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 274, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 319 + ], + "type": "text", + "content": "[51] Tong Zhao, Gang Liu, Daheng Wang, Wenhao Yu, and Meng Jiang. Learning from counterfactual links for link prediction. In Int. Conf. on Machine Learning, pages 26911-26926. PMLR, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "type": "text", + "content": "[52] Dave Zhenyu Chen, Qirui Wu, Matthias Niener, and Angel X Chang. D3net: A speaker-listener architecture for semi-supervised dense captionin and visual grounding in rgb-d scans. In ECCV, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 364, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 364, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 364, + 545, + 407 + ], + "type": "text", + "content": "[53] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17262" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/10ec9eda-ceb6-4b88-aa59-8a3b78a66785_content_list.json b/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/10ec9eda-ceb6-4b88-aa59-8a3b78a66785_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e0c42b29fed04349d36ab4512f39a905af512086 --- /dev/null +++ b/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/10ec9eda-ceb6-4b88-aa59-8a3b78a66785_content_list.json @@ -0,0 +1,1794 @@ +[ + { + "type": "text", + "text": "Multi-Level Neural Scene Graphs for Dynamic Urban Environments", + "text_level": 1, + "bbox": [ + 138, + 130, + 828, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tobias Fischer $^{1}$ Lorenzo Porzi $^{2}$ Samuel Rota Bulò $^{2}$", + "bbox": [ + 238, + 179, + 723, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Marc Pollefeys1 Peter Kontschieder2", + "bbox": [ + 316, + 198, + 647, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ETH Zürich $^{2}$ Meta Reality Labs", + "bbox": [ + 341, + 217, + 633, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://tobiasfshr.github.io/pub/ml-nsg/", + "bbox": [ + 300, + 239, + 658, + 253 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c1089944d4ef1d7a751ce36bcdde3419dad0203b973446f27dba488ce7b93109.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 270, + 151, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4efd39ef8f453b62cd2fbfa0f7f83c46b1af373aa59ecbbef8d58a61b8cd68d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 340, + 151, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ec4d11e7f858af17263a92b2c31121fe8384c43b33c6c751909a21267269ab1.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 81, + 407, + 151, + 478 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3ae6cb64d0fb026a45d3dd3cb46acf62d0c1bf05edb926c42ac5bf9d603bff6.jpg", + "image_caption": [ + "Multi-Level Neural Scene Graph" + ], + "image_footnote": [], + "bbox": [ + 160, + 275, + 480, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/27fe8028c93540aed69f136ab15fb7fdd7312b9260ea1028764672189019f51d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 270, + 620, + 372 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f686443d3ca75b8d9970d7c507f8c8aedd2aeaa29415ca17eadac66cb094e41d.jpg", + "image_caption": [ + "Figure 1. Overview. We represent sequences captured from moving vehicles in a shared geographic area with a multi-level scene graph. Each dynamic object $v_{o}$ is associated with a sequence node $v_{s}^{t}$ and time $t$ . The sequence nodes are registered in a common world frame at the root node $v_{r}$ through the vehicle poses $\\mathbf{P}_{s}^{t}$ , while the dynamic objects are localized w.r.t. the sequence node with pose $\\xi_{o}^{t}$ . Each camera $c$ is associated with an ego-vehicle position, i.e. node $v_{s}^{t}$ , through the extrinsic calibration $\\mathbf{T}_{c}$ . The sequence and object nodes hold latent codes $\\omega$ that condition the radiance field, synthesizing novel views in various conditions with distinct dynamic objects." + ], + "image_footnote": [], + "bbox": [ + 488, + 372, + 620, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/abacc3bd5b0ff214316c63989fb099052008ddc35f79c1b74cbd6cdd9b6697f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 270, + 753, + 372 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1b6b993264d94d48c329b911f10c0b384bb711a871081124b7777fcde8ebc5c2.jpg", + "image_caption": [ + "Novel Views" + ], + "image_footnote": [], + "bbox": [ + 622, + 372, + 754, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a25c0b68ad006e12ee21a69730b8a96cd8a93d1e8e69a017082c994e298e9c1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 270, + 887, + 372 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0992a2a06b4105a7447221a43e570a7bc0cc34ec362b568072beed7e664700f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 372, + 888, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 589, + 310, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We estimate the radiance field of large-scale dynamic areas from multiple vehicle captures under varying environmental conditions. Previous works in this domain are either restricted to static environments, do not scale to more than a single short video, or struggle to separately represent dynamic object instances. To this end, we present a novel, decomposable radiance field approach for dynamic urban environments. We propose a multi-level neural scene graph representation that scales to thousands of images from dozens of sequences with hundreds of fast-moving objects. To enable efficient training and rendering of our representation, we develop a fast composite ray sampling and rendering scheme. To test our approach in urban driving scenarios, we introduce a new, novel view synthesis benchmark. We show that our approach outperforms prior art by a significant margin on both established and our proposed benchmark while being faster in training and rendering.", + "bbox": [ + 73, + 628, + 470, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 589, + 630, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating the radiance field of a dynamic urban environment from data collected by sensor-equipped vehicles is a core challenge in closed-loop simulation for robotics and in mixed reality. It is particularly relevant for applications like autonomous driving and city-scale mapping, where capture vehicles can be frequently deployed. The growing amount of available data provides great opportunities for creating up-to-date digital twins of entire cities but also poses unique challenges as increasingly heterogeneous data sources must be processed. In particular, limited scene coverage, different lighting, weather, seasonal conditions, and varying geometry due to distinct dynamic and transient objects make radiance field reconstruction of dynamic urban environments extremely challenging.", + "bbox": [ + 496, + 614, + 890, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Neural Radiance Fields (NeRFs) have enabled significant progress in achieving realistic novel view synthesis from a set of input views [2, 31-33, 49]. These methods represent a static 3D scene with fully implicit [30, 31, 49] or low-level explicit structures [10, 13, 16, 23, 32, 43,", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21125", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "52, 65] such as voxels, points, surfels, meshes or 3D gaus-. sians. In parallel, implicit [19, 37, 50, 62, 63] and low-level explicit [9, 27, 36, 38] neural representations for dynamic 4D scenes have been investigated. However, fewer works have focused on decomposing scenes into higher-level entities [18, 26, 46, 56]. In computer graphics and 3D mapping, scene graphs have been used to represent complex scenes in a multi-level hierarchy [1, 44, 51]. In view synthesis, Ost et al. [35] apply this concept by describing actors in a dynamic scene as entities of a scene graph. However, their representation lacks the ability of [1, 44] to represent scenes at multiple levels, thereby limiting their representation to short, single sequences.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While earlier methods for view synthesis focused on object-centric scenes with controlled camera trajectories, recent works move towards radiance field reconstruction of large-scale environments from in-the-wild captures. Among these, many methods focus on static environments, removing dynamic actors from the input data [22, 24, 29, 42, 45, 53, 57]. A few works explicitly model dynamic actors [18, 35, 67], but either do not scale to more than a single, short video [18, 35, 67], or struggle to accurately represent individual object instances [58]. This complicates the evaluation of these methods for real-world applications because existing benchmarks neither scale to large urban areas [18, 35] nor reflect realistic capturing conditions [12].", + "bbox": [ + 75, + 289, + 472, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose a multi-level neural scene graph representation that spans large geographic areas with hundreds of dynamic objects. In contrast to previous works, our multi-level scene graph formulation allows us to distinguish and represent dynamic object instances effectively and further to represent a scene under varying conditions. To make our representation viable for large-scale dynamic environments, we develop a composite ray sampling and rendering scheme that enables fast training and rendering of our method. To test this hypothesis, we introduce a benchmark for radiance field reconstruction in dynamic urban environments based on [61]. We fuse data from dozens of vehicle captures under varying conditions amounting to more than ten thousand images with several hundreds of dynamic objects per reconstructed area. We summarize our contributions as follows:", + "bbox": [ + 75, + 488, + 472, + 728 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a multi-level neural scene graph formulation that scales to dozens of sequences with hundreds of fast-moving objects under varying environmental conditions.", + "- We develop an efficient composite ray sampling and rendering scheme that enables fast training and rendering of our representation.", + "- We present a benchmark that provides a realistic, application-driven evaluation of radiance field reconstruction in dynamic urban environments." + ], + "bbox": [ + 76, + 732, + 467, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show state-of-the-art view synthesis results on both established benchmarks [4, 14] and our proposed benchmark.", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D and 4D scene representations. Finding the right scene representation is a core issue in 3D computer vision and graphics [5]. Over the years, a wide variety of options have been explored [1, 3, 8, 15, 25, 30, 40, 46-48], and, more recently, neural rendering [33] has been used to enable a new generation of scene models that support photo-realistic novel view synthesis. Scene representations for neural rendering can be roughly classified as \"implicit\" [30, 31, 49], which store most of the information in the weights of a neural network, or \"explicit\", which use low-level spatial primitives such as voxels [10, 23, 32, 52], points [65], surfels [13], meshes [43], or 3D gaussians [16].", + "bbox": [ + 496, + 114, + 890, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Similarly, different 4D dynamic scene representations for view synthesis have been investigated, including both implicit [11, 19, 37, 41, 55, 58, 62, 63] and explicit [9, 27, 36, 67] approaches. Dynamics are generally modeled as deformations of a canonical volume [9, 37, 41, 55, 62], a separate scene motion function [11, 19, 20, 58, 63], or rigid transformations of local geometric primitives [27].", + "bbox": [ + 496, + 296, + 890, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another line of work investigates the decomposition of scenes into higher-level entities [18, 26, 46, 56]. To express the composition of different entities into a complex scene, classical computer graphics literature [7] uses scene graphs. In particular, entities are represented as nodes in a hierarchical graph and are connected through edges defined by coordinate frame transformations. Thus, global transformations can be acquired by traversing the graph from its root node. For indoor 3D mapping, this concept was proposed by Armeni et al. [1] to represent static scenes at multiple levels of hierarchy, i.e. buildings, rooms and objects, and was later extended to dynamic scenes by Rosinol et al. [44]. Recently, Ost et al. [35] have revisited this concept for view synthesis, describing multi-object scenes with a graph that represents the dynamic actors in the scene. However, their representation lacks the ability of [1, 44] to represent scenes at multiple levels of hierarchy and is thus inherently limited to single, short video clips. On the contrary, we present a scalable, multi-level scene graph representation that spans large geographic areas with hundreds of dynamic objects.", + "bbox": [ + 496, + 402, + 892, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Representing large-scale urban scenes. Compared to controlled captures of small scenes, in-the-wild captures of large-scale scenes pose distinct challenges. Limited viewpoint coverage, inaccurate camera poses, far-away buildings and sky, fast-moving objects, complex lighting, and auto exposure make radiance field reconstruction challenging. Therefore, previous works aid the reconstruction by using depth priors from e.g. LiDAR, refining camera parameters, adding information about camera exposure, and using specialized sky and light modeling components [22, 29, 42, 45, 53, 60, 64]. While many of these works simply remove dynamic actors, some methods explicitly model scene dynamics [18, 35, 58, 67]. However,", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21126", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cf6ac8ee88fa08ca1d431f0e43f8a28257acf473308ae33e24cb79da8e1b5dde.jpg", + "image_caption": [ + "Figure 2. Ego-vehicle trajectories of our benchmark. We show the residential (left) and the downtown (right) areas, trajectories superimposed on 2D maps obtained from OpenStreetMap [34]." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "these methods either struggle to scale to more than a single sequence [18, 35, 67] or to accurately represent and distinguish dynamic actors [58]. To address these issues, we present a decomposed, scalable scene representation for dynamic urban environments.", + "bbox": [ + 75, + 286, + 467, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient scene rendering. Especially for large-scale scenes, the efficiency of a scene representation in training and inference is crucial. To alleviate the burden of ray traversal in volumetric rendering, many techniques for efficient sampling [2, 10, 32, 57] have been proposed. Recently, researchers have exploited more efficient forms of rendering, e.g. rasterization of meshes [22, 24] or other 3D primitives [16]. While these approaches are focused on static scenes, we extend the approach of [2] to our scene graph representation and thus to dynamic scenes.", + "bbox": [ + 75, + 362, + 467, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Data", + "text_level": 1, + "bbox": [ + 76, + 527, + 140, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We investigate how to synthesize novel views of a dynamic urban environment from a set of captures taken from moving vehicles. Specifically, we are interested in captures spanning the same geographic area under varying conditions. This data presents unique challenges due to limited viewpoint coverage, different lighting, weather, and season, and even varying geometry due to distinct dynamic and transient objects.", + "bbox": [ + 75, + 551, + 467, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous benchmarks in this area [18, 35] are limited to short, single video clips with simple ego trajectories and a few dynamic objects. Therefore, we present a benchmark that better reflects the aforementioned challenges. We base our benchmark on Argoverse 2 [61] which provides a rich set of captures from a fleet of vehicles deployed in multiple US cities that span different weather, season, and time of day. The vehicles are equipped with a surround-view camera rig with seven global shutter cameras, a LiDAR sensor, and a GPS. Furthermore, the timings of the different sensors are provided, so that one can relate the LiDAR-based 3D bounding box annotations to camera timestamps.", + "bbox": [ + 75, + 674, + 467, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We leverage the GPS information to globally align the sequences and to identify regions that we are interested in mapping. We then extract the specific vehicle captures.", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/735ace0c8070dbc357097038deff46d0f5361f490277175df7f5127c9df5141f.jpg", + "image_caption": [ + "Figure 3. Sequence alignment visualization. The initial GPS-based alignment is imprecise, as evidenced by the duplicated structures in the overlaid LiDAR point clouds (left). After our ICP alignment, the area is well reconstructed (middle) according to its real geometry (right, from Argoverse 2 [61])." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since GPS-based localization accuracy is only coarsely precise in urban areas, we align the captures via a global, offline iterative-closest-point (ICP) procedure applied to the LiDAR point clouds of all sequences to achieve satisfactory alignment for novel view synthesis purposes (see Fig. 3).", + "bbox": [ + 496, + 285, + 890, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following this procedure, we build a benchmark that enables real-world evaluation of novel view synthesis from diverse vehicle captures. It is composed of 37 vehicle captures split into two geographic regions as illustrated in Fig. 2. The regions cover a residential and a downtown area to resemble the different characteristics of urban environments. The residential area spans 14 captures with 10493 training and 1162 testing images with more than 700 distinct moving objects. The downtown area spans 23 captures with 16933 training and 1876 testing images with more than 600 distinct moving objects.", + "bbox": [ + 496, + 361, + 890, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 537, + 589, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem setup. We are given a set of sequences $S$ captured from moving vehicles in different conditions. Each sequence $s \\in S$ consists of a set of images taken from cameras $C_s$ mounted on the vehicle at different timesteps indexed by $T_s$ . We assume the vehicle's sensors are calibrated with respect to a common vehicle and a global world frame. In particular, we assume given sensor extrinsic $\\mathbf{T}_c = [\\mathbf{R}_c|\\mathbf{t}_c] \\in \\mathbf{SE}(3)$ for each camera $c \\in C = \\bigcup_{s} C_s$ , and ego-vehicle poses $\\mathbf{P}_s^t = [\\mathbf{R}_s^t|\\mathbf{t}_s^t] \\in \\mathbf{SE}(3)$ for each timestamp $t \\in T_s$ and sequence $s \\in S$ . Furthermore, we assume all cameras to have known intrinsics $\\mathbf{K}_c$ . Each sequence $s$ entails a set $O_s$ of dynamic objects. For each object $o \\in O_s$ , we assume to have an estimated 3D bounding box track $\\mathcal{T}_o$ consisting of object poses $\\{\\xi_o^{t_0}, \\dots, \\xi_o^{t_n}\\} \\subset \\mathbf{SE}(3)$ w.r.t. the ego-vehicle frame, where $t_i \\in T_s$ , and the 3D object dimensions $\\mathbf{s}_o \\in \\mathbb{R}_{>0}^3$ . The images span a common geographic area for which we would like to estimate a radiance field", + "bbox": [ + 496, + 564, + 890, + 820 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} (\\mathbf {x}, \\mathbf {d}, t, s) = (\\sigma (\\mathbf {x}, t, s), \\mathbf {c} (\\mathbf {x}, \\mathbf {d}, t, s)) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 829, + 890, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that outputs volume density $\\sigma \\in \\mathbb{R}_{\\geq 0}$ and color $\\mathbf{c} \\in [0,1]^3$ conditioned on 3D location $\\mathbf{x}$ , viewing direction $\\mathbf{d}$ , time $t$ and sequence $s$ .", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21127", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Multi-Level Neural Scene Graph", + "text_level": 1, + "bbox": [ + 76, + 90, + 366, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. We illustrate our representation in Fig. 1. We decompose the scene into a graph $\\mathcal{G} = (\\mathcal{V},\\mathcal{E})$ where the set of nodes $\\nu$ is composed by a root node $v_{r}$ that defines the global coordinate system, camera nodes $\\{v_{c}\\}_{c\\in C}$ , and, for each sequence $s\\in S$ , sequence nodes $\\{v_s^t\\}_{t\\in T_s}$ and dynamic object nodes $\\{v_o\\}_{o\\in O_s}$ . The nodes are connected through oriented edges $e\\in \\mathcal{E}$ that represent rigid transformations between the coordinate frames of the nodes, consistently with the edge direction. In addition, each node $v\\in \\mathcal{V}$ can have an associated latent vector $\\omega$ . Given the scene graph $\\mathcal{G}$ , we model $f_{\\theta}$ using two radiance fields, namely $\\phi$ for largely static and $\\psi$ for highly dynamic scene parts. We use the latent vectors $\\omega$ to condition the radiance fields. In particular, we use latent vectors $\\omega_s^t$ of a sequence node $v_s^t$ to condition $\\phi$ :", + "bbox": [ + 75, + 114, + 473, + 340 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (\\mathbf {x}, \\mathbf {d}, \\omega_ {s} ^ {t}) = \\left(\\sigma_ {\\phi} \\left(\\mathbf {x}, \\omega_ {s} ^ {t}\\right), \\mathbf {c} _ {\\phi} \\left(\\mathbf {x}, \\mathbf {d}, \\omega_ {s} ^ {t}\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 354, + 468, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the graph $\\mathcal{G}$ has multiple levels, a node associated with a dynamic object will naturally fall into the sequence $s$ it appears in. Therefore, the radiance field $\\psi$ is conditioned on latent vectors $\\omega_{s}^{t}$ and $\\omega_{o}$ :", + "bbox": [ + 75, + 386, + 470, + 446 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\psi (\\mathbf {x}, \\mathbf {d}, \\omega_ {s} ^ {t}, \\omega_ {o}) = (\\sigma_ {\\psi} (\\mathbf {x}, \\omega_ {s} ^ {t}, \\omega_ {o}), \\mathbf {c} _ {\\psi} (\\mathbf {x}, \\mathbf {d}, \\omega_ {s} ^ {t}, \\omega_ {o})) (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 460, + 468, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the corresponding sequence and object nodes $v_{s}^{t}$ and $v_{o}$ . Appearance and geometry variation. Reconstructing an environment from multiple captures is challenging from two perspectives: in addition to varying dynamic objects, there is i) varying appearance across captures, and ii) slow-moving or static transient geometry such as tree leaves or construction sites. Since both transient geometry and appearance can vary across captures, but are usually smooth within a sequence, we model these phenomena as smooth functions over time", + "bbox": [ + 75, + 491, + 467, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {s} ^ {t} = \\left[ \\mathbf {A} _ {s} \\mathcal {F} (t), \\mathbf {G} _ {s} \\mathcal {F} (t) \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 657, + 468, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{A}_s$ is an appearance matrix, $\\mathbf{G}_s$ is a transient geometry matrix and $\\mathcal{F}(\\cdot)$ is a 1D basis function of sines and cosines with linearly increasing frequencies at log-scale [58, 66]. We normalize time $t$ into the interval $[-1,1]$ using the maximum sequence length $\\max_{s\\in S}|T_s|$ . Crucially, this allows to model near-static, but sequence-specific regions of the input, as well as appearance changes due to e.g. auto-exposure. The degree of variation across time can be controlled by the number of frequencies.", + "bbox": [ + 75, + 686, + 468, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sequence nodes. The sequence nodes $v_{s}^{t}$ are connected to the root node $v_{r}$ with an edge $e_{v_s^t v_r} = \\mathbf{P}_s^t$ , i.e. the sequences $S$ share a common global world frame. Each sequence node holds the latent vector $\\omega_{s}^{t}$ that conditions the radiance field $\\phi$ . We model $\\phi$ with a multi-scale 3D hash grid representa", + "bbox": [ + 75, + 824, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dc40757d267ad7d75e2719c0dcc3eaf408b19005a456657fd20d0e438ccfa412.jpg", + "image_caption": [ + "Figure 4. Modifying car appearance with scene appearance. We exchange $\\omega_{s}^{t}$ for different car instances. The car's appearance in a rendered, car-centric view (top) changes according to the environmental conditions visible in the sequence $s$ (bottom)." + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 893, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion [32] and lightweight MLP heads:", + "bbox": [ + 500, + 345, + 751, + 361 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\mathbf {x}} = \\mathrm {H} _ {3} (\\mathbf {x}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 736, + 369, + 890, + 386 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {\\sigma_ {\\phi}}, \\sigma_ {\\phi} = \\operatorname {M L P} _ {\\sigma_ {\\phi}} \\left(\\mathbf {f} _ {\\mathbf {x}}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 388, + 890, + 406 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {\\phi} = \\operatorname {M L P} _ {\\mathbf {c} _ {\\phi}} \\left(\\mathbf {h} _ {\\sigma_ {\\phi}}, \\gamma_ {\\mathrm {S H}} (\\mathbf {d}), \\mathbf {A} _ {s} \\mathcal {F} (t)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 407, + 890, + 425 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {\\mathbf {G}}, \\mathbf {c} _ {\\mathbf {G}} = \\operatorname {M L P} _ {\\mathbf {G}} \\left(\\mathbf {h} _ {\\sigma_ {\\phi}}, \\mathbf {G} _ {s} \\mathcal {F} (t)\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 426, + 890, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma_{\\mathrm{SH}}(\\cdot)$ is a spherical harmonics encoding [32]. The final colors and densities are computed as a mixture of static and transient output (analogous to Eq. 12, see supp. mat.).", + "bbox": [ + 498, + 452, + 890, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic nodes. The dynamic nodes $v_{o}$ are connected to the sequence nodes $v_{s}^{t}$ with edges $e_{v_{o}v_{s}^{t}} = \\xi_{o}^{t}$ . The dynamic node $v_{o}$ associated with object $o$ holds a latent vector $\\omega_{o}$ that conditions the radiance field $\\psi$ . We model $\\psi$ following [31] with an MLP conditioned on $\\omega_{o}$ to represent different instances with the same network [35, 67]", + "bbox": [ + 498, + 497, + 890, + 587 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {\\sigma_ {\\psi}}, \\sigma_ {\\psi} = \\operatorname {M L P} _ {\\sigma_ {\\psi}} \\left(\\gamma_ {\\mathrm {P E}} (\\mathbf {x}), \\omega_ {o}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 597, + 890, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {\\psi} = \\operatorname {M L P} _ {\\mathbf {c} _ {\\psi}} \\left(\\mathbf {h} _ {\\sigma_ {\\psi}}, \\gamma_ {\\mathrm {P E}} (\\mathbf {d}), \\omega_ {o}, \\omega_ {s} ^ {t}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 617, + 890, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma_{\\mathrm{PE}}(\\cdot)$ is a positional encoding [31]. Note that 3D position $\\mathbf{x}$ and viewing direction $\\mathbf{d}$ are transformed into the local object coordinate frame with $(\\mathbf{P}_s^t\\xi_o^t)^{-1}\\mathbf{I}_3(1 / \\max (\\mathbf{s}_o))$ . We condition $\\psi$ on both the scene and object-dependent latent vectors. This allows us to disentangle scene-dependent appearance from the object texture and thus to transfer objects across sequences. We illustrate this process in Fig. 4.", + "bbox": [ + 496, + 642, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camera nodes. The cameras $c \\in C_s$ are connected to the sequence nodes $v_s^t$ through edges $e_{v_c v_s^t} = \\mathbf{T}_c$ , i.e. the calibration of camera $c$ w.r.t. the ego-vehicle frame. This way, we can tie camera poses to a specific ego-vehicle pose.", + "bbox": [ + 496, + 750, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Scene Graph Rendering", + "text_level": 1, + "bbox": [ + 500, + 816, + 720, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We describe how we render our scene graph $\\mathcal{G}$ for a given set of rays $\\mathcal{R}$ . The sampling locations along a ray $(\\mathbf{r}, t, s) \\in \\mathcal{R}$ at time $t$ in sequence $s$ are defined as $\\mathbf{r}(u) = \\mathbf{o} + u\\mathbf{d}$ with $\\mathbf{o} = \\mathbf{R}_s^t\\mathbf{t}_c + \\mathbf{t}_s^t$ and $\\mathbf{d} = \\mathbf{R}_s^t\\mathbf{R}_c\\mathbf{K}_c^{-1}(p_x, p_y, 1)^T$ .", + "bbox": [ + 498, + 839, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21128", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Continuous-time pose. In order to realistically render videos at different frame rates, we treat the dynamic object poses $\\{\\xi_o^{t_0},\\dots,\\xi_o^{t_n}\\}$ as a continuous function of time $\\xi_o(t)$ . We compute $\\xi_o(t)$ by interpolating between the two nearest poses at $t_a\\leq t < t_b$ to time $t$ . This allows us also to synchronize estimated object poses originating from the LiDAR measurements with the camera timestamps.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ray-node intersection. To render the dynamic nodes, we measure the intersection of their 3D bounding boxes with each $(\\mathbf{r},t,s)\\in \\mathcal{R}$ . In particular, given the sequence $s$ and time $t$ the ray $\\mathbf{r}$ is associated with, we first traverse the graph $\\mathcal{G}$ to retrieve all relevant nodes $v_{o}$ and their 3D bounding boxes $\\mathbf{b}_o^t = [\\xi_o(t),\\mathbf{s}_o]$ at time $t$ . Then, we transform $\\mathbf{r}$ into each local node coordinate system and subsequently use AABB-ray intersection [28] to compute the entry and exit locations $u_{o}^{\\mathrm{in}},u_{o}^{\\mathrm{out}}$ along ray $\\mathbf{r}$ .", + "bbox": [ + 75, + 196, + 468, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Composite rendering. To render the color $\\hat{\\mathbf{C}}$ of ray $\\mathbf{r}$ in sequence $s$ at time $t$ , we use volumetric rendering [31]", + "bbox": [ + 76, + 332, + 468, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {C}} (\\mathbf {r}, t, s) = \\int_ {u _ {n}} ^ {u _ {f}} U (u) \\sigma (\\mathbf {r} (u), t, s) \\mathbf {c} (\\mathbf {r} (u), \\mathbf {d}, t, s) d u\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 375, + 450, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} U (u) = \\exp \\left(- \\int_ {u _ {n}} ^ {u} \\sigma \\left(u ^ {\\prime}\\right) d u ^ {\\prime}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 410, + 468, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We obtain density $\\sigma$ and color $\\mathbf{c}$ at sampling location $\\mathbf{r}(u)$ as mixture of the radiance fields $\\phi$ and $\\psi$", + "bbox": [ + 76, + 455, + 467, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma = \\sigma_ {\\phi} + \\sigma_ {\\psi}, \\mathbf {c} = \\frac {\\sigma_ {\\phi}}{\\sigma_ {\\phi} + \\sigma_ {\\psi}} \\mathbf {c} _ {\\phi} + \\frac {\\sigma_ {\\psi}}{\\sigma_ {\\phi} + \\sigma_ {\\psi}} \\mathbf {c} _ {\\psi}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 496, + 468, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Crucially, we set $\\sigma_{\\psi}$ to zero when $\\mathbf{r}(u)$ does not lie within a 3D bounding box of a dynamic node given the calculated entry and exit points $u^{\\mathrm{in}}, u^{\\mathrm{out}}$ .", + "bbox": [ + 75, + 537, + 468, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Composite ray sampling. Instead of densely sampling the space [18, 42] or leveraging separate ray sampling mechanisms per node [35], we use a composite ray sampling strategy illustrated in Fig. 5. We extend the proposal sampling mechanism introduced in [2] by joint sampling from a computationally efficient density field $\\sigma_{\\mathrm{prop}}$ and dynamic nodes $v_{o}$ at time $t$ . In particular, as in Eq. 12, we represent $\\sigma (\\mathbf{r}(u),t,s)$ by a mixture of $\\sigma_{\\mathrm{prop}}(\\mathbf{r}(u),\\omega_s^t)$ and $\\sigma_{\\psi}(\\mathbf{r}(u),\\omega_s^t,\\omega_o)$ . However, analogous to our rendering step, we constrain sampling from $\\sigma_{\\psi}$ to $[u^{\\mathrm{in}},u^{\\mathrm{out}}]$ . This allows us to skip empty space efficiently by distilling the static density $\\sigma_{\\phi}$ into $\\sigma_{\\mathrm{prop}}$ . At the same time, sparsely querying $\\sigma_{\\psi}$ when $\\mathbf{r}(u)$ falls into a dynamic node enables us to still accurately represent the full, dynamic $\\sigma$ .", + "bbox": [ + 75, + 583, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since there are only few samples that fall into dynamic nodes when performing uniform sampling, the computational overhead in the first ray sampling iteration is negligible. In the second iteration, we apply inverse transform sampling given the CDF $F(u) = 1 - U(u)$ along ray $\\mathbf{r}$ and thus the samples are concentrated at the first surface intersection. Hence, only for rays that fall in the line of sight of", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f6d9ca72be293397534baac9abb3b55225300c188e7f02bca0cef30689b151ac.jpg", + "image_caption": [ + "Figure 5. Composite ray sampling. If a ray intersects with an object $v_{o}$ , we sample from both proposal network $\\sigma_{\\mathrm{prop}}$ and radiance field $\\psi$ , and $\\sigma_{\\mathrm{prop}}$ otherwise. We condition each with the latents $\\omega$ of the respective nodes. The PDF is a mixture of all node densities that intersect with the ray. The transmittance $U$ drops at the first surface intersection (tree) where further samples will concentrate." + ], + "image_footnote": [], + "bbox": [ + 526, + 87, + 736, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5c4d60f029ed5277f343137817a12ee966bebe9263f7388265a021dda8eadf0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 88, + 870, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "an object surface will we sample $\\sigma_{\\psi}$ more than a few times. In total, we use two ray sampling iterations as in [54].", + "bbox": [ + 496, + 306, + 888, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Space contraction. Following [2, 54], we contract the unbounded scene space into a fixed-size bounding box, normalizing the scene with bounds computed from the LiDAR point clouds and the ego-vehicle poses $\\mathbf{P}$ into a unit cube.", + "bbox": [ + 496, + 338, + 890, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Optimization", + "text_level": 1, + "bbox": [ + 500, + 407, + 635, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We optimize the parameters $\\theta$ of the radiance field $f_{\\theta}$ with", + "bbox": [ + 500, + 429, + 883, + 445 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {(\\mathbf {r}, t, s) \\in \\mathcal {R}} \\mathcal {L} _ {\\mathrm {r g b}} (\\mathbf {r}, t, s) + \\lambda_ {\\mathrm {d i s t}} \\mathcal {L} _ {\\mathrm {d i s t}} (\\mathbf {r}, t, s) + \\lambda_ {\\mathrm {p r o p}}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 453, + 857, + 486 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r o p}} (\\mathbf {r}, t, s) + \\lambda_ {\\text {d e p}} \\mathcal {L} _ {\\text {d e p}} (\\mathbf {r}, t, s) + \\lambda_ {\\text {e n t r}} \\mathcal {L} _ {\\text {e n t r}} (\\mathbf {r}, t, s) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 489, + 890, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{dist}}$ and $\\mathcal{L}_{\\mathrm{prop}}$ follow [2]. We supervise $\\sigma_{\\mathrm{prop}}$ with $\\sigma_{\\phi}$ only to learn effective composite ray sampling.", + "bbox": [ + 500, + 532, + 890, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Photometric loss. We compare the rendered training rays with their ground truth color", + "bbox": [ + 500, + 563, + 888, + 592 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r g b}} (\\mathbf {r}, t, s) = \\left\\| \\mathbf {C} (\\mathbf {r}, t, s) - \\hat {\\mathbf {C}} (\\mathbf {r}, t, s) \\right\\| _ {2} ^ {2}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 601, + 890, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Expected depth loss. We render the expected depth values for each ray and compare it with the ground truth depth", + "bbox": [ + 500, + 628, + 888, + 659 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d e p}} (\\mathbf {r}, t, s) = \\left\\| \\mathbf {D} (\\mathbf {r}, t, s) - \\hat {\\mathbf {D}} (\\mathbf {r}, t, s) \\right\\| _ {2} ^ {2} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 667, + 890, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the expected depth is calculated via integrating the sampling values $\\hat{\\mathbf{D}} (\\mathbf{r},t,s) = \\int_{u^{\\prime}}^{u_{f}}uU(u)\\sigma (\\mathbf{r}(u),t,s)du$", + "bbox": [ + 496, + 694, + 888, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Entropy regularization. While static and dynamic scene parts can overlap, i.e. inside a 3D bounding box $\\mathbf{b}_o^t$ there could be a part of the street or sidewalk, their density should be strictly separated w.r.t. a single sampling location $\\mathbf{r}(u)$ . We leverage an entropy regularization loss [62] to encourage clear separation between entities in $\\phi$ and $\\psi$", + "bbox": [ + 496, + 724, + 890, + 814 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {e n t r}} (\\mathbf {r}, t, s) = \\int_ {u _ {n}} ^ {u _ {f}} \\mathcal {H} \\left(\\frac {\\sigma_ {\\psi} (\\mathbf {r} (u) , t , s)}{\\sigma_ {\\phi} (\\mathbf {r} (u) , s) + \\sigma_ {\\psi} (\\mathbf {r} (u) , t , s)}\\right) d u \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 823, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{H}(\\cdot)$ is the Shannon entropy and we use $t, s$ as a replacement for the vectors $\\omega_s^t, \\omega_o$ for ease of notation.", + "bbox": [ + 496, + 869, + 888, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21129", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9d7eed8e1f912b7c30285f605ca4fa3fa127454dd16945a035ec46f7e7e967a9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodResidentialDowntownMeanTrain time (h)
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerfacto + Emb.19.830.6370.56218.050.6550.62518.940.6460.5948.0
Nerfacto + Emb. + Time20.050.6410.56218.660.6560.60319.360.6540.58313.2
SUDS [58]21.760.6590.55619.910.6650.64520.840.6620.60154.8
Ours22.290.6780.52320.010.6810.58621.150.6800.55517.2
", + "bbox": [ + 116, + 89, + 854, + 165 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/04f350d262207905fdf5b7d913cfade7b6775a9806ce3ea8c9c56298f9f1a9ea.jpg", + "table_caption": [ + "Table 1. Novel View Synthesis on Argoverse 2. While the static Nerfacto baseline has the weakest performance, the dynamic variant Nerfacto + Time improves only marginally upon it. The state-of-the-art method SUDS exhibits stronger view synthesis results but takes more than $3\\times$ longer to train. Our method outperforms all methods across all metrics and exhibits competitive training speed." + ], + "table_footnote": [], + "table_body": "
MethodKITTI [75%]KITTI [50%]KITTI [25%]
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [31]18.560.5570.55419.120.5870.49718.610.5700.510
NeRF + Time21.010.6120.49221.340.6350.44819.550.5860.505
NSG [35]21.530.6730.25421.260.6590.26620.000.6320.281
Nerfacto + Emb.22.750.8010.15622.380.7930.16021.240.7580.178
Nerfacto + Emb. + Time23.190.8040.15523.180.8030.15521.980.7770.172
SUDS [58]22.770.7970.17123.120.8210.13520.760.7470.198
Ours28.380.9070.05227.510.8980.05526.510.8870.060
MethodVKITTI2 [75%]VKITTI2 [50%]VKITTI2 [25%]
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [31]18.670.5480.63418.580.5440.63518.170.5370.644
NeRF + Time19.030.5740.58718.900.5650.61018.040.5450.626
NSG [35]23.410.6890.31723.230.6790.32521.290.6660.317
Nerfacto + Emb.22.150.8470.14521.880.8430.14821.280.8270.155
Nerfacto + Emb. + Time22.110.8490.14421.780.8440.14721.000.8250.157
SUDS [58]23.870.8460.15023.780.8510.14222.180.8290.160
Ours29.730.9120.06529.190.9060.06628.290.9010.067
", + "bbox": [ + 156, + 213, + 815, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Novel View Synthesis on KITTI and VKITTI2. We compare our method to prior art, following the experimental protocol in [58]. We test the view synthesis performance of the methods with varying fractions of training views and observe that fewer training views generally result in lower performance. Our method outperforms previous works by a large margin on all settings.", + "bbox": [ + 75, + 436, + 893, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hierarchical pose optimization. Alongside scene geometry, we optimize edges $e_{v_s^t v_r}$ and $e_{v_o v_s^t}$ in our graph, i.e. we refine ego-vehicle poses $\\mathbf{P}_s^t\\in \\mathbf{SE}(3)$ and object poses $\\xi_o^t\\in \\mathbf{SE}(3)$ . In particular, we optimize for pose residuals $\\delta \\mathbf{P}_s^t\\in \\mathfrak{se}(3)$ and $\\delta \\xi_o^t\\in \\mathfrak{se}(2)$ . We constrain the object pose residual to $\\mathfrak{se}(2)$ since objects are usually upright and move along the ground plane. Given each residual, we update the ego-vehicle pose with $\\hat{\\mathbf{P}}_s^t = \\expmap(\\delta \\mathbf{P}_s^t)\\mathbf{P}_s^t$ and the object pose with $\\hat{\\xi}_o^t = \\expmap(\\delta \\xi_o^t)_{\\mathbf{SE}(2)\\to \\mathbf{SE}(3)}\\xi_o^t$ , where $\\expmap(\\cdot)$ is the exponential map of each lie group.", + "bbox": [ + 75, + 487, + 468, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to naively optimizing camera and object poses, optimizing the edges along our scene graph has two key advantages. First, we leverage multi-camera constraints, i.e. we keep the camera extrinsics $\\mathbf{T}_c$ fixed and optimize the ego-vehicle poses $\\mathbf{P}_s^t$ only. This is in contrast to prior art that generally treats each camera pose as independent. Second, the residual $\\delta \\mathbf{P}_s^t$ propagates to the object poses since they are defined w.r.t. the ego-vehicle coordinate frame instead of the global world frame. Given that optimizing a radiance field as well as camera and object poses jointly is notoriously difficult [21], these constraints are vital to view synthesis quality (see Tab. 6).", + "bbox": [ + 75, + 641, + 470, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 76, + 832, + 209, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. To evaluate against competing methods on our proposed benchmark on Argoverse 2 [61], we hold out every 10th sample in uniform time intervals where a sample", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "corresponds to seven ring-camera images. To compare our methods against prior art on KITTI [14] and VKITTI2 [4], we follow the experimental protocol and data splits in [58]. We use the provided 3D bounding box annotations for all datasets. We provide results using an off-the-shelf 3D tracker [68] in the supplemental material.", + "bbox": [ + 496, + 487, + 890, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Following [58], we measure image synthesis quality with PSNR, SSIM [59], and LPIPS (AlexNet) [69].", + "bbox": [ + 496, + 579, + 890, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We implement our method in PyTorch [39], accelerating the time-consuming ray-node intersection with a custom CUDA implementation. We train our model on a single RTX 3090 GPU for 250,000 steps on Argovvere 2 and 100,000 steps on KITTI and VKITTI2, with 8192 rays per batch. All model parameters are optimized using Adam [17] with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ , and an exponential decay learning rate schedule: from $10^{-5}$ to $10^{-6}$ for pose parameters, and from $10^{-3}$ to $10^{-4}$ for all others. In order to counter pose drift, we further apply weight decay with a factor $10^{-2}$ to $\\delta \\mathbf{P}$ and $\\delta \\xi$ . The static radiance field $\\phi$ is trained from scratch, while $\\psi$ is initialized from a semantic prior [6] following [18].", + "bbox": [ + 496, + 611, + 890, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method against prior work in dynamic outdoor scene representation, i.e. SUDS [58] and NSG [35]. Further, we include results obtained by augmenting Nerfacto [54], the closest state-of-the-art NeRF architecture to ours, with components designed to handle multicapture reconstruction and scene dynamics. In particular,", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21130", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/df396e5d2396063fdcca86347bed923f065020bbab0a001f7db96ac12891593d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SRN [49]NeRF [31]NeRF + TimeNSG [35]PNF [18]SUDS [58]Ours
PSNR ↑18.8323.3424.1826.6627.4828.3129.36
SSIM ↑0.5900.6620.6770.8060.8700.8760.911
", + "bbox": [ + 80, + 89, + 467, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8783d9847c2883eb37addca49651ab0b828c5bc165f9a085a3642881e6242f4f.jpg", + "image_caption": [ + "Figure 6. Qualitative results on KITTI. With only $25\\%$ of views (approx. 15-20) for training, we can still synthesize sharp and realistic novel views with dynamic objects rendered at high quality." + ], + "image_footnote": [], + "bbox": [ + 80, + 172, + 468, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we consider two variants: \"Nerfacto + Emb.\", where we incorporate our appearance embedding $\\mathbf{A}_s\\mathcal{F}(t)$ and the expected depth loss; and \"Nerfacto + Emb. + Time\", where we further include time modeling with 4D hash tables following [38, 63]. For both Nerfacto variants, we set hash table and MLP sizes to be aligned with our method.", + "bbox": [ + 75, + 323, + 468, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Comparison to state-of-the-art", + "text_level": 1, + "bbox": [ + 76, + 424, + 346, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first compare to prior work on our proposed benchmark in Tab 1. We observe that the static Nerfacto + Emb. has the weakest performance, which slightly increases by adding time modeling in Nerfacto + Emb. + Time. Meanwhile, the current state-of-the-art method SUDS [58] outperforms the Nerfacto variants, while being much slower to train. Our method performs the best on all metrics. The improvement is particularly pronounced in the perceptual quality metrics (SSIM and LPIPS). The training speed of our method is competitive to Nerfacto + Emb. + Time and more than $3 \\times$ faster than SUDS.", + "bbox": [ + 75, + 446, + 468, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we show a qualitative comparison in Fig. 7. We observe major differences in both the rendered images and depth maps. In particular, all other methods struggle to recover the dynamic objects in the scene, while our method produces realistic renderings and accurate depth maps for both static and dynamic areas. We also observe that, thanks to our transient geometry embedding $\\mathbf{G}_s\\mathcal{F}(t)$ , we are able to accurately recover the geometry of the trees and their leaves which notably are not present in every sequence of the area that is reconstructed due to seasonal changes. At the same time, other methods struggle to recover those details. Specifically, other methods exhibit artifacts in the depth maps and degraded rendering quality of the trees left and right of the street in columns five and six of Fig. 7, while our method produces accurate color and depth renderings.", + "bbox": [ + 75, + 613, + 468, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we evaluate our method on established benchmarks, namely KITTI and VKITTI2, in Tab. 2 following the experimental protocol in [58]. In particular, we compare novel view synthesis quality at different fractions of", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1d0882b14f8616a3ddfc5d4f5b2220b1f5fd17dc4d9b092b3b9bc3b31c3c1318.jpg", + "table_caption": [ + "Table 3. Image reconstruction on KITTI. We outperform prior art in image reconstruction, i.e. our method can better fit the training views. We follow the experimental protocol in [18, 35, 58]." + ], + "table_footnote": [], + "table_body": "
AsF(t)GsF(t)PSNR ↑SSIM ↑LPIPS ↓
--19.700.6530.588
✓-22.220.6700.546
✓✓22.490.6710.541
", + "bbox": [ + 545, + 89, + 846, + 143 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/281d98977513d4113854bd40d8f22abe911f6065656fcc6e6ef8f800693738f4.jpg", + "table_caption": [ + "Table 4. Ablation study on graph structure. We show that using the multi-level graph structure of sequences and objects is crucial, i.e. omitting sequence vectors $\\omega_{s}^{t}$ results in degraded quality since there is no conditioning on scene-specific appearance. Combining appearance and transient geometry embeddings performs best." + ], + "table_footnote": [], + "table_body": "
SamplingSamples per rayPSNR ↑SSIM ↑LPIPS ↓Rays / sec.
Uniform [18]19225.720.7300.45628K
Uniform [18]102425.840.7340.4494K
Separate [35]1024+64+(32+64)26.650.7620.3512.5K
Ours1024+6427.070.7590.36230K
", + "bbox": [ + 501, + 220, + 890, + 284 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/19f67193ed78db2061463f6c19284d05b068c2908a5b5688c0e9cdc7c436a492.jpg", + "table_caption": [ + "Table 5. Ray Sampling schemes. Our composite ray sampling is about $12 \\times$ more efficient to train than separate ray sampling [35] with comparable performance. Uniform sampling [18] exhibits lower performance and is slow when sampled more densely." + ], + "table_footnote": [], + "table_body": "
Pose optimizationPSNR ↑SSIM ↑LPIPS ↓
-22.490.6710.541
Naive21.280.6630.519
Hierarchical (Ours)22.290.6780.523
", + "bbox": [ + 555, + 347, + 834, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Pose optimization. Compared to naively optimizing camera and object poses, optimizing vehicle and object poses hierarchically on the edges of our graph benefits view synthesis quality more consistently, i.e. both in SSIM and LPIPS. Note that PSNR is very sensitive to pose drift, and thus does not improve.", + "bbox": [ + 496, + 407, + 890, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training views, i.e. testing different levels of view supervision. We run our method and our baselines and observe substantially better view synthesis quality for our method, both compared to our baselines and prior works. All methods exhibit a similar performance drop when the fraction of training views decreases. We illustrate the rendering quality of our method with only 15-20 training views in Fig. 6. Finally, we also report the results on the task of image reconstruction, i.e. reconstruction of images seen during training, following [18, 35, 58] in Tab. 3 on the KITTI dataset. We significantly outperform previous works in both metrics.", + "bbox": [ + 496, + 487, + 890, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Ablation studies", + "text_level": 1, + "bbox": [ + 500, + 665, + 660, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We verify our design through ablation studies. We perform these on the residential area of our benchmark unless otherwise noted. First, we ablate on the multi-level structure of our scene graph. In particular, in Tab. 4, we ablate the latent codes of the sequence nodes, reducing to a representation similar to [35] and observe that the quality of the synthesized views drops significantly. In contrast, our scene graph representation achieves better view synthesis quality, i.e. achieves almost three points higher PSNR. Further, adding our transient geometry embeddings in addition to sequence-level appearance embeddings improves the view synthesis quality. See supplemental material for a qualitative comparison. This adds to the fact that through our multi-level graph structure, we show the ability to modulate car appear", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21131", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b7faa7dc7233a4c56f0201e825efde6aefde940cee478759f2e25fa02cde53e.jpg", + "image_caption": [ + "Figure 7. Qualitative results on Argoverse 2. While prior art struggles with dynamic actors in the scene, our representation renders both realistic novel views and plausible depth maps. Further, our method models complex transient geometry subject to, e.g., seasonal changes such as tree leaves. Note that prior art produces depth artifacts and exhibits degraded view quality in those regions (columns 5 and 6)." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 888, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ance through sequence appearance in Fig. 4.", + "bbox": [ + 76, + 542, + 372, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 5, we compare our composite ray sampling scheme to uniformly sampled rays as in [18] and rays sampled separately per node as in [35]. We compare the schemes under the assumption that all other model parameters are equal, implementing them in our method. Ours is about $12 \\times$ more efficient to train than densely sampling a ray in uniform intervals or separately sampling the ray per node while producing similar results to the latter and being significantly better than the former. Sparsely sampling a ray at uniform intervals is similarly fast as our method, but yields degraded view synthesis quality. Note that we run this ablation study on only a single sequence of the residential area in our benchmark since densely sampling the rays and separately sampling the rays for all nodes are prohibitively expensive to train in large-scale urban areas.", + "bbox": [ + 75, + 563, + 468, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, in Tab. 6, we compare our hierarchical pose optimization to naive camera pose optimization employed in previous works [21, 54]. While naive camera pose optimization degrades the results significantly in pixel-wise metrics, our hierarchical pose optimization improves the SSIM and maintain a comparable PSNR. Meanwhile, our hierarchical pose optimization exhibits similar gains to", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "naive pose optimization in terms of the perceptual LPIPS metric. This shows that our pose optimization mitigates pose drift while also enabling a more accurate reconstruction. Pose drift usually causes a misalignment between the evaluation viewpoint and scene geometry, which degrades pixel-wise metrics in particular.", + "bbox": [ + 496, + 542, + 892, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 648, + 619, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce a novel multi-level scene graph representation for radiance field reconstruction in dynamic urban environments that scales to large geographic areas with more than ten thousand images from dozens of sequences and hundreds of dynamic objects. We train our representation with an efficient composite ray sampling and rendering scheme and introduce latent variables that enable modeling complex phenomena like varying environmental conditions and transient geometry present across different vehicle captures. We leverage our representation to refine camera and object poses hierarchically using multi-camera constraints. Finally, we propose a new view synthesis benchmark for dynamic urban driving scenarios. Our approach yields substantially improved results compared to prior art while allowing for flexible de- and recomposition of the scene.", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21132", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Zhi-Yang He, JunYoung Gwak, Amir R Zamir, Martin Fischer, Jitendra Malik, and Silvio Savarese. 3d scene graph: A structure for unified semantics, 3d space, and camera. In ICCV, 2019. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1, 3, 5", + "[3] Jules Bloomenthal and Brian Wyvill. Introduction to implicit surfaces. 1997. 2", + "[4] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual kitti 2, 2020. 2, 6", + "[5] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 2", + "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6", + "[7] Steve Cunningham and Michael J Bailey. Lessons from scene graphs: using scene graphs to teach hierarchical modeling. Computers & Graphics, 25(4):703-711, 2001. 2", + "[8] Nathaniel Fairfield, George A. Kantor, and David S. Wettergreen. Real-time slam with octree evidence grids for exploration in underwater tunnels. Journal of Field Robotics, 24, 2007. 2", + "[9] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, 2022. 2", + "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2, 3", + "[11] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In CVPR, 2021. 2", + "[12] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. NeurIPS, 2022. 2", + "[13] Yiming Gao, Yan-Pei Cao, and Ying Shan. Surfelnerf: Neural surfel radiance fields for online photorealistic reconstruction of indoor scenes. In CVPR, 2023. 1, 2", + "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2, 6", + "[15] Michael Kaess. Simultaneous localization and mapping with infinite planes. 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 4605-4611, 2015. 2", + "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[18] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, 2022. 2, 3, 5, 6, 7, 8", + "[19] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2", + "[20] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In CVPR, 2023. 2", + "[21] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 6, 8", + "[22] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In ICCV, 2023. 2, 3", + "[23] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020. 1, 2", + "[24] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In ICCV, 2023. 2, 3", + "[25] Yan Lu and Dezhen Song. Visual navigation using heterogeneous landmarks and unsupervised geometric constraints. IEEE Transactions on Robotics, 31:736-749, 2015. 2", + "[26] Jonathon Luiten, Tobias Fischer, and Bastian Leibe. Track to reconstruct and reconstruct to track. IEEE Robotics and Automation Letters, 5(2):1803-1810, 2020. 2", + "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2", + "[28] Alexander Majercik, Cyril Crassin, Peter Shirley, and Morgan McGuire. A ray-box intersection algorithm and efficient dynamic voxel rendering. Journal of Computer Graphics Techniques Vol, 7(3):66-81, 2018. 5", + "[29] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 2", + "[30] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 1, 2", + "[31] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 4, 5, 6, 7", + "[32] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 2, 3, 4" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21133", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 1, 2", + "[34] OpenStreetMap contributors. Planet dump retrieved from https://planet.osm.org . https : //www.openstreetmap.org, 2017. 3", + "[35] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In CVPR, 2021. 2, 3, 4, 5, 6, 7, 8", + "[36] Byeongjun Park and Changick Kim. Point-dynrf: Point-based dynamic radiance fields from a monocular video, 2023. 2", + "[37] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2", + "[38] Sungheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, and Nahiyup Kang. Temporal interpolation is all you need for dynamic neural radiance fields. In CVPR, 2023. 2, 7", + "[39] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 6", + "[40] Marc Pollefeys, David Nister, Jan-Michael Frahm, Amir Akbarzadeh, Philippos Mordohai, Brian Clipp, Chris Engels, David Gallup, Seon Joo Kim, Paul C. Merrell, C. Salmi, Sudipta N. Sinha, B. Talton, Liang Wang, Qingxiong Yang, Henrik Stewénius, Ruigang Yang, Greg Welch, and Herman Towles. Detailed real-time urban 3d reconstruction from video. IJCV, 78:143-167, 2007. 2", + "[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 2", + "[42] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2, 5", + "[43] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In CVPR, 2021. 1, 2", + "[44] Antoni Rosinol, Arjun Gupta, Marcus Abate, Jingnan Shi, and Luca Carlone. 3d dynamic scene graphs: Actionable spatial perception with places, objects, and humans. arXiv preprint arXiv:2002.06289, 2020. 2", + "[45] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In ECCV, 2022. 2", + "[46] Renato F Salas-Moreno, Richard A Newcombe, Hauke Strasdat, Paul HJ Kelly, and Andrew J Davison. Slam++: Simultaneous localisation and mapping at the level of objects. In CVPR, 2013. 2", + "[47] Aron Schmied, Tobias Fischer, Martin Danelljan, Marc Pollefeys, and Fisher Yu. R3d3: Dense 3d reconstruction of dynamic scenes from multiple cameras. In ICCV, 2023." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Chen Shen, James F. O'Brien, and Jonathan R. Shewchuk. Interpolating and approximating implicit surfaces from polygon soup. In ACM SIGGRAPH 2004 Papers, page 896-904, 2004. 2", + "[49] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. NeurIPS, 2019. 1, 2, 7", + "[50] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 2", + "[51] Henry Sowizral. Scene graphs in the new millennium. IEEE Computer Graphics and Applications, 20(1):56-57, 2000. 2", + "[52] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2", + "[53] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In CVPR, 2022. 2", + "[54] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 5, 6, 8", + "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 2", + "[56] Shubham Tulsiani, Saurabh Gupta, David F Fouhey, Alexei A Efros, and Jitendra Malik. Factoring shape, pose, and layout from the 2d image of a 3d scene. In CVPR, 2018. 2", + "[57] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In CVPR, 2022. 2, 3", + "[58] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In CVPR, 2023. 2, 3, 4, 6, 7, 8", + "[59] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[60] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 2", + "[61] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21134", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "self-driving perception and forecasting. arXiv preprint arXiv:2301.00493, 2023. 2, 3, 6", + "[62] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D' 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS, 2022. 2, 5", + "[63] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In CVPR, 2021. 2, 7", + "[64] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 2", + "[65] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In CVPR, 2022. 2", + "[66] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In CVPR, 2022. 4", + "[67] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, 2023. 2, 3, 4", + "[68] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, 2021. 6", + "[69] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + ], + "bbox": [ + 78, + 90, + 468, + 489 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21135", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/10ec9eda-ceb6-4b88-aa59-8a3b78a66785_model.json b/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/10ec9eda-ceb6-4b88-aa59-8a3b78a66785_model.json new file mode 100644 index 0000000000000000000000000000000000000000..521f9dfddebc81e93dac6f568748c0c970c472f4 --- /dev/null +++ b/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/10ec9eda-ceb6-4b88-aa59-8a3b78a66785_model.json @@ -0,0 +1,2576 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.131, + 0.83, + 0.152 + ], + "angle": 0, + "content": "Multi-Level Neural Scene Graphs for Dynamic Urban Environments" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.18, + 0.724, + 0.198 + ], + "angle": 0, + "content": "Tobias Fischer\\(^{1}\\) Lorenzo Porzi\\(^{2}\\) Samuel Rota Bulò\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.199, + 0.648, + 0.216 + ], + "angle": 0, + "content": "Marc Pollefeys1 Peter Kontschieder2" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.218, + 0.635, + 0.236 + ], + "angle": 0, + "content": "\\(^{1}\\) ETH Zürich \\(^{2}\\) Meta Reality Labs" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.24, + 0.66, + 0.255 + ], + "angle": 0, + "content": "https://tobiasfshr.github.io/pub/ml-nsg/" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.271, + 0.153, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.342, + 0.152, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.409, + 0.153, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.484, + 0.138, + 0.498 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.161, + 0.276, + 0.481, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.483, + 0.434, + 0.498 + ], + "angle": 0, + "content": "Multi-Level Neural Scene Graph" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.271, + 0.622, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.373, + 0.622, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.271, + 0.754, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.373, + 0.755, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.271, + 0.888, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.373, + 0.889, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.483, + 0.734, + 0.496 + ], + "angle": 0, + "content": "Novel Views" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.507, + 0.892, + 0.577 + ], + "angle": 0, + "content": "Figure 1. Overview. We represent sequences captured from moving vehicles in a shared geographic area with a multi-level scene graph. Each dynamic object \\( v_{o} \\) is associated with a sequence node \\( v_{s}^{t} \\) and time \\( t \\). The sequence nodes are registered in a common world frame at the root node \\( v_{r} \\) through the vehicle poses \\( \\mathbf{P}_{s}^{t} \\), while the dynamic objects are localized w.r.t. the sequence node with pose \\( \\xi_{o}^{t} \\). Each camera \\( c \\) is associated with an ego-vehicle position, i.e. node \\( v_{s}^{t} \\), through the extrinsic calibration \\( \\mathbf{T}_{c} \\). The sequence and object nodes hold latent codes \\( \\omega \\) that condition the radiance field, synthesizing novel views in various conditions with distinct dynamic objects." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.59, + 0.312, + 0.604 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.629, + 0.471, + 0.885 + ], + "angle": 0, + "content": "We estimate the radiance field of large-scale dynamic areas from multiple vehicle captures under varying environmental conditions. Previous works in this domain are either restricted to static environments, do not scale to more than a single short video, or struggle to separately represent dynamic object instances. To this end, we present a novel, decomposable radiance field approach for dynamic urban environments. We propose a multi-level neural scene graph representation that scales to thousands of images from dozens of sequences with hundreds of fast-moving objects. To enable efficient training and rendering of our representation, we develop a fast composite ray sampling and rendering scheme. To test our approach in urban driving scenarios, we introduce a new, novel view synthesis benchmark. We show that our approach outperforms prior art by a significant margin on both established and our proposed benchmark while being faster in training and rendering." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.59, + 0.631, + 0.604 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.827 + ], + "angle": 0, + "content": "Estimating the radiance field of a dynamic urban environment from data collected by sensor-equipped vehicles is a core challenge in closed-loop simulation for robotics and in mixed reality. It is particularly relevant for applications like autonomous driving and city-scale mapping, where capture vehicles can be frequently deployed. The growing amount of available data provides great opportunities for creating up-to-date digital twins of entire cities but also poses unique challenges as increasingly heterogeneous data sources must be processed. In particular, limited scene coverage, different lighting, weather, seasonal conditions, and varying geometry due to distinct dynamic and transient objects make radiance field reconstruction of dynamic urban environments extremely challenging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.827, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Recently, Neural Radiance Fields (NeRFs) have enabled significant progress in achieving realistic novel view synthesis from a set of input views [2, 31-33, 49]. These methods represent a static 3D scene with fully implicit [30, 31, 49] or low-level explicit structures [10, 13, 16, 23, 32, 43," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21125" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "52, 65] such as voxels, points, surfels, meshes or 3D gaus-. sians. In parallel, implicit [19, 37, 50, 62, 63] and low-level explicit [9, 27, 36, 38] neural representations for dynamic 4D scenes have been investigated. However, fewer works have focused on decomposing scenes into higher-level entities [18, 26, 46, 56]. In computer graphics and 3D mapping, scene graphs have been used to represent complex scenes in a multi-level hierarchy [1, 44, 51]. In view synthesis, Ost et al. [35] apply this concept by describing actors in a dynamic scene as entities of a scene graph. However, their representation lacks the ability of [1, 44] to represent scenes at multiple levels, thereby limiting their representation to short, single sequences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.473, + 0.488 + ], + "angle": 0, + "content": "While earlier methods for view synthesis focused on object-centric scenes with controlled camera trajectories, recent works move towards radiance field reconstruction of large-scale environments from in-the-wild captures. Among these, many methods focus on static environments, removing dynamic actors from the input data [22, 24, 29, 42, 45, 53, 57]. A few works explicitly model dynamic actors [18, 35, 67], but either do not scale to more than a single, short video [18, 35, 67], or struggle to accurately represent individual object instances [58]. This complicates the evaluation of these methods for real-world applications because existing benchmarks neither scale to large urban areas [18, 35] nor reflect realistic capturing conditions [12]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.473, + 0.729 + ], + "angle": 0, + "content": "To address these issues, we propose a multi-level neural scene graph representation that spans large geographic areas with hundreds of dynamic objects. In contrast to previous works, our multi-level scene graph formulation allows us to distinguish and represent dynamic object instances effectively and further to represent a scene under varying conditions. To make our representation viable for large-scale dynamic environments, we develop a composite ray sampling and rendering scheme that enables fast training and rendering of our method. To test this hypothesis, we introduce a benchmark for radiance field reconstruction in dynamic urban environments based on [61]. We fuse data from dozens of vehicle captures under varying conditions amounting to more than ten thousand images with several hundreds of dynamic objects per reconstructed area. We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.733, + 0.468, + 0.775 + ], + "angle": 0, + "content": "- We propose a multi-level neural scene graph formulation that scales to dozens of sequences with hundreds of fast-moving objects under varying environmental conditions." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.779, + 0.468, + 0.82 + ], + "angle": 0, + "content": "- We develop an efficient composite ray sampling and rendering scheme that enables fast training and rendering of our representation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.824, + 0.467, + 0.865 + ], + "angle": 0, + "content": "- We present a benchmark that provides a realistic, application-driven evaluation of radiance field reconstruction in dynamic urban environments." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.733, + 0.468, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We show state-of-the-art view synthesis results on both established benchmarks [4, 14] and our proposed benchmark." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.297 + ], + "angle": 0, + "content": "3D and 4D scene representations. Finding the right scene representation is a core issue in 3D computer vision and graphics [5]. Over the years, a wide variety of options have been explored [1, 3, 8, 15, 25, 30, 40, 46-48], and, more recently, neural rendering [33] has been used to enable a new generation of scene models that support photo-realistic novel view synthesis. Scene representations for neural rendering can be roughly classified as \"implicit\" [30, 31, 49], which store most of the information in the weights of a neural network, or \"explicit\", which use low-level spatial primitives such as voxels [10, 23, 32, 52], points [65], surfels [13], meshes [43], or 3D gaussians [16]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.892, + 0.403 + ], + "angle": 0, + "content": "Similarly, different 4D dynamic scene representations for view synthesis have been investigated, including both implicit [11, 19, 37, 41, 55, 58, 62, 63] and explicit [9, 27, 36, 67] approaches. Dynamics are generally modeled as deformations of a canonical volume [9, 37, 41, 55, 62], a separate scene motion function [11, 19, 20, 58, 63], or rigid transformations of local geometric primitives [27]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.893, + 0.705 + ], + "angle": 0, + "content": "Another line of work investigates the decomposition of scenes into higher-level entities [18, 26, 46, 56]. To express the composition of different entities into a complex scene, classical computer graphics literature [7] uses scene graphs. In particular, entities are represented as nodes in a hierarchical graph and are connected through edges defined by coordinate frame transformations. Thus, global transformations can be acquired by traversing the graph from its root node. For indoor 3D mapping, this concept was proposed by Armeni et al. [1] to represent static scenes at multiple levels of hierarchy, i.e. buildings, rooms and objects, and was later extended to dynamic scenes by Rosinol et al. [44]. Recently, Ost et al. [35] have revisited this concept for view synthesis, describing multi-object scenes with a graph that represents the dynamic actors in the scene. However, their representation lacks the ability of [1, 44] to represent scenes at multiple levels of hierarchy and is thus inherently limited to single, short video clips. On the contrary, we present a scalable, multi-level scene graph representation that spans large geographic areas with hundreds of dynamic objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Representing large-scale urban scenes. Compared to controlled captures of small scenes, in-the-wild captures of large-scale scenes pose distinct challenges. Limited viewpoint coverage, inaccurate camera poses, far-away buildings and sky, fast-moving objects, complex lighting, and auto exposure make radiance field reconstruction challenging. Therefore, previous works aid the reconstruction by using depth priors from e.g. LiDAR, refining camera parameters, adding information about camera exposure, and using specialized sky and light modeling components [22, 29, 42, 45, 53, 60, 64]. While many of these works simply remove dynamic actors, some methods explicitly model scene dynamics [18, 35, 58, 67]. However," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21126" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.218, + 0.47, + 0.26 + ], + "angle": 0, + "content": "Figure 2. Ego-vehicle trajectories of our benchmark. We show the residential (left) and the downtown (right) areas, trajectories superimposed on 2D maps obtained from OpenStreetMap [34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.468, + 0.361 + ], + "angle": 0, + "content": "these methods either struggle to scale to more than a single sequence [18, 35, 67] or to accurately represent and distinguish dynamic actors [58]. To address these issues, we present a decomposed, scalable scene representation for dynamic urban environments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.468, + 0.513 + ], + "angle": 0, + "content": "Efficient scene rendering. Especially for large-scale scenes, the efficiency of a scene representation in training and inference is crucial. To alleviate the burden of ray traversal in volumetric rendering, many techniques for efficient sampling [2, 10, 32, 57] have been proposed. Recently, researchers have exploited more efficient forms of rendering, e.g. rasterization of meshes [22, 24] or other 3D primitives [16]. While these approaches are focused on static scenes, we extend the approach of [2] to our scene graph representation and thus to dynamic scenes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.142, + 0.543 + ], + "angle": 0, + "content": "3. Data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.468, + 0.674 + ], + "angle": 0, + "content": "We investigate how to synthesize novel views of a dynamic urban environment from a set of captures taken from moving vehicles. Specifically, we are interested in captures spanning the same geographic area under varying conditions. This data presents unique challenges due to limited viewpoint coverage, different lighting, weather, and season, and even varying geometry due to distinct dynamic and transient objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.468, + 0.854 + ], + "angle": 0, + "content": "Previous benchmarks in this area [18, 35] are limited to short, single video clips with simple ego trajectories and a few dynamic objects. Therefore, we present a benchmark that better reflects the aforementioned challenges. We base our benchmark on Argoverse 2 [61] which provides a rich set of captures from a fleet of vehicles deployed in multiple US cities that span different weather, season, and time of day. The vehicles are equipped with a surround-view camera rig with seven global shutter cameras, a LiDAR sensor, and a GPS. Furthermore, the timings of the different sensors are provided, so that one can relate the LiDAR-based 3D bounding box annotations to camera timestamps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We leverage the GPS information to globally align the sequences and to identify regions that we are interested in mapping. We then extract the specific vehicle captures." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.892, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.192, + 0.892, + 0.262 + ], + "angle": 0, + "content": "Figure 3. Sequence alignment visualization. The initial GPS-based alignment is imprecise, as evidenced by the duplicated structures in the overlaid LiDAR point clouds (left). After our ICP alignment, the area is well reconstructed (middle) according to its real geometry (right, from Argoverse 2 [61])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.286, + 0.892, + 0.361 + ], + "angle": 0, + "content": "Since GPS-based localization accuracy is only coarsely precise in urban areas, we align the captures via a global, offline iterative-closest-point (ICP) procedure applied to the LiDAR point clouds of all sequences to achieve satisfactory alignment for novel view synthesis purposes (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.892, + 0.527 + ], + "angle": 0, + "content": "Following this procedure, we build a benchmark that enables real-world evaluation of novel view synthesis from diverse vehicle captures. It is composed of 37 vehicle captures split into two geographic regions as illustrated in Fig. 2. The regions cover a residential and a downtown area to resemble the different characteristics of urban environments. The residential area spans 14 captures with 10493 training and 1162 testing images with more than 700 distinct moving objects. The downtown area spans 23 captures with 16933 training and 1876 testing images with more than 600 distinct moving objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.539, + 0.59, + 0.555 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.821 + ], + "angle": 0, + "content": "Problem setup. We are given a set of sequences \\(S\\) captured from moving vehicles in different conditions. Each sequence \\(s \\in S\\) consists of a set of images taken from cameras \\(C_s\\) mounted on the vehicle at different timesteps indexed by \\(T_s\\). We assume the vehicle's sensors are calibrated with respect to a common vehicle and a global world frame. In particular, we assume given sensor extrinsic \\(\\mathbf{T}_c = [\\mathbf{R}_c|\\mathbf{t}_c] \\in \\mathbf{SE}(3)\\) for each camera \\(c \\in C = \\bigcup_{s} C_s\\), and ego-vehicle poses \\(\\mathbf{P}_s^t = [\\mathbf{R}_s^t|\\mathbf{t}_s^t] \\in \\mathbf{SE}(3)\\) for each timestamp \\(t \\in T_s\\) and sequence \\(s \\in S\\). Furthermore, we assume all cameras to have known intrinsics \\(\\mathbf{K}_c\\). Each sequence \\(s\\) entails a set \\(O_s\\) of dynamic objects. For each object \\(o \\in O_s\\), we assume to have an estimated 3D bounding box track \\(\\mathcal{T}_o\\) consisting of object poses \\(\\{\\xi_o^{t_0}, \\dots, \\xi_o^{t_n}\\} \\subset \\mathbf{SE}(3)\\) w.r.t. the ego-vehicle frame, where \\(t_i \\in T_s\\), and the 3D object dimensions \\(\\mathbf{s}_o \\in \\mathbb{R}_{>0}^3\\). The images span a common geographic area for which we would like to estimate a radiance field" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.83, + 0.891, + 0.847 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} (\\mathbf {x}, \\mathbf {d}, t, s) = (\\sigma (\\mathbf {x}, t, s), \\mathbf {c} (\\mathbf {x}, \\mathbf {d}, t, s)) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.892, + 0.901 + ], + "angle": 0, + "content": "that outputs volume density \\(\\sigma \\in \\mathbb{R}_{\\geq 0}\\) and color \\(\\mathbf{c} \\in [0,1]^3\\) conditioned on 3D location \\(\\mathbf{x}\\), viewing direction \\(\\mathbf{d}\\), time \\(t\\) and sequence \\(s\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21127" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.367, + 0.108 + ], + "angle": 0, + "content": "4.1. Multi-Level Neural Scene Graph" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.474, + 0.342 + ], + "angle": 0, + "content": "Overview. We illustrate our representation in Fig. 1. We decompose the scene into a graph \\(\\mathcal{G} = (\\mathcal{V},\\mathcal{E})\\) where the set of nodes \\(\\nu\\) is composed by a root node \\(v_{r}\\) that defines the global coordinate system, camera nodes \\(\\{v_{c}\\}_{c\\in C}\\), and, for each sequence \\(s\\in S\\), sequence nodes \\(\\{v_s^t\\}_{t\\in T_s}\\) and dynamic object nodes \\(\\{v_o\\}_{o\\in O_s}\\). The nodes are connected through oriented edges \\(e\\in \\mathcal{E}\\) that represent rigid transformations between the coordinate frames of the nodes, consistently with the edge direction. In addition, each node \\(v\\in \\mathcal{V}\\) can have an associated latent vector \\(\\omega\\). Given the scene graph \\(\\mathcal{G}\\), we model \\(f_{\\theta}\\) using two radiance fields, namely \\(\\phi\\) for largely static and \\(\\psi\\) for highly dynamic scene parts. We use the latent vectors \\(\\omega\\) to condition the radiance fields. In particular, we use latent vectors \\(\\omega_s^t\\) of a sequence node \\(v_s^t\\) to condition \\(\\phi\\):" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.355, + 0.47, + 0.373 + ], + "angle": 0, + "content": "\\[\n\\phi (\\mathbf {x}, \\mathbf {d}, \\omega_ {s} ^ {t}) = \\left(\\sigma_ {\\phi} \\left(\\mathbf {x}, \\omega_ {s} ^ {t}\\right), \\mathbf {c} _ {\\phi} \\left(\\mathbf {x}, \\mathbf {d}, \\omega_ {s} ^ {t}\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.471, + 0.448 + ], + "angle": 0, + "content": "Since the graph \\(\\mathcal{G}\\) has multiple levels, a node associated with a dynamic object will naturally fall into the sequence \\(s\\) it appears in. Therefore, the radiance field \\(\\psi\\) is conditioned on latent vectors \\(\\omega_{s}^{t}\\) and \\(\\omega_{o}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.461, + 0.47, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\psi (\\mathbf {x}, \\mathbf {d}, \\omega_ {s} ^ {t}, \\omega_ {o}) = (\\sigma_ {\\psi} (\\mathbf {x}, \\omega_ {s} ^ {t}, \\omega_ {o}), \\mathbf {c} _ {\\psi} (\\mathbf {x}, \\mathbf {d}, \\omega_ {s} ^ {t}, \\omega_ {o})) (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.468, + 0.643 + ], + "angle": 0, + "content": "of the corresponding sequence and object nodes \\( v_{s}^{t} \\) and \\( v_{o} \\). Appearance and geometry variation. Reconstructing an environment from multiple captures is challenging from two perspectives: in addition to varying dynamic objects, there is i) varying appearance across captures, and ii) slow-moving or static transient geometry such as tree leaves or construction sites. Since both transient geometry and appearance can vary across captures, but are usually smooth within a sequence, we model these phenomena as smooth functions over time" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.658, + 0.47, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\omega_ {s} ^ {t} = \\left[ \\mathbf {A} _ {s} \\mathcal {F} (t), \\mathbf {G} _ {s} \\mathcal {F} (t) \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.824 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}_s\\) is an appearance matrix, \\(\\mathbf{G}_s\\) is a transient geometry matrix and \\(\\mathcal{F}(\\cdot)\\) is a 1D basis function of sines and cosines with linearly increasing frequencies at log-scale [58, 66]. We normalize time \\(t\\) into the interval \\([-1,1]\\) using the maximum sequence length \\(\\max_{s\\in S}|T_s|\\). Crucially, this allows to model near-static, but sequence-specific regions of the input, as well as appearance changes due to e.g. auto-exposure. The degree of variation across time can be controlled by the number of frequencies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Sequence nodes. The sequence nodes \\( v_{s}^{t} \\) are connected to the root node \\( v_{r} \\) with an edge \\( e_{v_s^t v_r} = \\mathbf{P}_s^t \\), i.e. the sequences \\( S \\) share a common global world frame. Each sequence node holds the latent vector \\( \\omega_{s}^{t} \\) that conditions the radiance field \\( \\phi \\). We model \\( \\phi \\) with a multi-scale 3D hash grid representa" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.895, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.284, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Figure 4. Modifying car appearance with scene appearance. We exchange \\(\\omega_{s}^{t}\\) for different car instances. The car's appearance in a rendered, car-centric view (top) changes according to the environmental conditions visible in the sequence \\(s\\) (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.346, + 0.752, + 0.362 + ], + "angle": 0, + "content": "tion [32] and lightweight MLP heads:" + }, + { + "type": "equation", + "bbox": [ + 0.737, + 0.371, + 0.891, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\mathbf {x}} = \\mathrm {H} _ {3} (\\mathbf {x}) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.39, + 0.891, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {\\sigma_ {\\phi}}, \\sigma_ {\\phi} = \\operatorname {M L P} _ {\\sigma_ {\\phi}} \\left(\\mathbf {f} _ {\\mathbf {x}}\\right) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.409, + 0.891, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {\\phi} = \\operatorname {M L P} _ {\\mathbf {c} _ {\\phi}} \\left(\\mathbf {h} _ {\\sigma_ {\\phi}}, \\gamma_ {\\mathrm {S H}} (\\mathbf {d}), \\mathbf {A} _ {s} \\mathcal {F} (t)\\right) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.428, + 0.891, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {\\mathbf {G}}, \\mathbf {c} _ {\\mathbf {G}} = \\operatorname {M L P} _ {\\mathbf {G}} \\left(\\mathbf {h} _ {\\sigma_ {\\phi}}, \\mathbf {G} _ {s} \\mathcal {F} (t)\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.453, + 0.892, + 0.498 + ], + "angle": 0, + "content": "where \\(\\gamma_{\\mathrm{SH}}(\\cdot)\\) is a spherical harmonics encoding [32]. The final colors and densities are computed as a mixture of static and transient output (analogous to Eq. 12, see supp. mat.)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.498, + 0.892, + 0.588 + ], + "angle": 0, + "content": "Dynamic nodes. The dynamic nodes \\( v_{o} \\) are connected to the sequence nodes \\( v_{s}^{t} \\) with edges \\( e_{v_{o}v_{s}^{t}} = \\xi_{o}^{t} \\). The dynamic node \\( v_{o} \\) associated with object \\( o \\) holds a latent vector \\( \\omega_{o} \\) that conditions the radiance field \\( \\psi \\). We model \\( \\psi \\) following [31] with an MLP conditioned on \\( \\omega_{o} \\) to represent different instances with the same network [35, 67]" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.598, + 0.891, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {\\sigma_ {\\psi}}, \\sigma_ {\\psi} = \\operatorname {M L P} _ {\\sigma_ {\\psi}} \\left(\\gamma_ {\\mathrm {P E}} (\\mathbf {x}), \\omega_ {o}\\right) \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.618, + 0.891, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {\\psi} = \\operatorname {M L P} _ {\\mathbf {c} _ {\\psi}} \\left(\\mathbf {h} _ {\\sigma_ {\\psi}}, \\gamma_ {\\mathrm {P E}} (\\mathbf {d}), \\omega_ {o}, \\omega_ {s} ^ {t}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.75 + ], + "angle": 0, + "content": "where \\(\\gamma_{\\mathrm{PE}}(\\cdot)\\) is a positional encoding [31]. Note that 3D position \\(\\mathbf{x}\\) and viewing direction \\(\\mathbf{d}\\) are transformed into the local object coordinate frame with \\((\\mathbf{P}_s^t\\xi_o^t)^{-1}\\mathbf{I}_3(1 / \\max (\\mathbf{s}_o))\\). We condition \\(\\psi\\) on both the scene and object-dependent latent vectors. This allows us to disentangle scene-dependent appearance from the object texture and thus to transfer objects across sequences. We illustrate this process in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Camera nodes. The cameras \\(c \\in C_s\\) are connected to the sequence nodes \\(v_s^t\\) through edges \\(e_{v_c v_s^t} = \\mathbf{T}_c\\), i.e. the calibration of camera \\(c\\) w.r.t. the ego-vehicle frame. This way, we can tie camera poses to a specific ego-vehicle pose." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.722, + 0.835 + ], + "angle": 0, + "content": "4.2. Scene Graph Rendering" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We describe how we render our scene graph \\(\\mathcal{G}\\) for a given set of rays \\(\\mathcal{R}\\). The sampling locations along a ray \\((\\mathbf{r}, t, s) \\in \\mathcal{R}\\) at time \\(t\\) in sequence \\(s\\) are defined as \\(\\mathbf{r}(u) = \\mathbf{o} + u\\mathbf{d}\\) with \\(\\mathbf{o} = \\mathbf{R}_s^t\\mathbf{t}_c + \\mathbf{t}_s^t\\) and \\(\\mathbf{d} = \\mathbf{R}_s^t\\mathbf{R}_c\\mathbf{K}_c^{-1}(p_x, p_y, 1)^T\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21128" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "Continuous-time pose. In order to realistically render videos at different frame rates, we treat the dynamic object poses \\(\\{\\xi_o^{t_0},\\dots,\\xi_o^{t_n}\\}\\) as a continuous function of time \\(\\xi_o(t)\\). We compute \\(\\xi_o(t)\\) by interpolating between the two nearest poses at \\(t_a\\leq t < t_b\\) to time \\(t\\). This allows us also to synchronize estimated object poses originating from the LiDAR measurements with the camera timestamps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.469, + 0.333 + ], + "angle": 0, + "content": "Ray-node intersection. To render the dynamic nodes, we measure the intersection of their 3D bounding boxes with each \\((\\mathbf{r},t,s)\\in \\mathcal{R}\\). In particular, given the sequence \\(s\\) and time \\(t\\) the ray \\(\\mathbf{r}\\) is associated with, we first traverse the graph \\(\\mathcal{G}\\) to retrieve all relevant nodes \\(v_{o}\\) and their 3D bounding boxes \\(\\mathbf{b}_o^t = [\\xi_o(t),\\mathbf{s}_o]\\) at time \\(t\\). Then, we transform \\(\\mathbf{r}\\) into each local node coordinate system and subsequently use AABB-ray intersection [28] to compute the entry and exit locations \\(u_{o}^{\\mathrm{in}},u_{o}^{\\mathrm{out}}\\) along ray \\(\\mathbf{r}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.333, + 0.469, + 0.365 + ], + "angle": 0, + "content": "Composite rendering. To render the color \\(\\hat{\\mathbf{C}}\\) of ray \\(\\mathbf{r}\\) in sequence \\(s\\) at time \\(t\\), we use volumetric rendering [31]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.375, + 0.451, + 0.409 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {C}} (\\mathbf {r}, t, s) = \\int_ {u _ {n}} ^ {u _ {f}} U (u) \\sigma (\\mathbf {r} (u), t, s) \\mathbf {c} (\\mathbf {r} (u), \\mathbf {d}, t, s) d u\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.411, + 0.469, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} U (u) = \\exp \\left(- \\int_ {u _ {n}} ^ {u} \\sigma \\left(u ^ {\\prime}\\right) d u ^ {\\prime}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.456, + 0.468, + 0.486 + ], + "angle": 0, + "content": "We obtain density \\(\\sigma\\) and color \\(\\mathbf{c}\\) at sampling location \\(\\mathbf{r}(u)\\) as mixture of the radiance fields \\(\\phi\\) and \\(\\psi\\)" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.497, + 0.469, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\sigma = \\sigma_ {\\phi} + \\sigma_ {\\psi}, \\mathbf {c} = \\frac {\\sigma_ {\\phi}}{\\sigma_ {\\phi} + \\sigma_ {\\psi}} \\mathbf {c} _ {\\phi} + \\frac {\\sigma_ {\\psi}}{\\sigma_ {\\phi} + \\sigma_ {\\psi}} \\mathbf {c} _ {\\psi}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.469, + 0.583 + ], + "angle": 0, + "content": "Crucially, we set \\(\\sigma_{\\psi}\\) to zero when \\(\\mathbf{r}(u)\\) does not lie within a 3D bounding box of a dynamic node given the calculated entry and exit points \\(u^{\\mathrm{in}}, u^{\\mathrm{out}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.469, + 0.794 + ], + "angle": 0, + "content": "Composite ray sampling. Instead of densely sampling the space [18, 42] or leveraging separate ray sampling mechanisms per node [35], we use a composite ray sampling strategy illustrated in Fig. 5. We extend the proposal sampling mechanism introduced in [2] by joint sampling from a computationally efficient density field \\(\\sigma_{\\mathrm{prop}}\\) and dynamic nodes \\(v_{o}\\) at time \\(t\\). In particular, as in Eq. 12, we represent \\(\\sigma (\\mathbf{r}(u),t,s)\\) by a mixture of \\(\\sigma_{\\mathrm{prop}}(\\mathbf{r}(u),\\omega_s^t)\\) and \\(\\sigma_{\\psi}(\\mathbf{r}(u),\\omega_s^t,\\omega_o)\\). However, analogous to our rendering step, we constrain sampling from \\(\\sigma_{\\psi}\\) to \\([u^{\\mathrm{in}},u^{\\mathrm{out}}]\\). This allows us to skip empty space efficiently by distilling the static density \\(\\sigma_{\\phi}\\) into \\(\\sigma_{\\mathrm{prop}}\\). At the same time, sparsely querying \\(\\sigma_{\\psi}\\) when \\(\\mathbf{r}(u)\\) falls into a dynamic node enables us to still accurately represent the full, dynamic \\(\\sigma\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Since there are only few samples that fall into dynamic nodes when performing uniform sampling, the computational overhead in the first ray sampling iteration is negligible. In the second iteration, we apply inverse transform sampling given the CDF \\( F(u) = 1 - U(u) \\) along ray \\( \\mathbf{r} \\) and thus the samples are concentrated at the first surface intersection. Hence, only for rays that fall in the line of sight of" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.088, + 0.738, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.089, + 0.871, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.218, + 0.892, + 0.301 + ], + "angle": 0, + "content": "Figure 5. Composite ray sampling. If a ray intersects with an object \\( v_{o} \\), we sample from both proposal network \\( \\sigma_{\\mathrm{prop}} \\) and radiance field \\( \\psi \\), and \\( \\sigma_{\\mathrm{prop}} \\) otherwise. We condition each with the latents \\( \\omega \\) of the respective nodes. The PDF is a mixture of all node densities that intersect with the ray. The transmittance \\( U \\) drops at the first surface intersection (tree) where further samples will concentrate." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.89, + 0.339 + ], + "angle": 0, + "content": "an object surface will we sample \\(\\sigma_{\\psi}\\) more than a few times. In total, we use two ray sampling iterations as in [54]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.399 + ], + "angle": 0, + "content": "Space contraction. Following [2, 54], we contract the unbounded scene space into a fixed-size bounding box, normalizing the scene with bounds computed from the LiDAR point clouds and the ego-vehicle poses \\(\\mathbf{P}\\) into a unit cube." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.636, + 0.425 + ], + "angle": 0, + "content": "5. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.43, + 0.885, + 0.446 + ], + "angle": 0, + "content": "We optimize the parameters \\(\\theta\\) of the radiance field \\(f_{\\theta}\\) with" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.454, + 0.858, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {(\\mathbf {r}, t, s) \\in \\mathcal {R}} \\mathcal {L} _ {\\mathrm {r g b}} (\\mathbf {r}, t, s) + \\lambda_ {\\mathrm {d i s t}} \\mathcal {L} _ {\\mathrm {d i s t}} (\\mathbf {r}, t, s) + \\lambda_ {\\mathrm {p r o p}}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.491, + 0.891, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r o p}} (\\mathbf {r}, t, s) + \\lambda_ {\\text {d e p}} \\mathcal {L} _ {\\text {d e p}} (\\mathbf {r}, t, s) + \\lambda_ {\\text {e n t r}} \\mathcal {L} _ {\\text {e n t r}} (\\mathbf {r}, t, s) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.533, + 0.891, + 0.564 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{dist}}\\) and \\(\\mathcal{L}_{\\mathrm{prop}}\\) follow [2]. We supervise \\(\\sigma_{\\mathrm{prop}}\\) with \\(\\sigma_{\\phi}\\) only to learn effective composite ray sampling." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.564, + 0.89, + 0.593 + ], + "angle": 0, + "content": "Photometric loss. We compare the rendered training rays with their ground truth color" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.602, + 0.891, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r g b}} (\\mathbf {r}, t, s) = \\left\\| \\mathbf {C} (\\mathbf {r}, t, s) - \\hat {\\mathbf {C}} (\\mathbf {r}, t, s) \\right\\| _ {2} ^ {2}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.629, + 0.89, + 0.66 + ], + "angle": 0, + "content": "Expected depth loss. We render the expected depth values for each ray and compare it with the ground truth depth" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.668, + 0.891, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d e p}} (\\mathbf {r}, t, s) = \\left\\| \\mathbf {D} (\\mathbf {r}, t, s) - \\hat {\\mathbf {D}} (\\mathbf {r}, t, s) \\right\\| _ {2} ^ {2} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.89, + 0.725 + ], + "angle": 0, + "content": "where the expected depth is calculated via integrating the sampling values \\(\\hat{\\mathbf{D}} (\\mathbf{r},t,s) = \\int_{u^{\\prime}}^{u_{f}}uU(u)\\sigma (\\mathbf{r}(u),t,s)du\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.726, + 0.891, + 0.815 + ], + "angle": 0, + "content": "Entropy regularization. While static and dynamic scene parts can overlap, i.e. inside a 3D bounding box \\(\\mathbf{b}_o^t\\) there could be a part of the street or sidewalk, their density should be strictly separated w.r.t. a single sampling location \\(\\mathbf{r}(u)\\). We leverage an entropy regularization loss [62] to encourage clear separation between entities in \\(\\phi\\) and \\(\\psi\\)" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.824, + 0.892, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {e n t r}} (\\mathbf {r}, t, s) = \\int_ {u _ {n}} ^ {u _ {f}} \\mathcal {H} \\left(\\frac {\\sigma_ {\\psi} (\\mathbf {r} (u) , t , s)}{\\sigma_ {\\phi} (\\mathbf {r} (u) , s) + \\sigma_ {\\psi} (\\mathbf {r} (u) , t , s)}\\right) d u \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.89, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}(\\cdot)\\) is the Shannon entropy and we use \\(t, s\\) as a replacement for the vectors \\(\\omega_s^t, \\omega_o\\) for ease of notation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21129" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.09, + 0.855, + 0.166 + ], + "angle": 0, + "content": "
MethodResidentialDowntownMeanTrain time (h)
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerfacto + Emb.19.830.6370.56218.050.6550.62518.940.6460.5948.0
Nerfacto + Emb. + Time20.050.6410.56218.660.6560.60319.360.6540.58313.2
SUDS [58]21.760.6590.55619.910.6650.64520.840.6620.60154.8
Ours22.290.6780.52320.010.6810.58621.150.6800.55517.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.17, + 0.894, + 0.213 + ], + "angle": 0, + "content": "Table 1. Novel View Synthesis on Argoverse 2. While the static Nerfacto baseline has the weakest performance, the dynamic variant Nerfacto + Time improves only marginally upon it. The state-of-the-art method SUDS exhibits stronger view synthesis results but takes more than \\(3\\times\\) longer to train. Our method outperforms all methods across all metrics and exhibits competitive training speed." + }, + { + "type": "table", + "bbox": [ + 0.157, + 0.214, + 0.816, + 0.434 + ], + "angle": 0, + "content": "
MethodKITTI [75%]KITTI [50%]KITTI [25%]
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [31]18.560.5570.55419.120.5870.49718.610.5700.510
NeRF + Time21.010.6120.49221.340.6350.44819.550.5860.505
NSG [35]21.530.6730.25421.260.6590.26620.000.6320.281
Nerfacto + Emb.22.750.8010.15622.380.7930.16021.240.7580.178
Nerfacto + Emb. + Time23.190.8040.15523.180.8030.15521.980.7770.172
SUDS [58]22.770.7970.17123.120.8210.13520.760.7470.198
Ours28.380.9070.05227.510.8980.05526.510.8870.060
MethodVKITTI2 [75%]VKITTI2 [50%]VKITTI2 [25%]
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [31]18.670.5480.63418.580.5440.63518.170.5370.644
NeRF + Time19.030.5740.58718.900.5650.61018.040.5450.626
NSG [35]23.410.6890.31723.230.6790.32521.290.6660.317
Nerfacto + Emb.22.150.8470.14521.880.8430.14821.280.8270.155
Nerfacto + Emb. + Time22.110.8490.14421.780.8440.14721.000.8250.157
SUDS [58]23.870.8460.15023.780.8510.14222.180.8290.160
Ours29.730.9120.06529.190.9060.06628.290.9010.067
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.437, + 0.895, + 0.482 + ], + "angle": 0, + "content": "Table 2. Novel View Synthesis on KITTI and VKITTI2. We compare our method to prior art, following the experimental protocol in [58]. We test the view synthesis performance of the methods with varying fractions of training views and observe that fewer training views generally result in lower performance. Our method outperforms previous works by a large margin on all settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.47, + 0.641 + ], + "angle": 0, + "content": "Hierarchical pose optimization. Alongside scene geometry, we optimize edges \\(e_{v_s^t v_r}\\) and \\(e_{v_o v_s^t}\\) in our graph, i.e. we refine ego-vehicle poses \\(\\mathbf{P}_s^t\\in \\mathbf{SE}(3)\\) and object poses \\(\\xi_o^t\\in \\mathbf{SE}(3)\\). In particular, we optimize for pose residuals \\(\\delta \\mathbf{P}_s^t\\in \\mathfrak{se}(3)\\) and \\(\\delta \\xi_o^t\\in \\mathfrak{se}(2)\\). We constrain the object pose residual to \\(\\mathfrak{se}(2)\\) since objects are usually upright and move along the ground plane. Given each residual, we update the ego-vehicle pose with \\(\\hat{\\mathbf{P}}_s^t = \\expmap(\\delta \\mathbf{P}_s^t)\\mathbf{P}_s^t\\) and the object pose with \\(\\hat{\\xi}_o^t = \\expmap(\\delta \\xi_o^t)_{\\mathbf{SE}(2)\\to \\mathbf{SE}(3)}\\xi_o^t\\), where \\(\\expmap(\\cdot)\\) is the exponential map of each lie group." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.471, + 0.822 + ], + "angle": 0, + "content": "Compared to naively optimizing camera and object poses, optimizing the edges along our scene graph has two key advantages. First, we leverage multi-camera constraints, i.e. we keep the camera extrinsics \\(\\mathbf{T}_c\\) fixed and optimize the ego-vehicle poses \\(\\mathbf{P}_s^t\\) only. This is in contrast to prior art that generally treats each camera pose as independent. Second, the residual \\(\\delta \\mathbf{P}_s^t\\) propagates to the object poses since they are defined w.r.t. the ego-vehicle coordinate frame instead of the global world frame. Given that optimizing a radiance field as well as camera and object poses jointly is notoriously difficult [21], these constraints are vital to view synthesis quality (see Tab. 6)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.21, + 0.851 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Datasets. To evaluate against competing methods on our proposed benchmark on Argoverse 2 [61], we hold out every 10th sample in uniform time intervals where a sample" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.579 + ], + "angle": 0, + "content": "corresponds to seven ring-camera images. To compare our methods against prior art on KITTI [14] and VKITTI2 [4], we follow the experimental protocol and data splits in [58]. We use the provided 3D bounding box annotations for all datasets. We provide results using an off-the-shelf 3D tracker [68] in the supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Metrics. Following [58], we measure image synthesis quality with PSNR, SSIM [59], and LPIPS (AlexNet) [69]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Implementation details. We implement our method in PyTorch [39], accelerating the time-consuming ray-node intersection with a custom CUDA implementation. We train our model on a single RTX 3090 GPU for 250,000 steps on Argovvere 2 and 100,000 steps on KITTI and VKITTI2, with 8192 rays per batch. All model parameters are optimized using Adam [17] with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\), and an exponential decay learning rate schedule: from \\(10^{-5}\\) to \\(10^{-6}\\) for pose parameters, and from \\(10^{-3}\\) to \\(10^{-4}\\) for all others. In order to counter pose drift, we further apply weight decay with a factor \\(10^{-2}\\) to \\(\\delta \\mathbf{P}\\) and \\(\\delta \\xi\\). The static radiance field \\(\\phi\\) is trained from scratch, while \\(\\psi\\) is initialized from a semantic prior [6] following [18]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare our method against prior work in dynamic outdoor scene representation, i.e. SUDS [58] and NSG [35]. Further, we include results obtained by augmenting Nerfacto [54], the closest state-of-the-art NeRF architecture to ours, with components designed to handle multicapture reconstruction and scene dynamics. In particular," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21130" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.09, + 0.468, + 0.125 + ], + "angle": 0, + "content": "
SRN [49]NeRF [31]NeRF + TimeNSG [35]PNF [18]SUDS [58]Ours
PSNR ↑18.8323.3424.1826.6627.4828.3129.36
SSIM ↑0.5900.6620.6770.8060.8700.8760.911
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.129, + 0.468, + 0.171 + ], + "angle": 0, + "content": "Table 3. Image reconstruction on KITTI. We outperform prior art in image reconstruction, i.e. our method can better fit the training views. We follow the experimental protocol in [18, 35, 58]." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.174, + 0.47, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.469, + 0.316 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on KITTI. With only \\(25\\%\\) of views (approx. 15-20) for training, we can still synthesize sharp and realistic novel views with dynamic objects rendered at high quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.469, + 0.415 + ], + "angle": 0, + "content": "we consider two variants: \"Nerfacto + Emb.\", where we incorporate our appearance embedding \\(\\mathbf{A}_s\\mathcal{F}(t)\\) and the expected depth loss; and \"Nerfacto + Emb. + Time\", where we further include time modeling with 4D hash tables following [38, 63]. For both Nerfacto variants, we set hash table and MLP sizes to be aligned with our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.425, + 0.348, + 0.44 + ], + "angle": 0, + "content": "6.1. Comparison to state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.469, + 0.612 + ], + "angle": 0, + "content": "We first compare to prior work on our proposed benchmark in Tab 1. We observe that the static Nerfacto + Emb. has the weakest performance, which slightly increases by adding time modeling in Nerfacto + Emb. + Time. Meanwhile, the current state-of-the-art method SUDS [58] outperforms the Nerfacto variants, while being much slower to train. Our method performs the best on all metrics. The improvement is particularly pronounced in the perceptual quality metrics (SSIM and LPIPS). The training speed of our method is competitive to Nerfacto + Emb. + Time and more than \\(3 \\times\\) faster than SUDS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.84 + ], + "angle": 0, + "content": "In addition, we show a qualitative comparison in Fig. 7. We observe major differences in both the rendered images and depth maps. In particular, all other methods struggle to recover the dynamic objects in the scene, while our method produces realistic renderings and accurate depth maps for both static and dynamic areas. We also observe that, thanks to our transient geometry embedding \\(\\mathbf{G}_s\\mathcal{F}(t)\\), we are able to accurately recover the geometry of the trees and their leaves which notably are not present in every sequence of the area that is reconstructed due to seasonal changes. At the same time, other methods struggle to recover those details. Specifically, other methods exhibit artifacts in the depth maps and degraded rendering quality of the trees left and right of the street in columns five and six of Fig. 7, while our method produces accurate color and depth renderings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Next, we evaluate our method on established benchmarks, namely KITTI and VKITTI2, in Tab. 2 following the experimental protocol in [58]. In particular, we compare novel view synthesis quality at different fractions of" + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.09, + 0.848, + 0.145 + ], + "angle": 0, + "content": "
AsF(t)GsF(t)PSNR ↑SSIM ↑LPIPS ↓
--19.700.6530.588
✓-22.220.6700.546
✓✓22.490.6710.541
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.149, + 0.892, + 0.219 + ], + "angle": 0, + "content": "Table 4. Ablation study on graph structure. We show that using the multi-level graph structure of sequences and objects is crucial, i.e. omitting sequence vectors \\(\\omega_{s}^{t}\\) results in degraded quality since there is no conditioning on scene-specific appearance. Combining appearance and transient geometry embeddings performs best." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.285 + ], + "angle": 0, + "content": "
SamplingSamples per rayPSNR ↑SSIM ↑LPIPS ↓Rays / sec.
Uniform [18]19225.720.7300.45628K
Uniform [18]102425.840.7340.4494K
Separate [35]1024+64+(32+64)26.650.7620.3512.5K
Ours1024+6427.070.7590.36230K
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.29, + 0.892, + 0.346 + ], + "angle": 0, + "content": "Table 5. Ray Sampling schemes. Our composite ray sampling is about \\(12 \\times\\) more efficient to train than separate ray sampling [35] with comparable performance. Uniform sampling [18] exhibits lower performance and is slow when sampled more densely." + }, + { + "type": "table", + "bbox": [ + 0.556, + 0.348, + 0.836, + 0.404 + ], + "angle": 0, + "content": "
Pose optimizationPSNR ↑SSIM ↑LPIPS ↓
-22.490.6710.541
Naive21.280.6630.519
Hierarchical (Ours)22.290.6780.523
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.477 + ], + "angle": 0, + "content": "Table 6. Pose optimization. Compared to naively optimizing camera and object poses, optimizing vehicle and object poses hierarchically on the edges of our graph benefits view synthesis quality more consistently, i.e. both in SSIM and LPIPS. Note that PSNR is very sensitive to pose drift, and thus does not improve." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.653 + ], + "angle": 0, + "content": "training views, i.e. testing different levels of view supervision. We run our method and our baselines and observe substantially better view synthesis quality for our method, both compared to our baselines and prior works. All methods exhibit a similar performance drop when the fraction of training views decreases. We illustrate the rendering quality of our method with only 15-20 training views in Fig. 6. Finally, we also report the results on the task of image reconstruction, i.e. reconstruction of images seen during training, following [18, 35, 58] in Tab. 3 on the KITTI dataset. We significantly outperform previous works in both metrics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.661, + 0.68 + ], + "angle": 0, + "content": "6.2. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We verify our design through ablation studies. We perform these on the residential area of our benchmark unless otherwise noted. First, we ablate on the multi-level structure of our scene graph. In particular, in Tab. 4, we ablate the latent codes of the sequence nodes, reducing to a representation similar to [35] and observe that the quality of the synthesized views drops significantly. In contrast, our scene graph representation achieves better view synthesis quality, i.e. achieves almost three points higher PSNR. Further, adding our transient geometry embeddings in addition to sequence-level appearance embeddings improves the view synthesis quality. See supplemental material for a qualitative comparison. This adds to the fact that through our multi-level graph structure, we show the ability to modulate car appear" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21131" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.89, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.489, + 0.893, + 0.532 + ], + "angle": 0, + "content": "Figure 7. Qualitative results on Argoverse 2. While prior art struggles with dynamic actors in the scene, our representation renders both realistic novel views and plausible depth maps. Further, our method models complex transient geometry subject to, e.g., seasonal changes such as tree leaves. Note that prior art produces depth artifacts and exhibits degraded view quality in those regions (columns 5 and 6)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.544, + 0.373, + 0.559 + ], + "angle": 0, + "content": "ance through sequence appearance in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.47, + 0.79 + ], + "angle": 0, + "content": "In Tab. 5, we compare our composite ray sampling scheme to uniformly sampled rays as in [18] and rays sampled separately per node as in [35]. We compare the schemes under the assumption that all other model parameters are equal, implementing them in our method. Ours is about \\(12 \\times\\) more efficient to train than densely sampling a ray in uniform intervals or separately sampling the ray per node while producing similar results to the latter and being significantly better than the former. Sparsely sampling a ray at uniform intervals is similarly fast as our method, but yields degraded view synthesis quality. Note that we run this ablation study on only a single sequence of the residential area in our benchmark since densely sampling the rays and separately sampling the rays for all nodes are prohibitively expensive to train in large-scale urban areas." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Finally, in Tab. 6, we compare our hierarchical pose optimization to naive camera pose optimization employed in previous works [21, 54]. While naive camera pose optimization degrades the results significantly in pixel-wise metrics, our hierarchical pose optimization improves the SSIM and maintain a comparable PSNR. Meanwhile, our hierarchical pose optimization exhibits similar gains to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.893, + 0.635 + ], + "angle": 0, + "content": "naive pose optimization in terms of the perceptual LPIPS metric. This shows that our pose optimization mitigates pose drift while also enabling a more accurate reconstruction. Pose drift usually causes a misalignment between the evaluation viewpoint and scene geometry, which degrades pixel-wise metrics in particular." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.62, + 0.665 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We introduce a novel multi-level scene graph representation for radiance field reconstruction in dynamic urban environments that scales to large geographic areas with more than ten thousand images from dozens of sequences and hundreds of dynamic objects. We train our representation with an efficient composite ray sampling and rendering scheme and introduce latent variables that enable modeling complex phenomena like varying environmental conditions and transient geometry present across different vehicle captures. We leverage our representation to refine camera and object poses hierarchically using multi-camera constraints. Finally, we propose a new view synthesis benchmark for dynamic urban driving scenarios. Our approach yields substantially improved results compared to prior art while allowing for flexible de- and recomposition of the scene." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21132" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Iro Armeni, Zhi-Yang He, JunYoung Gwak, Amir R Zamir, Martin Fischer, Jitendra Malik, and Silvio Savarese. 3d scene graph: A structure for unified semantics, 3d space, and camera. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.211 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.24 + ], + "angle": 0, + "content": "[3] Jules Bloomenthal and Brian Wyvill. Introduction to implicit surfaces. 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.47, + 0.269 + ], + "angle": 0, + "content": "[4] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual kitti 2, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[5] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.47, + 0.408 + ], + "angle": 0, + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.47, + 0.451 + ], + "angle": 0, + "content": "[7] Steve Cunningham and Michael J Bailey. Lessons from scene graphs: using scene graphs to teach hierarchical modeling. Computers & Graphics, 25(4):703-711, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.506 + ], + "angle": 0, + "content": "[8] Nathaniel Fairfield, George A. Kantor, and David S. Wettergreen. Real-time slam with octree evidence grids for exploration in underwater tunnels. Journal of Field Robotics, 24, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[9] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.47, + 0.618 + ], + "angle": 0, + "content": "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.469, + 0.66 + ], + "angle": 0, + "content": "[11] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[12] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[13] Yiming Gao, Yan-Pei Cao, and Ying Shan. Surfelnerf: Neural surfel radiance fields for online photorealistic reconstruction of indoor scenes. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[15] Michael Kaess. Simultaneous localization and mapping with infinite planes. 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 4605-4611, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[18] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[19] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[20] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[21] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[22] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.411 + ], + "angle": 0, + "content": "[23] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[24] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[25] Yan Lu and Dezhen Song. Visual navigation using heterogeneous landmarks and unsupervised geometric constraints. IEEE Transactions on Robotics, 31:736-749, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[26] Jonathon Luiten, Tobias Fischer, and Bastian Leibe. Track to reconstruct and reconstruct to track. IEEE Robotics and Automation Letters, 5(2):1803-1810, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[28] Alexander Majercik, Cyril Crassin, Peter Shirley, and Morgan McGuire. A ray-box intersection algorithm and efficient dynamic voxel rendering. Journal of Computer Graphics Techniques Vol, 7(3):66-81, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[29] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[30] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[31] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[32] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21133" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[33] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[34] OpenStreetMap contributors. Planet dump retrieved from https://planet.osm.org . https : //www.openstreetmap.org, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.234 + ], + "angle": 0, + "content": "[35] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In CVPR, 2021. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.236, + 0.468, + 0.275 + ], + "angle": 0, + "content": "[36] Byeongjun Park and Changick Kim. Point-dynrf: Point-based dynamic radiance fields from a monocular video, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.468, + 0.347 + ], + "angle": 0, + "content": "[37] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[38] Sungheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, and Nahiyup Kang. Temporal interpolation is all you need for dynamic neural radiance fields. In CVPR, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.406, + 0.468, + 0.473 + ], + "angle": 0, + "content": "[39] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.476, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[40] Marc Pollefeys, David Nister, Jan-Michael Frahm, Amir Akbarzadeh, Philippos Mordohai, Brian Clipp, Chris Engels, David Gallup, Seon Joo Kim, Paul C. Merrell, C. Salmi, Sudipta N. Sinha, B. Talton, Liang Wang, Qingxiong Yang, Henrik Stewénius, Ruigang Yang, Greg Welch, and Herman Towles. Detailed real-time urban 3d reconstruction from video. IJCV, 78:143-167, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.575, + 0.468, + 0.616 + ], + "angle": 0, + "content": "[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.617, + 0.468, + 0.671 + ], + "angle": 0, + "content": "[42] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.674, + 0.468, + 0.701 + ], + "angle": 0, + "content": "[43] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.703, + 0.468, + 0.758 + ], + "angle": 0, + "content": "[44] Antoni Rosinol, Arjun Gupta, Marcus Abate, Jingnan Shi, and Luca Carlone. 3d dynamic scene graphs: Actionable spatial perception with places, objects, and humans. arXiv preprint arXiv:2002.06289, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.76, + 0.468, + 0.801 + ], + "angle": 0, + "content": "[45] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.803, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[46] Renato F Salas-Moreno, Richard A Newcombe, Hauke Strasdat, Paul HJ Kelly, and Andrew J Davison. Slam++: Simultaneous localisation and mapping at the level of objects. In CVPR, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[47] Aron Schmied, Tobias Fischer, Martin Danelljan, Marc Pollefeys, and Fisher Yu. R3d3: Dense 3d reconstruction of dynamic scenes from multiple cameras. In ICCV, 2023." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[48] Chen Shen, James F. O'Brien, and Jonathan R. Shewchuk. Interpolating and approximating implicit surfaces from polygon soup. In ACM SIGGRAPH 2004 Papers, page 896-904, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[49] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. NeurIPS, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[50] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[51] Henry Sowizral. Scene graphs in the new millennium. IEEE Computer Graphics and Applications, 20(1):56-57, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[52] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[53] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[54] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[56] Shubham Tulsiani, Saurabh Gupta, David F Fouhey, Alexei A Efros, and Jitendra Malik. Factoring shape, pose, and layout from the 2d image of a 3d scene. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[57] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[58] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In CVPR, 2023. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[59] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[60] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[61] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21134" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "self-driving perception and forecasting. arXiv preprint arXiv:2301.00493, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[62] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D' 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[63] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In CVPR, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[64] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.304 + ], + "angle": 0, + "content": "[65] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[66] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[67] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[68] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.49 + ], + "angle": 0, + "content": "[69] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21135" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/10ec9eda-ceb6-4b88-aa59-8a3b78a66785_origin.pdf b/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/10ec9eda-ceb6-4b88-aa59-8a3b78a66785_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fe57f642562a559ebc8f772757e9039eb15b1254 --- /dev/null +++ b/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/10ec9eda-ceb6-4b88-aa59-8a3b78a66785_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c1e279a5bb934583c3cede7d21509a7ed9a3345cd248027dfe7ca0cf3d05db7 +size 4010589 diff --git a/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/full.md b/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/full.md new file mode 100644 index 0000000000000000000000000000000000000000..798f378b4576e5cba1e6aeb100da7bece744b439 --- /dev/null +++ b/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/full.md @@ -0,0 +1,384 @@ +# Multi-Level Neural Scene Graphs for Dynamic Urban Environments + +Tobias Fischer $^{1}$ Lorenzo Porzi $^{2}$ Samuel Rota Bulò $^{2}$ + +Marc Pollefeys1 Peter Kontschieder2 + +$^{1}$ ETH Zürich $^{2}$ Meta Reality Labs + +https://tobiasfshr.github.io/pub/ml-nsg/ + +![](images/c1089944d4ef1d7a751ce36bcdde3419dad0203b973446f27dba488ce7b93109.jpg) + +![](images/4efd39ef8f453b62cd2fbfa0f7f83c46b1af373aa59ecbbef8d58a61b8cd68d5.jpg) + +![](images/5ec4d11e7f858af17263a92b2c31121fe8384c43b33c6c751909a21267269ab1.jpg) +Input + +![](images/f3ae6cb64d0fb026a45d3dd3cb46acf62d0c1bf05edb926c42ac5bf9d603bff6.jpg) +Multi-Level Neural Scene Graph + +![](images/27fe8028c93540aed69f136ab15fb7fdd7312b9260ea1028764672189019f51d.jpg) + +![](images/f686443d3ca75b8d9970d7c507f8c8aedd2aeaa29415ca17eadac66cb094e41d.jpg) +Figure 1. Overview. We represent sequences captured from moving vehicles in a shared geographic area with a multi-level scene graph. Each dynamic object $v_{o}$ is associated with a sequence node $v_{s}^{t}$ and time $t$ . The sequence nodes are registered in a common world frame at the root node $v_{r}$ through the vehicle poses $\mathbf{P}_{s}^{t}$ , while the dynamic objects are localized w.r.t. the sequence node with pose $\xi_{o}^{t}$ . Each camera $c$ is associated with an ego-vehicle position, i.e. node $v_{s}^{t}$ , through the extrinsic calibration $\mathbf{T}_{c}$ . The sequence and object nodes hold latent codes $\omega$ that condition the radiance field, synthesizing novel views in various conditions with distinct dynamic objects. + +![](images/abacc3bd5b0ff214316c63989fb099052008ddc35f79c1b74cbd6cdd9b6697f9.jpg) + +![](images/1b6b993264d94d48c329b911f10c0b384bb711a871081124b7777fcde8ebc5c2.jpg) +Novel Views + +![](images/a25c0b68ad006e12ee21a69730b8a96cd8a93d1e8e69a017082c994e298e9c1a.jpg) + +![](images/0992a2a06b4105a7447221a43e570a7bc0cc34ec362b568072beed7e664700f1.jpg) + +# Abstract + +We estimate the radiance field of large-scale dynamic areas from multiple vehicle captures under varying environmental conditions. Previous works in this domain are either restricted to static environments, do not scale to more than a single short video, or struggle to separately represent dynamic object instances. To this end, we present a novel, decomposable radiance field approach for dynamic urban environments. We propose a multi-level neural scene graph representation that scales to thousands of images from dozens of sequences with hundreds of fast-moving objects. To enable efficient training and rendering of our representation, we develop a fast composite ray sampling and rendering scheme. To test our approach in urban driving scenarios, we introduce a new, novel view synthesis benchmark. We show that our approach outperforms prior art by a significant margin on both established and our proposed benchmark while being faster in training and rendering. + +# 1. Introduction + +Estimating the radiance field of a dynamic urban environment from data collected by sensor-equipped vehicles is a core challenge in closed-loop simulation for robotics and in mixed reality. It is particularly relevant for applications like autonomous driving and city-scale mapping, where capture vehicles can be frequently deployed. The growing amount of available data provides great opportunities for creating up-to-date digital twins of entire cities but also poses unique challenges as increasingly heterogeneous data sources must be processed. In particular, limited scene coverage, different lighting, weather, seasonal conditions, and varying geometry due to distinct dynamic and transient objects make radiance field reconstruction of dynamic urban environments extremely challenging. + +Recently, Neural Radiance Fields (NeRFs) have enabled significant progress in achieving realistic novel view synthesis from a set of input views [2, 31-33, 49]. These methods represent a static 3D scene with fully implicit [30, 31, 49] or low-level explicit structures [10, 13, 16, 23, 32, 43, + +52, 65] such as voxels, points, surfels, meshes or 3D gaus-. sians. In parallel, implicit [19, 37, 50, 62, 63] and low-level explicit [9, 27, 36, 38] neural representations for dynamic 4D scenes have been investigated. However, fewer works have focused on decomposing scenes into higher-level entities [18, 26, 46, 56]. In computer graphics and 3D mapping, scene graphs have been used to represent complex scenes in a multi-level hierarchy [1, 44, 51]. In view synthesis, Ost et al. [35] apply this concept by describing actors in a dynamic scene as entities of a scene graph. However, their representation lacks the ability of [1, 44] to represent scenes at multiple levels, thereby limiting their representation to short, single sequences. + +While earlier methods for view synthesis focused on object-centric scenes with controlled camera trajectories, recent works move towards radiance field reconstruction of large-scale environments from in-the-wild captures. Among these, many methods focus on static environments, removing dynamic actors from the input data [22, 24, 29, 42, 45, 53, 57]. A few works explicitly model dynamic actors [18, 35, 67], but either do not scale to more than a single, short video [18, 35, 67], or struggle to accurately represent individual object instances [58]. This complicates the evaluation of these methods for real-world applications because existing benchmarks neither scale to large urban areas [18, 35] nor reflect realistic capturing conditions [12]. + +To address these issues, we propose a multi-level neural scene graph representation that spans large geographic areas with hundreds of dynamic objects. In contrast to previous works, our multi-level scene graph formulation allows us to distinguish and represent dynamic object instances effectively and further to represent a scene under varying conditions. To make our representation viable for large-scale dynamic environments, we develop a composite ray sampling and rendering scheme that enables fast training and rendering of our method. To test this hypothesis, we introduce a benchmark for radiance field reconstruction in dynamic urban environments based on [61]. We fuse data from dozens of vehicle captures under varying conditions amounting to more than ten thousand images with several hundreds of dynamic objects per reconstructed area. We summarize our contributions as follows: + +- We propose a multi-level neural scene graph formulation that scales to dozens of sequences with hundreds of fast-moving objects under varying environmental conditions. +- We develop an efficient composite ray sampling and rendering scheme that enables fast training and rendering of our representation. +- We present a benchmark that provides a realistic, application-driven evaluation of radiance field reconstruction in dynamic urban environments. + +We show state-of-the-art view synthesis results on both established benchmarks [4, 14] and our proposed benchmark. + +# 2. Related Work + +3D and 4D scene representations. Finding the right scene representation is a core issue in 3D computer vision and graphics [5]. Over the years, a wide variety of options have been explored [1, 3, 8, 15, 25, 30, 40, 46-48], and, more recently, neural rendering [33] has been used to enable a new generation of scene models that support photo-realistic novel view synthesis. Scene representations for neural rendering can be roughly classified as "implicit" [30, 31, 49], which store most of the information in the weights of a neural network, or "explicit", which use low-level spatial primitives such as voxels [10, 23, 32, 52], points [65], surfels [13], meshes [43], or 3D gaussians [16]. + +Similarly, different 4D dynamic scene representations for view synthesis have been investigated, including both implicit [11, 19, 37, 41, 55, 58, 62, 63] and explicit [9, 27, 36, 67] approaches. Dynamics are generally modeled as deformations of a canonical volume [9, 37, 41, 55, 62], a separate scene motion function [11, 19, 20, 58, 63], or rigid transformations of local geometric primitives [27]. + +Another line of work investigates the decomposition of scenes into higher-level entities [18, 26, 46, 56]. To express the composition of different entities into a complex scene, classical computer graphics literature [7] uses scene graphs. In particular, entities are represented as nodes in a hierarchical graph and are connected through edges defined by coordinate frame transformations. Thus, global transformations can be acquired by traversing the graph from its root node. For indoor 3D mapping, this concept was proposed by Armeni et al. [1] to represent static scenes at multiple levels of hierarchy, i.e. buildings, rooms and objects, and was later extended to dynamic scenes by Rosinol et al. [44]. Recently, Ost et al. [35] have revisited this concept for view synthesis, describing multi-object scenes with a graph that represents the dynamic actors in the scene. However, their representation lacks the ability of [1, 44] to represent scenes at multiple levels of hierarchy and is thus inherently limited to single, short video clips. On the contrary, we present a scalable, multi-level scene graph representation that spans large geographic areas with hundreds of dynamic objects. + +Representing large-scale urban scenes. Compared to controlled captures of small scenes, in-the-wild captures of large-scale scenes pose distinct challenges. Limited viewpoint coverage, inaccurate camera poses, far-away buildings and sky, fast-moving objects, complex lighting, and auto exposure make radiance field reconstruction challenging. Therefore, previous works aid the reconstruction by using depth priors from e.g. LiDAR, refining camera parameters, adding information about camera exposure, and using specialized sky and light modeling components [22, 29, 42, 45, 53, 60, 64]. While many of these works simply remove dynamic actors, some methods explicitly model scene dynamics [18, 35, 58, 67]. However, + +![](images/cf6ac8ee88fa08ca1d431f0e43f8a28257acf473308ae33e24cb79da8e1b5dde.jpg) +Figure 2. Ego-vehicle trajectories of our benchmark. We show the residential (left) and the downtown (right) areas, trajectories superimposed on 2D maps obtained from OpenStreetMap [34]. + +these methods either struggle to scale to more than a single sequence [18, 35, 67] or to accurately represent and distinguish dynamic actors [58]. To address these issues, we present a decomposed, scalable scene representation for dynamic urban environments. + +Efficient scene rendering. Especially for large-scale scenes, the efficiency of a scene representation in training and inference is crucial. To alleviate the burden of ray traversal in volumetric rendering, many techniques for efficient sampling [2, 10, 32, 57] have been proposed. Recently, researchers have exploited more efficient forms of rendering, e.g. rasterization of meshes [22, 24] or other 3D primitives [16]. While these approaches are focused on static scenes, we extend the approach of [2] to our scene graph representation and thus to dynamic scenes. + +# 3. Data + +We investigate how to synthesize novel views of a dynamic urban environment from a set of captures taken from moving vehicles. Specifically, we are interested in captures spanning the same geographic area under varying conditions. This data presents unique challenges due to limited viewpoint coverage, different lighting, weather, and season, and even varying geometry due to distinct dynamic and transient objects. + +Previous benchmarks in this area [18, 35] are limited to short, single video clips with simple ego trajectories and a few dynamic objects. Therefore, we present a benchmark that better reflects the aforementioned challenges. We base our benchmark on Argoverse 2 [61] which provides a rich set of captures from a fleet of vehicles deployed in multiple US cities that span different weather, season, and time of day. The vehicles are equipped with a surround-view camera rig with seven global shutter cameras, a LiDAR sensor, and a GPS. Furthermore, the timings of the different sensors are provided, so that one can relate the LiDAR-based 3D bounding box annotations to camera timestamps. + +We leverage the GPS information to globally align the sequences and to identify regions that we are interested in mapping. We then extract the specific vehicle captures. + +![](images/735ace0c8070dbc357097038deff46d0f5361f490277175df7f5127c9df5141f.jpg) +Figure 3. Sequence alignment visualization. The initial GPS-based alignment is imprecise, as evidenced by the duplicated structures in the overlaid LiDAR point clouds (left). After our ICP alignment, the area is well reconstructed (middle) according to its real geometry (right, from Argoverse 2 [61]). + +Since GPS-based localization accuracy is only coarsely precise in urban areas, we align the captures via a global, offline iterative-closest-point (ICP) procedure applied to the LiDAR point clouds of all sequences to achieve satisfactory alignment for novel view synthesis purposes (see Fig. 3). + +Following this procedure, we build a benchmark that enables real-world evaluation of novel view synthesis from diverse vehicle captures. It is composed of 37 vehicle captures split into two geographic regions as illustrated in Fig. 2. The regions cover a residential and a downtown area to resemble the different characteristics of urban environments. The residential area spans 14 captures with 10493 training and 1162 testing images with more than 700 distinct moving objects. The downtown area spans 23 captures with 16933 training and 1876 testing images with more than 600 distinct moving objects. + +# 4. Method + +Problem setup. We are given a set of sequences $S$ captured from moving vehicles in different conditions. Each sequence $s \in S$ consists of a set of images taken from cameras $C_s$ mounted on the vehicle at different timesteps indexed by $T_s$ . We assume the vehicle's sensors are calibrated with respect to a common vehicle and a global world frame. In particular, we assume given sensor extrinsic $\mathbf{T}_c = [\mathbf{R}_c|\mathbf{t}_c] \in \mathbf{SE}(3)$ for each camera $c \in C = \bigcup_{s} C_s$ , and ego-vehicle poses $\mathbf{P}_s^t = [\mathbf{R}_s^t|\mathbf{t}_s^t] \in \mathbf{SE}(3)$ for each timestamp $t \in T_s$ and sequence $s \in S$ . Furthermore, we assume all cameras to have known intrinsics $\mathbf{K}_c$ . Each sequence $s$ entails a set $O_s$ of dynamic objects. For each object $o \in O_s$ , we assume to have an estimated 3D bounding box track $\mathcal{T}_o$ consisting of object poses $\{\xi_o^{t_0}, \dots, \xi_o^{t_n}\} \subset \mathbf{SE}(3)$ w.r.t. the ego-vehicle frame, where $t_i \in T_s$ , and the 3D object dimensions $\mathbf{s}_o \in \mathbb{R}_{>0}^3$ . The images span a common geographic area for which we would like to estimate a radiance field + +$$ +f _ {\theta} (\mathbf {x}, \mathbf {d}, t, s) = (\sigma (\mathbf {x}, t, s), \mathbf {c} (\mathbf {x}, \mathbf {d}, t, s)) \tag {1} +$$ + +that outputs volume density $\sigma \in \mathbb{R}_{\geq 0}$ and color $\mathbf{c} \in [0,1]^3$ conditioned on 3D location $\mathbf{x}$ , viewing direction $\mathbf{d}$ , time $t$ and sequence $s$ . + +# 4.1. Multi-Level Neural Scene Graph + +Overview. We illustrate our representation in Fig. 1. We decompose the scene into a graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ where the set of nodes $\nu$ is composed by a root node $v_{r}$ that defines the global coordinate system, camera nodes $\{v_{c}\}_{c\in C}$ , and, for each sequence $s\in S$ , sequence nodes $\{v_s^t\}_{t\in T_s}$ and dynamic object nodes $\{v_o\}_{o\in O_s}$ . The nodes are connected through oriented edges $e\in \mathcal{E}$ that represent rigid transformations between the coordinate frames of the nodes, consistently with the edge direction. In addition, each node $v\in \mathcal{V}$ can have an associated latent vector $\omega$ . Given the scene graph $\mathcal{G}$ , we model $f_{\theta}$ using two radiance fields, namely $\phi$ for largely static and $\psi$ for highly dynamic scene parts. We use the latent vectors $\omega$ to condition the radiance fields. In particular, we use latent vectors $\omega_s^t$ of a sequence node $v_s^t$ to condition $\phi$ : + +$$ +\phi (\mathbf {x}, \mathbf {d}, \omega_ {s} ^ {t}) = \left(\sigma_ {\phi} \left(\mathbf {x}, \omega_ {s} ^ {t}\right), \mathbf {c} _ {\phi} \left(\mathbf {x}, \mathbf {d}, \omega_ {s} ^ {t}\right)\right). \tag {2} +$$ + +Since the graph $\mathcal{G}$ has multiple levels, a node associated with a dynamic object will naturally fall into the sequence $s$ it appears in. Therefore, the radiance field $\psi$ is conditioned on latent vectors $\omega_{s}^{t}$ and $\omega_{o}$ : + +$$ +\psi (\mathbf {x}, \mathbf {d}, \omega_ {s} ^ {t}, \omega_ {o}) = (\sigma_ {\psi} (\mathbf {x}, \omega_ {s} ^ {t}, \omega_ {o}), \mathbf {c} _ {\psi} (\mathbf {x}, \mathbf {d}, \omega_ {s} ^ {t}, \omega_ {o})) (3) +$$ + +of the corresponding sequence and object nodes $v_{s}^{t}$ and $v_{o}$ . Appearance and geometry variation. Reconstructing an environment from multiple captures is challenging from two perspectives: in addition to varying dynamic objects, there is i) varying appearance across captures, and ii) slow-moving or static transient geometry such as tree leaves or construction sites. Since both transient geometry and appearance can vary across captures, but are usually smooth within a sequence, we model these phenomena as smooth functions over time + +$$ +\omega_ {s} ^ {t} = \left[ \mathbf {A} _ {s} \mathcal {F} (t), \mathbf {G} _ {s} \mathcal {F} (t) \right] \tag {4} +$$ + +where $\mathbf{A}_s$ is an appearance matrix, $\mathbf{G}_s$ is a transient geometry matrix and $\mathcal{F}(\cdot)$ is a 1D basis function of sines and cosines with linearly increasing frequencies at log-scale [58, 66]. We normalize time $t$ into the interval $[-1,1]$ using the maximum sequence length $\max_{s\in S}|T_s|$ . Crucially, this allows to model near-static, but sequence-specific regions of the input, as well as appearance changes due to e.g. auto-exposure. The degree of variation across time can be controlled by the number of frequencies. + +Sequence nodes. The sequence nodes $v_{s}^{t}$ are connected to the root node $v_{r}$ with an edge $e_{v_s^t v_r} = \mathbf{P}_s^t$ , i.e. the sequences $S$ share a common global world frame. Each sequence node holds the latent vector $\omega_{s}^{t}$ that conditions the radiance field $\phi$ . We model $\phi$ with a multi-scale 3D hash grid representa + +![](images/dc40757d267ad7d75e2719c0dcc3eaf408b19005a456657fd20d0e438ccfa412.jpg) +Figure 4. Modifying car appearance with scene appearance. We exchange $\omega_{s}^{t}$ for different car instances. The car's appearance in a rendered, car-centric view (top) changes according to the environmental conditions visible in the sequence $s$ (bottom). + +tion [32] and lightweight MLP heads: + +$$ +\mathbf {f} _ {\mathbf {x}} = \mathrm {H} _ {3} (\mathbf {x}) \tag {5} +$$ + +$$ +\mathbf {h} _ {\sigma_ {\phi}}, \sigma_ {\phi} = \operatorname {M L P} _ {\sigma_ {\phi}} \left(\mathbf {f} _ {\mathbf {x}}\right) \tag {6} +$$ + +$$ +\mathbf {c} _ {\phi} = \operatorname {M L P} _ {\mathbf {c} _ {\phi}} \left(\mathbf {h} _ {\sigma_ {\phi}}, \gamma_ {\mathrm {S H}} (\mathbf {d}), \mathbf {A} _ {s} \mathcal {F} (t)\right) \tag {7} +$$ + +$$ +\sigma_ {\mathbf {G}}, \mathbf {c} _ {\mathbf {G}} = \operatorname {M L P} _ {\mathbf {G}} \left(\mathbf {h} _ {\sigma_ {\phi}}, \mathbf {G} _ {s} \mathcal {F} (t)\right) \tag {8} +$$ + +where $\gamma_{\mathrm{SH}}(\cdot)$ is a spherical harmonics encoding [32]. The final colors and densities are computed as a mixture of static and transient output (analogous to Eq. 12, see supp. mat.). + +Dynamic nodes. The dynamic nodes $v_{o}$ are connected to the sequence nodes $v_{s}^{t}$ with edges $e_{v_{o}v_{s}^{t}} = \xi_{o}^{t}$ . The dynamic node $v_{o}$ associated with object $o$ holds a latent vector $\omega_{o}$ that conditions the radiance field $\psi$ . We model $\psi$ following [31] with an MLP conditioned on $\omega_{o}$ to represent different instances with the same network [35, 67] + +$$ +\mathbf {h} _ {\sigma_ {\psi}}, \sigma_ {\psi} = \operatorname {M L P} _ {\sigma_ {\psi}} \left(\gamma_ {\mathrm {P E}} (\mathbf {x}), \omega_ {o}\right) \tag {9} +$$ + +$$ +\mathbf {c} _ {\psi} = \operatorname {M L P} _ {\mathbf {c} _ {\psi}} \left(\mathbf {h} _ {\sigma_ {\psi}}, \gamma_ {\mathrm {P E}} (\mathbf {d}), \omega_ {o}, \omega_ {s} ^ {t}\right) \tag {10} +$$ + +where $\gamma_{\mathrm{PE}}(\cdot)$ is a positional encoding [31]. Note that 3D position $\mathbf{x}$ and viewing direction $\mathbf{d}$ are transformed into the local object coordinate frame with $(\mathbf{P}_s^t\xi_o^t)^{-1}\mathbf{I}_3(1 / \max (\mathbf{s}_o))$ . We condition $\psi$ on both the scene and object-dependent latent vectors. This allows us to disentangle scene-dependent appearance from the object texture and thus to transfer objects across sequences. We illustrate this process in Fig. 4. + +Camera nodes. The cameras $c \in C_s$ are connected to the sequence nodes $v_s^t$ through edges $e_{v_c v_s^t} = \mathbf{T}_c$ , i.e. the calibration of camera $c$ w.r.t. the ego-vehicle frame. This way, we can tie camera poses to a specific ego-vehicle pose. + +# 4.2. Scene Graph Rendering + +We describe how we render our scene graph $\mathcal{G}$ for a given set of rays $\mathcal{R}$ . The sampling locations along a ray $(\mathbf{r}, t, s) \in \mathcal{R}$ at time $t$ in sequence $s$ are defined as $\mathbf{r}(u) = \mathbf{o} + u\mathbf{d}$ with $\mathbf{o} = \mathbf{R}_s^t\mathbf{t}_c + \mathbf{t}_s^t$ and $\mathbf{d} = \mathbf{R}_s^t\mathbf{R}_c\mathbf{K}_c^{-1}(p_x, p_y, 1)^T$ . + +Continuous-time pose. In order to realistically render videos at different frame rates, we treat the dynamic object poses $\{\xi_o^{t_0},\dots,\xi_o^{t_n}\}$ as a continuous function of time $\xi_o(t)$ . We compute $\xi_o(t)$ by interpolating between the two nearest poses at $t_a\leq t < t_b$ to time $t$ . This allows us also to synchronize estimated object poses originating from the LiDAR measurements with the camera timestamps. + +Ray-node intersection. To render the dynamic nodes, we measure the intersection of their 3D bounding boxes with each $(\mathbf{r},t,s)\in \mathcal{R}$ . In particular, given the sequence $s$ and time $t$ the ray $\mathbf{r}$ is associated with, we first traverse the graph $\mathcal{G}$ to retrieve all relevant nodes $v_{o}$ and their 3D bounding boxes $\mathbf{b}_o^t = [\xi_o(t),\mathbf{s}_o]$ at time $t$ . Then, we transform $\mathbf{r}$ into each local node coordinate system and subsequently use AABB-ray intersection [28] to compute the entry and exit locations $u_{o}^{\mathrm{in}},u_{o}^{\mathrm{out}}$ along ray $\mathbf{r}$ . + +Composite rendering. To render the color $\hat{\mathbf{C}}$ of ray $\mathbf{r}$ in sequence $s$ at time $t$ , we use volumetric rendering [31] + +$$ +\hat {\mathbf {C}} (\mathbf {r}, t, s) = \int_ {u _ {n}} ^ {u _ {f}} U (u) \sigma (\mathbf {r} (u), t, s) \mathbf {c} (\mathbf {r} (u), \mathbf {d}, t, s) d u +$$ + +$$ +\text {w h e r e} U (u) = \exp \left(- \int_ {u _ {n}} ^ {u} \sigma \left(u ^ {\prime}\right) d u ^ {\prime}\right). \tag {11} +$$ + +We obtain density $\sigma$ and color $\mathbf{c}$ at sampling location $\mathbf{r}(u)$ as mixture of the radiance fields $\phi$ and $\psi$ + +$$ +\sigma = \sigma_ {\phi} + \sigma_ {\psi}, \mathbf {c} = \frac {\sigma_ {\phi}}{\sigma_ {\phi} + \sigma_ {\psi}} \mathbf {c} _ {\phi} + \frac {\sigma_ {\psi}}{\sigma_ {\phi} + \sigma_ {\psi}} \mathbf {c} _ {\psi}. \tag {12} +$$ + +Crucially, we set $\sigma_{\psi}$ to zero when $\mathbf{r}(u)$ does not lie within a 3D bounding box of a dynamic node given the calculated entry and exit points $u^{\mathrm{in}}, u^{\mathrm{out}}$ . + +Composite ray sampling. Instead of densely sampling the space [18, 42] or leveraging separate ray sampling mechanisms per node [35], we use a composite ray sampling strategy illustrated in Fig. 5. We extend the proposal sampling mechanism introduced in [2] by joint sampling from a computationally efficient density field $\sigma_{\mathrm{prop}}$ and dynamic nodes $v_{o}$ at time $t$ . In particular, as in Eq. 12, we represent $\sigma (\mathbf{r}(u),t,s)$ by a mixture of $\sigma_{\mathrm{prop}}(\mathbf{r}(u),\omega_s^t)$ and $\sigma_{\psi}(\mathbf{r}(u),\omega_s^t,\omega_o)$ . However, analogous to our rendering step, we constrain sampling from $\sigma_{\psi}$ to $[u^{\mathrm{in}},u^{\mathrm{out}}]$ . This allows us to skip empty space efficiently by distilling the static density $\sigma_{\phi}$ into $\sigma_{\mathrm{prop}}$ . At the same time, sparsely querying $\sigma_{\psi}$ when $\mathbf{r}(u)$ falls into a dynamic node enables us to still accurately represent the full, dynamic $\sigma$ . + +Since there are only few samples that fall into dynamic nodes when performing uniform sampling, the computational overhead in the first ray sampling iteration is negligible. In the second iteration, we apply inverse transform sampling given the CDF $F(u) = 1 - U(u)$ along ray $\mathbf{r}$ and thus the samples are concentrated at the first surface intersection. Hence, only for rays that fall in the line of sight of + +![](images/f6d9ca72be293397534baac9abb3b55225300c188e7f02bca0cef30689b151ac.jpg) +Figure 5. Composite ray sampling. If a ray intersects with an object $v_{o}$ , we sample from both proposal network $\sigma_{\mathrm{prop}}$ and radiance field $\psi$ , and $\sigma_{\mathrm{prop}}$ otherwise. We condition each with the latents $\omega$ of the respective nodes. The PDF is a mixture of all node densities that intersect with the ray. The transmittance $U$ drops at the first surface intersection (tree) where further samples will concentrate. + +![](images/5c4d60f029ed5277f343137817a12ee966bebe9263f7388265a021dda8eadf0b.jpg) + +an object surface will we sample $\sigma_{\psi}$ more than a few times. In total, we use two ray sampling iterations as in [54]. + +Space contraction. Following [2, 54], we contract the unbounded scene space into a fixed-size bounding box, normalizing the scene with bounds computed from the LiDAR point clouds and the ego-vehicle poses $\mathbf{P}$ into a unit cube. + +# 5. Optimization + +We optimize the parameters $\theta$ of the radiance field $f_{\theta}$ with + +$$ +\mathcal {L} = \sum_ {(\mathbf {r}, t, s) \in \mathcal {R}} \mathcal {L} _ {\mathrm {r g b}} (\mathbf {r}, t, s) + \lambda_ {\mathrm {d i s t}} \mathcal {L} _ {\mathrm {d i s t}} (\mathbf {r}, t, s) + \lambda_ {\mathrm {p r o p}} +$$ + +$$ +\mathcal {L} _ {\text {p r o p}} (\mathbf {r}, t, s) + \lambda_ {\text {d e p}} \mathcal {L} _ {\text {d e p}} (\mathbf {r}, t, s) + \lambda_ {\text {e n t r}} \mathcal {L} _ {\text {e n t r}} (\mathbf {r}, t, s) \tag {13} +$$ + +where $\mathcal{L}_{\mathrm{dist}}$ and $\mathcal{L}_{\mathrm{prop}}$ follow [2]. We supervise $\sigma_{\mathrm{prop}}$ with $\sigma_{\phi}$ only to learn effective composite ray sampling. + +Photometric loss. We compare the rendered training rays with their ground truth color + +$$ +\mathcal {L} _ {\mathrm {r g b}} (\mathbf {r}, t, s) = \left\| \mathbf {C} (\mathbf {r}, t, s) - \hat {\mathbf {C}} (\mathbf {r}, t, s) \right\| _ {2} ^ {2}. \tag {14} +$$ + +Expected depth loss. We render the expected depth values for each ray and compare it with the ground truth depth + +$$ +\mathcal {L} _ {\mathrm {d e p}} (\mathbf {r}, t, s) = \left\| \mathbf {D} (\mathbf {r}, t, s) - \hat {\mathbf {D}} (\mathbf {r}, t, s) \right\| _ {2} ^ {2} \tag {15} +$$ + +where the expected depth is calculated via integrating the sampling values $\hat{\mathbf{D}} (\mathbf{r},t,s) = \int_{u^{\prime}}^{u_{f}}uU(u)\sigma (\mathbf{r}(u),t,s)du$ + +Entropy regularization. While static and dynamic scene parts can overlap, i.e. inside a 3D bounding box $\mathbf{b}_o^t$ there could be a part of the street or sidewalk, their density should be strictly separated w.r.t. a single sampling location $\mathbf{r}(u)$ . We leverage an entropy regularization loss [62] to encourage clear separation between entities in $\phi$ and $\psi$ + +$$ +\mathcal {L} _ {\mathrm {e n t r}} (\mathbf {r}, t, s) = \int_ {u _ {n}} ^ {u _ {f}} \mathcal {H} \left(\frac {\sigma_ {\psi} (\mathbf {r} (u) , t , s)}{\sigma_ {\phi} (\mathbf {r} (u) , s) + \sigma_ {\psi} (\mathbf {r} (u) , t , s)}\right) d u \tag {16} +$$ + +where $\mathcal{H}(\cdot)$ is the Shannon entropy and we use $t, s$ as a replacement for the vectors $\omega_s^t, \omega_o$ for ease of notation. + +
MethodResidentialDowntownMeanTrain time (h)
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerfacto + Emb.19.830.6370.56218.050.6550.62518.940.6460.5948.0
Nerfacto + Emb. + Time20.050.6410.56218.660.6560.60319.360.6540.58313.2
SUDS [58]21.760.6590.55619.910.6650.64520.840.6620.60154.8
Ours22.290.6780.52320.010.6810.58621.150.6800.55517.2
+ +Table 1. Novel View Synthesis on Argoverse 2. While the static Nerfacto baseline has the weakest performance, the dynamic variant Nerfacto + Time improves only marginally upon it. The state-of-the-art method SUDS exhibits stronger view synthesis results but takes more than $3\times$ longer to train. Our method outperforms all methods across all metrics and exhibits competitive training speed. + +
MethodKITTI [75%]KITTI [50%]KITTI [25%]
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [31]18.560.5570.55419.120.5870.49718.610.5700.510
NeRF + Time21.010.6120.49221.340.6350.44819.550.5860.505
NSG [35]21.530.6730.25421.260.6590.26620.000.6320.281
Nerfacto + Emb.22.750.8010.15622.380.7930.16021.240.7580.178
Nerfacto + Emb. + Time23.190.8040.15523.180.8030.15521.980.7770.172
SUDS [58]22.770.7970.17123.120.8210.13520.760.7470.198
Ours28.380.9070.05227.510.8980.05526.510.8870.060
MethodVKITTI2 [75%]VKITTI2 [50%]VKITTI2 [25%]
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [31]18.670.5480.63418.580.5440.63518.170.5370.644
NeRF + Time19.030.5740.58718.900.5650.61018.040.5450.626
NSG [35]23.410.6890.31723.230.6790.32521.290.6660.317
Nerfacto + Emb.22.150.8470.14521.880.8430.14821.280.8270.155
Nerfacto + Emb. + Time22.110.8490.14421.780.8440.14721.000.8250.157
SUDS [58]23.870.8460.15023.780.8510.14222.180.8290.160
Ours29.730.9120.06529.190.9060.06628.290.9010.067
+ +Table 2. Novel View Synthesis on KITTI and VKITTI2. We compare our method to prior art, following the experimental protocol in [58]. We test the view synthesis performance of the methods with varying fractions of training views and observe that fewer training views generally result in lower performance. Our method outperforms previous works by a large margin on all settings. + +Hierarchical pose optimization. Alongside scene geometry, we optimize edges $e_{v_s^t v_r}$ and $e_{v_o v_s^t}$ in our graph, i.e. we refine ego-vehicle poses $\mathbf{P}_s^t\in \mathbf{SE}(3)$ and object poses $\xi_o^t\in \mathbf{SE}(3)$ . In particular, we optimize for pose residuals $\delta \mathbf{P}_s^t\in \mathfrak{se}(3)$ and $\delta \xi_o^t\in \mathfrak{se}(2)$ . We constrain the object pose residual to $\mathfrak{se}(2)$ since objects are usually upright and move along the ground plane. Given each residual, we update the ego-vehicle pose with $\hat{\mathbf{P}}_s^t = \expmap(\delta \mathbf{P}_s^t)\mathbf{P}_s^t$ and the object pose with $\hat{\xi}_o^t = \expmap(\delta \xi_o^t)_{\mathbf{SE}(2)\to \mathbf{SE}(3)}\xi_o^t$ , where $\expmap(\cdot)$ is the exponential map of each lie group. + +Compared to naively optimizing camera and object poses, optimizing the edges along our scene graph has two key advantages. First, we leverage multi-camera constraints, i.e. we keep the camera extrinsics $\mathbf{T}_c$ fixed and optimize the ego-vehicle poses $\mathbf{P}_s^t$ only. This is in contrast to prior art that generally treats each camera pose as independent. Second, the residual $\delta \mathbf{P}_s^t$ propagates to the object poses since they are defined w.r.t. the ego-vehicle coordinate frame instead of the global world frame. Given that optimizing a radiance field as well as camera and object poses jointly is notoriously difficult [21], these constraints are vital to view synthesis quality (see Tab. 6). + +# 6. Experiments + +Datasets. To evaluate against competing methods on our proposed benchmark on Argoverse 2 [61], we hold out every 10th sample in uniform time intervals where a sample + +corresponds to seven ring-camera images. To compare our methods against prior art on KITTI [14] and VKITTI2 [4], we follow the experimental protocol and data splits in [58]. We use the provided 3D bounding box annotations for all datasets. We provide results using an off-the-shelf 3D tracker [68] in the supplemental material. + +Metrics. Following [58], we measure image synthesis quality with PSNR, SSIM [59], and LPIPS (AlexNet) [69]. + +Implementation details. We implement our method in PyTorch [39], accelerating the time-consuming ray-node intersection with a custom CUDA implementation. We train our model on a single RTX 3090 GPU for 250,000 steps on Argovvere 2 and 100,000 steps on KITTI and VKITTI2, with 8192 rays per batch. All model parameters are optimized using Adam [17] with $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ , and an exponential decay learning rate schedule: from $10^{-5}$ to $10^{-6}$ for pose parameters, and from $10^{-3}$ to $10^{-4}$ for all others. In order to counter pose drift, we further apply weight decay with a factor $10^{-2}$ to $\delta \mathbf{P}$ and $\delta \xi$ . The static radiance field $\phi$ is trained from scratch, while $\psi$ is initialized from a semantic prior [6] following [18]. + +Baselines. We compare our method against prior work in dynamic outdoor scene representation, i.e. SUDS [58] and NSG [35]. Further, we include results obtained by augmenting Nerfacto [54], the closest state-of-the-art NeRF architecture to ours, with components designed to handle multicapture reconstruction and scene dynamics. In particular, + +
SRN [49]NeRF [31]NeRF + TimeNSG [35]PNF [18]SUDS [58]Ours
PSNR ↑18.8323.3424.1826.6627.4828.3129.36
SSIM ↑0.5900.6620.6770.8060.8700.8760.911
+ +![](images/8783d9847c2883eb37addca49651ab0b828c5bc165f9a085a3642881e6242f4f.jpg) +Figure 6. Qualitative results on KITTI. With only $25\%$ of views (approx. 15-20) for training, we can still synthesize sharp and realistic novel views with dynamic objects rendered at high quality. + +we consider two variants: "Nerfacto + Emb.", where we incorporate our appearance embedding $\mathbf{A}_s\mathcal{F}(t)$ and the expected depth loss; and "Nerfacto + Emb. + Time", where we further include time modeling with 4D hash tables following [38, 63]. For both Nerfacto variants, we set hash table and MLP sizes to be aligned with our method. + +# 6.1. Comparison to state-of-the-art + +We first compare to prior work on our proposed benchmark in Tab 1. We observe that the static Nerfacto + Emb. has the weakest performance, which slightly increases by adding time modeling in Nerfacto + Emb. + Time. Meanwhile, the current state-of-the-art method SUDS [58] outperforms the Nerfacto variants, while being much slower to train. Our method performs the best on all metrics. The improvement is particularly pronounced in the perceptual quality metrics (SSIM and LPIPS). The training speed of our method is competitive to Nerfacto + Emb. + Time and more than $3 \times$ faster than SUDS. + +In addition, we show a qualitative comparison in Fig. 7. We observe major differences in both the rendered images and depth maps. In particular, all other methods struggle to recover the dynamic objects in the scene, while our method produces realistic renderings and accurate depth maps for both static and dynamic areas. We also observe that, thanks to our transient geometry embedding $\mathbf{G}_s\mathcal{F}(t)$ , we are able to accurately recover the geometry of the trees and their leaves which notably are not present in every sequence of the area that is reconstructed due to seasonal changes. At the same time, other methods struggle to recover those details. Specifically, other methods exhibit artifacts in the depth maps and degraded rendering quality of the trees left and right of the street in columns five and six of Fig. 7, while our method produces accurate color and depth renderings. + +Next, we evaluate our method on established benchmarks, namely KITTI and VKITTI2, in Tab. 2 following the experimental protocol in [58]. In particular, we compare novel view synthesis quality at different fractions of + +Table 3. Image reconstruction on KITTI. We outperform prior art in image reconstruction, i.e. our method can better fit the training views. We follow the experimental protocol in [18, 35, 58]. + +
AsF(t)GsF(t)PSNR ↑SSIM ↑LPIPS ↓
--19.700.6530.588
✓-22.220.6700.546
✓✓22.490.6710.541
+ +Table 4. Ablation study on graph structure. We show that using the multi-level graph structure of sequences and objects is crucial, i.e. omitting sequence vectors $\omega_{s}^{t}$ results in degraded quality since there is no conditioning on scene-specific appearance. Combining appearance and transient geometry embeddings performs best. + +
SamplingSamples per rayPSNR ↑SSIM ↑LPIPS ↓Rays / sec.
Uniform [18]19225.720.7300.45628K
Uniform [18]102425.840.7340.4494K
Separate [35]1024+64+(32+64)26.650.7620.3512.5K
Ours1024+6427.070.7590.36230K
+ +Table 5. Ray Sampling schemes. Our composite ray sampling is about $12 \times$ more efficient to train than separate ray sampling [35] with comparable performance. Uniform sampling [18] exhibits lower performance and is slow when sampled more densely. + +
Pose optimizationPSNR ↑SSIM ↑LPIPS ↓
-22.490.6710.541
Naive21.280.6630.519
Hierarchical (Ours)22.290.6780.523
+ +Table 6. Pose optimization. Compared to naively optimizing camera and object poses, optimizing vehicle and object poses hierarchically on the edges of our graph benefits view synthesis quality more consistently, i.e. both in SSIM and LPIPS. Note that PSNR is very sensitive to pose drift, and thus does not improve. + +training views, i.e. testing different levels of view supervision. We run our method and our baselines and observe substantially better view synthesis quality for our method, both compared to our baselines and prior works. All methods exhibit a similar performance drop when the fraction of training views decreases. We illustrate the rendering quality of our method with only 15-20 training views in Fig. 6. Finally, we also report the results on the task of image reconstruction, i.e. reconstruction of images seen during training, following [18, 35, 58] in Tab. 3 on the KITTI dataset. We significantly outperform previous works in both metrics. + +# 6.2. Ablation studies + +We verify our design through ablation studies. We perform these on the residential area of our benchmark unless otherwise noted. First, we ablate on the multi-level structure of our scene graph. In particular, in Tab. 4, we ablate the latent codes of the sequence nodes, reducing to a representation similar to [35] and observe that the quality of the synthesized views drops significantly. In contrast, our scene graph representation achieves better view synthesis quality, i.e. achieves almost three points higher PSNR. Further, adding our transient geometry embeddings in addition to sequence-level appearance embeddings improves the view synthesis quality. See supplemental material for a qualitative comparison. This adds to the fact that through our multi-level graph structure, we show the ability to modulate car appear + +![](images/3b7faa7dc7233a4c56f0201e825efde6aefde940cee478759f2e25fa02cde53e.jpg) +Figure 7. Qualitative results on Argoverse 2. While prior art struggles with dynamic actors in the scene, our representation renders both realistic novel views and plausible depth maps. Further, our method models complex transient geometry subject to, e.g., seasonal changes such as tree leaves. Note that prior art produces depth artifacts and exhibits degraded view quality in those regions (columns 5 and 6). + +ance through sequence appearance in Fig. 4. + +In Tab. 5, we compare our composite ray sampling scheme to uniformly sampled rays as in [18] and rays sampled separately per node as in [35]. We compare the schemes under the assumption that all other model parameters are equal, implementing them in our method. Ours is about $12 \times$ more efficient to train than densely sampling a ray in uniform intervals or separately sampling the ray per node while producing similar results to the latter and being significantly better than the former. Sparsely sampling a ray at uniform intervals is similarly fast as our method, but yields degraded view synthesis quality. Note that we run this ablation study on only a single sequence of the residential area in our benchmark since densely sampling the rays and separately sampling the rays for all nodes are prohibitively expensive to train in large-scale urban areas. + +Finally, in Tab. 6, we compare our hierarchical pose optimization to naive camera pose optimization employed in previous works [21, 54]. While naive camera pose optimization degrades the results significantly in pixel-wise metrics, our hierarchical pose optimization improves the SSIM and maintain a comparable PSNR. Meanwhile, our hierarchical pose optimization exhibits similar gains to + +naive pose optimization in terms of the perceptual LPIPS metric. This shows that our pose optimization mitigates pose drift while also enabling a more accurate reconstruction. Pose drift usually causes a misalignment between the evaluation viewpoint and scene geometry, which degrades pixel-wise metrics in particular. + +# 7. Conclusion + +We introduce a novel multi-level scene graph representation for radiance field reconstruction in dynamic urban environments that scales to large geographic areas with more than ten thousand images from dozens of sequences and hundreds of dynamic objects. We train our representation with an efficient composite ray sampling and rendering scheme and introduce latent variables that enable modeling complex phenomena like varying environmental conditions and transient geometry present across different vehicle captures. We leverage our representation to refine camera and object poses hierarchically using multi-camera constraints. Finally, we propose a new view synthesis benchmark for dynamic urban driving scenarios. Our approach yields substantially improved results compared to prior art while allowing for flexible de- and recomposition of the scene. + +# References + +[1] Iro Armeni, Zhi-Yang He, JunYoung Gwak, Amir R Zamir, Martin Fischer, Jitendra Malik, and Silvio Savarese. 3d scene graph: A structure for unified semantics, 3d space, and camera. In ICCV, 2019. 2 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1, 3, 5 +[3] Jules Bloomenthal and Brian Wyvill. Introduction to implicit surfaces. 1997. 2 +[4] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual kitti 2, 2020. 2, 6 +[5] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 2 +[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6 +[7] Steve Cunningham and Michael J Bailey. Lessons from scene graphs: using scene graphs to teach hierarchical modeling. Computers & Graphics, 25(4):703-711, 2001. 2 +[8] Nathaniel Fairfield, George A. Kantor, and David S. Wettergreen. Real-time slam with octree evidence grids for exploration in underwater tunnels. Journal of Field Robotics, 24, 2007. 2 +[9] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, 2022. 2 +[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2, 3 +[11] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In CVPR, 2021. 2 +[12] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. NeurIPS, 2022. 2 +[13] Yiming Gao, Yan-Pei Cao, and Ying Shan. Surfelnerf: Neural surfel radiance fields for online photorealistic reconstruction of indoor scenes. In CVPR, 2023. 1, 2 +[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2, 6 +[15] Michael Kaess. Simultaneous localization and mapping with infinite planes. 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 4605-4611, 2015. 2 +[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3 + +[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[18] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, 2022. 2, 3, 5, 6, 7, 8 +[19] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2 +[20] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In CVPR, 2023. 2 +[21] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 6, 8 +[22] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In ICCV, 2023. 2, 3 +[23] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020. 1, 2 +[24] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In ICCV, 2023. 2, 3 +[25] Yan Lu and Dezhen Song. Visual navigation using heterogeneous landmarks and unsupervised geometric constraints. IEEE Transactions on Robotics, 31:736-749, 2015. 2 +[26] Jonathon Luiten, Tobias Fischer, and Bastian Leibe. Track to reconstruct and reconstruct to track. IEEE Robotics and Automation Letters, 5(2):1803-1810, 2020. 2 +[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2 +[28] Alexander Majercik, Cyril Crassin, Peter Shirley, and Morgan McGuire. A ray-box intersection algorithm and efficient dynamic voxel rendering. Journal of Computer Graphics Techniques Vol, 7(3):66-81, 2018. 5 +[29] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 2 +[30] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 1, 2 +[31] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 4, 5, 6, 7 +[32] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 2, 3, 4 + +[33] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 1, 2 +[34] OpenStreetMap contributors. Planet dump retrieved from https://planet.osm.org . https : //www.openstreetmap.org, 2017. 3 +[35] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In CVPR, 2021. 2, 3, 4, 5, 6, 7, 8 +[36] Byeongjun Park and Changick Kim. Point-dynrf: Point-based dynamic radiance fields from a monocular video, 2023. 2 +[37] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2 +[38] Sungheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, and Nahiyup Kang. Temporal interpolation is all you need for dynamic neural radiance fields. In CVPR, 2023. 2, 7 +[39] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 6 +[40] Marc Pollefeys, David Nister, Jan-Michael Frahm, Amir Akbarzadeh, Philippos Mordohai, Brian Clipp, Chris Engels, David Gallup, Seon Joo Kim, Paul C. Merrell, C. Salmi, Sudipta N. Sinha, B. Talton, Liang Wang, Qingxiong Yang, Henrik Stewénius, Ruigang Yang, Greg Welch, and Herman Towles. Detailed real-time urban 3d reconstruction from video. IJCV, 78:143-167, 2007. 2 +[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 2 +[42] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2, 5 +[43] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In CVPR, 2021. 1, 2 +[44] Antoni Rosinol, Arjun Gupta, Marcus Abate, Jingnan Shi, and Luca Carlone. 3d dynamic scene graphs: Actionable spatial perception with places, objects, and humans. arXiv preprint arXiv:2002.06289, 2020. 2 +[45] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In ECCV, 2022. 2 +[46] Renato F Salas-Moreno, Richard A Newcombe, Hauke Strasdat, Paul HJ Kelly, and Andrew J Davison. Slam++: Simultaneous localisation and mapping at the level of objects. In CVPR, 2013. 2 +[47] Aron Schmied, Tobias Fischer, Martin Danelljan, Marc Pollefeys, and Fisher Yu. R3d3: Dense 3d reconstruction of dynamic scenes from multiple cameras. In ICCV, 2023. + +[48] Chen Shen, James F. O'Brien, and Jonathan R. Shewchuk. Interpolating and approximating implicit surfaces from polygon soup. In ACM SIGGRAPH 2004 Papers, page 896-904, 2004. 2 +[49] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. NeurIPS, 2019. 1, 2, 7 +[50] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 2 +[51] Henry Sowizral. Scene graphs in the new millennium. IEEE Computer Graphics and Applications, 20(1):56-57, 2000. 2 +[52] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2 +[53] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In CVPR, 2022. 2 +[54] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 5, 6, 8 +[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 2 +[56] Shubham Tulsiani, Saurabh Gupta, David F Fouhey, Alexei A Efros, and Jitendra Malik. Factoring shape, pose, and layout from the 2d image of a 3d scene. In CVPR, 2018. 2 +[57] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In CVPR, 2022. 2, 3 +[58] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In CVPR, 2023. 2, 3, 4, 6, 7, 8 +[59] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[60] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 2 +[61] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for + +self-driving perception and forecasting. arXiv preprint arXiv:2301.00493, 2023. 2, 3, 6 +[62] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D' 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS, 2022. 2, 5 +[63] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In CVPR, 2021. 2, 7 +[64] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 2 +[65] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In CVPR, 2022. 2 +[66] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In CVPR, 2022. 4 +[67] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, 2023. 2, 3, 4 +[68] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, 2021. 6 +[69] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 \ No newline at end of file diff --git a/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/images.zip b/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f4f07ca824ffc0cdc0551dca9c1f8cf49bdb3e65 --- /dev/null +++ b/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b59833c299844d7680bcdd61a817fde57c9642ce18ee768b1af418e2d058a395 +size 736863 diff --git a/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/layout.json b/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..73e3d3371338476b8cbc6d24106875f1cf61ee90 --- /dev/null +++ b/2024/Multi-Level Neural Scene Graphs for Dynamic Urban Environments/layout.json @@ -0,0 +1,11959 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 507, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 507, + 120 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 507, + 120 + ], + "type": "text", + "content": "Multi-Level Neural Scene Graphs for Dynamic Urban Environments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 142, + 443, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 142, + 443, + 156 + ], + "spans": [ + { + "bbox": [ + 146, + 142, + 443, + 156 + ], + "type": "text", + "content": "Tobias Fischer" + }, + { + "bbox": [ + 146, + 142, + 443, + 156 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 142, + 443, + 156 + ], + "type": "text", + "content": " Lorenzo Porzi" + }, + { + "bbox": [ + 146, + 142, + 443, + 156 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 142, + 443, + 156 + ], + "type": "text", + "content": " Samuel Rota Bulò" + }, + { + "bbox": [ + 146, + 142, + 443, + 156 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 194, + 157, + 396, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 157, + 396, + 171 + ], + "spans": [ + { + "bbox": [ + 194, + 157, + 396, + 171 + ], + "type": "text", + "content": "Marc Pollefeys1 Peter Kontschieder2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 172, + 388, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 172, + 388, + 186 + ], + "spans": [ + { + "bbox": [ + 209, + 172, + 388, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 209, + 172, + 388, + 186 + ], + "type": "text", + "content": " ETH Zürich " + }, + { + "bbox": [ + 209, + 172, + 388, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 209, + 172, + 388, + 186 + ], + "type": "text", + "content": " Meta Reality Labs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 184, + 190, + 403, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 190, + 403, + 201 + ], + "spans": [ + { + "bbox": [ + 184, + 190, + 403, + 201 + ], + "type": "text", + "content": "https://tobiasfshr.github.io/pub/ml-nsg/" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 50, + 214, + 93, + 270 + ], + "blocks": [ + { + "bbox": [ + 50, + 214, + 93, + 270 + ], + "lines": [ + { + "bbox": [ + 50, + 214, + 93, + 270 + ], + "spans": [ + { + "bbox": [ + 50, + 214, + 93, + 270 + ], + "type": "image", + "image_path": "c1089944d4ef1d7a751ce36bcdde3419dad0203b973446f27dba488ce7b93109.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 50, + 270, + 93, + 322 + ], + "blocks": [ + { + "bbox": [ + 50, + 270, + 93, + 322 + ], + "lines": [ + { + "bbox": [ + 50, + 270, + 93, + 322 + ], + "spans": [ + { + "bbox": [ + 50, + 270, + 93, + 322 + ], + "type": "image", + "image_path": "4efd39ef8f453b62cd2fbfa0f7f83c46b1af373aa59ecbbef8d58a61b8cd68d5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 323, + 93, + 379 + ], + "blocks": [ + { + "bbox": [ + 50, + 323, + 93, + 379 + ], + "lines": [ + { + "bbox": [ + 50, + 323, + 93, + 379 + ], + "spans": [ + { + "bbox": [ + 50, + 323, + 93, + 379 + ], + "type": "image", + "image_path": "5ec4d11e7f858af17263a92b2c31121fe8384c43b33c6c751909a21267269ab1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 383, + 84, + 394 + ], + "lines": [ + { + "bbox": [ + 59, + 383, + 84, + 394 + ], + "spans": [ + { + "bbox": [ + 59, + 383, + 84, + 394 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 98, + 218, + 294, + 373 + ], + "blocks": [ + { + "bbox": [ + 98, + 218, + 294, + 373 + ], + "lines": [ + { + "bbox": [ + 98, + 218, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 98, + 218, + 294, + 373 + ], + "type": "image", + "image_path": "f3ae6cb64d0fb026a45d3dd3cb46acf62d0c1bf05edb926c42ac5bf9d603bff6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 382, + 265, + 394 + ], + "lines": [ + { + "bbox": [ + 128, + 382, + 265, + 394 + ], + "spans": [ + { + "bbox": [ + 128, + 382, + 265, + 394 + ], + "type": "text", + "content": "Multi-Level Neural Scene Graph" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 299, + 214, + 380, + 295 + ], + "blocks": [ + { + "bbox": [ + 299, + 214, + 380, + 295 + ], + "lines": [ + { + "bbox": [ + 299, + 214, + 380, + 295 + ], + "spans": [ + { + "bbox": [ + 299, + 214, + 380, + 295 + ], + "type": "image", + "image_path": "27fe8028c93540aed69f136ab15fb7fdd7312b9260ea1028764672189019f51d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 299, + 295, + 380, + 376 + ], + "blocks": [ + { + "bbox": [ + 299, + 295, + 380, + 376 + ], + "lines": [ + { + "bbox": [ + 299, + 295, + 380, + 376 + ], + "spans": [ + { + "bbox": [ + 299, + 295, + 380, + 376 + ], + "type": "image", + "image_path": "f686443d3ca75b8d9970d7c507f8c8aedd2aeaa29415ca17eadac66cb094e41d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "lines": [ + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "text", + "content": "Figure 1. Overview. We represent sequences captured from moving vehicles in a shared geographic area with a multi-level scene graph. Each dynamic object " + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "inline_equation", + "content": "v_{o}" + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "text", + "content": " is associated with a sequence node " + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "inline_equation", + "content": "v_{s}^{t}" + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "text", + "content": ". The sequence nodes are registered in a common world frame at the root node " + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "text", + "content": " through the vehicle poses " + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{s}^{t}" + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "text", + "content": ", while the dynamic objects are localized w.r.t. the sequence node with pose " + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\xi_{o}^{t}" + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "text", + "content": ". Each camera " + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "text", + "content": " is associated with an ego-vehicle position, i.e. node " + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "inline_equation", + "content": "v_{s}^{t}" + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "text", + "content": ", through the extrinsic calibration " + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{c}" + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "text", + "content": ". The sequence and object nodes hold latent codes " + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 401, + 545, + 456 + ], + "type": "text", + "content": " that condition the radiance field, synthesizing novel views in various conditions with distinct dynamic objects." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 381, + 214, + 461, + 295 + ], + "blocks": [ + { + "bbox": [ + 381, + 214, + 461, + 295 + ], + "lines": [ + { + "bbox": [ + 381, + 214, + 461, + 295 + ], + "spans": [ + { + "bbox": [ + 381, + 214, + 461, + 295 + ], + "type": "image", + "image_path": "abacc3bd5b0ff214316c63989fb099052008ddc35f79c1b74cbd6cdd9b6697f9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 381, + 295, + 462, + 376 + ], + "blocks": [ + { + "bbox": [ + 381, + 295, + 462, + 376 + ], + "lines": [ + { + "bbox": [ + 381, + 295, + 462, + 376 + ], + "spans": [ + { + "bbox": [ + 381, + 295, + 462, + 376 + ], + "type": "image", + "image_path": "1b6b993264d94d48c329b911f10c0b384bb711a871081124b7777fcde8ebc5c2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 382, + 449, + 392 + ], + "lines": [ + { + "bbox": [ + 394, + 382, + 449, + 392 + ], + "spans": [ + { + "bbox": [ + 394, + 382, + 449, + 392 + ], + "type": "text", + "content": "Novel Views" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 463, + 214, + 543, + 295 + ], + "blocks": [ + { + "bbox": [ + 463, + 214, + 543, + 295 + ], + "lines": [ + { + "bbox": [ + 463, + 214, + 543, + 295 + ], + "spans": [ + { + "bbox": [ + 463, + 214, + 543, + 295 + ], + "type": "image", + "image_path": "a25c0b68ad006e12ee21a69730b8a96cd8a93d1e8e69a017082c994e298e9c1a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 463, + 295, + 544, + 376 + ], + "blocks": [ + { + "bbox": [ + 463, + 295, + 544, + 376 + ], + "lines": [ + { + "bbox": [ + 463, + 295, + 544, + 376 + ], + "spans": [ + { + "bbox": [ + 463, + 295, + 544, + 376 + ], + "type": "image", + "image_path": "0992a2a06b4105a7447221a43e570a7bc0cc34ec362b568072beed7e664700f1.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 143, + 467, + 190, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 467, + 190, + 478 + ], + "spans": [ + { + "bbox": [ + 143, + 467, + 190, + 478 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 45, + 498, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 498, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 45, + 498, + 288, + 700 + ], + "type": "text", + "content": "We estimate the radiance field of large-scale dynamic areas from multiple vehicle captures under varying environmental conditions. Previous works in this domain are either restricted to static environments, do not scale to more than a single short video, or struggle to separately represent dynamic object instances. To this end, we present a novel, decomposable radiance field approach for dynamic urban environments. We propose a multi-level neural scene graph representation that scales to thousands of images from dozens of sequences with hundreds of fast-moving objects. To enable efficient training and rendering of our representation, we develop a fast composite ray sampling and rendering scheme. To test our approach in urban driving scenarios, we introduce a new, novel view synthesis benchmark. We show that our approach outperforms prior art by a significant margin on both established and our proposed benchmark while being faster in training and rendering." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 467, + 386, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 386, + 478 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 386, + 478 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 487, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 654 + ], + "type": "text", + "content": "Estimating the radiance field of a dynamic urban environment from data collected by sensor-equipped vehicles is a core challenge in closed-loop simulation for robotics and in mixed reality. It is particularly relevant for applications like autonomous driving and city-scale mapping, where capture vehicles can be frequently deployed. The growing amount of available data provides great opportunities for creating up-to-date digital twins of entire cities but also poses unique challenges as increasingly heterogeneous data sources must be processed. In particular, limited scene coverage, different lighting, weather, seasonal conditions, and varying geometry due to distinct dynamic and transient objects make radiance field reconstruction of dynamic urban environments extremely challenging." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "Recently, Neural Radiance Fields (NeRFs) have enabled significant progress in achieving realistic novel view synthesis from a set of input views [2, 31-33, 49]. These methods represent a static 3D scene with fully implicit [30, 31, 49] or low-level explicit structures [10, 13, 16, 23, 32, 43," + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21125" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "52, 65] such as voxels, points, surfels, meshes or 3D gaus-. sians. In parallel, implicit [19, 37, 50, 62, 63] and low-level explicit [9, 27, 36, 38] neural representations for dynamic 4D scenes have been investigated. However, fewer works have focused on decomposing scenes into higher-level entities [18, 26, 46, 56]. In computer graphics and 3D mapping, scene graphs have been used to represent complex scenes in a multi-level hierarchy [1, 44, 51]. In view synthesis, Ost et al. [35] apply this concept by describing actors in a dynamic scene as entities of a scene graph. However, their representation lacks the ability of [1, 44] to represent scenes at multiple levels, thereby limiting their representation to short, single sequences." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 229, + 289, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 289, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 289, + 386 + ], + "type": "text", + "content": "While earlier methods for view synthesis focused on object-centric scenes with controlled camera trajectories, recent works move towards radiance field reconstruction of large-scale environments from in-the-wild captures. Among these, many methods focus on static environments, removing dynamic actors from the input data [22, 24, 29, 42, 45, 53, 57]. A few works explicitly model dynamic actors [18, 35, 67], but either do not scale to more than a single, short video [18, 35, 67], or struggle to accurately represent individual object instances [58]. This complicates the evaluation of these methods for real-world applications because existing benchmarks neither scale to large urban areas [18, 35] nor reflect realistic capturing conditions [12]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 387, + 289, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 289, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 289, + 577 + ], + "type": "text", + "content": "To address these issues, we propose a multi-level neural scene graph representation that spans large geographic areas with hundreds of dynamic objects. In contrast to previous works, our multi-level scene graph formulation allows us to distinguish and represent dynamic object instances effectively and further to represent a scene under varying conditions. To make our representation viable for large-scale dynamic environments, we develop a composite ray sampling and rendering scheme that enables fast training and rendering of our method. To test this hypothesis, we introduce a benchmark for radiance field reconstruction in dynamic urban environments based on [61]. We fuse data from dozens of vehicle captures under varying conditions amounting to more than ten thousand images with several hundreds of dynamic objects per reconstructed area. We summarize our contributions as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 580, + 286, + 685 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 580, + 286, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 286, + 613 + ], + "type": "text", + "content": "- We propose a multi-level neural scene graph formulation that scales to dozens of sequences with hundreds of fast-moving objects under varying environmental conditions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 616, + 286, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 286, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 286, + 649 + ], + "type": "text", + "content": "- We develop an efficient composite ray sampling and rendering scheme that enables fast training and rendering of our representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 652, + 285, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 285, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 285, + 685 + ], + "type": "text", + "content": "- We present a benchmark that provides a realistic, application-driven evaluation of radiance field reconstruction in dynamic urban environments." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "We show state-of-the-art view synthesis results on both established benchmarks [4, 14] and our proposed benchmark." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 91, + 545, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 235 + ], + "type": "text", + "content": "3D and 4D scene representations. Finding the right scene representation is a core issue in 3D computer vision and graphics [5]. Over the years, a wide variety of options have been explored [1, 3, 8, 15, 25, 30, 40, 46-48], and, more recently, neural rendering [33] has been used to enable a new generation of scene models that support photo-realistic novel view synthesis. Scene representations for neural rendering can be roughly classified as \"implicit\" [30, 31, 49], which store most of the information in the weights of a neural network, or \"explicit\", which use low-level spatial primitives such as voxels [10, 23, 32, 52], points [65], surfels [13], meshes [43], or 3D gaussians [16]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 235, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 545, + 319 + ], + "type": "text", + "content": "Similarly, different 4D dynamic scene representations for view synthesis have been investigated, including both implicit [11, 19, 37, 41, 55, 58, 62, 63] and explicit [9, 27, 36, 67] approaches. Dynamics are generally modeled as deformations of a canonical volume [9, 37, 41, 55, 62], a separate scene motion function [11, 19, 20, 58, 63], or rigid transformations of local geometric primitives [27]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 319, + 546, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 546, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 546, + 558 + ], + "type": "text", + "content": "Another line of work investigates the decomposition of scenes into higher-level entities [18, 26, 46, 56]. To express the composition of different entities into a complex scene, classical computer graphics literature [7] uses scene graphs. In particular, entities are represented as nodes in a hierarchical graph and are connected through edges defined by coordinate frame transformations. Thus, global transformations can be acquired by traversing the graph from its root node. For indoor 3D mapping, this concept was proposed by Armeni et al. [1] to represent static scenes at multiple levels of hierarchy, i.e. buildings, rooms and objects, and was later extended to dynamic scenes by Rosinol et al. [44]. Recently, Ost et al. [35] have revisited this concept for view synthesis, describing multi-object scenes with a graph that represents the dynamic actors in the scene. However, their representation lacks the ability of [1, 44] to represent scenes at multiple levels of hierarchy and is thus inherently limited to single, short video clips. On the contrary, we present a scalable, multi-level scene graph representation that spans large geographic areas with hundreds of dynamic objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Representing large-scale urban scenes. Compared to controlled captures of small scenes, in-the-wild captures of large-scale scenes pose distinct challenges. Limited viewpoint coverage, inaccurate camera poses, far-away buildings and sky, fast-moving objects, complex lighting, and auto exposure make radiance field reconstruction challenging. Therefore, previous works aid the reconstruction by using depth priors from e.g. LiDAR, refining camera parameters, adding information about camera exposure, and using specialized sky and light modeling components [22, 29, 42, 45, 53, 60, 64]. While many of these works simply remove dynamic actors, some methods explicitly model scene dynamics [18, 35, 58, 67]. However," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21126" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 163 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 163 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 163 + ], + "type": "image", + "image_path": "cf6ac8ee88fa08ca1d431f0e43f8a28257acf473308ae33e24cb79da8e1b5dde.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 172, + 287, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 172, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 287, + 205 + ], + "type": "text", + "content": "Figure 2. Ego-vehicle trajectories of our benchmark. We show the residential (left) and the downtown (right) areas, trajectories superimposed on 2D maps obtained from OpenStreetMap [34]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 227, + 286, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 286, + 285 + ], + "type": "text", + "content": "these methods either struggle to scale to more than a single sequence [18, 35, 67] or to accurately represent and distinguish dynamic actors [58]. To address these issues, we present a decomposed, scalable scene representation for dynamic urban environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 287, + 286, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 286, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 286, + 406 + ], + "type": "text", + "content": "Efficient scene rendering. Especially for large-scale scenes, the efficiency of a scene representation in training and inference is crucial. To alleviate the burden of ray traversal in volumetric rendering, many techniques for efficient sampling [2, 10, 32, 57] have been proposed. Recently, researchers have exploited more efficient forms of rendering, e.g. rasterization of meshes [22, 24] or other 3D primitives [16]. While these approaches are focused on static scenes, we extend the approach of [2] to our scene graph representation and thus to dynamic scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 418, + 86, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 86, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 86, + 430 + ], + "type": "text", + "content": "3. Data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 286, + 533 + ], + "type": "text", + "content": "We investigate how to synthesize novel views of a dynamic urban environment from a set of captures taken from moving vehicles. Specifically, we are interested in captures spanning the same geographic area under varying conditions. This data presents unique challenges due to limited viewpoint coverage, different lighting, weather, and season, and even varying geometry due to distinct dynamic and transient objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 534, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 286, + 676 + ], + "type": "text", + "content": "Previous benchmarks in this area [18, 35] are limited to short, single video clips with simple ego trajectories and a few dynamic objects. Therefore, we present a benchmark that better reflects the aforementioned challenges. We base our benchmark on Argoverse 2 [61] which provides a rich set of captures from a fleet of vehicles deployed in multiple US cities that span different weather, season, and time of day. The vehicles are equipped with a surround-view camera rig with seven global shutter cameras, a LiDAR sensor, and a GPS. Furthermore, the timings of the different sensors are provided, so that one can relate the LiDAR-based 3D bounding box annotations to camera timestamps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "We leverage the GPS information to globally align the sequences and to identify regions that we are interested in mapping. We then extract the specific vehicle captures." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 545, + 144 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 144 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 144 + ], + "type": "image", + "image_path": "735ace0c8070dbc357097038deff46d0f5361f490277175df7f5127c9df5141f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 152, + 545, + 207 + ], + "lines": [ + { + "bbox": [ + 304, + 152, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 545, + 207 + ], + "type": "text", + "content": "Figure 3. Sequence alignment visualization. The initial GPS-based alignment is imprecise, as evidenced by the duplicated structures in the overlaid LiDAR point clouds (left). After our ICP alignment, the area is well reconstructed (middle) according to its real geometry (right, from Argoverse 2 [61])." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 226, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 226, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 226, + 545, + 285 + ], + "type": "text", + "content": "Since GPS-based localization accuracy is only coarsely precise in urban areas, we align the captures via a global, offline iterative-closest-point (ICP) procedure applied to the LiDAR point clouds of all sequences to achieve satisfactory alignment for novel view synthesis purposes (see Fig. 3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 286, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 417 + ], + "type": "text", + "content": "Following this procedure, we build a benchmark that enables real-world evaluation of novel view synthesis from diverse vehicle captures. It is composed of 37 vehicle captures split into two geographic regions as illustrated in Fig. 2. The regions cover a residential and a downtown area to resemble the different characteristics of urban environments. The residential area spans 14 captures with 10493 training and 1162 testing images with more than 700 distinct moving objects. The downtown area spans 23 captures with 16933 training and 1876 testing images with more than 600 distinct moving objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 426, + 361, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 361, + 439 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 361, + 439 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": "Problem setup. We are given a set of sequences " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": " captured from moving vehicles in different conditions. Each sequence " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "s \\in S" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": " consists of a set of images taken from cameras " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "C_s" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": " mounted on the vehicle at different timesteps indexed by " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "T_s" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": ". We assume the vehicle's sensors are calibrated with respect to a common vehicle and a global world frame. In particular, we assume given sensor extrinsic " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_c = [\\mathbf{R}_c|\\mathbf{t}_c] \\in \\mathbf{SE}(3)" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": " for each camera " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "c \\in C = \\bigcup_{s} C_s" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": ", and ego-vehicle poses " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_s^t = [\\mathbf{R}_s^t|\\mathbf{t}_s^t] \\in \\mathbf{SE}(3)" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": " for each timestamp " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "t \\in T_s" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": " and sequence " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "s \\in S" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": ". Furthermore, we assume all cameras to have known intrinsics " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_c" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": ". Each sequence " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": " entails a set " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "O_s" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": " of dynamic objects. For each object " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "o \\in O_s" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": ", we assume to have an estimated 3D bounding box track " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_o" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": " consisting of object poses " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\{\\xi_o^{t_0}, \\dots, \\xi_o^{t_n}\\} \\subset \\mathbf{SE}(3)" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": " w.r.t. the ego-vehicle frame, where " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "t_i \\in T_s" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": ", and the 3D object dimensions " + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_o \\in \\mathbb{R}_{>0}^3" + }, + { + "bbox": [ + 304, + 447, + 545, + 650 + ], + "type": "text", + "content": ". The images span a common geographic area for which we would like to estimate a radiance field" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 343, + 657, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 657, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 343, + 657, + 545, + 670 + ], + "type": "interline_equation", + "content": "f _ {\\theta} (\\mathbf {x}, \\mathbf {d}, t, s) = (\\sigma (\\mathbf {x}, t, s), \\mathbf {c} (\\mathbf {x}, \\mathbf {d}, t, s)) \\tag {1}", + "image_path": "5d9414b46f3edeafe05af1b11b571411464e096106b01683f1a4c45cdc9cbdf1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "that outputs volume density " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathbb{R}_{\\geq 0}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in [0,1]^3" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " conditioned on 3D location " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ", viewing direction " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ", time " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and sequence " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21127" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 224, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 224, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 224, + 85 + ], + "type": "text", + "content": "4.1. Multi-Level Neural Scene Graph" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": "Overview. We illustrate our representation in Fig. 1. We decompose the scene into a graph " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{V},\\mathcal{E})" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": " where the set of nodes " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": " is composed by a root node " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": " that defines the global coordinate system, camera nodes " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "\\{v_{c}\\}_{c\\in C}" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": ", and, for each sequence " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "s\\in S" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": ", sequence nodes " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "\\{v_s^t\\}_{t\\in T_s}" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": " and dynamic object nodes " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "\\{v_o\\}_{o\\in O_s}" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": ". The nodes are connected through oriented edges " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "e\\in \\mathcal{E}" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": " that represent rigid transformations between the coordinate frames of the nodes, consistently with the edge direction. In addition, each node " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "v\\in \\mathcal{V}" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": " can have an associated latent vector " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": ". Given the scene graph " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": ", we model " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": " using two radiance fields, namely " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": " for largely static and " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": " for highly dynamic scene parts. We use the latent vectors " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": " to condition the radiance fields. In particular, we use latent vectors " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "\\omega_s^t" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": " of a sequence node " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "v_s^t" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": " to condition " + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 91, + 290, + 270 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 83, + 281, + 287, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 281, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 83, + 281, + 287, + 295 + ], + "type": "interline_equation", + "content": "\\phi (\\mathbf {x}, \\mathbf {d}, \\omega_ {s} ^ {t}) = \\left(\\sigma_ {\\phi} \\left(\\mathbf {x}, \\omega_ {s} ^ {t}\\right), \\mathbf {c} _ {\\phi} \\left(\\mathbf {x}, \\mathbf {d}, \\omega_ {s} ^ {t}\\right)\\right). \\tag {2}", + "image_path": "ec62bda188145a3528558ef34a6f9fe06396699a689098714780bfc296532ddf.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 306, + 288, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 288, + 354 + ], + "type": "text", + "content": "Since the graph " + }, + { + "bbox": [ + 46, + 306, + 288, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 306, + 288, + 354 + ], + "type": "text", + "content": " has multiple levels, a node associated with a dynamic object will naturally fall into the sequence " + }, + { + "bbox": [ + 46, + 306, + 288, + 354 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 306, + 288, + 354 + ], + "type": "text", + "content": " it appears in. Therefore, the radiance field " + }, + { + "bbox": [ + 46, + 306, + 288, + 354 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 306, + 288, + 354 + ], + "type": "text", + "content": " is conditioned on latent vectors " + }, + { + "bbox": [ + 46, + 306, + 288, + 354 + ], + "type": "inline_equation", + "content": "\\omega_{s}^{t}" + }, + { + "bbox": [ + 46, + 306, + 288, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 306, + 288, + 354 + ], + "type": "inline_equation", + "content": "\\omega_{o}" + }, + { + "bbox": [ + 46, + 306, + 288, + 354 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 365, + 287, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 287, + 380 + ], + "type": "interline_equation", + "content": "\\psi (\\mathbf {x}, \\mathbf {d}, \\omega_ {s} ^ {t}, \\omega_ {o}) = (\\sigma_ {\\psi} (\\mathbf {x}, \\omega_ {s} ^ {t}, \\omega_ {o}), \\mathbf {c} _ {\\psi} (\\mathbf {x}, \\mathbf {d}, \\omega_ {s} ^ {t}, \\omega_ {o})) (3)", + "image_path": "0a12a745bdba877b1d8bef141614889c3552d8afebe6dde88ed35db90c4ddbbc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 389, + 286, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 286, + 509 + ], + "type": "text", + "content": "of the corresponding sequence and object nodes " + }, + { + "bbox": [ + 46, + 389, + 286, + 509 + ], + "type": "inline_equation", + "content": "v_{s}^{t}" + }, + { + "bbox": [ + 46, + 389, + 286, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 389, + 286, + 509 + ], + "type": "inline_equation", + "content": "v_{o}" + }, + { + "bbox": [ + 46, + 389, + 286, + 509 + ], + "type": "text", + "content": ". Appearance and geometry variation. Reconstructing an environment from multiple captures is challenging from two perspectives: in addition to varying dynamic objects, there is i) varying appearance across captures, and ii) slow-moving or static transient geometry such as tree leaves or construction sites. Since both transient geometry and appearance can vary across captures, but are usually smooth within a sequence, we model these phenomena as smooth functions over time" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 521, + 287, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 521, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 115, + 521, + 287, + 534 + ], + "type": "interline_equation", + "content": "\\omega_ {s} ^ {t} = \\left[ \\mathbf {A} _ {s} \\mathcal {F} (t), \\mathbf {G} _ {s} \\mathcal {F} (t) \\right] \\tag {4}", + "image_path": "6ca317985abc03433883dde58962f6e8c448fdafa67ca1498df7bcd0418bd441.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_s" + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": " is an appearance matrix, " + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_s" + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": " is a transient geometry matrix and " + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": " is a 1D basis function of sines and cosines with linearly increasing frequencies at log-scale [58, 66]. We normalize time " + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": " into the interval " + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": " using the maximum sequence length " + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\max_{s\\in S}|T_s|" + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": ". Crucially, this allows to model near-static, but sequence-specific regions of the input, as well as appearance changes due to e.g. auto-exposure. The degree of variation across time can be controlled by the number of frequencies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": "Sequence nodes. The sequence nodes " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "v_{s}^{t}" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " are connected to the root node " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " with an edge " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "e_{v_s^t v_r} = \\mathbf{P}_s^t" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": ", i.e. the sequences " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " share a common global world frame. Each sequence node holds the latent vector " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\omega_{s}^{t}" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " that conditions the radiance field " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": ". We model " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " with a multi-scale 3D hash grid representa" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 547, + 220 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 547, + 220 + ], + "type": "image", + "image_path": "dc40757d267ad7d75e2719c0dcc3eaf408b19005a456657fd20d0e438ccfa412.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 224, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 269 + ], + "type": "text", + "content": "Figure 4. Modifying car appearance with scene appearance. We exchange " + }, + { + "bbox": [ + 305, + 224, + 545, + 269 + ], + "type": "inline_equation", + "content": "\\omega_{s}^{t}" + }, + { + "bbox": [ + 305, + 224, + 545, + 269 + ], + "type": "text", + "content": " for different car instances. The car's appearance in a rendered, car-centric view (top) changes according to the environmental conditions visible in the sequence " + }, + { + "bbox": [ + 305, + 224, + 545, + 269 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 224, + 545, + 269 + ], + "type": "text", + "content": " (bottom)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 274, + 460, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 460, + 286 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 460, + 286 + ], + "type": "text", + "content": "tion [32] and lightweight MLP heads:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 451, + 293, + 545, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 293, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 451, + 293, + 545, + 306 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\mathbf {x}} = \\mathrm {H} _ {3} (\\mathbf {x}) \\tag {5}", + "image_path": "453676230dba7711666d0cf561f5e441875e7557c40faaedf8a3fbcb932ff0f1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 405, + 308, + 545, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 308, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 405, + 308, + 545, + 322 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {\\sigma_ {\\phi}}, \\sigma_ {\\phi} = \\operatorname {M L P} _ {\\sigma_ {\\phi}} \\left(\\mathbf {f} _ {\\mathbf {x}}\\right) \\tag {6}", + "image_path": "361e74bbe9d1176f21ad187cc0015eeb04e6fe6e2675da09b71c7900e95f134d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 350, + 323, + 545, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 323, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 350, + 323, + 545, + 337 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {\\phi} = \\operatorname {M L P} _ {\\mathbf {c} _ {\\phi}} \\left(\\mathbf {h} _ {\\sigma_ {\\phi}}, \\gamma_ {\\mathrm {S H}} (\\mathbf {d}), \\mathbf {A} _ {s} \\mathcal {F} (t)\\right) \\tag {7}", + "image_path": "3561efb840acdcaeec52016d1c42995d59362ea8893666e864f705a8c01400b0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 364, + 338, + 545, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 338, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 364, + 338, + 545, + 352 + ], + "type": "interline_equation", + "content": "\\sigma_ {\\mathbf {G}}, \\mathbf {c} _ {\\mathbf {G}} = \\operatorname {M L P} _ {\\mathbf {G}} \\left(\\mathbf {h} _ {\\sigma_ {\\phi}}, \\mathbf {G} _ {s} \\mathcal {F} (t)\\right) \\tag {8}", + "image_path": "c68408315bf0df0b54b4ceb0f1726e9302b5efe7124dc5f685b83bfe2d61fa10.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 358, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 358, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 545, + 394 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 358, + 545, + 394 + ], + "type": "inline_equation", + "content": "\\gamma_{\\mathrm{SH}}(\\cdot)" + }, + { + "bbox": [ + 305, + 358, + 545, + 394 + ], + "type": "text", + "content": " is a spherical harmonics encoding [32]. The final colors and densities are computed as a mixture of static and transient output (analogous to Eq. 12, see supp. mat.)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "text", + "content": "Dynamic nodes. The dynamic nodes " + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "inline_equation", + "content": "v_{o}" + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "text", + "content": " are connected to the sequence nodes " + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "inline_equation", + "content": "v_{s}^{t}" + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "text", + "content": " with edges " + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "inline_equation", + "content": "e_{v_{o}v_{s}^{t}} = \\xi_{o}^{t}" + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "text", + "content": ". The dynamic node " + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "inline_equation", + "content": "v_{o}" + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "text", + "content": " associated with object " + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "text", + "content": " holds a latent vector " + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\omega_{o}" + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "text", + "content": " that conditions the radiance field " + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "text", + "content": ". We model " + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "text", + "content": " following [31] with an MLP conditioned on " + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\omega_{o}" + }, + { + "bbox": [ + 305, + 394, + 545, + 465 + ], + "type": "text", + "content": " to represent different instances with the same network [35, 67]" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 365, + 473, + 545, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 473, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 365, + 473, + 545, + 487 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {\\sigma_ {\\psi}}, \\sigma_ {\\psi} = \\operatorname {M L P} _ {\\sigma_ {\\psi}} \\left(\\gamma_ {\\mathrm {P E}} (\\mathbf {x}), \\omega_ {o}\\right) \\tag {9}", + "image_path": "a5293461ba30d808f71879023bca244af4e6a76675d20d547ce8f5ac8f503b08.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 353, + 489, + 545, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 489, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 353, + 489, + 545, + 503 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {\\psi} = \\operatorname {M L P} _ {\\mathbf {c} _ {\\psi}} \\left(\\mathbf {h} _ {\\sigma_ {\\psi}}, \\gamma_ {\\mathrm {P E}} (\\mathbf {d}), \\omega_ {o}, \\omega_ {s} ^ {t}\\right) \\tag {10}", + "image_path": "245a8e8546d67a74d130bd54d7cb2feba061132dc3eae380335353481c512e9e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 509, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 509, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\gamma_{\\mathrm{PE}}(\\cdot)" + }, + { + "bbox": [ + 304, + 509, + 545, + 594 + ], + "type": "text", + "content": " is a positional encoding [31]. Note that 3D position " + }, + { + "bbox": [ + 304, + 509, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 509, + 545, + 594 + ], + "type": "text", + "content": " and viewing direction " + }, + { + "bbox": [ + 304, + 509, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 509, + 545, + 594 + ], + "type": "text", + "content": " are transformed into the local object coordinate frame with " + }, + { + "bbox": [ + 304, + 509, + 545, + 594 + ], + "type": "inline_equation", + "content": "(\\mathbf{P}_s^t\\xi_o^t)^{-1}\\mathbf{I}_3(1 / \\max (\\mathbf{s}_o))" + }, + { + "bbox": [ + 304, + 509, + 545, + 594 + ], + "type": "text", + "content": ". We condition " + }, + { + "bbox": [ + 304, + 509, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 509, + 545, + 594 + ], + "type": "text", + "content": " on both the scene and object-dependent latent vectors. This allows us to disentangle scene-dependent appearance from the object texture and thus to transfer objects across sequences. We illustrate this process in Fig. 4." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "type": "text", + "content": "Camera nodes. The cameras " + }, + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "type": "inline_equation", + "content": "c \\in C_s" + }, + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "type": "text", + "content": " are connected to the sequence nodes " + }, + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "type": "inline_equation", + "content": "v_s^t" + }, + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "type": "text", + "content": " through edges " + }, + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "type": "inline_equation", + "content": "e_{v_c v_s^t} = \\mathbf{T}_c" + }, + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "type": "text", + "content": ", i.e. the calibration of camera " + }, + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "type": "text", + "content": " w.r.t. the ego-vehicle frame. This way, we can tie camera poses to a specific ego-vehicle pose." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 647, + 441, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 441, + 661 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 441, + 661 + ], + "type": "text", + "content": "4.2. Scene Graph Rendering" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": "We describe how we render our scene graph " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": " for a given set of rays " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": ". The sampling locations along a ray " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{r}, t, s) \\in \\mathcal{R}" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": " in sequence " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": " are defined as " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(u) = \\mathbf{o} + u\\mathbf{d}" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{o} = \\mathbf{R}_s^t\\mathbf{t}_c + \\mathbf{t}_s^t" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{d} = \\mathbf{R}_s^t\\mathbf{R}_c\\mathbf{K}_c^{-1}(p_x, p_y, 1)^T" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21128" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "Continuous-time pose. In order to realistically render videos at different frame rates, we treat the dynamic object poses " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\{\\xi_o^{t_0},\\dots,\\xi_o^{t_n}\\}" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " as a continuous function of time " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\xi_o(t)" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ". We compute " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\xi_o(t)" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " by interpolating between the two nearest poses at " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "t_a\\leq t < t_b" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " to time " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ". This allows us also to synchronize estimated object poses originating from the LiDAR measurements with the camera timestamps." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "content": "Ray-node intersection. To render the dynamic nodes, we measure the intersection of their 3D bounding boxes with each " + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "inline_equation", + "content": "(\\mathbf{r},t,s)\\in \\mathcal{R}" + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "content": ". In particular, given the sequence " + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "content": " the ray " + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "content": " is associated with, we first traverse the graph " + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "content": " to retrieve all relevant nodes " + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "inline_equation", + "content": "v_{o}" + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "content": " and their 3D bounding boxes " + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_o^t = [\\xi_o(t),\\mathbf{s}_o]" + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "content": ". Then, we transform " + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "content": " into each local node coordinate system and subsequently use AABB-ray intersection [28] to compute the entry and exit locations " + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "inline_equation", + "content": "u_{o}^{\\mathrm{in}},u_{o}^{\\mathrm{out}}" + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "content": " along ray " + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 263, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 287, + 289 + ], + "type": "text", + "content": "Composite rendering. To render the color " + }, + { + "bbox": [ + 47, + 263, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{C}}" + }, + { + "bbox": [ + 47, + 263, + 287, + 289 + ], + "type": "text", + "content": " of ray " + }, + { + "bbox": [ + 47, + 263, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 47, + 263, + 287, + 289 + ], + "type": "text", + "content": " in sequence " + }, + { + "bbox": [ + 47, + 263, + 287, + 289 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 263, + 287, + 289 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 47, + 263, + 287, + 289 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 263, + 287, + 289 + ], + "type": "text", + "content": ", we use volumetric rendering [31]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 297, + 276, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 297, + 276, + 323 + ], + "spans": [ + { + "bbox": [ + 58, + 297, + 276, + 323 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {C}} (\\mathbf {r}, t, s) = \\int_ {u _ {n}} ^ {u _ {f}} U (u) \\sigma (\\mathbf {r} (u), t, s) \\mathbf {c} (\\mathbf {r} (u), \\mathbf {d}, t, s) d u", + "image_path": "3b44f4719325f2bac8baff8f4d11911e51e5a6339af61c633cdeb2084a6a7ea4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 325, + 287, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 325, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 82, + 325, + 287, + 352 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} U (u) = \\exp \\left(- \\int_ {u _ {n}} ^ {u} \\sigma \\left(u ^ {\\prime}\\right) d u ^ {\\prime}\\right). \\tag {11}", + "image_path": "a009dc0643d0a00ecde025fefd5b55d7eb45cea825ea80d2669b1e437e5abbc9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 361, + 286, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 286, + 384 + ], + "type": "text", + "content": "We obtain density " + }, + { + "bbox": [ + 47, + 361, + 286, + 384 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 361, + 286, + 384 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 47, + 361, + 286, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 47, + 361, + 286, + 384 + ], + "type": "text", + "content": " at sampling location " + }, + { + "bbox": [ + 47, + 361, + 286, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(u)" + }, + { + "bbox": [ + 47, + 361, + 286, + 384 + ], + "type": "text", + "content": " as mixture of the radiance fields " + }, + { + "bbox": [ + 47, + 361, + 286, + 384 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 361, + 286, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 361, + 286, + 384 + ], + "type": "inline_equation", + "content": "\\psi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 393, + 287, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 393, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 61, + 393, + 287, + 417 + ], + "type": "interline_equation", + "content": "\\sigma = \\sigma_ {\\phi} + \\sigma_ {\\psi}, \\mathbf {c} = \\frac {\\sigma_ {\\phi}}{\\sigma_ {\\phi} + \\sigma_ {\\psi}} \\mathbf {c} _ {\\phi} + \\frac {\\sigma_ {\\psi}}{\\sigma_ {\\phi} + \\sigma_ {\\psi}} \\mathbf {c} _ {\\psi}. \\tag {12}", + "image_path": "5d36e05e9d78ddf1de8803199254a72a3a800b818bc2cadc41683ac19e1f8cc7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 426, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 461 + ], + "type": "text", + "content": "Crucially, we set " + }, + { + "bbox": [ + 46, + 426, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\sigma_{\\psi}" + }, + { + "bbox": [ + 46, + 426, + 287, + 461 + ], + "type": "text", + "content": " to zero when " + }, + { + "bbox": [ + 46, + 426, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(u)" + }, + { + "bbox": [ + 46, + 426, + 287, + 461 + ], + "type": "text", + "content": " does not lie within a 3D bounding box of a dynamic node given the calculated entry and exit points " + }, + { + "bbox": [ + 46, + 426, + 287, + 461 + ], + "type": "inline_equation", + "content": "u^{\\mathrm{in}}, u^{\\mathrm{out}}" + }, + { + "bbox": [ + 46, + 426, + 287, + 461 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": "Composite ray sampling. Instead of densely sampling the space [18, 42] or leveraging separate ray sampling mechanisms per node [35], we use a composite ray sampling strategy illustrated in Fig. 5. We extend the proposal sampling mechanism introduced in [2] by joint sampling from a computationally efficient density field " + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{prop}}" + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": " and dynamic nodes " + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "inline_equation", + "content": "v_{o}" + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": ". In particular, as in Eq. 12, we represent " + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\sigma (\\mathbf{r}(u),t,s)" + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": " by a mixture of " + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{prop}}(\\mathbf{r}(u),\\omega_s^t)" + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\sigma_{\\psi}(\\mathbf{r}(u),\\omega_s^t,\\omega_o)" + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": ". However, analogous to our rendering step, we constrain sampling from " + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\sigma_{\\psi}" + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "inline_equation", + "content": "[u^{\\mathrm{in}},u^{\\mathrm{out}}]" + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": ". This allows us to skip empty space efficiently by distilling the static density " + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\sigma_{\\phi}" + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{prop}}" + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": ". At the same time, sparsely querying " + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\sigma_{\\psi}" + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(u)" + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": " falls into a dynamic node enables us to still accurately represent the full, dynamic " + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 462, + 287, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Since there are only few samples that fall into dynamic nodes when performing uniform sampling, the computational overhead in the first ray sampling iteration is negligible. In the second iteration, we apply inverse transform sampling given the CDF " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "F(u) = 1 - U(u)" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " along ray " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " and thus the samples are concentrated at the first surface intersection. Hence, only for rays that fall in the line of sight of" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 322, + 69, + 451, + 168 + ], + "blocks": [ + { + "bbox": [ + 322, + 69, + 451, + 168 + ], + "lines": [ + { + "bbox": [ + 322, + 69, + 451, + 168 + ], + "spans": [ + { + "bbox": [ + 322, + 69, + 451, + 168 + ], + "type": "image", + "image_path": "f6d9ca72be293397534baac9abb3b55225300c188e7f02bca0cef30689b151ac.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "lines": [ + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "text", + "content": "Figure 5. Composite ray sampling. If a ray intersects with an object " + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "inline_equation", + "content": "v_{o}" + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "text", + "content": ", we sample from both proposal network " + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{prop}}" + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "text", + "content": " and radiance field " + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{prop}}" + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "text", + "content": " otherwise. We condition each with the latents " + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "text", + "content": " of the respective nodes. The PDF is a mixture of all node densities that intersect with the ray. The transmittance " + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "text", + "content": " drops at the first surface intersection (tree) where further samples will concentrate." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 455, + 70, + 533, + 168 + ], + "blocks": [ + { + "bbox": [ + 455, + 70, + 533, + 168 + ], + "lines": [ + { + "bbox": [ + 455, + 70, + 533, + 168 + ], + "spans": [ + { + "bbox": [ + 455, + 70, + 533, + 168 + ], + "type": "image", + "image_path": "5c4d60f029ed5277f343137817a12ee966bebe9263f7388265a021dda8eadf0b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 243, + 544, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 544, + 268 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 544, + 268 + ], + "type": "text", + "content": "an object surface will we sample " + }, + { + "bbox": [ + 304, + 243, + 544, + 268 + ], + "type": "inline_equation", + "content": "\\sigma_{\\psi}" + }, + { + "bbox": [ + 304, + 243, + 544, + 268 + ], + "type": "text", + "content": " more than a few times. In total, we use two ray sampling iterations as in [54]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "text", + "content": "Space contraction. Following [2, 54], we contract the unbounded scene space into a fixed-size bounding box, normalizing the scene with bounds computed from the LiDAR point clouds and the ego-vehicle poses " + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "text", + "content": " into a unit cube." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 323, + 389, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 389, + 336 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 389, + 336 + ], + "type": "text", + "content": "5. Optimization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 340, + 541, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 340, + 541, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 340, + 541, + 353 + ], + "type": "text", + "content": "We optimize the parameters " + }, + { + "bbox": [ + 306, + 340, + 541, + 353 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 306, + 340, + 541, + 353 + ], + "type": "text", + "content": " of the radiance field " + }, + { + "bbox": [ + 306, + 340, + 541, + 353 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 306, + 340, + 541, + 353 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 359, + 525, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 525, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 525, + 385 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {(\\mathbf {r}, t, s) \\in \\mathcal {R}} \\mathcal {L} _ {\\mathrm {r g b}} (\\mathbf {r}, t, s) + \\lambda_ {\\mathrm {d i s t}} \\mathcal {L} _ {\\mathrm {d i s t}} (\\mathbf {r}, t, s) + \\lambda_ {\\mathrm {p r o p}}", + "image_path": "677a18388569a723b15d0dc6e531d8732f89ce0f6afabd8e87146e5fe100b2f0.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 332, + 388, + 545, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 388, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 332, + 388, + 545, + 414 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r o p}} (\\mathbf {r}, t, s) + \\lambda_ {\\text {d e p}} \\mathcal {L} _ {\\text {d e p}} (\\mathbf {r}, t, s) + \\lambda_ {\\text {e n t r}} \\mathcal {L} _ {\\text {e n t r}} (\\mathbf {r}, t, s) \\tag {13}", + "image_path": "01ac410a1b694f1d38ee119ef6acb9ba3956bd6689bc4cb287ae016197f6ca66.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 422, + 545, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 545, + 446 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 422, + 545, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dist}}" + }, + { + "bbox": [ + 306, + 422, + 545, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 422, + 545, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{prop}}" + }, + { + "bbox": [ + 306, + 422, + 545, + 446 + ], + "type": "text", + "content": " follow [2]. We supervise " + }, + { + "bbox": [ + 306, + 422, + 545, + 446 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{prop}}" + }, + { + "bbox": [ + 306, + 422, + 545, + 446 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 306, + 422, + 545, + 446 + ], + "type": "inline_equation", + "content": "\\sigma_{\\phi}" + }, + { + "bbox": [ + 306, + 422, + 545, + 446 + ], + "type": "text", + "content": " only to learn effective composite ray sampling." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 446, + 544, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 446, + 544, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 446, + 544, + 469 + ], + "type": "text", + "content": "Photometric loss. We compare the rendered training rays with their ground truth color" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 341, + 476, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 476, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 341, + 476, + 545, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r g b}} (\\mathbf {r}, t, s) = \\left\\| \\mathbf {C} (\\mathbf {r}, t, s) - \\hat {\\mathbf {C}} (\\mathbf {r}, t, s) \\right\\| _ {2} ^ {2}. \\tag {14}", + "image_path": "f06225ad29afcfba900e13bd14919f1c26df95167abaa808033cfc90bd019fcf.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 498, + 544, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 544, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 544, + 522 + ], + "type": "text", + "content": "Expected depth loss. We render the expected depth values for each ray and compare it with the ground truth depth" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 341, + 529, + 545, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 529, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 341, + 529, + 545, + 543 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d e p}} (\\mathbf {r}, t, s) = \\left\\| \\mathbf {D} (\\mathbf {r}, t, s) - \\hat {\\mathbf {D}} (\\mathbf {r}, t, s) \\right\\| _ {2} ^ {2} \\tag {15}", + "image_path": "ee6a2419d0e8937b5423cb5315ed532a2bb0aa8ebc19c7bcd2fb1679665ecc3e.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 550, + 544, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 544, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 544, + 574 + ], + "type": "text", + "content": "where the expected depth is calculated via integrating the sampling values " + }, + { + "bbox": [ + 304, + 550, + 544, + 574 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}} (\\mathbf{r},t,s) = \\int_{u^{\\prime}}^{u_{f}}uU(u)\\sigma (\\mathbf{r}(u),t,s)du" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 574, + 545, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 545, + 645 + ], + "type": "text", + "content": "Entropy regularization. While static and dynamic scene parts can overlap, i.e. inside a 3D bounding box " + }, + { + "bbox": [ + 304, + 574, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_o^t" + }, + { + "bbox": [ + 304, + 574, + 545, + 645 + ], + "type": "text", + "content": " there could be a part of the street or sidewalk, their density should be strictly separated w.r.t. a single sampling location " + }, + { + "bbox": [ + 304, + 574, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(u)" + }, + { + "bbox": [ + 304, + 574, + 545, + 645 + ], + "type": "text", + "content": ". We leverage an entropy regularization loss [62] to encourage clear separation between entities in " + }, + { + "bbox": [ + 304, + 574, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 574, + 545, + 645 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 574, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\psi" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 652, + 545, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 652, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 306, + 652, + 545, + 689 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {e n t r}} (\\mathbf {r}, t, s) = \\int_ {u _ {n}} ^ {u _ {f}} \\mathcal {H} \\left(\\frac {\\sigma_ {\\psi} (\\mathbf {r} (u) , t , s)}{\\sigma_ {\\phi} (\\mathbf {r} (u) , s) + \\sigma_ {\\psi} (\\mathbf {r} (u) , t , s)}\\right) d u \\tag {16}", + "image_path": "b7da631a4580b847af33a26fb1de1ad032a705efeb22f8b6edee05abbe12b63f.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 689, + 544, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 544, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 544, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot)" + }, + { + "bbox": [ + 304, + 689, + 544, + 713 + ], + "type": "text", + "content": " is the Shannon entropy and we use " + }, + { + "bbox": [ + 304, + 689, + 544, + 713 + ], + "type": "inline_equation", + "content": "t, s" + }, + { + "bbox": [ + 304, + 689, + 544, + 713 + ], + "type": "text", + "content": " as a replacement for the vectors " + }, + { + "bbox": [ + 304, + 689, + 544, + 713 + ], + "type": "inline_equation", + "content": "\\omega_s^t, \\omega_o" + }, + { + "bbox": [ + 304, + 689, + 544, + 713 + ], + "type": "text", + "content": " for ease of notation." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21129" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 71, + 523, + 131 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 523, + 131 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 523, + 131 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 523, + 131 + ], + "type": "table", + "html": "
MethodResidentialDowntownMeanTrain time (h)
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerfacto + Emb.19.830.6370.56218.050.6550.62518.940.6460.5948.0
Nerfacto + Emb. + Time20.050.6410.56218.660.6560.60319.360.6540.58313.2
SUDS [58]21.760.6590.55619.910.6650.64520.840.6620.60154.8
Ours22.290.6780.52320.010.6810.58621.150.6800.55517.2
", + "image_path": "9d7eed8e1f912b7c30285f605ca4fa3fa127454dd16945a035ec46f7e7e967a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 96, + 169, + 499, + 343 + ], + "blocks": [ + { + "bbox": [ + 46, + 134, + 547, + 168 + ], + "lines": [ + { + "bbox": [ + 46, + 134, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 547, + 168 + ], + "type": "text", + "content": "Table 1. Novel View Synthesis on Argoverse 2. While the static Nerfacto baseline has the weakest performance, the dynamic variant Nerfacto + Time improves only marginally upon it. The state-of-the-art method SUDS exhibits stronger view synthesis results but takes more than " + }, + { + "bbox": [ + 46, + 134, + 547, + 168 + ], + "type": "inline_equation", + "content": "3\\times" + }, + { + "bbox": [ + 46, + 134, + 547, + 168 + ], + "type": "text", + "content": " longer to train. Our method outperforms all methods across all metrics and exhibits competitive training speed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 96, + 169, + 499, + 343 + ], + "lines": [ + { + "bbox": [ + 96, + 169, + 499, + 343 + ], + "spans": [ + { + "bbox": [ + 96, + 169, + 499, + 343 + ], + "type": "table", + "html": "
MethodKITTI [75%]KITTI [50%]KITTI [25%]
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [31]18.560.5570.55419.120.5870.49718.610.5700.510
NeRF + Time21.010.6120.49221.340.6350.44819.550.5860.505
NSG [35]21.530.6730.25421.260.6590.26620.000.6320.281
Nerfacto + Emb.22.750.8010.15622.380.7930.16021.240.7580.178
Nerfacto + Emb. + Time23.190.8040.15523.180.8030.15521.980.7770.172
SUDS [58]22.770.7970.17123.120.8210.13520.760.7470.198
Ours28.380.9070.05227.510.8980.05526.510.8870.060
MethodVKITTI2 [75%]VKITTI2 [50%]VKITTI2 [25%]
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [31]18.670.5480.63418.580.5440.63518.170.5370.644
NeRF + Time19.030.5740.58718.900.5650.61018.040.5450.626
NSG [35]23.410.6890.31723.230.6790.32521.290.6660.317
Nerfacto + Emb.22.150.8470.14521.880.8430.14821.280.8270.155
Nerfacto + Emb. + Time22.110.8490.14421.780.8440.14721.000.8250.157
SUDS [58]23.870.8460.15023.780.8510.14222.180.8290.160
Ours29.730.9120.06529.190.9060.06628.290.9010.067
", + "image_path": "04f350d262207905fdf5b7d913cfade7b6775a9806ce3ea8c9c56298f9f1a9ea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 346, + 547, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 547, + 381 + ], + "type": "text", + "content": "Table 2. Novel View Synthesis on KITTI and VKITTI2. We compare our method to prior art, following the experimental protocol in [58]. We test the view synthesis performance of the methods with varying fractions of training views and observe that fewer training views generally result in lower performance. Our method outperforms previous works by a large margin on all settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "text", + "content": "Hierarchical pose optimization. Alongside scene geometry, we optimize edges " + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "inline_equation", + "content": "e_{v_s^t v_r}" + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "inline_equation", + "content": "e_{v_o v_s^t}" + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "text", + "content": " in our graph, i.e. we refine ego-vehicle poses " + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_s^t\\in \\mathbf{SE}(3)" + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "text", + "content": " and object poses " + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\xi_o^t\\in \\mathbf{SE}(3)" + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "text", + "content": ". In particular, we optimize for pose residuals " + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{P}_s^t\\in \\mathfrak{se}(3)" + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\delta \\xi_o^t\\in \\mathfrak{se}(2)" + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "text", + "content": ". We constrain the object pose residual to " + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\mathfrak{se}(2)" + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "text", + "content": " since objects are usually upright and move along the ground plane. Given each residual, we update the ego-vehicle pose with " + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_s^t = \\expmap(\\delta \\mathbf{P}_s^t)\\mathbf{P}_s^t" + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "text", + "content": " and the object pose with " + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\hat{\\xi}_o^t = \\expmap(\\delta \\xi_o^t)_{\\mathbf{SE}(2)\\to \\mathbf{SE}(3)}\\xi_o^t" + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\expmap(\\cdot)" + }, + { + "bbox": [ + 46, + 386, + 287, + 507 + ], + "type": "text", + "content": " is the exponential map of each lie group." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 508, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 288, + 651 + ], + "type": "text", + "content": "Compared to naively optimizing camera and object poses, optimizing the edges along our scene graph has two key advantages. First, we leverage multi-camera constraints, i.e. we keep the camera extrinsics " + }, + { + "bbox": [ + 46, + 508, + 288, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_c" + }, + { + "bbox": [ + 46, + 508, + 288, + 651 + ], + "type": "text", + "content": " fixed and optimize the ego-vehicle poses " + }, + { + "bbox": [ + 46, + 508, + 288, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_s^t" + }, + { + "bbox": [ + 46, + 508, + 288, + 651 + ], + "type": "text", + "content": " only. This is in contrast to prior art that generally treats each camera pose as independent. Second, the residual " + }, + { + "bbox": [ + 46, + 508, + 288, + 651 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{P}_s^t" + }, + { + "bbox": [ + 46, + 508, + 288, + 651 + ], + "type": "text", + "content": " propagates to the object poses since they are defined w.r.t. the ego-vehicle coordinate frame instead of the global world frame. Given that optimizing a radiance field as well as camera and object poses jointly is notoriously difficult [21], these constraints are vital to view synthesis quality (see Tab. 6)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 659, + 128, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 128, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 128, + 673 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Datasets. To evaluate against competing methods on our proposed benchmark on Argoverse 2 [61], we hold out every 10th sample in uniform time intervals where a sample" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 386, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 458 + ], + "type": "text", + "content": "corresponds to seven ring-camera images. To compare our methods against prior art on KITTI [14] and VKITTI2 [4], we follow the experimental protocol and data splits in [58]. We use the provided 3D bounding box annotations for all datasets. We provide results using an off-the-shelf 3D tracker [68] in the supplemental material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 459, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 484 + ], + "type": "text", + "content": "Metrics. Following [58], we measure image synthesis quality with PSNR, SSIM [59], and LPIPS (AlexNet) [69]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "text", + "content": "Implementation details. We implement our method in PyTorch [39], accelerating the time-consuming ray-node intersection with a custom CUDA implementation. We train our model on a single RTX 3090 GPU for 250,000 steps on Argovvere 2 and 100,000 steps on KITTI and VKITTI2, with 8192 rays per batch. All model parameters are optimized using Adam [17] with " + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "text", + "content": ", and an exponential decay learning rate schedule: from " + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "text", + "content": " for pose parameters, and from " + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "text", + "content": " for all others. In order to counter pose drift, we further apply weight decay with a factor " + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{P}" + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\delta \\xi" + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "text", + "content": ". The static radiance field " + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "text", + "content": " is trained from scratch, while " + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 484, + 545, + 640 + ], + "type": "text", + "content": " is initialized from a semantic prior [6] following [18]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Baselines. We compare our method against prior work in dynamic outdoor scene representation, i.e. SUDS [58] and NSG [35]. Further, we include results obtained by augmenting Nerfacto [54], the closest state-of-the-art NeRF architecture to ours, with components designed to handle multicapture reconstruction and scene dynamics. In particular," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21130" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 71, + 286, + 99 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 286, + 99 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 286, + 99 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 286, + 99 + ], + "type": "table", + "html": "
SRN [49]NeRF [31]NeRF + TimeNSG [35]PNF [18]SUDS [58]Ours
PSNR ↑18.8323.3424.1826.6627.4828.3129.36
SSIM ↑0.5900.6620.6770.8060.8700.8760.911
", + "image_path": "df396e5d2396063fdcca86347bed923f065020bbab0a001f7db96ac12891593d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 137, + 287, + 212 + ], + "blocks": [ + { + "bbox": [ + 49, + 137, + 287, + 212 + ], + "lines": [ + { + "bbox": [ + 49, + 137, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 49, + 137, + 287, + 212 + ], + "type": "image", + "image_path": "8783d9847c2883eb37addca49651ab0b828c5bc165f9a085a3642881e6242f4f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 216, + 287, + 250 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 287, + 250 + ], + "type": "text", + "content": "Figure 6. Qualitative results on KITTI. With only " + }, + { + "bbox": [ + 46, + 216, + 287, + 250 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 216, + 287, + 250 + ], + "type": "text", + "content": " of views (approx. 15-20) for training, we can still synthesize sharp and realistic novel views with dynamic objects rendered at high quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 256, + 287, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 328 + ], + "type": "text", + "content": "we consider two variants: \"Nerfacto + Emb.\", where we incorporate our appearance embedding " + }, + { + "bbox": [ + 46, + 256, + 287, + 328 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_s\\mathcal{F}(t)" + }, + { + "bbox": [ + 46, + 256, + 287, + 328 + ], + "type": "text", + "content": " and the expected depth loss; and \"Nerfacto + Emb. + Time\", where we further include time modeling with 4D hash tables following [38, 63]. For both Nerfacto variants, we set hash table and MLP sizes to be aligned with our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 336, + 212, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 212, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 212, + 348 + ], + "type": "text", + "content": "6.1. Comparison to state-of-the-art" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 354, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 484 + ], + "type": "text", + "content": "We first compare to prior work on our proposed benchmark in Tab 1. We observe that the static Nerfacto + Emb. has the weakest performance, which slightly increases by adding time modeling in Nerfacto + Emb. + Time. Meanwhile, the current state-of-the-art method SUDS [58] outperforms the Nerfacto variants, while being much slower to train. Our method performs the best on all metrics. The improvement is particularly pronounced in the perceptual quality metrics (SSIM and LPIPS). The training speed of our method is competitive to Nerfacto + Emb. + Time and more than " + }, + { + "bbox": [ + 46, + 354, + 287, + 484 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 354, + 287, + 484 + ], + "type": "text", + "content": " faster than SUDS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 665 + ], + "type": "text", + "content": "In addition, we show a qualitative comparison in Fig. 7. We observe major differences in both the rendered images and depth maps. In particular, all other methods struggle to recover the dynamic objects in the scene, while our method produces realistic renderings and accurate depth maps for both static and dynamic areas. We also observe that, thanks to our transient geometry embedding " + }, + { + "bbox": [ + 46, + 486, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_s\\mathcal{F}(t)" + }, + { + "bbox": [ + 46, + 486, + 287, + 665 + ], + "type": "text", + "content": ", we are able to accurately recover the geometry of the trees and their leaves which notably are not present in every sequence of the area that is reconstructed due to seasonal changes. At the same time, other methods struggle to recover those details. Specifically, other methods exhibit artifacts in the depth maps and degraded rendering quality of the trees left and right of the street in columns five and six of Fig. 7, while our method produces accurate color and depth renderings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Next, we evaluate our method on established benchmarks, namely KITTI and VKITTI2, in Tab. 2 following the experimental protocol in [58]. In particular, we compare novel view synthesis quality at different fractions of" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 334, + 71, + 518, + 114 + ], + "blocks": [ + { + "bbox": [ + 47, + 102, + 286, + 135 + ], + "lines": [ + { + "bbox": [ + 47, + 102, + 286, + 135 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 286, + 135 + ], + "type": "text", + "content": "Table 3. Image reconstruction on KITTI. We outperform prior art in image reconstruction, i.e. our method can better fit the training views. We follow the experimental protocol in [18, 35, 58]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 71, + 518, + 114 + ], + "lines": [ + { + "bbox": [ + 334, + 71, + 518, + 114 + ], + "spans": [ + { + "bbox": [ + 334, + 71, + 518, + 114 + ], + "type": "table", + "html": "
AsF(t)GsF(t)PSNR ↑SSIM ↑LPIPS ↓
--19.700.6530.588
✓-22.220.6700.546
✓✓22.490.6710.541
", + "image_path": "1d0882b14f8616a3ddfc5d4f5b2220b1f5fd17dc4d9b092b3b9bc3b31c3c1318.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 175, + 545, + 225 + ], + "blocks": [ + { + "bbox": [ + 305, + 118, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 305, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 305, + 118, + 545, + 173 + ], + "type": "text", + "content": "Table 4. Ablation study on graph structure. We show that using the multi-level graph structure of sequences and objects is crucial, i.e. omitting sequence vectors " + }, + { + "bbox": [ + 305, + 118, + 545, + 173 + ], + "type": "inline_equation", + "content": "\\omega_{s}^{t}" + }, + { + "bbox": [ + 305, + 118, + 545, + 173 + ], + "type": "text", + "content": " results in degraded quality since there is no conditioning on scene-specific appearance. Combining appearance and transient geometry embeddings performs best." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 175, + 545, + 225 + ], + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 225 + ], + "type": "table", + "html": "
SamplingSamples per rayPSNR ↑SSIM ↑LPIPS ↓Rays / sec.
Uniform [18]19225.720.7300.45628K
Uniform [18]102425.840.7340.4494K
Separate [35]1024+64+(32+64)26.650.7620.3512.5K
Ours1024+6427.070.7590.36230K
", + "image_path": "281d98977513d4113854bd40d8f22abe911f6065656fcc6e6ef8f800693738f4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 340, + 275, + 511, + 319 + ], + "blocks": [ + { + "bbox": [ + 305, + 229, + 545, + 274 + ], + "lines": [ + { + "bbox": [ + 305, + 229, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 305, + 229, + 545, + 274 + ], + "type": "text", + "content": "Table 5. Ray Sampling schemes. Our composite ray sampling is about " + }, + { + "bbox": [ + 305, + 229, + 545, + 274 + ], + "type": "inline_equation", + "content": "12 \\times" + }, + { + "bbox": [ + 305, + 229, + 545, + 274 + ], + "type": "text", + "content": " more efficient to train than separate ray sampling [35] with comparable performance. Uniform sampling [18] exhibits lower performance and is slow when sampled more densely." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 275, + 511, + 319 + ], + "lines": [ + { + "bbox": [ + 340, + 275, + 511, + 319 + ], + "spans": [ + { + "bbox": [ + 340, + 275, + 511, + 319 + ], + "type": "table", + "html": "
Pose optimizationPSNR ↑SSIM ↑LPIPS ↓
-22.490.6710.541
Naive21.280.6630.519
Hierarchical (Ours)22.290.6780.523
", + "image_path": "19f67193ed78db2061463f6c19284d05b068c2908a5b5688c0e9cdc7c436a492.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 323, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 377 + ], + "type": "text", + "content": "Table 6. Pose optimization. Compared to naively optimizing camera and object poses, optimizing vehicle and object poses hierarchically on the edges of our graph benefits view synthesis quality more consistently, i.e. both in SSIM and LPIPS. Note that PSNR is very sensitive to pose drift, and thus does not improve." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 386, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 517 + ], + "type": "text", + "content": "training views, i.e. testing different levels of view supervision. We run our method and our baselines and observe substantially better view synthesis quality for our method, both compared to our baselines and prior works. All methods exhibit a similar performance drop when the fraction of training views decreases. We illustrate the rendering quality of our method with only 15-20 training views in Fig. 6. Finally, we also report the results on the task of image reconstruction, i.e. reconstruction of images seen during training, following [18, 35, 58] in Tab. 3 on the KITTI dataset. We significantly outperform previous works in both metrics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 527, + 404, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 404, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 404, + 538 + ], + "type": "text", + "content": "6.2. Ablation studies" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "We verify our design through ablation studies. We perform these on the residential area of our benchmark unless otherwise noted. First, we ablate on the multi-level structure of our scene graph. In particular, in Tab. 4, we ablate the latent codes of the sequence nodes, reducing to a representation similar to [35] and observe that the quality of the synthesized views drops significantly. In contrast, our scene graph representation achieves better view synthesis quality, i.e. achieves almost three points higher PSNR. Further, adding our transient geometry embeddings in addition to sequence-level appearance embeddings improves the view synthesis quality. See supplemental material for a qualitative comparison. This adds to the fact that through our multi-level graph structure, we show the ability to modulate car appear" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21131" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 544, + 383 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 544, + 383 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 544, + 383 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 544, + 383 + ], + "type": "image", + "image_path": "3b7faa7dc7233a4c56f0201e825efde6aefde940cee478759f2e25fa02cde53e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 387, + 546, + 421 + ], + "lines": [ + { + "bbox": [ + 46, + 387, + 546, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 546, + 421 + ], + "type": "text", + "content": "Figure 7. Qualitative results on Argoverse 2. While prior art struggles with dynamic actors in the scene, our representation renders both realistic novel views and plausible depth maps. Further, our method models complex transient geometry subject to, e.g., seasonal changes such as tree leaves. Note that prior art produces depth artifacts and exhibits degraded view quality in those regions (columns 5 and 6)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 430, + 228, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 228, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 228, + 442 + ], + "type": "text", + "content": "ance through sequence appearance in Fig. 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 446, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 625 + ], + "type": "text", + "content": "In Tab. 5, we compare our composite ray sampling scheme to uniformly sampled rays as in [18] and rays sampled separately per node as in [35]. We compare the schemes under the assumption that all other model parameters are equal, implementing them in our method. Ours is about " + }, + { + "bbox": [ + 46, + 446, + 287, + 625 + ], + "type": "inline_equation", + "content": "12 \\times" + }, + { + "bbox": [ + 46, + 446, + 287, + 625 + ], + "type": "text", + "content": " more efficient to train than densely sampling a ray in uniform intervals or separately sampling the ray per node while producing similar results to the latter and being significantly better than the former. Sparsely sampling a ray at uniform intervals is similarly fast as our method, but yields degraded view synthesis quality. Note that we run this ablation study on only a single sequence of the residential area in our benchmark since densely sampling the rays and separately sampling the rays for all nodes are prohibitively expensive to train in large-scale urban areas." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Finally, in Tab. 6, we compare our hierarchical pose optimization to naive camera pose optimization employed in previous works [21, 54]. While naive camera pose optimization degrades the results significantly in pixel-wise metrics, our hierarchical pose optimization improves the SSIM and maintain a comparable PSNR. Meanwhile, our hierarchical pose optimization exhibits similar gains to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 430, + 546, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 546, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 546, + 502 + ], + "type": "text", + "content": "naive pose optimization in terms of the perceptual LPIPS metric. This shows that our pose optimization mitigates pose drift while also enabling a more accurate reconstruction. Pose drift usually causes a misalignment between the evaluation viewpoint and scene geometry, which degrades pixel-wise metrics in particular." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 514, + 379, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 379, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 379, + 526 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "We introduce a novel multi-level scene graph representation for radiance field reconstruction in dynamic urban environments that scales to large geographic areas with more than ten thousand images from dozens of sequences and hundreds of dynamic objects. We train our representation with an efficient composite ray sampling and rendering scheme and introduce latent variables that enable modeling complex phenomena like varying environmental conditions and transient geometry present across different vehicle captures. We leverage our representation to refine camera and object poses hierarchically using multi-camera constraints. Finally, we propose a new view synthesis benchmark for dynamic urban driving scenarios. Our approach yields substantially improved results compared to prior art while allowing for flexible de- and recomposition of the scene." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21132" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Iro Armeni, Zhi-Yang He, JunYoung Gwak, Amir R Zamir, Martin Fischer, Jitendra Malik, and Silvio Savarese. 3d scene graph: A structure for unified semantics, 3d space, and camera. In ICCV, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 167 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 190 + ], + "type": "text", + "content": "[3] Jules Bloomenthal and Brian Wyvill. Introduction to implicit surfaces. 1997. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 192, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 287, + 213 + ], + "type": "text", + "content": "[4] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual kitti 2, 2020. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 214, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 267 + ], + "type": "text", + "content": "[5] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "text", + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 325, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 287, + 357 + ], + "type": "text", + "content": "[7] Steve Cunningham and Michael J Bailey. Lessons from scene graphs: using scene graphs to teach hierarchical modeling. Computers & Graphics, 25(4):703-711, 2001. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 358, + 287, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 400 + ], + "type": "text", + "content": "[8] Nathaniel Fairfield, George A. Kantor, and David S. Wettergreen. Real-time slam with octree evidence grids for exploration in underwater tunnels. Journal of Field Robotics, 24, 2007. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "text", + "content": "[9] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 447, + 287, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 287, + 489 + ], + "type": "text", + "content": "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 522 + ], + "type": "text", + "content": "[11] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In CVPR, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 525, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 556 + ], + "type": "text", + "content": "[12] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 559, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 590 + ], + "type": "text", + "content": "[13] Yiming Gao, Yan-Pei Cao, and Ying Shan. Surfelnerf: Neural surfel radiance fields for online photorealistic reconstruction of indoor scenes. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 592, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 623 + ], + "type": "text", + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 625, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 666 + ], + "type": "text", + "content": "[15] Michael Kaess. Simultaneous localization and mapping with infinite planes. 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 4605-4611, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 160 + ], + "type": "text", + "content": "[18] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 194 + ], + "type": "text", + "content": "[19] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "type": "text", + "content": "[20] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In CVPR, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "type": "text", + "content": "[21] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "text", + "content": "[22] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 294, + 545, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 325 + ], + "type": "text", + "content": "[23] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 327, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 369 + ], + "type": "text", + "content": "[24] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "type": "text", + "content": "[25] Yan Lu and Dezhen Song. Visual navigation using heterogeneous landmarks and unsupervised geometric constraints. IEEE Transactions on Robotics, 31:736-749, 2015. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 405, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 436 + ], + "type": "text", + "content": "[26] Jonathon Luiten, Tobias Fischer, and Bastian Leibe. Track to reconstruct and reconstruct to track. IEEE Robotics and Automation Letters, 5(2):1803-1810, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 438, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 480 + ], + "type": "text", + "content": "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "text", + "content": "[28] Alexander Majercik, Cyril Crassin, Peter Shirley, and Morgan McGuire. A ray-box intersection algorithm and efficient dynamic voxel rendering. Journal of Computer Graphics Techniques Vol, 7(3):66-81, 2018. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 526, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 569 + ], + "type": "text", + "content": "[29] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 571, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 612 + ], + "type": "text", + "content": "[30] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 614, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 667 + ], + "type": "text", + "content": "[31] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[32] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21133" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[33] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "text", + "content": "[34] OpenStreetMap contributors. Planet dump retrieved from https://planet.osm.org . https : //www.openstreetmap.org, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "type": "text", + "content": "[35] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In CVPR, 2021. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 217 + ], + "type": "text", + "content": "[36] Byeongjun Park and Changick Kim. Point-dynrf: Point-based dynamic radiance fields from a monocular video, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 286, + 274 + ], + "type": "text", + "content": "[37] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 276, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 286, + 319 + ], + "type": "text", + "content": "[38] Sungheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, and Nahiyup Kang. Temporal interpolation is all you need for dynamic neural radiance fields. In CVPR, 2023. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 321, + 286, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 321, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 49, + 321, + 286, + 374 + ], + "type": "text", + "content": "[39] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 376, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 376, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 376, + 286, + 453 + ], + "type": "text", + "content": "[40] Marc Pollefeys, David Nister, Jan-Michael Frahm, Amir Akbarzadeh, Philippos Mordohai, Brian Clipp, Chris Engels, David Gallup, Seon Joo Kim, Paul C. Merrell, C. Salmi, Sudipta N. Sinha, B. Talton, Liang Wang, Qingxiong Yang, Henrik Stewénius, Ruigang Yang, Greg Welch, and Herman Towles. Detailed real-time urban 3d reconstruction from video. IJCV, 78:143-167, 2007. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 455, + 286, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 455, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 49, + 455, + 286, + 487 + ], + "type": "text", + "content": "[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 488, + 286, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 488, + 286, + 531 + ], + "spans": [ + { + "bbox": [ + 49, + 488, + 286, + 531 + ], + "type": "text", + "content": "[42] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 533, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 533, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 533, + 286, + 555 + ], + "type": "text", + "content": "[43] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 556, + 286, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 556, + 286, + 600 + ], + "spans": [ + { + "bbox": [ + 49, + 556, + 286, + 600 + ], + "type": "text", + "content": "[44] Antoni Rosinol, Arjun Gupta, Marcus Abate, Jingnan Shi, and Luca Carlone. 3d dynamic scene graphs: Actionable spatial perception with places, objects, and humans. arXiv preprint arXiv:2002.06289, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 601, + 286, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 286, + 634 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 286, + 634 + ], + "type": "text", + "content": "[45] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In ECCV, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 635, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 286, + 678 + ], + "type": "text", + "content": "[46] Renato F Salas-Moreno, Richard A Newcombe, Hauke Strasdat, Paul HJ Kelly, and Andrew J Davison. Slam++: Simultaneous localisation and mapping at the level of objects. In CVPR, 2013. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 680, + 286, + 713 + ], + "type": "text", + "content": "[47] Aron Schmied, Tobias Fischer, Martin Danelljan, Marc Pollefeys, and Fisher Yu. R3d3: Dense 3d reconstruction of dynamic scenes from multiple cameras. In ICCV, 2023." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[48] Chen Shen, James F. O'Brien, and Jonathan R. Shewchuk. Interpolating and approximating implicit surfaces from polygon soup. In ACM SIGGRAPH 2004 Papers, page 896-904, 2004. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "text", + "content": "[49] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. NeurIPS, 2019. 1, 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "type": "text", + "content": "[50] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 230, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 251 + ], + "type": "text", + "content": "[51] Henry Sowizral. Scene graphs in the new millennium. IEEE Computer Graphics and Applications, 20(1):56-57, 2000. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "type": "text", + "content": "[52] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "text", + "content": "[53] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In CVPR, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 332, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 397 + ], + "type": "text", + "content": "[54] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 5, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "text", + "content": "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 455, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 497 + ], + "type": "text", + "content": "[56] Shubham Tulsiani, Saurabh Gupta, David F Fouhey, Alexei A Efros, and Jitendra Malik. Factoring shape, pose, and layout from the 2d image of a 3d scene. In CVPR, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 500, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 533 + ], + "type": "text", + "content": "[57] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 535, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 566 + ], + "type": "text", + "content": "[58] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In CVPR, 2023. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 568, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 611 + ], + "type": "text", + "content": "[59] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "type": "text", + "content": "[60] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[61] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21134" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 388 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "self-driving perception and forecasting. arXiv preprint arXiv:2301.00493, 2023. 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[62] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D' 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS, 2022. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "type": "text", + "content": "[63] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In CVPR, 2021. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[64] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "type": "text", + "content": "[65] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In CVPR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "type": "text", + "content": "[66] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In CVPR, 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "type": "text", + "content": "[67] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, 2023. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 332, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 354 + ], + "type": "text", + "content": "[68] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "type": "text", + "content": "[69] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21135" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-Modal Hallucination Control by Visual Information Grounding/1e5cae18-a58b-4415-90f8-bc4858438f19_content_list.json b/2024/Multi-Modal Hallucination Control by Visual Information Grounding/1e5cae18-a58b-4415-90f8-bc4858438f19_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..28be1e36d549a22c3b492d6a948e92cc636050fa --- /dev/null +++ b/2024/Multi-Modal Hallucination Control by Visual Information Grounding/1e5cae18-a58b-4415-90f8-bc4858438f19_content_list.json @@ -0,0 +1,1525 @@ +[ + { + "type": "text", + "text": "Multi-Modal Hallucination Control by Visual Information Grounding", + "text_level": 1, + "bbox": [ + 133, + 130, + 836, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alessandro Favero* \nPramuditha Perera", + "bbox": [ + 129, + 181, + 307, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Luca Zancato Alessandro Achille", + "bbox": [ + 315, + 181, + 468, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matthew Trager Sid Ashwin Swaminathan", + "bbox": [ + 480, + 181, + 679, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siddharth Choudhary \nthan Stefano Soatto", + "bbox": [ + 645, + 181, + 833, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "AWS AI Labs", + "text_level": 1, + "bbox": [ + 424, + 223, + 540, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "alessandro.favero@epfl.ch", + "bbox": [ + 372, + 247, + 596, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zancato,mttrager,sidchoud,pramudi,aachille,swashwin,soattos}@amazon.com", + "bbox": [ + 166, + 263, + 803, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 314, + 313, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative Vision-Language Models (VLMs) are prone to generate plausible-sounding textual answers that, however, are not always grounded in the input image. We investigate this phenomenon, usually referred to as \"hallucination\" and show that it stems from an excessive reliance on the language prior. In particular, we show that as more tokens are generated, the reliance on the visual prompt decreases, and this behavior strongly correlates with the emergence of hallucinations. To reduce hallucinations, we introduce Multi-Modal Mutual-Information Decoding (M3ID), a new sampling method for prompt amplification. M3ID amplifies the influence of the reference image over the language prior, hence favoring the generation of tokens with higher mutual information with the visual prompt. M3ID can be applied to any pre-trained autoregressive VLM at inference time without necessitating further training and with minimal computational overhead. If training is an option, we show that M3ID can be paired with Direct Preference Optimization (DPO) to improve the model's reliance on the prompt image without requiring any labels. Our empirical findings show that our algorithms maintain the fluency and linguistic capabilities of pre-trained VLMs while reducing hallucinations by mitigating visually ungrounded answers. Specifically, for the LLaVA 13B model, M3ID and M3ID+DPO reduce the percentage of hallucinated objects in captioning tasks by $25\\%$ and $28\\%$ , respectively, and improve the accuracy on VQA benchmarks such as POPE by $21\\%$ and $24\\%$ .", + "bbox": [ + 76, + 347, + 472, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 786, + 209, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent autoregressive Vision-Language Models (VLMs) have shown remarkable multimodal capabilities [4, 13]. However, VLMs, similarly to large language models (LLMs), are prone to \"hallucinations\" - generating", + "bbox": [ + 75, + 811, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6418e547ac52e90379698ef9281e967ecfdd611c091bd8937d57aa9add76bce7.jpg", + "image_caption": [ + "Prompt: Describe the image." + ], + "image_footnote": [ + "$\\mathbf{y}_{< t}$ : The image depicts a kitchen with an oven, a" + ], + "bbox": [ + 580, + 316, + 799, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c527d128499359f3de06d1ab5b5d55ca3e31b434594184bb42fe87d978530ce1.jpg", + "image_caption": [ + "Figure 1. Multi-Modal Mutual Information Decoding (M3ID). Given a VLM $p$ , an image $c$ , and a text prompt, M3ID intervenes in the generative distribution by finding tokens that \"surprise\" the unconditioned VLM (i.e., the VLM without the image prompt). M3ID amplifies the conditioned directions that are not already predicted by the unconditioned model more as new tokens are generated by leveraging a progressively smaller $\\gamma_{t}$ . In the example, the VLM assigns a high likelihood to the hallucinated object fridge, over-relying on its unconditioned language prior. Instead, M3ID assigns a high likelihood to toaster, which is present in the image." + ], + "image_footnote": [], + "bbox": [ + 509, + 455, + 887, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plausible-sounding answers without factual basis, leading to potentially ungrounded or fabricated information [7]. Consequently, the community has been developing ever so complex and expensive alignment algorithms involving direct human supervision [17] and often brittle prompt engineering methods (\"model begging\") [28].", + "bbox": [ + 496, + 731, + 890, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose to investigate hallucinations in VLMs through a quantifiable measure of visual prompt dependency. We assess whether a model output is ungrounded with respect to a visual prompt by comparing its likelihood with the likelihood of generating the same output", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at AWS AI Labs.", + "bbox": [ + 94, + 886, + 362, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14303", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "without the visual information. In general, our measure of visual prompt dependency quantifies the extent to which a token is generic or specific to a visual signal. Importantly, low visual prompt dependency does not necessarily imply that a token is hallucinated, as it may occur in linguistically essential elements like conjunctions and prepositions. Yet, unlike the notion of hallucinations, which requires human judgment, our prompt dependency measure is always well-defined even without any ground truth annotations.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our first contribution is to empirically demonstrate that the visual prompt dependency measure decreases as more tokens are generated. In Fig. 3, we show that, as more tokens are generated, the conditioning information gets diluted and \"forgotten\" or ignored by the model, possibly leading to more hallucinations. We refer to this effect as conditioning dilution or fading memory effect.", + "bbox": [ + 75, + 228, + 472, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To counteract this, we propose an intervention on the generative distribution of a VLM that maximizes visual prompt dependency at inference time. We show that our method for prompt amplification maximizes the mutual information between the text output tokens and the visual prompt, effectively rescaling the image-conditioned component against the unconditioned distribution. We name our inference-time intervention on the generative distribution Multi-Modal Mutual Information Decoding (M3ID). M3ID is applicable to any off-the-shelf model without additional training or access to model weights, offering a low computational overhead alternative to standard decoding algorithms [13, 24]. Our results show that M3ID enhances the dependence on the visual prompt and reduces the number of hallucinations across various benchmarks while preserving the linguistic fluency of the original model.", + "bbox": [ + 75, + 335, + 472, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, for users with access to model weights, we propose a training objective to further ground the model outputs on visual prompts. This approach involves recasting our prompt amplification objective as a preference optimization problem, using Direct Preference Optimization (DPO) [19] to align the pre-trained VLM. This method aims to increase the likelihood of continuations with a higher visual prompt dependency measure and allows to learn a better generation policy, further reducing the conditioning dilution effect and hallucinations.", + "bbox": [ + 75, + 580, + 472, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 96, + 732, + 434, + 747 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a visual prompt dependency measure (PDM) to assess whether a model output is ungrounded with respect to the visual input. Empirically, we demonstrate that PDM decreases as more tokens are generated making the generations more likely to be hallucinated.", + "2. We introduce M3ID, a training-free intervention on the generative distribution of autoregressive VLMs which improves visual grounding and reduces hallucinations by amplifying the importance of the visual prompt over the language prior. In addition, we extend M3ID with" + ], + "bbox": [ + 76, + 750, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DPO for multi-modal preference optimization to further improve visual grounding.", + "bbox": [ + 517, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. We show that applying M3ID or DPO reduces the percentage of hallucinated objects in captioning tasks [20] by $25\\%$ and by $28\\%$ , respectively and improves accuracy on the POPE [9] VQA hallucination benchmark by $21\\%$ and $24\\%$ over the base model.", + "bbox": [ + 500, + 121, + 893, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 210, + 637, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hallucinations in VLMs. A recent line of work introduced VLMs obtained by grafting visual encoders into pre-trained Large Language Models (LLMs) [4, 13, 32]. Forcing a pre-trained LLM to learn to \"see\" by reading vision tokens from a pre-trained vision backbone has proven successful and the resulting models show remarkable vision-language understanding capabilities on many multi-modal tasks [13, 23]. However, while inheriting strong linguistic capabilities and fluency from their base LLM, grafted VLMs also inherit the tendency to produce ungrounded or fabricated information [1, 6]. This is commonly referred to as \"hallucination\" in the recent machine learning literature [1, 6, 25]. In particular, several recent works empirically reported a sharp tendency of grafted VLMs to report objects not grounded on the visual information when probed with questions on the image content [1, 6, 9, 27]. The authors of [9] observed that this phenomenon is especially pronounced with objects that are either common or frequently appear together in the datasets used at training time. Furthermore, the authors of [11] suggest that VLMs often fail to correctly follow instructions involving absent objects and propose a new instruction-following training objective to enhance model alignment and robustness in the face of uncertainty. Parallel to our work, [31] delves into the factors contributing to object hallucinations in VLMs, exploring aspects like object co-occurrence, model uncertainty, and spatial position of hallucinations in the sentence. They propose a post-hoc algorithm to identify and correct hallucinations in VLM-generated content.", + "bbox": [ + 496, + 243, + 893, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Context-dependent decodings. Decoding algorithms can be classified as either search or sampling algorithms [8]. Current search methods (e.g., greedy and beam search) can produce factually accurate generations but usually suffer from tedious and repetitive continuations. Sampling methods, on the other hand, (like nucleus [5], or typical decoding [15]) produce more diverse text, but suffer from topic drift. To solve these shortcomings, various decoding methods have been proposed to enhance reasoning and factual accuracy in generative autoregressive language models. Approaches such as context-aware decoding [21] and Pointwise Mutual Information (PMI) decoding [16, 25] focus on aligning the outputs of language models with the intended context or factual data. These strategies aim to reduce hal", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14304", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2e23e32c7b76705df7ff11bf12fb56e3da31459f752a9f5ffbf8c1db06c867ef.jpg", + "image_caption": [ + "Figure 2. How much does the conditioning prompt \"surprise\" the unconditioned predictor? We report the likelihood that the conditioned and unconditioned models assign to each token in the string \"The image ...\" The probability gap increases on tokens for which visual information is required (e.g. objects and attributes), while it decreases for punctuation and articles. Note that both models mostly agree when sufficient visual information has already been extracted and is present in the caption (e.g. \"kitchen\" is very likely for the unconditioned model as well). See Sec. D.1 for a plot with our prompt dependency measure." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 464, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lucinations in standard natural language processing tasks such as summarization. Mutual-information-based decoding techniques have proven to be helpful in steering the generation of LLMs to stay faithful to the given input text [25], or to promote diversity or relevance in neural dialogue models [7]. Our work is the first to use mutual information to improve multi-modal grounding and to vary the penalization of marginally likely outputs in a time-varying fashion to counteract the progressive forgetting of the visual prompt.", + "bbox": [ + 75, + 422, + 472, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Analysis of hallucinations in VLMs", + "text_level": 1, + "bbox": [ + 76, + 571, + 397, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start by investigating hallucinations in VLMs. We introduce a visual prompt dependency measure (PDM), which will later motivate our algorithms (Sec. 4). The main aim of this section is to assess whether a model output, such as a token, is hallucinated or not by comparing its likelihood with the likelihood of that same output being generated without the relevant conditioning factor, in this case, the image.", + "bbox": [ + 75, + 595, + 468, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider decoding approaches for open-ended language generation of VLMs, where the model receives an input image and an input textual prompt and aims to generate a fluent and coherent continuation conditioned on both the image and the text. We denote probabilistic generative VLMs over a vocabulary of text tokens $\\mathcal{V}$ with $p$ and denote the probability of a token $y\\in \\mathcal{V}$ given a textual prompt $\\mathbf{x}$ and an image context $c$ as $p(y|\\mathbf{x},c)$ . The goal of the VLM is to leverage the information contained in the input image $c$ to provide a continuation $\\mathbf{y} = [y_0,\\dots ,y_T]$ of the input prompt $\\mathbf{x}$ , e.g., \"Describe this image in detail\". The probability of a valid sequence $\\mathbf{y}$ can be computed as $p(\\mathbf{y}|\\mathbf{x},c) = \\prod_{t = 1}^{T}p(y_{t}|\\mathbf{y}_{< t},\\mathbf{x},c)$ , where $\\mathbf{y}_{< t}\\triangleq [y_0,\\dots ,y_{t - 1}]$ .", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a4798a083b13589972e8daec3c766f035778c91c998e9c02277996eb41b63b83.jpg", + "image_caption": [ + "Figure 3. The influence of the image conditioning decreases as we generate more tokens. (Left) Conditioning dilution. We report the average of the prompt dependency measure (PDM-H) and its standard deviation for different synthetic captions generated by LLaVA on MS COCO's validation split. We see that the influence of the image over the next token prediction decreases as we generate more. This suggests that the information within VLM's visual prompt gets diluted and fades away as more tokens are generated. (Right) Frequency of hallucinated objects as a function of the token position. We report the number of non-existent objects present on the same synthetic captions as a function of the number of generated tokens. Note that very few objects are hallucinated for tokens near the visual prompt, while their number increases as more tokens are generated and with a smaller PDM." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 694, + 198 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bddc9be237dcad3e5725c9e6b78dde8a00d2157155c044e6c59eb9a564d34428.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 90, + 885, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose to study hallucinations on VLMs using PDMs defined as follows", + "bbox": [ + 498, + 422, + 892, + 452 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P D M} (\\mathbf {y} _ {< t}; c | \\mathbf {x}) \\triangleq \\operatorname {d i s t} \\left(p (\\cdot | \\mathbf {y} _ {< t}, \\mathbf {x}, c), p (\\cdot | \\mathbf {y} _ {< t}, \\mathbf {x})\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 459, + 867, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where dist is any distance measure between probability distributions, such as Hellinger, total variation, or KL. PDMs quantify how generic or context-specific a language model's output is. In particular, a high $\\mathrm{PDM}(\\mathbf{y}_{< t};c|\\mathbf{x})$ indicates that the token $y_{t}$ is strongly associated with a specific input prompt, while a low PDM suggests that the token is more prompt-neutral or prompt-agnostic. Depending on the choice of the distance function, PDMs highlight different aspects of the generative distribution. We will mainly use PDM-H based on the Hellinger distance defined as $H(p,q) = 2^{-1 / 2}\\sqrt{\\sum_{i = 1}^{k}(\\sqrt{p_i} - \\sqrt{q_i})^2}$ , where $p = (p_i)_{i\\in [k]}$ and $q = (q_i)_{i\\in [k]}$ are discrete probability distributions. We refer the reader to the Supplementary Material for an analysis of the impact of other distances.", + "bbox": [ + 496, + 493, + 892, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contextual pressure. Fig. 2 shows the conditional and unconditional prediction likelihoods on a single caption. First, we note that tokens that are required for linguistic fluency, like prepositions and conjunctions, are highly predictable by both the conditioned and the unconditioned models. This pattern also emerges when the model generates tokens to compose \"fine-grained\" objects. For example, when the model generates \"Peanut butter\", both the VLM and the LLM can correctly predict the last token even without looking at the image and only having access to a sufficiently long truncation (e.g. \"Peanut bu\"). Therefore, in these circumstances, $p(y_{t}|\\mathbf{y}_{CaptioningCHAIRi ↓CHAIRs ↓Cover ↑LLaVA7B8.117.553.3LLaVA7B PMI [25]6.716.251.5LLaVA7B Contrastive [8]6.314.855.2LLaVA7B M3ID5.913.855.1LLaVA13B7.418.555.2LLaVA13B PMI [25]6.516.654.7LLaVA13B Contrastive [8]6.514.554.7LLaVA13B M3ID5.513.254.0LLaVA13B + LURE [31]6.427.1-mPLUG-Owl + LURE [31]5.418.8-LLaVA7B M3ID + DPO5.713.555.8LLaVA13B M3ID + DPO5.312.654.2", + "bbox": [ + 81, + 196, + 467, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "eration by increasing the likelihood of generating tokens that are related to the text to be summarized. Differently from our method, the weight assigned to the realization of marginally likely continuations is time-independent. Similarly, Contrastive Decoding [7, 8] has been developed to foster generations of an expert LLM to contain plausible and fluent text with the textual input prompt by amplifying its prediction difference with respect to a weaker model. To ensure a fair comparison, we replace the text to be summarized with the input image for PMI and set the weaker model in Contrastive Decoding as the unconditioned model (see Sec. B for more details). We also compare with training-based methods that have been proposed to improve the grounding of VLMs on the visual prompt like LLaVARLHF [17], Robust mPLUG-Owl [11] and LURE [31]. However, all these methods require some form of annotations, like preference data [17] or positive/negative examples [11, 31], which M3ID+DPO does not require.", + "bbox": [ + 75, + 417, + 468, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. We evaluate our algorithm both on captioning and VQA benchmarks. On captioning, we measure object hallucinations by using the CHAIR (Captioning Hallucination Assessment with Image Relevance) metrics [20]. These metrics assess the quality of captions by comparing the mentioned objects to the annotated objects present in an image. In particular, CHAIR comprises two variants: one computes the fraction of hallucinated objects in the whole caption (CHAIRi) and the other evaluates the fraction of captions with at least one object hallucination (CHAIRs),", + "bbox": [ + 75, + 700, + 468, + 853 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C H A I R i} = \\frac {\\# \\text {h a l l u c i n a t e d o b j e c t s}}{\\# \\text {g e n e r a t e d o b j e c t s}},\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 871, + 392, + 905 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {C H A I R s} = \\frac {\\# \\text {h a l l u c i n a t e d c a p t i o n s}}{\\# \\text {g e n e r a t e d c a p t i o n s}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 87, + 818, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we introduce the Cover metric to quantify the comprehensiveness of the captions by measuring the fraction of ground-truth annotated objects mentioned by the model,", + "bbox": [ + 498, + 125, + 890, + 184 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {C o v e r} = \\frac {\\left| \\{\\text {c o r r e c t m e n t i o n e d o b j e c t s} \\} \\right|}{\\left| \\{\\text {a n n o t a t e d o b j e c t s} \\} \\right|}.\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 194, + 830, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate object perception, we use the POPE (Polling-based Object Probing Evaluation) VQA benchmark [9]. In particular, POPE recasts the evaluation of object hallucinations into a binary classification problem with yes/no questions of the type \"Is a present in the image?\"", + "bbox": [ + 498, + 237, + 890, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DPO training details. We use Hugging Face's DPO implementation in the TRL library [26] and train LLaVA for 5 epochs on 10,000 self-generated preference pairs with LoRA and cosine decaying learning rate with $2 \\times 10^{-5}$ peak. Following [19], we set $\\beta = 0.1$ in the DPO loss.", + "bbox": [ + 498, + 319, + 890, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. VLM grounding on captioning", + "text_level": 1, + "bbox": [ + 500, + 404, + 769, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1 we test how much our method can prevent the generation of hallucinated objects in image captions. In particular, we compare against training-free and training-based baselines. We prompt all baseline methods with the instruction \"Describe the image.\" and let the models generate until the EOS token is obtained.", + "bbox": [ + 496, + 426, + 890, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we compare M3ID with other decoding strategies, PMI [25] and Contrastive Decoding [8]. The main difference between M3ID and these baselines is that M3ID increasingly counteracts the language prior as more tokens are generated. As such, M3ID reduces ungrounded generations compared to all other training-free baselines both on the large LLaVA13B and on the smaller LLaVA7B. Furthermore, it has uniformly good improvements for different model sizes over standard multinomial decoding, in particular, M3ID achieves $27\\% / 21\\%$ relative improvement over LLaVA7B and $26\\% / 29\\%$ over LLaVA13B on the CHAIRi and CHAIRs metrics. Importantly, this improvement does not come at the price of high reductions of the Cover metric, which actually improves on the 7B model and decreases by less than $2.2\\%$ for the larger 13B model1. Lastly, we note that M3ID shows an improvement in absolute performance that correlates with model size, suggesting that further gains could be obtained as larger models are used. In Sec. D.2, we also show that M3ID still improves grounding when longer captions (twice as long on average) are generated and where other decodings do not perform as well.", + "bbox": [ + 496, + 518, + 892, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1 we also report training-based results and show that pairing DPO with M3ID leads to a smaller number", + "bbox": [ + 498, + 835, + 890, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1Note that CHAIR metrics can be hacked by simply returning shorter captions that do not attempt to predict any object in the image.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14308", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ed56b1b58a63c0e0640907694d9a3c01d8ac2a5e88fd0393c799bce90642287a.jpg", + "table_caption": [ + "Table 2. Evaluation on the POPE VQA hallucination benchmark [9]. *;† indicate results taken from [11] and [23] respectively. SUP. Indicates that the method involves supervised training on annotated data, in contrast to our approaches. POPE is divided into three subsets, Random, Popular, and Adversarial. We report the VQA accuracy of the model using the following template: “Is a present in the image?”, where we sample randomly (Random), among the most frequent objects in the dataset (Popular), or among the objects that frequently co-occur with (Adversarial). Acc. is the binary classification accuracy and Yes is the percentage of “Yes” answers." + ], + "table_footnote": [], + "table_body": "
POPE
RandomPopularAdversarialAll
Acc. ↑Yes (%)Acc. ↑Yes (%)Acc. ↑Yes (%)Acc. ↑Yes (%)
Robust mPLUG-Owl-7B* [11]sup.LLaVA-RLHF†7B [23]sup.86.0-73.0-65.0-74.7-
84.839.683.341.880.744.082.941.8
mPLUG-Owl-7B* [29]52.0-57.0-60.0-67.3-
LLaVA7B74.875.161.886.758.190.164.984.0
LLaVA7B M3ID76.067.769.373.365.877.670.372.9
LLaVA7B M3ID + DPO81.265.673.967.368.275.474.469.4
LLaVA-RLHF†13B [23]sup.85.238.483.938.082.340.583.839.0
MiniGPT4-13B* [32]73.0-67.0-62.0-74.7-
LLaVA13B67.980.663.883.259.887.363.883.7
LLaVA13B M3ID84.355.677.061.671.368.277.561.8
LLaVA13B M3ID + DPO85.253.479.157.573.267.579.251.1
", + "bbox": [ + 80, + 170, + 883, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of hallucinated objects and improved Cover numbers compared both to our training-free approach and LLaVA+LURE [31] a concurrent training based method to improve the grounding of VLMs that relies on GPT-3.5 annotations.", + "bbox": [ + 75, + 438, + 470, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. VLM grounding on VQA", + "text_level": 1, + "bbox": [ + 76, + 507, + 308, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Differently from the MS COCO captioning task, the POPE VQA benchmark only requires the generation of \"Yes/No\" tokens. As such, one does not expect the dilution effect to play a key role. However, we highlight that the image tokens and the output of the VLM are separated by the input template used to prime the model for the VQA task, which, as we show in Sec. D.4, introduces a non-negligible dilution effect. To address this, we take an offset into account when using M3ID on POPE, and select $t = t_0$ where $t_0$ is the number of tokens in between the output and the image.", + "bbox": [ + 75, + 530, + 468, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 2, we report the results on the POPE VQA hallucination benchmark [9]. Hallucinations (wrong answers) tend to correlate with the tendency of the VLM to reply using the \"Yes\" token (see the disproportionately high percentage rate, $84\\% / 83.7\\%$ , of \"Yes\" answers for the LLaVAv1.3 base model). This tendency can be counteracted at inference time and without any training by simply using M3ID. In our experiments, M3ID reduces the Yes ratio to $72.9\\% / 61.8\\%$ for 7B and 13B models, respectively, which leads to relative accuracy improvements over standard LLaVA decoding by $8\\%$ and $21\\%$ , respectively.", + "bbox": [ + 75, + 681, + 468, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To test whether our training-based approach increases the VLM's grounding on the visual prompt we trained a", + "bbox": [ + 75, + 847, + 468, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "model with our DPO objective on the MS COCO captioning dataset and then tested it on the POPE benchmark. Note that improving caption generation does not directly imply improvements on the POPE VQA benchmark, in fact, even if the images are both from MS COCO, the output format is quite different: open-ended generation on the former and binary classification on the latter. However, intuitively, one expects that by improving the reliance on the visual prompt, the model could make better use of the visual information regardless of the output format required to solve the task at hand. Tab. 2 shows that M3ID+DPO further improves performance over M3ID's inference time intervention. Specifically, M3ID+DPO achieves $15\\%$ and $24\\%$ accuracy improvements over the LLaVA 7B and 13B models respectively. For completeness, we also compare with other training-based baselines that are fine-tuned on labeled [11] and preference data [17]. While M3ID+DPO does not have access to any labeled information we show that it is close to these baselines without requiring any annotations.", + "bbox": [ + 500, + 438, + 893, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablations", + "text_level": 1, + "bbox": [ + 500, + 734, + 609, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conditioning dilution and \"overcompensation\". In Fig. 4 and Tab. 3 we show that when the forgetting factor $\\lambda$ is high, corresponding to the assumption that the input image gets forgotten quickly, M3ID effectively maintains a large PDM-H throughout the generation. However, while deviating from the unconditioned probability is often a desired behavior, an excessive deviation from it across the whole generation could result in \"overcompensation\", which, as we show in Tab. 3 can be counterproductive and can lead to", + "bbox": [ + 498, + 763, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2We use [Img] [Model Prompt] [Question], see Sec. D.4.", + "bbox": [ + 94, + 886, + 447, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14309", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e1f2967b7693e9ee7f78ba4cd1ce0834178b64515ac247f7c9b52bfe9d53f7a.jpg", + "image_caption": [ + "Figure 4. Sensitivity analysis on hyper-parameters. We vary both the forgetting factor $\\lambda$ and the thresholding parameter $\\alpha$ . Small $\\lambda$ and $\\alpha$ result in minor effects. Instead, either a high forgetting factor $\\lambda$ or a high $\\alpha$ results in a stronger intervention which leads to higher PDM-H." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 269, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da35ae2648ffb20772286d7e259bfff710fadbe0823dcf0c22df29a15592a3a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 90, + 464, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "higher hallucination rates. In particular, observe that higher corrections detrimentally impact the cover metric. This suggests that while M3ID tries to find tokens that diverge from the language prior the most, it is more likely to fail in captioning elements that are inherently predictable by the language prior alone. For instance, in describing an image of a dog on a leash accompanied by a man, M3ID might overlook mentioning the presence of the man, a token that the language prior could anticipate without necessitating any visual information only from context clues. We report examples of this behavior in the Sec. D.2.", + "bbox": [ + 75, + 292, + 468, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importance of the confidence threshold. When we set a high threshold $\\alpha$ , the indicator function remains active more frequently along the generation, leading M3ID to consistently prioritize maximizing the distance of the generation from the language prior. Hence, similarly to the scenario with a high $\\lambda$ , M3ID is likely to overcompensate and increase the rate of multi-modal hallucinations. However, too high $\\alpha$ disrupts linguistic fluency. We report qualitative examples of this in Sec. D.2. In Tab. 4 we also show that adding the term to compensate for contextual pressure does not significantly change the results over LLaVA decoding in the POPE benchmark since it is essentially framed as a binary classification over the \"Yes\"/ \"No\" tokens and not open-ended generation like captioning.", + "bbox": [ + 75, + 465, + 468, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 693, + 204, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced M3ID, a new approach designed to combat multi-modal hallucinations by maximizing the mutual information between the text generated by VLMs and the corresponding visual context. M3ID operates at inference time and can be seamlessly integrated with any pre-trained autoregressive VLM. This makes M3ID a cost-effective and flexible solution to enhance vision-language grounding. Furthermore, for settings where model training is feasible and higher visual grounding is expected, we also paired M3ID with Direct Preference Optimization (DPO), showcasing reduced hallucinatory behaviors. Interestingly, our findings suggest that object hallucinations in VLMs result", + "bbox": [ + 75, + 719, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8125663bf5adccdee9e5c4cad2eefbe263dcf4c6e3b3cce3901b0a5b0e09a369.jpg", + "table_caption": [ + "Table 3. Sensitivity to hyper-parameters. $\\lambda$ is the forgetting factor used to counteract conditioning dilution while $\\alpha$ is the threshold used to mitigate overcompensation when the contextual pressure is high. We report results on LLaVA-7B." + ], + "table_footnote": [], + "table_body": "
Captioning
CHAIRi ↓CHAIRs ↓Cover ↑
LLaVA7B8.017.853.5
M3ID α = 0.3, λ = 0.025.813.655.0
M3ID α = 0.3, λ = 0.0017.317.453.7
M3ID α = 0.3, λ = 0.036.214.752.0
M3ID α = 0.3, λ = 0.17.216.345.5
M3ID α = 0.5, λ = 0.026.114.954.6
M3ID α = 0.1, λ = 0.026.915.453.1
M3ID α = 0.01, λ = 0.026.916.652.5
", + "bbox": [ + 504, + 152, + 888, + 309 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/74a4789cfbdb90bf6d8b67a23ed3d583bcc033a120a64b503ee4222a3feae6da.jpg", + "table_caption": [ + "Table 4. Ablation studies. We ablate the components of M3ID on both model sizes. Removing the amplification term to counteract conditioning dilution leads to more hallucinations." + ], + "table_footnote": [], + "table_body": "
POPE\nAccuracy ↑CHAIRi ↓Captioning\nCHAIRs ↓Cover ↑
LLaVA7B64.98.017.853.5
w/ context pressure65.56.916.754.5
w/ conditioning dilution77.56.414.153.9
LLaVA7B M3ID77.55.813.655.0
LLaVA13B63.815.318.255.3
w/ context pressure64.96.514.455.0
w/ conditioning dilution70.35.914.853.7
LLaVA13B M3ID70.35.413.054.0
", + "bbox": [ + 503, + 369, + 888, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from excessive reliance on the language prior rather than a poor understanding of the visual modality. In fact, M3ID at inference time is sufficient to significantly reduce the amount of generated hallucinations without any training.", + "bbox": [ + 496, + 507, + 890, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A limitation of M3ID is that it requires two forward passes at inference time, one for the conditioned and one for the unconditioned prediction. A possible solution to not increase inference time, but at the expense of higher memory consumption, is to use two batched queries with one having masked visual tokens. Furthermore, as we observed in Tab. 1, sometimes M3ID may prevent the generation of objects that are highly likely under the unprompted language prior (due to context clues). Interestingly, a similar observation has been reported in [22], where, people asked to provide a 10-word list of objects contained in a given image often failed to report the most obvious objects while mainly focusing on secondary ones. While we already showed that this can be mitigated with proper hyperparameter selection Tab. 3 we believe that an interesting avenue for future research is to directly encode this preference within the model's weights by favoring the generation of structured captions that get progressively more detailed while still being grounded to the image.", + "bbox": [ + 496, + 568, + 892, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In general, integrating human- or AI-annotated preference pairs, assessed based on their level of grounding, constitutes a promising avenue for future investigation.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14310", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, et al. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023, 2023. 2", + "[2] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 12", + "[3] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 5", + "[4] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 5, 16", + "[5] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. In International Conference on Learning Representations, 2020. 2", + "[6] Yichong Huang, Xiachong Feng, Xiaocheng Feng, and Bing Qin. The factual inconsistency problem in abstractive text summarization: A survey. arXiv preprint arXiv:2104.14839, 2021. 2", + "[7] Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110–119, San Diego, California, 2016. Association for Computational Linguistics. 1, 3, 4, 6, 11", + "[8] Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. Contrastive decoding: Open-ended text generation as optimization, 2023. 2, 5, 6, 11, 14", + "[9] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 2, 6, 7, 11, 17", + "[10] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 4, 5, 6, 11, 14, 15", + "[11] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 2, 6, 7", + "[12] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 5" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 2, 4, 5, 11", + "[14] M.B. Matthews and G.S. Moschytz. The identification of nonlinear discrete-time fading-memory systems using neural network models. IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing, 41(11):740-751, 1994. 4", + "[15] Clara Meister, Tiago Pimentel, Gian Wiher, and Ryan Coterell. Locally typical sampling. Transactions of the Association for Computational Linguistics, 11:102-121, 2023. 2", + "[16] Yatin Nandwani, Vineet Kumar, Dinesh Raghu, Sachindra Joshi, and Luis A Lastras. Pointwise mutual information based metric and decoding strategy for faithful generation in document grounded dialogs. arXiv preprint arXiv:2305.12191, 2023. 2", + "[17] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 1, 6, 7", + "[18] 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. 5", + "[19] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. arXiv preprint arXiv:2305.18290, 2023. 2, 5, 6, 11, 12", + "[20] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4035-4045, Brussels, Belgium, 2018. Association for Computational Linguistics. 2, 6, 11, 14, 17", + "[21] Weijia Shi, Xiaochuang Han, Mike Lewis, Yulia Tsvetkov, Luke Zettlemoyer, and Scott Wen-tau Yih. Trusting your evidence: Hallucinate less with context-aware decoding. arXiv preprint arXiv:2305.14739, 2023. 2", + "[22] Merrielle Spain and Pietro Perona. Measuring and predicting object importance. Int. J. Comput. Vis., 91(1):59-76, 2011. 8", + "[23] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 2, 7", + "[24] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 5", + "[25] Liam van der Poel, Ryan Cotterell, and Clara Meister. Mutual information alleviates hallucinations in abstractive sum" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14311", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "marization. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5956-5965, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 2, 3, 4, 5, 6, 11, 14", + "[26] Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, and Shengyi Huang. Trl: Transformer reinforcement learning. https://github.com/huggingface/trl, 2020.6", + "[27] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, et al. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 2", + "[28] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 1", + "[29] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 7", + "[30] Luca Zancato and Alessandro Chiuso. A novel deep neural network architecture for non-linear system identification. IFAC-PapersOnLine, 54(7):186-191, 2021. 19th IFAC Symposium on System Identification SYSID 2021. 4", + "[31] Yiyang Zhou, Chenhang Cui, Jaehong Yoon, Linjun Zhang, Zhun Deng, Chelsea Finn, Mohit Bansal, and Huaxiu Yao. Analyzing and mitigating object hallucination in large vision-language models. arXiv preprint arXiv:2310.00754, 2023. 2, 6, 7", + "[32] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 7" + ], + "bbox": [ + 78, + 90, + 470, + 599 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14312", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Multi-Modal Hallucination Control by Visual Information Grounding/1e5cae18-a58b-4415-90f8-bc4858438f19_model.json b/2024/Multi-Modal Hallucination Control by Visual Information Grounding/1e5cae18-a58b-4415-90f8-bc4858438f19_model.json new file mode 100644 index 0000000000000000000000000000000000000000..82364e9e2ee4cf4759f1149d9a96f247535a6762 --- /dev/null +++ b/2024/Multi-Modal Hallucination Control by Visual Information Grounding/1e5cae18-a58b-4415-90f8-bc4858438f19_model.json @@ -0,0 +1,1925 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.131, + 0.838, + 0.154 + ], + "angle": 0, + "content": "Multi-Modal Hallucination Control by Visual Information Grounding" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.182, + 0.308, + 0.216 + ], + "angle": 0, + "content": "Alessandro Favero* \nPramuditha Perera" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.182, + 0.47, + 0.215 + ], + "angle": 0, + "content": "Luca Zancato Alessandro Achille" + }, + { + "type": "text", + "bbox": [ + 0.482, + 0.182, + 0.68, + 0.216 + ], + "angle": 0, + "content": "Matthew Trager Sid Ashwin Swaminathan" + }, + { + "type": "text", + "bbox": [ + 0.647, + 0.182, + 0.834, + 0.216 + ], + "angle": 0, + "content": "Siddharth Choudhary \nthan Stefano Soatto" + }, + { + "type": "title", + "bbox": [ + 0.425, + 0.224, + 0.541, + 0.24 + ], + "angle": 0, + "content": "AWS AI Labs" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.248, + 0.597, + 0.262 + ], + "angle": 0, + "content": "alessandro.favero@epfl.ch" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.265, + 0.804, + 0.28 + ], + "angle": 0, + "content": "{zancato,mttrager,sidchoud,pramudi,aachille,swashwin,soattos}@amazon.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.315, + 0.314, + 0.33 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.348, + 0.473, + 0.756 + ], + "angle": 0, + "content": "Generative Vision-Language Models (VLMs) are prone to generate plausible-sounding textual answers that, however, are not always grounded in the input image. We investigate this phenomenon, usually referred to as \"hallucination\" and show that it stems from an excessive reliance on the language prior. In particular, we show that as more tokens are generated, the reliance on the visual prompt decreases, and this behavior strongly correlates with the emergence of hallucinations. To reduce hallucinations, we introduce Multi-Modal Mutual-Information Decoding (M3ID), a new sampling method for prompt amplification. M3ID amplifies the influence of the reference image over the language prior, hence favoring the generation of tokens with higher mutual information with the visual prompt. M3ID can be applied to any pre-trained autoregressive VLM at inference time without necessitating further training and with minimal computational overhead. If training is an option, we show that M3ID can be paired with Direct Preference Optimization (DPO) to improve the model's reliance on the prompt image without requiring any labels. Our empirical findings show that our algorithms maintain the fluency and linguistic capabilities of pre-trained VLMs while reducing hallucinations by mitigating visually ungrounded answers. Specifically, for the LLaVA 13B model, M3ID and M3ID+DPO reduce the percentage of hallucinated objects in captioning tasks by \\(25\\%\\) and \\(28\\%\\), respectively, and improve the accuracy on VQA benchmarks such as POPE by \\(21\\%\\) and \\(24\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.21, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.813, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Recent autoregressive Vision-Language Models (VLMs) have shown remarkable multimodal capabilities [4, 13]. However, VLMs, similarly to large language models (LLMs), are prone to \"hallucinations\" - generating" + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.318, + 0.8, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.423, + 0.776, + 0.434 + ], + "angle": 0, + "content": "Prompt: Describe the image." + }, + { + "type": "image_footnote", + "bbox": [ + 0.561, + 0.437, + 0.832, + 0.447 + ], + "angle": 0, + "content": "\\(\\mathbf{y}_{< t}\\): The image depicts a kitchen with an oven, a" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.456, + 0.888, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Figure 1. Multi-Modal Mutual Information Decoding (M3ID). Given a VLM \\( p \\), an image \\( c \\), and a text prompt, M3ID intervenes in the generative distribution by finding tokens that \"surprise\" the unconditioned VLM (i.e., the VLM without the image prompt). M3ID amplifies the conditioned directions that are not already predicted by the unconditioned model more as new tokens are generated by leveraging a progressively smaller \\( \\gamma_{t} \\). In the example, the VLM assigns a high likelihood to the hallucinated object fridge, over-relying on its unconditioned language prior. Instead, M3ID assigns a high likelihood to toaster, which is present in the image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.824 + ], + "angle": 0, + "content": "plausible-sounding answers without factual basis, leading to potentially ungrounded or fabricated information [7]. Consequently, the community has been developing ever so complex and expensive alignment algorithms involving direct human supervision [17] and often brittle prompt engineering methods (\"model begging\") [28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose to investigate hallucinations in VLMs through a quantifiable measure of visual prompt dependency. We assess whether a model output is ungrounded with respect to a visual prompt by comparing its likelihood with the likelihood of generating the same output" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.364, + 0.9 + ], + "angle": 0, + "content": "*Work done during an internship at AWS AI Labs." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14303" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "without the visual information. In general, our measure of visual prompt dependency quantifies the extent to which a token is generic or specific to a visual signal. Importantly, low visual prompt dependency does not necessarily imply that a token is hallucinated, as it may occur in linguistically essential elements like conjunctions and prepositions. Yet, unlike the notion of hallucinations, which requires human judgment, our prompt dependency measure is always well-defined even without any ground truth annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.473, + 0.336 + ], + "angle": 0, + "content": "Our first contribution is to empirically demonstrate that the visual prompt dependency measure decreases as more tokens are generated. In Fig. 3, we show that, as more tokens are generated, the conditioning information gets diluted and \"forgotten\" or ignored by the model, possibly leading to more hallucinations. We refer to this effect as conditioning dilution or fading memory effect." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.473, + 0.579 + ], + "angle": 0, + "content": "To counteract this, we propose an intervention on the generative distribution of a VLM that maximizes visual prompt dependency at inference time. We show that our method for prompt amplification maximizes the mutual information between the text output tokens and the visual prompt, effectively rescaling the image-conditioned component against the unconditioned distribution. We name our inference-time intervention on the generative distribution Multi-Modal Mutual Information Decoding (M3ID). M3ID is applicable to any off-the-shelf model without additional training or access to model weights, offering a low computational overhead alternative to standard decoding algorithms [13, 24]. Our results show that M3ID enhances the dependence on the visual prompt and reduces the number of hallucinations across various benchmarks while preserving the linguistic fluency of the original model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.473, + 0.732 + ], + "angle": 0, + "content": "Additionally, for users with access to model weights, we propose a training objective to further ground the model outputs on visual prompts. This approach involves recasting our prompt amplification objective as a preference optimization problem, using Direct Preference Optimization (DPO) [19] to align the pre-trained VLM. This method aims to increase the likelihood of continuations with a higher visual prompt dependency measure and allows to learn a better generation policy, further reducing the conditioning dilution effect and hallucinations." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.733, + 0.435, + 0.748 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.75, + 0.47, + 0.825 + ], + "angle": 0, + "content": "1. We propose a visual prompt dependency measure (PDM) to assess whether a model output is ungrounded with respect to the visual input. Empirically, we demonstrate that PDM decreases as more tokens are generated making the generations more likely to be hallucinated." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "2. We introduce M3ID, a training-free intervention on the generative distribution of autoregressive VLMs which improves visual grounding and reduces hallucinations by amplifying the importance of the visual prompt over the language prior. In addition, we extend M3ID with" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "DPO for multi-modal preference optimization to further improve visual grounding." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.894, + 0.198 + ], + "angle": 0, + "content": "3. We show that applying M3ID or DPO reduces the percentage of hallucinated objects in captioning tasks [20] by \\(25\\%\\) and by \\(28\\%\\), respectively and improves accuracy on the POPE [9] VQA hallucination benchmark by \\(21\\%\\) and \\(24\\%\\) over the base model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.212, + 0.638, + 0.228 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.895, + 0.683 + ], + "angle": 0, + "content": "Hallucinations in VLMs. A recent line of work introduced VLMs obtained by grafting visual encoders into pre-trained Large Language Models (LLMs) [4, 13, 32]. Forcing a pre-trained LLM to learn to \"see\" by reading vision tokens from a pre-trained vision backbone has proven successful and the resulting models show remarkable vision-language understanding capabilities on many multi-modal tasks [13, 23]. However, while inheriting strong linguistic capabilities and fluency from their base LLM, grafted VLMs also inherit the tendency to produce ungrounded or fabricated information [1, 6]. This is commonly referred to as \"hallucination\" in the recent machine learning literature [1, 6, 25]. In particular, several recent works empirically reported a sharp tendency of grafted VLMs to report objects not grounded on the visual information when probed with questions on the image content [1, 6, 9, 27]. The authors of [9] observed that this phenomenon is especially pronounced with objects that are either common or frequently appear together in the datasets used at training time. Furthermore, the authors of [11] suggest that VLMs often fail to correctly follow instructions involving absent objects and propose a new instruction-following training objective to enhance model alignment and robustness in the face of uncertainty. Parallel to our work, [31] delves into the factors contributing to object hallucinations in VLMs, exploring aspects like object co-occurrence, model uncertainty, and spatial position of hallucinations in the sentence. They propose a post-hoc algorithm to identify and correct hallucinations in VLM-generated content." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Context-dependent decodings. Decoding algorithms can be classified as either search or sampling algorithms [8]. Current search methods (e.g., greedy and beam search) can produce factually accurate generations but usually suffer from tedious and repetitive continuations. Sampling methods, on the other hand, (like nucleus [5], or typical decoding [15]) produce more diverse text, but suffer from topic drift. To solve these shortcomings, various decoding methods have been proposed to enhance reasoning and factual accuracy in generative autoregressive language models. Approaches such as context-aware decoding [21] and Pointwise Mutual Information (PMI) decoding [16, 25] focus on aligning the outputs of language models with the intended context or factual data. These strategies aim to reduce hal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14304" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.465, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.473, + 0.41 + ], + "angle": 0, + "content": "Figure 2. How much does the conditioning prompt \"surprise\" the unconditioned predictor? We report the likelihood that the conditioned and unconditioned models assign to each token in the string \"The image ...\" The probability gap increases on tokens for which visual information is required (e.g. objects and attributes), while it decreases for punctuation and articles. Note that both models mostly agree when sufficient visual information has already been extracted and is present in the caption (e.g. \"kitchen\" is very likely for the unconditioned model as well). See Sec. D.1 for a plot with our prompt dependency measure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.473, + 0.561 + ], + "angle": 0, + "content": "lucinations in standard natural language processing tasks such as summarization. Mutual-information-based decoding techniques have proven to be helpful in steering the generation of LLMs to stay faithful to the given input text [25], or to promote diversity or relevance in neural dialogue models [7]. Our work is the first to use mutual information to improve multi-modal grounding and to vary the penalization of marginally likely outputs in a time-varying fashion to counteract the progressive forgetting of the visual prompt." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.398, + 0.589 + ], + "angle": 0, + "content": "3. Analysis of hallucinations in VLMs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.703 + ], + "angle": 0, + "content": "We start by investigating hallucinations in VLMs. We introduce a visual prompt dependency measure (PDM), which will later motivate our algorithms (Sec. 4). The main aim of this section is to assess whether a model output, such as a token, is hallucinated or not by comparing its likelihood with the likelihood of that same output being generated without the relevant conditioning factor, in this case, the image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.904 + ], + "angle": 0, + "content": "We consider decoding approaches for open-ended language generation of VLMs, where the model receives an input image and an input textual prompt and aims to generate a fluent and coherent continuation conditioned on both the image and the text. We denote probabilistic generative VLMs over a vocabulary of text tokens \\(\\mathcal{V}\\) with \\(p\\) and denote the probability of a token \\(y\\in \\mathcal{V}\\) given a textual prompt \\(\\mathbf{x}\\) and an image context \\(c\\) as \\(p(y|\\mathbf{x},c)\\). The goal of the VLM is to leverage the information contained in the input image \\(c\\) to provide a continuation \\(\\mathbf{y} = [y_0,\\dots ,y_T]\\) of the input prompt \\(\\mathbf{x}\\), e.g., \"Describe this image in detail\". The probability of a valid sequence \\(\\mathbf{y}\\) can be computed as \\(p(\\mathbf{y}|\\mathbf{x},c) = \\prod_{t = 1}^{T}p(y_{t}|\\mathbf{y}_{< t},\\mathbf{x},c)\\), where \\(\\mathbf{y}_{< t}\\triangleq [y_0,\\dots ,y_{t - 1}]\\)." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.092, + 0.696, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.092, + 0.887, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.214, + 0.895, + 0.41 + ], + "angle": 0, + "content": "Figure 3. The influence of the image conditioning decreases as we generate more tokens. (Left) Conditioning dilution. We report the average of the prompt dependency measure (PDM-H) and its standard deviation for different synthetic captions generated by LLaVA on MS COCO's validation split. We see that the influence of the image over the next token prediction decreases as we generate more. This suggests that the information within VLM's visual prompt gets diluted and fades away as more tokens are generated. (Right) Frequency of hallucinated objects as a function of the token position. We report the number of non-existent objects present on the same synthetic captions as a function of the number of generated tokens. Note that very few objects are hallucinated for tokens near the visual prompt, while their number increases as more tokens are generated and with a smaller PDM." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.423, + 0.893, + 0.453 + ], + "angle": 0, + "content": "We propose to study hallucinations on VLMs using PDMs defined as follows" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.46, + 0.868, + 0.486 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P D M} (\\mathbf {y} _ {< t}; c | \\mathbf {x}) \\triangleq \\operatorname {d i s t} \\left(p (\\cdot | \\mathbf {y} _ {< t}, \\mathbf {x}, c), p (\\cdot | \\mathbf {y} _ {< t}, \\mathbf {x})\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.893, + 0.715 + ], + "angle": 0, + "content": "where dist is any distance measure between probability distributions, such as Hellinger, total variation, or KL. PDMs quantify how generic or context-specific a language model's output is. In particular, a high \\(\\mathrm{PDM}(\\mathbf{y}_{< t};c|\\mathbf{x})\\) indicates that the token \\(y_{t}\\) is strongly associated with a specific input prompt, while a low PDM suggests that the token is more prompt-neutral or prompt-agnostic. Depending on the choice of the distance function, PDMs highlight different aspects of the generative distribution. We will mainly use PDM-H based on the Hellinger distance defined as \\(H(p,q) = 2^{-1 / 2}\\sqrt{\\sum_{i = 1}^{k}(\\sqrt{p_i} - \\sqrt{q_i})^2}\\), where \\(p = (p_i)_{i\\in [k]}\\) and \\(q = (q_i)_{i\\in [k]}\\) are discrete probability distributions. We refer the reader to the Supplementary Material for an analysis of the impact of other distances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Contextual pressure. Fig. 2 shows the conditional and unconditional prediction likelihoods on a single caption. First, we note that tokens that are required for linguistic fluency, like prepositions and conjunctions, are highly predictable by both the conditioned and the unconditioned models. This pattern also emerges when the model generates tokens to compose \"fine-grained\" objects. For example, when the model generates \"Peanut butter\", both the VLM and the LLM can correctly predict the last token even without looking at the image and only having access to a sufficiently long truncation (e.g. \"Peanut bu\"). Therefore, in these circumstances, \\( p(y_{t}|\\mathbf{y}_{CaptioningCHAIRi ↓CHAIRs ↓Cover ↑LLaVA7B8.117.553.3LLaVA7B PMI [25]6.716.251.5LLaVA7B Contrastive [8]6.314.855.2LLaVA7B M3ID5.913.855.1LLaVA13B7.418.555.2LLaVA13B PMI [25]6.516.654.7LLaVA13B Contrastive [8]6.514.554.7LLaVA13B M3ID5.513.254.0LLaVA13B + LURE [31]6.427.1-mPLUG-Owl + LURE [31]5.418.8-LLaVA7B M3ID + DPO5.713.555.8LLaVA13B M3ID + DPO5.312.654.2" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.691 + ], + "angle": 0, + "content": "eration by increasing the likelihood of generating tokens that are related to the text to be summarized. Differently from our method, the weight assigned to the realization of marginally likely continuations is time-independent. Similarly, Contrastive Decoding [7, 8] has been developed to foster generations of an expert LLM to contain plausible and fluent text with the textual input prompt by amplifying its prediction difference with respect to a weaker model. To ensure a fair comparison, we replace the text to be summarized with the input image for PMI and set the weaker model in Contrastive Decoding as the unconditioned model (see Sec. B for more details). We also compare with training-based methods that have been proposed to improve the grounding of VLMs on the visual prompt like LLaVARLHF [17], Robust mPLUG-Owl [11] and LURE [31]. However, all these methods require some form of annotations, like preference data [17] or positive/negative examples [11, 31], which M3ID+DPO does not require." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Evaluation. We evaluate our algorithm both on captioning and VQA benchmarks. On captioning, we measure object hallucinations by using the CHAIR (Captioning Hallucination Assessment with Image Relevance) metrics [20]. These metrics assess the quality of captions by comparing the mentioned objects to the annotated objects present in an image. In particular, CHAIR comprises two variants: one computes the fraction of hallucinated objects in the whole caption (CHAIRi) and the other evaluates the fraction of captions with at least one object hallucination (CHAIRs)," + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.872, + 0.393, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C H A I R i} = \\frac {\\# \\text {h a l l u c i n a t e d o b j e c t s}}{\\# \\text {g e n e r a t e d o b j e c t s}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.088, + 0.819, + 0.121 + ], + "angle": 0, + "content": "\\[\n\\text {C H A I R s} = \\frac {\\# \\text {h a l l u c i n a t e d c a p t i o n s}}{\\# \\text {g e n e r a t e d c a p t i o n s}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.125, + 0.892, + 0.185 + ], + "angle": 0, + "content": "Furthermore, we introduce the Cover metric to quantify the comprehensiveness of the captions by measuring the fraction of ground-truth annotated objects mentioned by the model," + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.195, + 0.831, + 0.23 + ], + "angle": 0, + "content": "\\[\n\\text {C o v e r} = \\frac {\\left| \\{\\text {c o r r e c t m e n t i o n e d o b j e c t s} \\} \\right|}{\\left| \\{\\text {a n n o t a t e d o b j e c t s} \\} \\right|}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.238, + 0.892, + 0.315 + ], + "angle": 0, + "content": "To evaluate object perception, we use the POPE (Polling-based Object Probing Evaluation) VQA benchmark [9]. In particular, POPE recasts the evaluation of object hallucinations into a binary classification problem with yes/no questions of the type \"Is a present in the image?\"" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.32, + 0.892, + 0.396 + ], + "angle": 0, + "content": "DPO training details. We use Hugging Face's DPO implementation in the TRL library [26] and train LLaVA for 5 epochs on 10,000 self-generated preference pairs with LoRA and cosine decaying learning rate with \\(2 \\times 10^{-5}\\) peak. Following [19], we set \\(\\beta = 0.1\\) in the DPO loss." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.405, + 0.771, + 0.422 + ], + "angle": 0, + "content": "5.1. VLM grounding on captioning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.892, + 0.517 + ], + "angle": 0, + "content": "In Tab. 1 we test how much our method can prevent the generation of hallucinated objects in image captions. In particular, we compare against training-free and training-based baselines. We prompt all baseline methods with the instruction \"Describe the image.\" and let the models generate until the EOS token is obtained." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.893, + 0.836 + ], + "angle": 0, + "content": "First, we compare M3ID with other decoding strategies, PMI [25] and Contrastive Decoding [8]. The main difference between M3ID and these baselines is that M3ID increasingly counteracts the language prior as more tokens are generated. As such, M3ID reduces ungrounded generations compared to all other training-free baselines both on the large LLaVA13B and on the smaller LLaVA7B. Furthermore, it has uniformly good improvements for different model sizes over standard multinomial decoding, in particular, M3ID achieves \\(27\\% / 21\\%\\) relative improvement over LLaVA7B and \\(26\\% / 29\\%\\) over LLaVA13B on the CHAIRi and CHAIRs metrics. Importantly, this improvement does not come at the price of high reductions of the Cover metric, which actually improves on the 7B model and decreases by less than \\(2.2\\%\\) for the larger 13B model1. Lastly, we note that M3ID shows an improvement in absolute performance that correlates with model size, suggesting that further gains could be obtained as larger models are used. In Sec. D.2, we also show that M3ID still improves grounding when longer captions (twice as long on average) are generated and where other decodings do not perform as well." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.836, + 0.892, + 0.867 + ], + "angle": 0, + "content": "In Tab. 1 we also report training-based results and show that pairing DPO with M3ID leads to a smaller number" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1Note that CHAIR metrics can be hacked by simply returning shorter captions that do not attempt to predict any object in the image." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14308" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.161 + ], + "angle": 0, + "content": "Table 2. Evaluation on the POPE VQA hallucination benchmark [9]. *;† indicate results taken from [11] and [23] respectively. SUP. Indicates that the method involves supervised training on annotated data, in contrast to our approaches. POPE is divided into three subsets, Random, Popular, and Adversarial. We report the VQA accuracy of the model using the following template: “Is a present in the image?”, where we sample randomly (Random), among the most frequent objects in the dataset (Popular), or among the objects that frequently co-occur with (Adversarial). Acc. is the binary classification accuracy and Yes is the percentage of “Yes” answers." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.171, + 0.885, + 0.427 + ], + "angle": 0, + "content": "
POPE
RandomPopularAdversarialAll
Acc. ↑Yes (%)Acc. ↑Yes (%)Acc. ↑Yes (%)Acc. ↑Yes (%)
Robust mPLUG-Owl-7B* [11]sup.LLaVA-RLHF†7B [23]sup.86.0-73.0-65.0-74.7-
84.839.683.341.880.744.082.941.8
mPLUG-Owl-7B* [29]52.0-57.0-60.0-67.3-
LLaVA7B74.875.161.886.758.190.164.984.0
LLaVA7B M3ID76.067.769.373.365.877.670.372.9
LLaVA7B M3ID + DPO81.265.673.967.368.275.474.469.4
LLaVA-RLHF†13B [23]sup.85.238.483.938.082.340.583.839.0
MiniGPT4-13B* [32]73.0-67.0-62.0-74.7-
LLaVA13B67.980.663.883.259.887.363.883.7
LLaVA13B M3ID84.355.677.061.671.368.277.561.8
LLaVA13B M3ID + DPO85.253.479.157.573.267.579.251.1
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.471, + 0.5 + ], + "angle": 0, + "content": "of hallucinated objects and improved Cover numbers compared both to our training-free approach and LLaVA+LURE [31] a concurrent training based method to improve the grounding of VLMs that relies on GPT-3.5 annotations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.508, + 0.31, + 0.525 + ], + "angle": 0, + "content": "5.2. VLM grounding on VQA" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.47, + 0.682 + ], + "angle": 0, + "content": "Differently from the MS COCO captioning task, the POPE VQA benchmark only requires the generation of \"Yes/No\" tokens. As such, one does not expect the dilution effect to play a key role. However, we highlight that the image tokens and the output of the VLM are separated by the input template used to prime the model for the VQA task, which, as we show in Sec. D.4, introduces a non-negligible dilution effect. To address this, we take an offset into account when using M3ID on POPE, and select \\( t = t_0 \\) where \\( t_0 \\) is the number of tokens in between the output and the image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.848 + ], + "angle": 0, + "content": "In Tab. 2, we report the results on the POPE VQA hallucination benchmark [9]. Hallucinations (wrong answers) tend to correlate with the tendency of the VLM to reply using the \"Yes\" token (see the disproportionately high percentage rate, \\(84\\% / 83.7\\%\\), of \"Yes\" answers for the LLaVAv1.3 base model). This tendency can be counteracted at inference time and without any training by simply using M3ID. In our experiments, M3ID reduces the Yes ratio to \\(72.9\\% / 61.8\\%\\) for 7B and 13B models, respectively, which leads to relative accuracy improvements over standard LLaVA decoding by \\(8\\%\\) and \\(21\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.848, + 0.47, + 0.88 + ], + "angle": 0, + "content": "To test whether our training-based approach increases the VLM's grounding on the visual prompt we trained a" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.439, + 0.895, + 0.726 + ], + "angle": 0, + "content": "model with our DPO objective on the MS COCO captioning dataset and then tested it on the POPE benchmark. Note that improving caption generation does not directly imply improvements on the POPE VQA benchmark, in fact, even if the images are both from MS COCO, the output format is quite different: open-ended generation on the former and binary classification on the latter. However, intuitively, one expects that by improving the reliance on the visual prompt, the model could make better use of the visual information regardless of the output format required to solve the task at hand. Tab. 2 shows that M3ID+DPO further improves performance over M3ID's inference time intervention. Specifically, M3ID+DPO achieves \\(15\\%\\) and \\(24\\%\\) accuracy improvements over the LLaVA 7B and 13B models respectively. For completeness, we also compare with other training-based baselines that are fine-tuned on labeled [11] and preference data [17]. While M3ID+DPO does not have access to any labeled information we show that it is close to these baselines without requiring any annotations." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.735, + 0.611, + 0.75 + ], + "angle": 0, + "content": "5.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Conditioning dilution and \"overcompensation\". In Fig. 4 and Tab. 3 we show that when the forgetting factor \\(\\lambda\\) is high, corresponding to the assumption that the input image gets forgotten quickly, M3ID effectively maintains a large PDM-H throughout the generation. However, while deviating from the unconditioned probability is often a desired behavior, an excessive deviation from it across the whole generation could result in \"overcompensation\", which, as we show in Tab. 3 can be counterproductive and can lead to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.448, + 0.902 + ], + "angle": 0, + "content": "2We use [Img] [Model Prompt] [Question], see Sec. D.4." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14309" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.271, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.091, + 0.465, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.214, + 0.47, + 0.284 + ], + "angle": 0, + "content": "Figure 4. Sensitivity analysis on hyper-parameters. We vary both the forgetting factor \\(\\lambda\\) and the thresholding parameter \\(\\alpha\\). Small \\(\\lambda\\) and \\(\\alpha\\) result in minor effects. Instead, either a high forgetting factor \\(\\lambda\\) or a high \\(\\alpha\\) results in a stronger intervention which leads to higher PDM-H." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.47, + 0.46 + ], + "angle": 0, + "content": "higher hallucination rates. In particular, observe that higher corrections detrimentally impact the cover metric. This suggests that while M3ID tries to find tokens that diverge from the language prior the most, it is more likely to fail in captioning elements that are inherently predictable by the language prior alone. For instance, in describing an image of a dog on a leash accompanied by a man, M3ID might overlook mentioning the presence of the man, a token that the language prior could anticipate without necessitating any visual information only from context clues. We report examples of this behavior in the Sec. D.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.679 + ], + "angle": 0, + "content": "Importance of the confidence threshold. When we set a high threshold \\(\\alpha\\), the indicator function remains active more frequently along the generation, leading M3ID to consistently prioritize maximizing the distance of the generation from the language prior. Hence, similarly to the scenario with a high \\(\\lambda\\), M3ID is likely to overcompensate and increase the rate of multi-modal hallucinations. However, too high \\(\\alpha\\) disrupts linguistic fluency. We report qualitative examples of this in Sec. D.2. In Tab. 4 we also show that adding the term to compensate for contextual pressure does not significantly change the results over LLaVA decoding in the POPE benchmark since it is essentially framed as a binary classification over the \"Yes\"/ \"No\" tokens and not open-ended generation like captioning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.694, + 0.205, + 0.71 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We introduced M3ID, a new approach designed to combat multi-modal hallucinations by maximizing the mutual information between the text generated by VLMs and the corresponding visual context. M3ID operates at inference time and can be seamlessly integrated with any pre-trained autoregressive VLM. This makes M3ID a cost-effective and flexible solution to enhance vision-language grounding. Furthermore, for settings where model training is feasible and higher visual grounding is expected, we also paired M3ID with Direct Preference Optimization (DPO), showcasing reduced hallucinatory behaviors. Interestingly, our findings suggest that object hallucinations in VLMs result" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.145 + ], + "angle": 0, + "content": "Table 3. Sensitivity to hyper-parameters. \\(\\lambda\\) is the forgetting factor used to counteract conditioning dilution while \\(\\alpha\\) is the threshold used to mitigate overcompensation when the contextual pressure is high. We report results on LLaVA-7B." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.153, + 0.89, + 0.31 + ], + "angle": 0, + "content": "
Captioning
CHAIRi ↓CHAIRs ↓Cover ↑
LLaVA7B8.017.853.5
M3ID α = 0.3, λ = 0.025.813.655.0
M3ID α = 0.3, λ = 0.0017.317.453.7
M3ID α = 0.3, λ = 0.036.214.752.0
M3ID α = 0.3, λ = 0.17.216.345.5
M3ID α = 0.5, λ = 0.026.114.954.6
M3ID α = 0.1, λ = 0.026.915.453.1
M3ID α = 0.01, λ = 0.026.916.652.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.32, + 0.892, + 0.362 + ], + "angle": 0, + "content": "Table 4. Ablation studies. We ablate the components of M3ID on both model sizes. Removing the amplification term to counteract conditioning dilution leads to more hallucinations." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.37, + 0.89, + 0.499 + ], + "angle": 0, + "content": "
POPE\nAccuracy ↑CHAIRi ↓Captioning\nCHAIRs ↓Cover ↑
LLaVA7B64.98.017.853.5
w/ context pressure65.56.916.754.5
w/ conditioning dilution77.56.414.153.9
LLaVA7B M3ID77.55.813.655.0
LLaVA13B63.815.318.255.3
w/ context pressure64.96.514.455.0
w/ conditioning dilution70.35.914.853.7
LLaVA13B M3ID70.35.413.054.0
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.569 + ], + "angle": 0, + "content": "from excessive reliance on the language prior rather than a poor understanding of the visual modality. In fact, M3ID at inference time is sufficient to significantly reduce the amount of generated hallucinations without any training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.893, + 0.856 + ], + "angle": 0, + "content": "A limitation of M3ID is that it requires two forward passes at inference time, one for the conditioned and one for the unconditioned prediction. A possible solution to not increase inference time, but at the expense of higher memory consumption, is to use two batched queries with one having masked visual tokens. Furthermore, as we observed in Tab. 1, sometimes M3ID may prevent the generation of objects that are highly likely under the unprompted language prior (due to context clues). Interestingly, a similar observation has been reported in [22], where, people asked to provide a 10-word list of objects contained in a given image often failed to report the most obvious objects while mainly focusing on secondary ones. While we already showed that this can be mitigated with proper hyperparameter selection Tab. 3 we believe that an interesting avenue for future research is to directly encode this preference within the model's weights by favoring the generation of structured captions that get progressively more detailed while still being grounded to the image." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In general, integrating human- or AI-annotated preference pairs, assessed based on their level of grounding, constitutes a promising avenue for future investigation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14310" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, et al. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[3] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[4] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 5, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.471, + 0.45 + ], + "angle": 0, + "content": "[5] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[6] Yichong Huang, Xiachong Feng, Xiaocheng Feng, and Bing Qin. The factual inconsistency problem in abstractive text summarization: A survey. arXiv preprint arXiv:2104.14839, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.606 + ], + "angle": 0, + "content": "[7] Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110–119, San Diego, California, 2016. Association for Computational Linguistics. 1, 3, 4, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.471, + 0.662 + ], + "angle": 0, + "content": "[8] Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. Contrastive decoding: Open-ended text generation as optimization, 2023. 2, 5, 6, 11, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.663, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[9] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 2, 6, 7, 11, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[10] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 4, 5, 6, 11, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.804, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[11] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[12] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[13] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 2, 4, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[14] M.B. Matthews and G.S. Moschytz. The identification of nonlinear discrete-time fading-memory systems using neural network models. IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing, 41(11):740-751, 1994. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.204, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[15] Clara Meister, Tiago Pimentel, Gian Wiher, and Ryan Coterell. Locally typical sampling. Transactions of the Association for Computational Linguistics, 11:102-121, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[16] Yatin Nandwani, Vineet Kumar, Dinesh Raghu, Sachindra Joshi, and Luis A Lastras. Pointwise mutual information based metric and decoding strategy for faithful generation in document grounded dialogs. arXiv preprint arXiv:2305.12191, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[17] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[18] 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. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[19] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. arXiv preprint arXiv:2305.18290, 2023. 2, 5, 6, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[20] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4035-4045, Brussels, Belgium, 2018. Association for Computational Linguistics. 2, 6, 11, 14, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[21] Weijia Shi, Xiaochuang Han, Mike Lewis, Yulia Tsvetkov, Luke Zettlemoyer, and Scott Wen-tau Yih. Trusting your evidence: Hallucinate less with context-aware decoding. arXiv preprint arXiv:2305.14739, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[22] Merrielle Spain and Pietro Perona. Measuring and predicting object importance. Int. J. Comput. Vis., 91(1):59-76, 2011. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[23] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.873 + ], + "angle": 0, + "content": "[24] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[25] Liam van der Poel, Ryan Cotterell, and Clara Meister. Mutual information alleviates hallucinations in abstractive sum" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14311" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "marization. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5956-5965, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 2, 3, 4, 5, 6, 11, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[26] Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, and Shengyi Huang. Trl: Transformer reinforcement learning. https://github.com/huggingface/trl, 2020.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.274 + ], + "angle": 0, + "content": "[27] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, et al. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.345 + ], + "angle": 0, + "content": "[28] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.47, + 0.415 + ], + "angle": 0, + "content": "[29] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[30] Luca Zancato and Alessandro Chiuso. A novel deep neural network architecture for non-linear system identification. IFAC-PapersOnLine, 54(7):186-191, 2021. 19th IFAC Symposium on System Identification SYSID 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.475, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[31] Yiyang Zhou, Chenhang Cui, Jaehong Yoon, Linjun Zhang, Zhun Deng, Chelsea Finn, Mohit Bansal, and Huaxiu Yao. Analyzing and mitigating object hallucination in large vision-language models. arXiv preprint arXiv:2310.00754, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.47, + 0.6 + ], + "angle": 0, + "content": "[32] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14312" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-Modal Hallucination Control by Visual Information Grounding/1e5cae18-a58b-4415-90f8-bc4858438f19_origin.pdf b/2024/Multi-Modal Hallucination Control by Visual Information Grounding/1e5cae18-a58b-4415-90f8-bc4858438f19_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bd7b3c7028b9a2b46521822dd821d81ae0f1b29f --- /dev/null +++ b/2024/Multi-Modal Hallucination Control by Visual Information Grounding/1e5cae18-a58b-4415-90f8-bc4858438f19_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9170d73b003956feed087ed70c475529b993ffc39a73092c29adf6b5fb7b2d6 +size 2620216 diff --git a/2024/Multi-Modal Hallucination Control by Visual Information Grounding/full.md b/2024/Multi-Modal Hallucination Control by Visual Information Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..39130042d38bb53df836f007d603264b41e6b67f --- /dev/null +++ b/2024/Multi-Modal Hallucination Control by Visual Information Grounding/full.md @@ -0,0 +1,294 @@ +# Multi-Modal Hallucination Control by Visual Information Grounding + +Alessandro Favero* +Pramuditha Perera + +Luca Zancato Alessandro Achille + +Matthew Trager Sid Ashwin Swaminathan + +Siddharth Choudhary +than Stefano Soatto + +# AWS AI Labs + +alessandro.favero@epfl.ch + +{zancato,mttrager,sidchoud,pramudi,aachille,swashwin,soattos}@amazon.com + +# Abstract + +Generative Vision-Language Models (VLMs) are prone to generate plausible-sounding textual answers that, however, are not always grounded in the input image. We investigate this phenomenon, usually referred to as "hallucination" and show that it stems from an excessive reliance on the language prior. In particular, we show that as more tokens are generated, the reliance on the visual prompt decreases, and this behavior strongly correlates with the emergence of hallucinations. To reduce hallucinations, we introduce Multi-Modal Mutual-Information Decoding (M3ID), a new sampling method for prompt amplification. M3ID amplifies the influence of the reference image over the language prior, hence favoring the generation of tokens with higher mutual information with the visual prompt. M3ID can be applied to any pre-trained autoregressive VLM at inference time without necessitating further training and with minimal computational overhead. If training is an option, we show that M3ID can be paired with Direct Preference Optimization (DPO) to improve the model's reliance on the prompt image without requiring any labels. Our empirical findings show that our algorithms maintain the fluency and linguistic capabilities of pre-trained VLMs while reducing hallucinations by mitigating visually ungrounded answers. Specifically, for the LLaVA 13B model, M3ID and M3ID+DPO reduce the percentage of hallucinated objects in captioning tasks by $25\%$ and $28\%$ , respectively, and improve the accuracy on VQA benchmarks such as POPE by $21\%$ and $24\%$ . + +# 1. Introduction + +Recent autoregressive Vision-Language Models (VLMs) have shown remarkable multimodal capabilities [4, 13]. However, VLMs, similarly to large language models (LLMs), are prone to "hallucinations" - generating + +Prompt: Describe the image. +![](images/6418e547ac52e90379698ef9281e967ecfdd611c091bd8937d57aa9add76bce7.jpg) +$\mathbf{y}_{< t}$ : The image depicts a kitchen with an oven, a + +![](images/c527d128499359f3de06d1ab5b5d55ca3e31b434594184bb42fe87d978530ce1.jpg) +Figure 1. Multi-Modal Mutual Information Decoding (M3ID). Given a VLM $p$ , an image $c$ , and a text prompt, M3ID intervenes in the generative distribution by finding tokens that "surprise" the unconditioned VLM (i.e., the VLM without the image prompt). M3ID amplifies the conditioned directions that are not already predicted by the unconditioned model more as new tokens are generated by leveraging a progressively smaller $\gamma_{t}$ . In the example, the VLM assigns a high likelihood to the hallucinated object fridge, over-relying on its unconditioned language prior. Instead, M3ID assigns a high likelihood to toaster, which is present in the image. + +plausible-sounding answers without factual basis, leading to potentially ungrounded or fabricated information [7]. Consequently, the community has been developing ever so complex and expensive alignment algorithms involving direct human supervision [17] and often brittle prompt engineering methods ("model begging") [28]. + +In this work, we propose to investigate hallucinations in VLMs through a quantifiable measure of visual prompt dependency. We assess whether a model output is ungrounded with respect to a visual prompt by comparing its likelihood with the likelihood of generating the same output + +without the visual information. In general, our measure of visual prompt dependency quantifies the extent to which a token is generic or specific to a visual signal. Importantly, low visual prompt dependency does not necessarily imply that a token is hallucinated, as it may occur in linguistically essential elements like conjunctions and prepositions. Yet, unlike the notion of hallucinations, which requires human judgment, our prompt dependency measure is always well-defined even without any ground truth annotations. + +Our first contribution is to empirically demonstrate that the visual prompt dependency measure decreases as more tokens are generated. In Fig. 3, we show that, as more tokens are generated, the conditioning information gets diluted and "forgotten" or ignored by the model, possibly leading to more hallucinations. We refer to this effect as conditioning dilution or fading memory effect. + +To counteract this, we propose an intervention on the generative distribution of a VLM that maximizes visual prompt dependency at inference time. We show that our method for prompt amplification maximizes the mutual information between the text output tokens and the visual prompt, effectively rescaling the image-conditioned component against the unconditioned distribution. We name our inference-time intervention on the generative distribution Multi-Modal Mutual Information Decoding (M3ID). M3ID is applicable to any off-the-shelf model without additional training or access to model weights, offering a low computational overhead alternative to standard decoding algorithms [13, 24]. Our results show that M3ID enhances the dependence on the visual prompt and reduces the number of hallucinations across various benchmarks while preserving the linguistic fluency of the original model. + +Additionally, for users with access to model weights, we propose a training objective to further ground the model outputs on visual prompts. This approach involves recasting our prompt amplification objective as a preference optimization problem, using Direct Preference Optimization (DPO) [19] to align the pre-trained VLM. This method aims to increase the likelihood of continuations with a higher visual prompt dependency measure and allows to learn a better generation policy, further reducing the conditioning dilution effect and hallucinations. + +In summary, our main contributions are as follows: + +1. We propose a visual prompt dependency measure (PDM) to assess whether a model output is ungrounded with respect to the visual input. Empirically, we demonstrate that PDM decreases as more tokens are generated making the generations more likely to be hallucinated. +2. We introduce M3ID, a training-free intervention on the generative distribution of autoregressive VLMs which improves visual grounding and reduces hallucinations by amplifying the importance of the visual prompt over the language prior. In addition, we extend M3ID with + +DPO for multi-modal preference optimization to further improve visual grounding. + +3. We show that applying M3ID or DPO reduces the percentage of hallucinated objects in captioning tasks [20] by $25\%$ and by $28\%$ , respectively and improves accuracy on the POPE [9] VQA hallucination benchmark by $21\%$ and $24\%$ over the base model. + +# 2. Related work + +Hallucinations in VLMs. A recent line of work introduced VLMs obtained by grafting visual encoders into pre-trained Large Language Models (LLMs) [4, 13, 32]. Forcing a pre-trained LLM to learn to "see" by reading vision tokens from a pre-trained vision backbone has proven successful and the resulting models show remarkable vision-language understanding capabilities on many multi-modal tasks [13, 23]. However, while inheriting strong linguistic capabilities and fluency from their base LLM, grafted VLMs also inherit the tendency to produce ungrounded or fabricated information [1, 6]. This is commonly referred to as "hallucination" in the recent machine learning literature [1, 6, 25]. In particular, several recent works empirically reported a sharp tendency of grafted VLMs to report objects not grounded on the visual information when probed with questions on the image content [1, 6, 9, 27]. The authors of [9] observed that this phenomenon is especially pronounced with objects that are either common or frequently appear together in the datasets used at training time. Furthermore, the authors of [11] suggest that VLMs often fail to correctly follow instructions involving absent objects and propose a new instruction-following training objective to enhance model alignment and robustness in the face of uncertainty. Parallel to our work, [31] delves into the factors contributing to object hallucinations in VLMs, exploring aspects like object co-occurrence, model uncertainty, and spatial position of hallucinations in the sentence. They propose a post-hoc algorithm to identify and correct hallucinations in VLM-generated content. + +Context-dependent decodings. Decoding algorithms can be classified as either search or sampling algorithms [8]. Current search methods (e.g., greedy and beam search) can produce factually accurate generations but usually suffer from tedious and repetitive continuations. Sampling methods, on the other hand, (like nucleus [5], or typical decoding [15]) produce more diverse text, but suffer from topic drift. To solve these shortcomings, various decoding methods have been proposed to enhance reasoning and factual accuracy in generative autoregressive language models. Approaches such as context-aware decoding [21] and Pointwise Mutual Information (PMI) decoding [16, 25] focus on aligning the outputs of language models with the intended context or factual data. These strategies aim to reduce hal + +![](images/2e23e32c7b76705df7ff11bf12fb56e3da31459f752a9f5ffbf8c1db06c867ef.jpg) +Figure 2. How much does the conditioning prompt "surprise" the unconditioned predictor? We report the likelihood that the conditioned and unconditioned models assign to each token in the string "The image ..." The probability gap increases on tokens for which visual information is required (e.g. objects and attributes), while it decreases for punctuation and articles. Note that both models mostly agree when sufficient visual information has already been extracted and is present in the caption (e.g. "kitchen" is very likely for the unconditioned model as well). See Sec. D.1 for a plot with our prompt dependency measure. + +lucinations in standard natural language processing tasks such as summarization. Mutual-information-based decoding techniques have proven to be helpful in steering the generation of LLMs to stay faithful to the given input text [25], or to promote diversity or relevance in neural dialogue models [7]. Our work is the first to use mutual information to improve multi-modal grounding and to vary the penalization of marginally likely outputs in a time-varying fashion to counteract the progressive forgetting of the visual prompt. + +# 3. Analysis of hallucinations in VLMs + +We start by investigating hallucinations in VLMs. We introduce a visual prompt dependency measure (PDM), which will later motivate our algorithms (Sec. 4). The main aim of this section is to assess whether a model output, such as a token, is hallucinated or not by comparing its likelihood with the likelihood of that same output being generated without the relevant conditioning factor, in this case, the image. + +We consider decoding approaches for open-ended language generation of VLMs, where the model receives an input image and an input textual prompt and aims to generate a fluent and coherent continuation conditioned on both the image and the text. We denote probabilistic generative VLMs over a vocabulary of text tokens $\mathcal{V}$ with $p$ and denote the probability of a token $y\in \mathcal{V}$ given a textual prompt $\mathbf{x}$ and an image context $c$ as $p(y|\mathbf{x},c)$ . The goal of the VLM is to leverage the information contained in the input image $c$ to provide a continuation $\mathbf{y} = [y_0,\dots ,y_T]$ of the input prompt $\mathbf{x}$ , e.g., "Describe this image in detail". The probability of a valid sequence $\mathbf{y}$ can be computed as $p(\mathbf{y}|\mathbf{x},c) = \prod_{t = 1}^{T}p(y_{t}|\mathbf{y}_{< t},\mathbf{x},c)$ , where $\mathbf{y}_{< t}\triangleq [y_0,\dots ,y_{t - 1}]$ . + +![](images/a4798a083b13589972e8daec3c766f035778c91c998e9c02277996eb41b63b83.jpg) +Figure 3. The influence of the image conditioning decreases as we generate more tokens. (Left) Conditioning dilution. We report the average of the prompt dependency measure (PDM-H) and its standard deviation for different synthetic captions generated by LLaVA on MS COCO's validation split. We see that the influence of the image over the next token prediction decreases as we generate more. This suggests that the information within VLM's visual prompt gets diluted and fades away as more tokens are generated. (Right) Frequency of hallucinated objects as a function of the token position. We report the number of non-existent objects present on the same synthetic captions as a function of the number of generated tokens. Note that very few objects are hallucinated for tokens near the visual prompt, while their number increases as more tokens are generated and with a smaller PDM. + +![](images/bddc9be237dcad3e5725c9e6b78dde8a00d2157155c044e6c59eb9a564d34428.jpg) + +We propose to study hallucinations on VLMs using PDMs defined as follows + +$$ +\mathrm {P D M} (\mathbf {y} _ {< t}; c | \mathbf {x}) \triangleq \operatorname {d i s t} \left(p (\cdot | \mathbf {y} _ {< t}, \mathbf {x}, c), p (\cdot | \mathbf {y} _ {< t}, \mathbf {x})\right) +$$ + +where dist is any distance measure between probability distributions, such as Hellinger, total variation, or KL. PDMs quantify how generic or context-specific a language model's output is. In particular, a high $\mathrm{PDM}(\mathbf{y}_{< t};c|\mathbf{x})$ indicates that the token $y_{t}$ is strongly associated with a specific input prompt, while a low PDM suggests that the token is more prompt-neutral or prompt-agnostic. Depending on the choice of the distance function, PDMs highlight different aspects of the generative distribution. We will mainly use PDM-H based on the Hellinger distance defined as $H(p,q) = 2^{-1 / 2}\sqrt{\sum_{i = 1}^{k}(\sqrt{p_i} - \sqrt{q_i})^2}$ , where $p = (p_i)_{i\in [k]}$ and $q = (q_i)_{i\in [k]}$ are discrete probability distributions. We refer the reader to the Supplementary Material for an analysis of the impact of other distances. + +Contextual pressure. Fig. 2 shows the conditional and unconditional prediction likelihoods on a single caption. First, we note that tokens that are required for linguistic fluency, like prepositions and conjunctions, are highly predictable by both the conditioned and the unconditioned models. This pattern also emerges when the model generates tokens to compose "fine-grained" objects. For example, when the model generates "Peanut butter", both the VLM and the LLM can correctly predict the last token even without looking at the image and only having access to a sufficiently long truncation (e.g. "Peanut bu"). Therefore, in these circumstances, $p(y_{t}|\mathbf{y}_{CaptioningCHAIRi ↓CHAIRs ↓Cover ↑LLaVA7B8.117.553.3LLaVA7B PMI [25]6.716.251.5LLaVA7B Contrastive [8]6.314.855.2LLaVA7B M3ID5.913.855.1LLaVA13B7.418.555.2LLaVA13B PMI [25]6.516.654.7LLaVA13B Contrastive [8]6.514.554.7LLaVA13B M3ID5.513.254.0LLaVA13B + LURE [31]6.427.1-mPLUG-Owl + LURE [31]5.418.8-LLaVA7B M3ID + DPO5.713.555.8LLaVA13B M3ID + DPO5.312.654.2 + +eration by increasing the likelihood of generating tokens that are related to the text to be summarized. Differently from our method, the weight assigned to the realization of marginally likely continuations is time-independent. Similarly, Contrastive Decoding [7, 8] has been developed to foster generations of an expert LLM to contain plausible and fluent text with the textual input prompt by amplifying its prediction difference with respect to a weaker model. To ensure a fair comparison, we replace the text to be summarized with the input image for PMI and set the weaker model in Contrastive Decoding as the unconditioned model (see Sec. B for more details). We also compare with training-based methods that have been proposed to improve the grounding of VLMs on the visual prompt like LLaVARLHF [17], Robust mPLUG-Owl [11] and LURE [31]. However, all these methods require some form of annotations, like preference data [17] or positive/negative examples [11, 31], which M3ID+DPO does not require. + +Evaluation. We evaluate our algorithm both on captioning and VQA benchmarks. On captioning, we measure object hallucinations by using the CHAIR (Captioning Hallucination Assessment with Image Relevance) metrics [20]. These metrics assess the quality of captions by comparing the mentioned objects to the annotated objects present in an image. In particular, CHAIR comprises two variants: one computes the fraction of hallucinated objects in the whole caption (CHAIRi) and the other evaluates the fraction of captions with at least one object hallucination (CHAIRs), + +$$ +\mathrm {C H A I R i} = \frac {\# \text {h a l l u c i n a t e d o b j e c t s}}{\# \text {g e n e r a t e d o b j e c t s}}, +$$ + +$$ +\text {C H A I R s} = \frac {\# \text {h a l l u c i n a t e d c a p t i o n s}}{\# \text {g e n e r a t e d c a p t i o n s}}. +$$ + +Furthermore, we introduce the Cover metric to quantify the comprehensiveness of the captions by measuring the fraction of ground-truth annotated objects mentioned by the model, + +$$ +\text {C o v e r} = \frac {\left| \{\text {c o r r e c t m e n t i o n e d o b j e c t s} \} \right|}{\left| \{\text {a n n o t a t e d o b j e c t s} \} \right|}. +$$ + +To evaluate object perception, we use the POPE (Polling-based Object Probing Evaluation) VQA benchmark [9]. In particular, POPE recasts the evaluation of object hallucinations into a binary classification problem with yes/no questions of the type "Is a present in the image?" + +DPO training details. We use Hugging Face's DPO implementation in the TRL library [26] and train LLaVA for 5 epochs on 10,000 self-generated preference pairs with LoRA and cosine decaying learning rate with $2 \times 10^{-5}$ peak. Following [19], we set $\beta = 0.1$ in the DPO loss. + +# 5.1. VLM grounding on captioning + +In Tab. 1 we test how much our method can prevent the generation of hallucinated objects in image captions. In particular, we compare against training-free and training-based baselines. We prompt all baseline methods with the instruction "Describe the image." and let the models generate until the EOS token is obtained. + +First, we compare M3ID with other decoding strategies, PMI [25] and Contrastive Decoding [8]. The main difference between M3ID and these baselines is that M3ID increasingly counteracts the language prior as more tokens are generated. As such, M3ID reduces ungrounded generations compared to all other training-free baselines both on the large LLaVA13B and on the smaller LLaVA7B. Furthermore, it has uniformly good improvements for different model sizes over standard multinomial decoding, in particular, M3ID achieves $27\% / 21\%$ relative improvement over LLaVA7B and $26\% / 29\%$ over LLaVA13B on the CHAIRi and CHAIRs metrics. Importantly, this improvement does not come at the price of high reductions of the Cover metric, which actually improves on the 7B model and decreases by less than $2.2\%$ for the larger 13B model1. Lastly, we note that M3ID shows an improvement in absolute performance that correlates with model size, suggesting that further gains could be obtained as larger models are used. In Sec. D.2, we also show that M3ID still improves grounding when longer captions (twice as long on average) are generated and where other decodings do not perform as well. + +In Tab. 1 we also report training-based results and show that pairing DPO with M3ID leads to a smaller number + +Table 2. Evaluation on the POPE VQA hallucination benchmark [9]. *;† indicate results taken from [11] and [23] respectively. SUP. Indicates that the method involves supervised training on annotated data, in contrast to our approaches. POPE is divided into three subsets, Random, Popular, and Adversarial. We report the VQA accuracy of the model using the following template: “Is a present in the image?”, where we sample randomly (Random), among the most frequent objects in the dataset (Popular), or among the objects that frequently co-occur with (Adversarial). Acc. is the binary classification accuracy and Yes is the percentage of “Yes” answers. + +
POPE
RandomPopularAdversarialAll
Acc. ↑Yes (%)Acc. ↑Yes (%)Acc. ↑Yes (%)Acc. ↑Yes (%)
Robust mPLUG-Owl-7B* [11]sup.LLaVA-RLHF†7B [23]sup.86.0-73.0-65.0-74.7-
84.839.683.341.880.744.082.941.8
mPLUG-Owl-7B* [29]52.0-57.0-60.0-67.3-
LLaVA7B74.875.161.886.758.190.164.984.0
LLaVA7B M3ID76.067.769.373.365.877.670.372.9
LLaVA7B M3ID + DPO81.265.673.967.368.275.474.469.4
LLaVA-RLHF†13B [23]sup.85.238.483.938.082.340.583.839.0
MiniGPT4-13B* [32]73.0-67.0-62.0-74.7-
LLaVA13B67.980.663.883.259.887.363.883.7
LLaVA13B M3ID84.355.677.061.671.368.277.561.8
LLaVA13B M3ID + DPO85.253.479.157.573.267.579.251.1
+ +of hallucinated objects and improved Cover numbers compared both to our training-free approach and LLaVA+LURE [31] a concurrent training based method to improve the grounding of VLMs that relies on GPT-3.5 annotations. + +# 5.2. VLM grounding on VQA + +Differently from the MS COCO captioning task, the POPE VQA benchmark only requires the generation of "Yes/No" tokens. As such, one does not expect the dilution effect to play a key role. However, we highlight that the image tokens and the output of the VLM are separated by the input template used to prime the model for the VQA task, which, as we show in Sec. D.4, introduces a non-negligible dilution effect. To address this, we take an offset into account when using M3ID on POPE, and select $t = t_0$ where $t_0$ is the number of tokens in between the output and the image. + +In Tab. 2, we report the results on the POPE VQA hallucination benchmark [9]. Hallucinations (wrong answers) tend to correlate with the tendency of the VLM to reply using the "Yes" token (see the disproportionately high percentage rate, $84\% / 83.7\%$ , of "Yes" answers for the LLaVAv1.3 base model). This tendency can be counteracted at inference time and without any training by simply using M3ID. In our experiments, M3ID reduces the Yes ratio to $72.9\% / 61.8\%$ for 7B and 13B models, respectively, which leads to relative accuracy improvements over standard LLaVA decoding by $8\%$ and $21\%$ , respectively. + +To test whether our training-based approach increases the VLM's grounding on the visual prompt we trained a + +model with our DPO objective on the MS COCO captioning dataset and then tested it on the POPE benchmark. Note that improving caption generation does not directly imply improvements on the POPE VQA benchmark, in fact, even if the images are both from MS COCO, the output format is quite different: open-ended generation on the former and binary classification on the latter. However, intuitively, one expects that by improving the reliance on the visual prompt, the model could make better use of the visual information regardless of the output format required to solve the task at hand. Tab. 2 shows that M3ID+DPO further improves performance over M3ID's inference time intervention. Specifically, M3ID+DPO achieves $15\%$ and $24\%$ accuracy improvements over the LLaVA 7B and 13B models respectively. For completeness, we also compare with other training-based baselines that are fine-tuned on labeled [11] and preference data [17]. While M3ID+DPO does not have access to any labeled information we show that it is close to these baselines without requiring any annotations. + +# 5.3. Ablations + +Conditioning dilution and "overcompensation". In Fig. 4 and Tab. 3 we show that when the forgetting factor $\lambda$ is high, corresponding to the assumption that the input image gets forgotten quickly, M3ID effectively maintains a large PDM-H throughout the generation. However, while deviating from the unconditioned probability is often a desired behavior, an excessive deviation from it across the whole generation could result in "overcompensation", which, as we show in Tab. 3 can be counterproductive and can lead to + +![](images/0e1f2967b7693e9ee7f78ba4cd1ce0834178b64515ac247f7c9b52bfe9d53f7a.jpg) +Figure 4. Sensitivity analysis on hyper-parameters. We vary both the forgetting factor $\lambda$ and the thresholding parameter $\alpha$ . Small $\lambda$ and $\alpha$ result in minor effects. Instead, either a high forgetting factor $\lambda$ or a high $\alpha$ results in a stronger intervention which leads to higher PDM-H. + +![](images/da35ae2648ffb20772286d7e259bfff710fadbe0823dcf0c22df29a15592a3a9.jpg) + +higher hallucination rates. In particular, observe that higher corrections detrimentally impact the cover metric. This suggests that while M3ID tries to find tokens that diverge from the language prior the most, it is more likely to fail in captioning elements that are inherently predictable by the language prior alone. For instance, in describing an image of a dog on a leash accompanied by a man, M3ID might overlook mentioning the presence of the man, a token that the language prior could anticipate without necessitating any visual information only from context clues. We report examples of this behavior in the Sec. D.2. + +Importance of the confidence threshold. When we set a high threshold $\alpha$ , the indicator function remains active more frequently along the generation, leading M3ID to consistently prioritize maximizing the distance of the generation from the language prior. Hence, similarly to the scenario with a high $\lambda$ , M3ID is likely to overcompensate and increase the rate of multi-modal hallucinations. However, too high $\alpha$ disrupts linguistic fluency. We report qualitative examples of this in Sec. D.2. In Tab. 4 we also show that adding the term to compensate for contextual pressure does not significantly change the results over LLaVA decoding in the POPE benchmark since it is essentially framed as a binary classification over the "Yes"/ "No" tokens and not open-ended generation like captioning. + +# 6. Conclusions + +We introduced M3ID, a new approach designed to combat multi-modal hallucinations by maximizing the mutual information between the text generated by VLMs and the corresponding visual context. M3ID operates at inference time and can be seamlessly integrated with any pre-trained autoregressive VLM. This makes M3ID a cost-effective and flexible solution to enhance vision-language grounding. Furthermore, for settings where model training is feasible and higher visual grounding is expected, we also paired M3ID with Direct Preference Optimization (DPO), showcasing reduced hallucinatory behaviors. Interestingly, our findings suggest that object hallucinations in VLMs result + +Table 3. Sensitivity to hyper-parameters. $\lambda$ is the forgetting factor used to counteract conditioning dilution while $\alpha$ is the threshold used to mitigate overcompensation when the contextual pressure is high. We report results on LLaVA-7B. + +
Captioning
CHAIRi ↓CHAIRs ↓Cover ↑
LLaVA7B8.017.853.5
M3ID α = 0.3, λ = 0.025.813.655.0
M3ID α = 0.3, λ = 0.0017.317.453.7
M3ID α = 0.3, λ = 0.036.214.752.0
M3ID α = 0.3, λ = 0.17.216.345.5
M3ID α = 0.5, λ = 0.026.114.954.6
M3ID α = 0.1, λ = 0.026.915.453.1
M3ID α = 0.01, λ = 0.026.916.652.5
+ +Table 4. Ablation studies. We ablate the components of M3ID on both model sizes. Removing the amplification term to counteract conditioning dilution leads to more hallucinations. + +
POPE +Accuracy ↑CHAIRi ↓Captioning +CHAIRs ↓Cover ↑
LLaVA7B64.98.017.853.5
w/ context pressure65.56.916.754.5
w/ conditioning dilution77.56.414.153.9
LLaVA7B M3ID77.55.813.655.0
LLaVA13B63.815.318.255.3
w/ context pressure64.96.514.455.0
w/ conditioning dilution70.35.914.853.7
LLaVA13B M3ID70.35.413.054.0
+ +from excessive reliance on the language prior rather than a poor understanding of the visual modality. In fact, M3ID at inference time is sufficient to significantly reduce the amount of generated hallucinations without any training. + +A limitation of M3ID is that it requires two forward passes at inference time, one for the conditioned and one for the unconditioned prediction. A possible solution to not increase inference time, but at the expense of higher memory consumption, is to use two batched queries with one having masked visual tokens. Furthermore, as we observed in Tab. 1, sometimes M3ID may prevent the generation of objects that are highly likely under the unprompted language prior (due to context clues). Interestingly, a similar observation has been reported in [22], where, people asked to provide a 10-word list of objects contained in a given image often failed to report the most obvious objects while mainly focusing on secondary ones. While we already showed that this can be mitigated with proper hyperparameter selection Tab. 3 we believe that an interesting avenue for future research is to directly encode this preference within the model's weights by favoring the generation of structured captions that get progressively more detailed while still being grounded to the image. + +In general, integrating human- or AI-annotated preference pairs, assessed based on their level of grounding, constitutes a promising avenue for future investigation. + +# References + +[1] Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, et al. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023, 2023. 2 +[2] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 12 +[3] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 5 +[4] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 5, 16 +[5] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. In International Conference on Learning Representations, 2020. 2 +[6] Yichong Huang, Xiachong Feng, Xiaocheng Feng, and Bing Qin. The factual inconsistency problem in abstractive text summarization: A survey. arXiv preprint arXiv:2104.14839, 2021. 2 +[7] Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110–119, San Diego, California, 2016. Association for Computational Linguistics. 1, 3, 4, 6, 11 +[8] Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. Contrastive decoding: Open-ended text generation as optimization, 2023. 2, 5, 6, 11, 14 +[9] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 2, 6, 7, 11, 17 +[10] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 4, 5, 6, 11, 14, 15 +[11] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 2, 6, 7 +[12] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 5 + +[13] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 2, 4, 5, 11 +[14] M.B. Matthews and G.S. Moschytz. The identification of nonlinear discrete-time fading-memory systems using neural network models. IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing, 41(11):740-751, 1994. 4 +[15] Clara Meister, Tiago Pimentel, Gian Wiher, and Ryan Coterell. Locally typical sampling. Transactions of the Association for Computational Linguistics, 11:102-121, 2023. 2 +[16] Yatin Nandwani, Vineet Kumar, Dinesh Raghu, Sachindra Joshi, and Luis A Lastras. Pointwise mutual information based metric and decoding strategy for faithful generation in document grounded dialogs. arXiv preprint arXiv:2305.12191, 2023. 2 +[17] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 1, 6, 7 +[18] 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. 5 +[19] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. arXiv preprint arXiv:2305.18290, 2023. 2, 5, 6, 11, 12 +[20] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4035-4045, Brussels, Belgium, 2018. Association for Computational Linguistics. 2, 6, 11, 14, 17 +[21] Weijia Shi, Xiaochuang Han, Mike Lewis, Yulia Tsvetkov, Luke Zettlemoyer, and Scott Wen-tau Yih. Trusting your evidence: Hallucinate less with context-aware decoding. arXiv preprint arXiv:2305.14739, 2023. 2 +[22] Merrielle Spain and Pietro Perona. Measuring and predicting object importance. Int. J. Comput. Vis., 91(1):59-76, 2011. 8 +[23] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 2, 7 +[24] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 5 +[25] Liam van der Poel, Ryan Cotterell, and Clara Meister. Mutual information alleviates hallucinations in abstractive sum + +marization. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5956-5965, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 2, 3, 4, 5, 6, 11, 14 +[26] Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, and Shengyi Huang. Trl: Transformer reinforcement learning. https://github.com/huggingface/trl, 2020.6 +[27] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, et al. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 2 +[28] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 1 +[29] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 7 +[30] Luca Zancato and Alessandro Chiuso. A novel deep neural network architecture for non-linear system identification. IFAC-PapersOnLine, 54(7):186-191, 2021. 19th IFAC Symposium on System Identification SYSID 2021. 4 +[31] Yiyang Zhou, Chenhang Cui, Jaehong Yoon, Linjun Zhang, Zhun Deng, Chelsea Finn, Mohit Bansal, and Huaxiu Yao. Analyzing and mitigating object hallucination in large vision-language models. arXiv preprint arXiv:2310.00754, 2023. 2, 6, 7 +[32] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 7 \ No newline at end of file diff --git a/2024/Multi-Modal Hallucination Control by Visual Information Grounding/images.zip b/2024/Multi-Modal Hallucination Control by Visual Information Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eb39875fc1305e3d11594fa0b6bf8efa3f3a12a2 --- /dev/null +++ b/2024/Multi-Modal Hallucination Control by Visual Information Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c84573b34e6e4f1379163aff550ef4c941267afe0e2376f131071959807980e +size 366599 diff --git a/2024/Multi-Modal Hallucination Control by Visual Information Grounding/layout.json b/2024/Multi-Modal Hallucination Control by Visual Information Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d4b281f00f4dc7d456acddd9b12ce464446b6325 --- /dev/null +++ b/2024/Multi-Modal Hallucination Control by Visual Information Grounding/layout.json @@ -0,0 +1,8395 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "type": "text", + "content": "Multi-Modal Hallucination Control by Visual Information Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 144, + 188, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 144, + 188, + 171 + ], + "spans": [ + { + "bbox": [ + 79, + 144, + 188, + 171 + ], + "type": "text", + "content": "Alessandro Favero* \nPramuditha Perera" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 193, + 144, + 287, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 144, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 193, + 144, + 287, + 170 + ], + "type": "text", + "content": "Luca Zancato Alessandro Achille" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 294, + 144, + 416, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 144, + 416, + 171 + ], + "spans": [ + { + "bbox": [ + 294, + 144, + 416, + 171 + ], + "type": "text", + "content": "Matthew Trager Sid Ashwin Swaminathan" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 395, + 144, + 510, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 144, + 510, + 171 + ], + "spans": [ + { + "bbox": [ + 395, + 144, + 510, + 171 + ], + "type": "text", + "content": "Siddharth Choudhary \nthan Stefano Soatto" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 260, + 177, + 331, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 177, + 331, + 190 + ], + "spans": [ + { + "bbox": [ + 260, + 177, + 331, + 190 + ], + "type": "text", + "content": "AWS AI Labs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 228, + 196, + 365, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 196, + 365, + 207 + ], + "spans": [ + { + "bbox": [ + 228, + 196, + 365, + 207 + ], + "type": "text", + "content": "alessandro.favero@epfl.ch" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 102, + 209, + 492, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 209, + 492, + 221 + ], + "spans": [ + { + "bbox": [ + 102, + 209, + 492, + 221 + ], + "type": "text", + "content": "{zancato,mttrager,sidchoud,pramudi,aachille,swashwin,soattos}@amazon.com" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 249, + 192, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 249, + 192, + 261 + ], + "spans": [ + { + "bbox": [ + 143, + 249, + 192, + 261 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 275, + 289, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 289, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 289, + 598 + ], + "type": "text", + "content": "Generative Vision-Language Models (VLMs) are prone to generate plausible-sounding textual answers that, however, are not always grounded in the input image. We investigate this phenomenon, usually referred to as \"hallucination\" and show that it stems from an excessive reliance on the language prior. In particular, we show that as more tokens are generated, the reliance on the visual prompt decreases, and this behavior strongly correlates with the emergence of hallucinations. To reduce hallucinations, we introduce Multi-Modal Mutual-Information Decoding (M3ID), a new sampling method for prompt amplification. M3ID amplifies the influence of the reference image over the language prior, hence favoring the generation of tokens with higher mutual information with the visual prompt. M3ID can be applied to any pre-trained autoregressive VLM at inference time without necessitating further training and with minimal computational overhead. If training is an option, we show that M3ID can be paired with Direct Preference Optimization (DPO) to improve the model's reliance on the prompt image without requiring any labels. Our empirical findings show that our algorithms maintain the fluency and linguistic capabilities of pre-trained VLMs while reducing hallucinations by mitigating visually ungrounded answers. Specifically, for the LLaVA 13B model, M3ID and M3ID+DPO reduce the percentage of hallucinated objects in captioning tasks by " + }, + { + "bbox": [ + 47, + 275, + 289, + 598 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 47, + 275, + 289, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 275, + 289, + 598 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 47, + 275, + 289, + 598 + ], + "type": "text", + "content": ", respectively, and improve the accuracy on VQA benchmarks such as POPE by " + }, + { + "bbox": [ + 47, + 275, + 289, + 598 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 47, + 275, + 289, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 275, + 289, + 598 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 47, + 275, + 289, + 598 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "type": "text", + "content": "Recent autoregressive Vision-Language Models (VLMs) have shown remarkable multimodal capabilities [4, 13]. However, VLMs, similarly to large language models (LLMs), are prone to \"hallucinations\" - generating" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 355, + 251, + 489, + 331 + ], + "blocks": [ + { + "bbox": [ + 355, + 251, + 489, + 331 + ], + "lines": [ + { + "bbox": [ + 355, + 251, + 489, + 331 + ], + "spans": [ + { + "bbox": [ + 355, + 251, + 489, + 331 + ], + "type": "image", + "image_path": "6418e547ac52e90379698ef9281e967ecfdd611c091bd8937d57aa9add76bce7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 335, + 474, + 343 + ], + "lines": [ + { + "bbox": [ + 381, + 335, + 474, + 343 + ], + "spans": [ + { + "bbox": [ + 381, + 335, + 474, + 343 + ], + "type": "text", + "content": "Prompt: Describe the image." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 346, + 509, + 354 + ], + "lines": [ + { + "bbox": [ + 343, + 346, + 509, + 354 + ], + "spans": [ + { + "bbox": [ + 343, + 346, + 509, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{< t}" + }, + { + "bbox": [ + 343, + 346, + 509, + 354 + ], + "type": "text", + "content": ": The image depicts a kitchen with an oven, a" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 312, + 361, + 543, + 445 + ], + "blocks": [ + { + "bbox": [ + 312, + 361, + 543, + 445 + ], + "lines": [ + { + "bbox": [ + 312, + 361, + 543, + 445 + ], + "spans": [ + { + "bbox": [ + 312, + 361, + 543, + 445 + ], + "type": "image", + "image_path": "c527d128499359f3de06d1ab5b5d55ca3e31b434594184bb42fe87d978530ce1.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 458, + 545, + 569 + ], + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 569 + ], + "type": "text", + "content": "Figure 1. Multi-Modal Mutual Information Decoding (M3ID). Given a VLM " + }, + { + "bbox": [ + 304, + 458, + 545, + 569 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 458, + 545, + 569 + ], + "type": "text", + "content": ", an image " + }, + { + "bbox": [ + 304, + 458, + 545, + 569 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 458, + 545, + 569 + ], + "type": "text", + "content": ", and a text prompt, M3ID intervenes in the generative distribution by finding tokens that \"surprise\" the unconditioned VLM (i.e., the VLM without the image prompt). M3ID amplifies the conditioned directions that are not already predicted by the unconditioned model more as new tokens are generated by leveraging a progressively smaller " + }, + { + "bbox": [ + 304, + 458, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 304, + 458, + 545, + 569 + ], + "type": "text", + "content": ". In the example, the VLM assigns a high likelihood to the hallucinated object fridge, over-relying on its unconditioned language prior. Instead, M3ID assigns a high likelihood to toaster, which is present in the image." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 579, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 652 + ], + "type": "text", + "content": "plausible-sounding answers without factual basis, leading to potentially ungrounded or fabricated information [7]. Consequently, the community has been developing ever so complex and expensive alignment algorithms involving direct human supervision [17] and often brittle prompt engineering methods (\"model begging\") [28]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "In this work, we propose to investigate hallucinations in VLMs through a quantifiable measure of visual prompt dependency. We assess whether a model output is ungrounded with respect to a visual prompt by comparing its likelihood with the likelihood of generating the same output" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 222, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 222, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 222, + 712 + ], + "type": "text", + "content": "*Work done during an internship at AWS AI Labs." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14303" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "without the visual information. In general, our measure of visual prompt dependency quantifies the extent to which a token is generic or specific to a visual signal. Importantly, low visual prompt dependency does not necessarily imply that a token is hallucinated, as it may occur in linguistically essential elements like conjunctions and prepositions. Yet, unlike the notion of hallucinations, which requires human judgment, our prompt dependency measure is always well-defined even without any ground truth annotations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 181, + 289, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 289, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 289, + 266 + ], + "type": "text", + "content": "Our first contribution is to empirically demonstrate that the visual prompt dependency measure decreases as more tokens are generated. In Fig. 3, we show that, as more tokens are generated, the conditioning information gets diluted and \"forgotten\" or ignored by the model, possibly leading to more hallucinations. We refer to this effect as conditioning dilution or fading memory effect." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 266, + 289, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 458 + ], + "type": "text", + "content": "To counteract this, we propose an intervention on the generative distribution of a VLM that maximizes visual prompt dependency at inference time. We show that our method for prompt amplification maximizes the mutual information between the text output tokens and the visual prompt, effectively rescaling the image-conditioned component against the unconditioned distribution. We name our inference-time intervention on the generative distribution Multi-Modal Mutual Information Decoding (M3ID). M3ID is applicable to any off-the-shelf model without additional training or access to model weights, offering a low computational overhead alternative to standard decoding algorithms [13, 24]. Our results show that M3ID enhances the dependence on the visual prompt and reduces the number of hallucinations across various benchmarks while preserving the linguistic fluency of the original model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 289, + 579 + ], + "type": "text", + "content": "Additionally, for users with access to model weights, we propose a training objective to further ground the model outputs on visual prompts. This approach involves recasting our prompt amplification objective as a preference optimization problem, using Direct Preference Optimization (DPO) [19] to align the pre-trained VLM. This method aims to increase the likelihood of continuations with a higher visual prompt dependency measure and allows to learn a better generation policy, further reducing the conditioning dilution effect and hallucinations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 580, + 266, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 580, + 266, + 592 + ], + "spans": [ + { + "bbox": [ + 59, + 580, + 266, + 592 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 594, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 653 + ], + "type": "text", + "content": "1. We propose a visual prompt dependency measure (PDM) to assess whether a model output is ungrounded with respect to the visual input. Empirically, we demonstrate that PDM decreases as more tokens are generated making the generations more likely to be hallucinated." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "text", + "content": "2. We introduce M3ID, a training-free intervention on the generative distribution of autoregressive VLMs which improves visual grounding and reduces hallucinations by amplifying the importance of the visual prompt over the language prior. In addition, we extend M3ID with" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 95 + ], + "type": "text", + "content": "DPO for multi-modal preference optimization to further improve visual grounding." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 96, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 547, + 156 + ], + "type": "text", + "content": "3. We show that applying M3ID or DPO reduces the percentage of hallucinated objects in captioning tasks [20] by " + }, + { + "bbox": [ + 306, + 96, + 547, + 156 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 306, + 96, + 547, + 156 + ], + "type": "text", + "content": " and by " + }, + { + "bbox": [ + 306, + 96, + 547, + 156 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 306, + 96, + 547, + 156 + ], + "type": "text", + "content": ", respectively and improves accuracy on the POPE [9] VQA hallucination benchmark by " + }, + { + "bbox": [ + 306, + 96, + 547, + 156 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 306, + 96, + 547, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 96, + 547, + 156 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 306, + 96, + 547, + 156 + ], + "type": "text", + "content": " over the base model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 167, + 390, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 390, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 390, + 180 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 193, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 547, + 540 + ], + "type": "text", + "content": "Hallucinations in VLMs. A recent line of work introduced VLMs obtained by grafting visual encoders into pre-trained Large Language Models (LLMs) [4, 13, 32]. Forcing a pre-trained LLM to learn to \"see\" by reading vision tokens from a pre-trained vision backbone has proven successful and the resulting models show remarkable vision-language understanding capabilities on many multi-modal tasks [13, 23]. However, while inheriting strong linguistic capabilities and fluency from their base LLM, grafted VLMs also inherit the tendency to produce ungrounded or fabricated information [1, 6]. This is commonly referred to as \"hallucination\" in the recent machine learning literature [1, 6, 25]. In particular, several recent works empirically reported a sharp tendency of grafted VLMs to report objects not grounded on the visual information when probed with questions on the image content [1, 6, 9, 27]. The authors of [9] observed that this phenomenon is especially pronounced with objects that are either common or frequently appear together in the datasets used at training time. Furthermore, the authors of [11] suggest that VLMs often fail to correctly follow instructions involving absent objects and propose a new instruction-following training objective to enhance model alignment and robustness in the face of uncertainty. Parallel to our work, [31] delves into the factors contributing to object hallucinations in VLMs, exploring aspects like object co-occurrence, model uncertainty, and spatial position of hallucinations in the sentence. They propose a post-hoc algorithm to identify and correct hallucinations in VLM-generated content." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "Context-dependent decodings. Decoding algorithms can be classified as either search or sampling algorithms [8]. Current search methods (e.g., greedy and beam search) can produce factually accurate generations but usually suffer from tedious and repetitive continuations. Sampling methods, on the other hand, (like nucleus [5], or typical decoding [15]) produce more diverse text, but suffer from topic drift. To solve these shortcomings, various decoding methods have been proposed to enhance reasoning and factual accuracy in generative autoregressive language models. Approaches such as context-aware decoding [21] and Pointwise Mutual Information (PMI) decoding [16, 25] focus on aligning the outputs of language models with the intended context or factual data. These strategies aim to reduce hal" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14304" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 284, + 210 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 284, + 210 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 284, + 210 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 284, + 210 + ], + "type": "image", + "image_path": "2e23e32c7b76705df7ff11bf12fb56e3da31459f752a9f5ffbf8c1db06c867ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 289, + 324 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 289, + 324 + ], + "type": "text", + "content": "Figure 2. How much does the conditioning prompt \"surprise\" the unconditioned predictor? We report the likelihood that the conditioned and unconditioned models assign to each token in the string \"The image ...\" The probability gap increases on tokens for which visual information is required (e.g. objects and attributes), while it decreases for punctuation and articles. Note that both models mostly agree when sufficient visual information has already been extracted and is present in the caption (e.g. \"kitchen\" is very likely for the unconditioned model as well). See Sec. D.1 for a plot with our prompt dependency measure." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 335, + 289, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 289, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 289, + 444 + ], + "type": "text", + "content": "lucinations in standard natural language processing tasks such as summarization. Mutual-information-based decoding techniques have proven to be helpful in steering the generation of LLMs to stay faithful to the given input text [25], or to promote diversity or relevance in neural dialogue models [7]. Our work is the first to use mutual information to improve multi-modal grounding and to vary the penalization of marginally likely outputs in a time-varying fashion to counteract the progressive forgetting of the visual prompt." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 453, + 243, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 243, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 243, + 466 + ], + "type": "text", + "content": "3. Analysis of hallucinations in VLMs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 556 + ], + "type": "text", + "content": "We start by investigating hallucinations in VLMs. We introduce a visual prompt dependency measure (PDM), which will later motivate our algorithms (Sec. 4). The main aim of this section is to assess whether a model output, such as a token, is hallucinated or not by comparing its likelihood with the likelihood of that same output being generated without the relevant conditioning factor, in this case, the image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "We consider decoding approaches for open-ended language generation of VLMs, where the model receives an input image and an input textual prompt and aims to generate a fluent and coherent continuation conditioned on both the image and the text. We denote probabilistic generative VLMs over a vocabulary of text tokens " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " and denote the probability of a token " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "y\\in \\mathcal{V}" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " given a textual prompt " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " and an image context " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "p(y|\\mathbf{x},c)" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": ". The goal of the VLM is to leverage the information contained in the input image " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " to provide a continuation " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = [y_0,\\dots ,y_T]" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " of the input prompt " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": ", e.g., \"Describe this image in detail\". The probability of a valid sequence " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " can be computed as " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y}|\\mathbf{x},c) = \\prod_{t = 1}^{T}p(y_{t}|\\mathbf{y}_{< t},\\mathbf{x},c)" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{< t}\\triangleq [y_0,\\dots ,y_{t - 1}]" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 425, + 157 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 425, + 157 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 425, + 157 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 425, + 157 + ], + "type": "image", + "image_path": "a4798a083b13589972e8daec3c766f035778c91c998e9c02277996eb41b63b83.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 169, + 547, + 324 + ], + "lines": [ + { + "bbox": [ + 304, + 169, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 547, + 324 + ], + "type": "text", + "content": "Figure 3. The influence of the image conditioning decreases as we generate more tokens. (Left) Conditioning dilution. We report the average of the prompt dependency measure (PDM-H) and its standard deviation for different synthetic captions generated by LLaVA on MS COCO's validation split. We see that the influence of the image over the next token prediction decreases as we generate more. This suggests that the information within VLM's visual prompt gets diluted and fades away as more tokens are generated. (Right) Frequency of hallucinated objects as a function of the token position. We report the number of non-existent objects present on the same synthetic captions as a function of the number of generated tokens. Note that very few objects are hallucinated for tokens near the visual prompt, while their number increases as more tokens are generated and with a smaller PDM." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 542, + 158 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 542, + 158 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 542, + 158 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 542, + 158 + ], + "type": "image", + "image_path": "bddc9be237dcad3e5725c9e6b78dde8a00d2157155c044e6c59eb9a564d34428.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 335, + 546, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 546, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 546, + 358 + ], + "type": "text", + "content": "We propose to study hallucinations on VLMs using PDMs defined as follows" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 364, + 531, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 364, + 531, + 384 + ], + "spans": [ + { + "bbox": [ + 319, + 364, + 531, + 384 + ], + "type": "interline_equation", + "content": "\\mathrm {P D M} (\\mathbf {y} _ {< t}; c | \\mathbf {x}) \\triangleq \\operatorname {d i s t} \\left(p (\\cdot | \\mathbf {y} _ {< t}, \\mathbf {x}, c), p (\\cdot | \\mathbf {y} _ {< t}, \\mathbf {x})\\right)", + "image_path": "f494b13afca7ff8efa2c0f1b70468d44d047d19702dd70bc07ea4a01bf5dc1b8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 391, + 546, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 546, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 546, + 566 + ], + "type": "text", + "content": "where dist is any distance measure between probability distributions, such as Hellinger, total variation, or KL. PDMs quantify how generic or context-specific a language model's output is. In particular, a high " + }, + { + "bbox": [ + 304, + 391, + 546, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{PDM}(\\mathbf{y}_{< t};c|\\mathbf{x})" + }, + { + "bbox": [ + 304, + 391, + 546, + 566 + ], + "type": "text", + "content": " indicates that the token " + }, + { + "bbox": [ + 304, + 391, + 546, + 566 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 304, + 391, + 546, + 566 + ], + "type": "text", + "content": " is strongly associated with a specific input prompt, while a low PDM suggests that the token is more prompt-neutral or prompt-agnostic. Depending on the choice of the distance function, PDMs highlight different aspects of the generative distribution. We will mainly use PDM-H based on the Hellinger distance defined as " + }, + { + "bbox": [ + 304, + 391, + 546, + 566 + ], + "type": "inline_equation", + "content": "H(p,q) = 2^{-1 / 2}\\sqrt{\\sum_{i = 1}^{k}(\\sqrt{p_i} - \\sqrt{q_i})^2}" + }, + { + "bbox": [ + 304, + 391, + 546, + 566 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 391, + 546, + 566 + ], + "type": "inline_equation", + "content": "p = (p_i)_{i\\in [k]}" + }, + { + "bbox": [ + 304, + 391, + 546, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 391, + 546, + 566 + ], + "type": "inline_equation", + "content": "q = (q_i)_{i\\in [k]}" + }, + { + "bbox": [ + 304, + 391, + 546, + 566 + ], + "type": "text", + "content": " are discrete probability distributions. We refer the reader to the Supplementary Material for an analysis of the impact of other distances." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "Contextual pressure. Fig. 2 shows the conditional and unconditional prediction likelihoods on a single caption. First, we note that tokens that are required for linguistic fluency, like prepositions and conjunctions, are highly predictable by both the conditioned and the unconditioned models. This pattern also emerges when the model generates tokens to compose \"fine-grained\" objects. For example, when the model generates \"Peanut butter\", both the VLM and the LLM can correctly predict the last token even without looking at the image and only having access to a sufficiently long truncation (e.g. \"Peanut bu\"). Therefore, in these circumstances, " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "p(y_{t}|\\mathbf{y}_{CaptioningCHAIRi ↓CHAIRs ↓Cover ↑LLaVA7B8.117.553.3LLaVA7B PMI [25]6.716.251.5LLaVA7B Contrastive [8]6.314.855.2LLaVA7B M3ID5.913.855.1LLaVA13B7.418.555.2LLaVA13B PMI [25]6.516.654.7LLaVA13B Contrastive [8]6.514.554.7LLaVA13B M3ID5.513.254.0LLaVA13B + LURE [31]6.427.1-mPLUG-Owl + LURE [31]5.418.8-LLaVA7B M3ID + DPO5.713.555.8LLaVA13B M3ID + DPO5.312.654.2", + "image_path": "c18823db2d36b0425e30aa79e354067961278236921f3ed377dec8e774604001.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 331, + 287, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 547 + ], + "type": "text", + "content": "eration by increasing the likelihood of generating tokens that are related to the text to be summarized. Differently from our method, the weight assigned to the realization of marginally likely continuations is time-independent. Similarly, Contrastive Decoding [7, 8] has been developed to foster generations of an expert LLM to contain plausible and fluent text with the textual input prompt by amplifying its prediction difference with respect to a weaker model. To ensure a fair comparison, we replace the text to be summarized with the input image for PMI and set the weaker model in Contrastive Decoding as the unconditioned model (see Sec. B for more details). We also compare with training-based methods that have been proposed to improve the grounding of VLMs on the visual prompt like LLaVARLHF [17], Robust mPLUG-Owl [11] and LURE [31]. However, all these methods require some form of annotations, like preference data [17] or positive/negative examples [11, 31], which M3ID+DPO does not require." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 555, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 676 + ], + "type": "text", + "content": "Evaluation. We evaluate our algorithm both on captioning and VQA benchmarks. On captioning, we measure object hallucinations by using the CHAIR (Captioning Hallucination Assessment with Image Relevance) metrics [20]. These metrics assess the quality of captions by comparing the mentioned objects to the annotated objects present in an image. In particular, CHAIR comprises two variants: one computes the fraction of hallucinated objects in the whole caption (CHAIRi) and the other evaluates the fraction of captions with at least one object hallucination (CHAIRs)," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 690, + 240, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 690, + 240, + 717 + ], + "spans": [ + { + "bbox": [ + 94, + 690, + 240, + 717 + ], + "type": "interline_equation", + "content": "\\mathrm {C H A I R i} = \\frac {\\# \\text {h a l l u c i n a t e d o b j e c t s}}{\\# \\text {g e n e r a t e d o b j e c t s}},", + "image_path": "7fa626b8dab50f3b28e44fd65cfbbbd4ed2381e833a58dc655e3ee27a93c7849.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 350, + 69, + 501, + 95 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 69, + 501, + 95 + ], + "spans": [ + { + "bbox": [ + 350, + 69, + 501, + 95 + ], + "type": "interline_equation", + "content": "\\text {C H A I R s} = \\frac {\\# \\text {h a l l u c i n a t e d c a p t i o n s}}{\\# \\text {g e n e r a t e d c a p t i o n s}}.", + "image_path": "c046247fa1175c607ea8efa49a8807147b07f112e80cf4b91eab6becab77cdf2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 99, + 545, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 99, + 545, + 146 + ], + "spans": [ + { + "bbox": [ + 305, + 99, + 545, + 146 + ], + "type": "text", + "content": "Furthermore, we introduce the Cover metric to quantify the comprehensiveness of the captions by measuring the fraction of ground-truth annotated objects mentioned by the model," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 342, + 154, + 508, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 154, + 508, + 182 + ], + "spans": [ + { + "bbox": [ + 342, + 154, + 508, + 182 + ], + "type": "interline_equation", + "content": "\\text {C o v e r} = \\frac {\\left| \\{\\text {c o r r e c t m e n t i o n e d o b j e c t s} \\} \\right|}{\\left| \\{\\text {a n n o t a t e d o b j e c t s} \\} \\right|}.", + "image_path": "abc0eef6137e0c14b583d27e6b181a068b92a16723b13850b6cfe1c196a73fe4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 188, + 545, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 188, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 305, + 188, + 545, + 249 + ], + "type": "text", + "content": "To evaluate object perception, we use the POPE (Polling-based Object Probing Evaluation) VQA benchmark [9]. In particular, POPE recasts the evaluation of object hallucinations into a binary classification problem with yes/no questions of the type \"Is a present in the image?\"" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 253, + 545, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 253, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 305, + 253, + 545, + 313 + ], + "type": "text", + "content": "DPO training details. We use Hugging Face's DPO implementation in the TRL library [26] and train LLaVA for 5 epochs on 10,000 self-generated preference pairs with LoRA and cosine decaying learning rate with " + }, + { + "bbox": [ + 305, + 253, + 545, + 313 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 305, + 253, + 545, + 313 + ], + "type": "text", + "content": " peak. Following [19], we set " + }, + { + "bbox": [ + 305, + 253, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\beta = 0.1" + }, + { + "bbox": [ + 305, + 253, + 545, + 313 + ], + "type": "text", + "content": " in the DPO loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 320, + 471, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 320, + 471, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 471, + 334 + ], + "type": "text", + "content": "5.1. VLM grounding on captioning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 338, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 409 + ], + "type": "text", + "content": "In Tab. 1 we test how much our method can prevent the generation of hallucinated objects in image captions. In particular, we compare against training-free and training-based baselines. We prompt all baseline methods with the instruction \"Describe the image.\" and let the models generate until the EOS token is obtained." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 411, + 546, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 546, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 546, + 662 + ], + "type": "text", + "content": "First, we compare M3ID with other decoding strategies, PMI [25] and Contrastive Decoding [8]. The main difference between M3ID and these baselines is that M3ID increasingly counteracts the language prior as more tokens are generated. As such, M3ID reduces ungrounded generations compared to all other training-free baselines both on the large LLaVA13B and on the smaller LLaVA7B. Furthermore, it has uniformly good improvements for different model sizes over standard multinomial decoding, in particular, M3ID achieves " + }, + { + "bbox": [ + 304, + 411, + 546, + 662 + ], + "type": "inline_equation", + "content": "27\\% / 21\\%" + }, + { + "bbox": [ + 304, + 411, + 546, + 662 + ], + "type": "text", + "content": " relative improvement over LLaVA7B and " + }, + { + "bbox": [ + 304, + 411, + 546, + 662 + ], + "type": "inline_equation", + "content": "26\\% / 29\\%" + }, + { + "bbox": [ + 304, + 411, + 546, + 662 + ], + "type": "text", + "content": " over LLaVA13B on the CHAIRi and CHAIRs metrics. Importantly, this improvement does not come at the price of high reductions of the Cover metric, which actually improves on the 7B model and decreases by less than " + }, + { + "bbox": [ + 304, + 411, + 546, + 662 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 304, + 411, + 546, + 662 + ], + "type": "text", + "content": " for the larger 13B model1. Lastly, we note that M3ID shows an improvement in absolute performance that correlates with model size, suggesting that further gains could be obtained as larger models are used. In Sec. D.2, we also show that M3ID still improves grounding when longer captions (twice as long on average) are generated and where other decodings do not perform as well." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 662, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 662, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 305, + 662, + 545, + 686 + ], + "type": "text", + "content": "In Tab. 1 we also report training-based results and show that pairing DPO with M3ID leads to a smaller number" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1Note that CHAIR metrics can be hacked by simply returning shorter captions that do not attempt to predict any object in the image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14308" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 135, + 541, + 338 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "type": "text", + "content": "Table 2. Evaluation on the POPE VQA hallucination benchmark [9]. *;† indicate results taken from [11] and [23] respectively. SUP. Indicates that the method involves supervised training on annotated data, in contrast to our approaches. POPE is divided into three subsets, Random, Popular, and Adversarial. We report the VQA accuracy of the model using the following template: “Is a present in the image?”, where we sample randomly (Random), among the most frequent objects in the dataset (Popular), or among the objects that frequently co-occur with (Adversarial). Acc. is the binary classification accuracy and Yes is the percentage of “Yes” answers." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 135, + 541, + 338 + ], + "lines": [ + { + "bbox": [ + 49, + 135, + 541, + 338 + ], + "spans": [ + { + "bbox": [ + 49, + 135, + 541, + 338 + ], + "type": "table", + "html": "
POPE
RandomPopularAdversarialAll
Acc. ↑Yes (%)Acc. ↑Yes (%)Acc. ↑Yes (%)Acc. ↑Yes (%)
Robust mPLUG-Owl-7B* [11]sup.LLaVA-RLHF†7B [23]sup.86.0-73.0-65.0-74.7-
84.839.683.341.880.744.082.941.8
mPLUG-Owl-7B* [29]52.0-57.0-60.0-67.3-
LLaVA7B74.875.161.886.758.190.164.984.0
LLaVA7B M3ID76.067.769.373.365.877.670.372.9
LLaVA7B M3ID + DPO81.265.673.967.368.275.474.469.4
LLaVA-RLHF†13B [23]sup.85.238.483.938.082.340.583.839.0
MiniGPT4-13B* [32]73.0-67.0-62.0-74.7-
LLaVA13B67.980.663.883.259.887.363.883.7
LLaVA13B M3ID84.355.677.061.671.368.277.561.8
LLaVA13B M3ID + DPO85.253.479.157.573.267.579.251.1
", + "image_path": "ed56b1b58a63c0e0640907694d9a3c01d8ac2a5e88fd0393c799bce90642287a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 347, + 288, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 288, + 396 + ], + "type": "text", + "content": "of hallucinated objects and improved Cover numbers compared both to our training-free approach and LLaVA+LURE [31] a concurrent training based method to improve the grounding of VLMs that relies on GPT-3.5 annotations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 402, + 189, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 402, + 189, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 189, + 415 + ], + "type": "text", + "content": "5.2. VLM grounding on VQA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 420, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 540 + ], + "type": "text", + "content": "Differently from the MS COCO captioning task, the POPE VQA benchmark only requires the generation of \"Yes/No\" tokens. As such, one does not expect the dilution effect to play a key role. However, we highlight that the image tokens and the output of the VLM are separated by the input template used to prime the model for the VQA task, which, as we show in Sec. D.4, introduces a non-negligible dilution effect. To address this, we take an offset into account when using M3ID on POPE, and select " + }, + { + "bbox": [ + 46, + 420, + 287, + 540 + ], + "type": "inline_equation", + "content": "t = t_0" + }, + { + "bbox": [ + 46, + 420, + 287, + 540 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 420, + 287, + 540 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 46, + 420, + 287, + 540 + ], + "type": "text", + "content": " is the number of tokens in between the output and the image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 540, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 671 + ], + "type": "text", + "content": "In Tab. 2, we report the results on the POPE VQA hallucination benchmark [9]. Hallucinations (wrong answers) tend to correlate with the tendency of the VLM to reply using the \"Yes\" token (see the disproportionately high percentage rate, " + }, + { + "bbox": [ + 46, + 540, + 287, + 671 + ], + "type": "inline_equation", + "content": "84\\% / 83.7\\%" + }, + { + "bbox": [ + 46, + 540, + 287, + 671 + ], + "type": "text", + "content": ", of \"Yes\" answers for the LLaVAv1.3 base model). This tendency can be counteracted at inference time and without any training by simply using M3ID. In our experiments, M3ID reduces the Yes ratio to " + }, + { + "bbox": [ + 46, + 540, + 287, + 671 + ], + "type": "inline_equation", + "content": "72.9\\% / 61.8\\%" + }, + { + "bbox": [ + 46, + 540, + 287, + 671 + ], + "type": "text", + "content": " for 7B and 13B models, respectively, which leads to relative accuracy improvements over standard LLaVA decoding by " + }, + { + "bbox": [ + 46, + 540, + 287, + 671 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 46, + 540, + 287, + 671 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 540, + 287, + 671 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 46, + 540, + 287, + 671 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 671, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 671, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 671, + 287, + 696 + ], + "type": "text", + "content": "To test whether our training-based approach increases the VLM's grounding on the visual prompt we trained a" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 347, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 347, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 547, + 574 + ], + "type": "text", + "content": "model with our DPO objective on the MS COCO captioning dataset and then tested it on the POPE benchmark. Note that improving caption generation does not directly imply improvements on the POPE VQA benchmark, in fact, even if the images are both from MS COCO, the output format is quite different: open-ended generation on the former and binary classification on the latter. However, intuitively, one expects that by improving the reliance on the visual prompt, the model could make better use of the visual information regardless of the output format required to solve the task at hand. Tab. 2 shows that M3ID+DPO further improves performance over M3ID's inference time intervention. Specifically, M3ID+DPO achieves " + }, + { + "bbox": [ + 306, + 347, + 547, + 574 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 306, + 347, + 547, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 347, + 547, + 574 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 306, + 347, + 547, + 574 + ], + "type": "text", + "content": " accuracy improvements over the LLaVA 7B and 13B models respectively. For completeness, we also compare with other training-based baselines that are fine-tuned on labeled [11] and preference data [17]. While M3ID+DPO does not have access to any labeled information we show that it is close to these baselines without requiring any annotations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 582, + 373, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 373, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 373, + 594 + ], + "type": "text", + "content": "5.3. Ablations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 547, + 715 + ], + "type": "text", + "content": "Conditioning dilution and \"overcompensation\". In Fig. 4 and Tab. 3 we show that when the forgetting factor " + }, + { + "bbox": [ + 305, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 605, + 547, + 715 + ], + "type": "text", + "content": " is high, corresponding to the assumption that the input image gets forgotten quickly, M3ID effectively maintains a large PDM-H throughout the generation. However, while deviating from the unconditioned probability is often a desired behavior, an excessive deviation from it across the whole generation could result in \"overcompensation\", which, as we show in Tab. 3 can be counterproductive and can lead to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 274, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 274, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 274, + 714 + ], + "type": "text", + "content": "2We use [Img] [Model Prompt] [Question], see Sec. D.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14309" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 165, + 157 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 165, + 157 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 165, + 157 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 165, + 157 + ], + "type": "image", + "image_path": "0e1f2967b7693e9ee7f78ba4cd1ce0834178b64515ac247f7c9b52bfe9d53f7a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "type": "text", + "content": "Figure 4. Sensitivity analysis on hyper-parameters. We vary both the forgetting factor " + }, + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "type": "text", + "content": " and the thresholding parameter " + }, + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "type": "text", + "content": ". Small " + }, + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "type": "text", + "content": " result in minor effects. Instead, either a high forgetting factor " + }, + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "type": "text", + "content": " or a high " + }, + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 169, + 287, + 224 + ], + "type": "text", + "content": " results in a stronger intervention which leads to higher PDM-H." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 72, + 284, + 157 + ], + "blocks": [ + { + "bbox": [ + 168, + 72, + 284, + 157 + ], + "lines": [ + { + "bbox": [ + 168, + 72, + 284, + 157 + ], + "spans": [ + { + "bbox": [ + 168, + 72, + 284, + 157 + ], + "type": "image", + "image_path": "da35ae2648ffb20772286d7e259bfff710fadbe0823dcf0c22df29a15592a3a9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 232, + 287, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 287, + 364 + ], + "type": "text", + "content": "higher hallucination rates. In particular, observe that higher corrections detrimentally impact the cover metric. This suggests that while M3ID tries to find tokens that diverge from the language prior the most, it is more likely to fail in captioning elements that are inherently predictable by the language prior alone. For instance, in describing an image of a dog on a leash accompanied by a man, M3ID might overlook mentioning the presence of the man, a token that the language prior could anticipate without necessitating any visual information only from context clues. We report examples of this behavior in the Sec. D.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 537 + ], + "type": "text", + "content": "Importance of the confidence threshold. When we set a high threshold " + }, + { + "bbox": [ + 46, + 369, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 369, + 287, + 537 + ], + "type": "text", + "content": ", the indicator function remains active more frequently along the generation, leading M3ID to consistently prioritize maximizing the distance of the generation from the language prior. Hence, similarly to the scenario with a high " + }, + { + "bbox": [ + 46, + 369, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 369, + 287, + 537 + ], + "type": "text", + "content": ", M3ID is likely to overcompensate and increase the rate of multi-modal hallucinations. However, too high " + }, + { + "bbox": [ + 46, + 369, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 369, + 287, + 537 + ], + "type": "text", + "content": " disrupts linguistic fluency. We report qualitative examples of this in Sec. D.2. In Tab. 4 we also show that adding the term to compensate for contextual pressure does not significantly change the results over LLaVA decoding in the POPE benchmark since it is essentially framed as a binary classification over the \"Yes\"/ \"No\" tokens and not open-ended generation like captioning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 549, + 125, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 125, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 125, + 562 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": "We introduced M3ID, a new approach designed to combat multi-modal hallucinations by maximizing the mutual information between the text generated by VLMs and the corresponding visual context. M3ID operates at inference time and can be seamlessly integrated with any pre-trained autoregressive VLM. This makes M3ID a cost-effective and flexible solution to enhance vision-language grounding. Furthermore, for settings where model training is feasible and higher visual grounding is expected, we also paired M3ID with Direct Preference Optimization (DPO), showcasing reduced hallucinatory behaviors. Interestingly, our findings suggest that object hallucinations in VLMs result" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 121, + 544, + 245 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "type": "text", + "content": "Table 3. Sensitivity to hyper-parameters. " + }, + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "type": "text", + "content": " is the forgetting factor used to counteract conditioning dilution while " + }, + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "type": "text", + "content": " is the threshold used to mitigate overcompensation when the contextual pressure is high. We report results on LLaVA-7B." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 121, + 544, + 245 + ], + "lines": [ + { + "bbox": [ + 309, + 121, + 544, + 245 + ], + "spans": [ + { + "bbox": [ + 309, + 121, + 544, + 245 + ], + "type": "table", + "html": "
Captioning
CHAIRi ↓CHAIRs ↓Cover ↑
LLaVA7B8.017.853.5
M3ID α = 0.3, λ = 0.025.813.655.0
M3ID α = 0.3, λ = 0.0017.317.453.7
M3ID α = 0.3, λ = 0.036.214.752.0
M3ID α = 0.3, λ = 0.17.216.345.5
M3ID α = 0.5, λ = 0.026.114.954.6
M3ID α = 0.1, λ = 0.026.915.453.1
M3ID α = 0.01, λ = 0.026.916.652.5
", + "image_path": "8125663bf5adccdee9e5c4cad2eefbe263dcf4c6e3b3cce3901b0a5b0e09a369.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 293, + 544, + 395 + ], + "blocks": [ + { + "bbox": [ + 305, + 253, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 305, + 253, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 305, + 253, + 545, + 286 + ], + "type": "text", + "content": "Table 4. Ablation studies. We ablate the components of M3ID on both model sizes. Removing the amplification term to counteract conditioning dilution leads to more hallucinations." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 293, + 544, + 395 + ], + "lines": [ + { + "bbox": [ + 308, + 293, + 544, + 395 + ], + "spans": [ + { + "bbox": [ + 308, + 293, + 544, + 395 + ], + "type": "table", + "html": "
POPE\nAccuracy ↑CHAIRi ↓Captioning\nCHAIRs ↓Cover ↑
LLaVA7B64.98.017.853.5
w/ context pressure65.56.916.754.5
w/ conditioning dilution77.56.414.153.9
LLaVA7B M3ID77.55.813.655.0
LLaVA13B63.815.318.255.3
w/ context pressure64.96.514.455.0
w/ conditioning dilution70.35.914.853.7
LLaVA13B M3ID70.35.413.054.0
", + "image_path": "74a4789cfbdb90bf6d8b67a23ed3d583bcc033a120a64b503ee4222a3feae6da.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 402, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 450 + ], + "type": "text", + "content": "from excessive reliance on the language prior rather than a poor understanding of the visual modality. In fact, M3ID at inference time is sufficient to significantly reduce the amount of generated hallucinations without any training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 450, + 546, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 546, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 546, + 677 + ], + "type": "text", + "content": "A limitation of M3ID is that it requires two forward passes at inference time, one for the conditioned and one for the unconditioned prediction. A possible solution to not increase inference time, but at the expense of higher memory consumption, is to use two batched queries with one having masked visual tokens. Furthermore, as we observed in Tab. 1, sometimes M3ID may prevent the generation of objects that are highly likely under the unprompted language prior (due to context clues). Interestingly, a similar observation has been reported in [22], where, people asked to provide a 10-word list of objects contained in a given image often failed to report the most obvious objects while mainly focusing on secondary ones. While we already showed that this can be mitigated with proper hyperparameter selection Tab. 3 we believe that an interesting avenue for future research is to directly encode this preference within the model's weights by favoring the generation of structured captions that get progressively more detailed while still being grounded to the image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "In general, integrating human- or AI-annotated preference pairs, assessed based on their level of grounding, constitutes a promising avenue for future investigation." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14310" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, et al. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 179 + ], + "type": "text", + "content": "[2] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 245 + ], + "type": "text", + "content": "[3] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 53, + 180, + 288, + 245 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 53, + 180, + 288, + 245 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 288, + 312 + ], + "type": "text", + "content": "[4] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 5, 16" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 313, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 356 + ], + "type": "text", + "content": "[5] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "type": "text", + "content": "[6] Yichong Huang, Xiachong Feng, Xiaocheng Feng, and Bing Qin. The factual inconsistency problem in abstractive text summarization: A survey. arXiv preprint arXiv:2104.14839, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 479 + ], + "type": "text", + "content": "[7] Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110–119, San Diego, California, 2016. Association for Computational Linguistics. 1, 3, 4, 6, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 480, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 524 + ], + "type": "text", + "content": "[8] Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. Contrastive decoding: Open-ended text generation as optimization, 2023. 2, 5, 6, 11, 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 525, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 288, + 568 + ], + "type": "text", + "content": "[9] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 2, 6, 7, 11, 17" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "type": "text", + "content": "[10] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 4, 5, 6, 11, 14, 15" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 679 + ], + "type": "text", + "content": "[11] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[12] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[13] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 2, 4, 5, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 106, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 547, + 160 + ], + "type": "text", + "content": "[14] M.B. Matthews and G.S. Moschytz. The identification of nonlinear discrete-time fading-memory systems using neural network models. IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing, 41(11):740-751, 1994. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 161, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 161, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 161, + 545, + 195 + ], + "type": "text", + "content": "[15] Clara Meister, Tiago Pimentel, Gian Wiher, and Ryan Coterell. Locally typical sampling. Transactions of the Association for Computational Linguistics, 11:102-121, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "type": "text", + "content": "[16] Yatin Nandwani, Vineet Kumar, Dinesh Raghu, Sachindra Joshi, and Luis A Lastras. Pointwise mutual information based metric and decoding strategy for faithful generation in document grounded dialogs. arXiv preprint arXiv:2305.12191, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 250, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 316 + ], + "type": "text", + "content": "[17] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 1, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 316, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 381 + ], + "type": "text", + "content": "[18] 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. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "type": "text", + "content": "[19] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. arXiv preprint arXiv:2305.18290, 2023. 2, 5, 6, 11, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 437, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 503 + ], + "type": "text", + "content": "[20] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4035-4045, Brussels, Belgium, 2018. Association for Computational Linguistics. 2, 6, 11, 14, 17" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 503, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 545, + 548 + ], + "type": "text", + "content": "[21] Weijia Shi, Xiaochuang Han, Mike Lewis, Yulia Tsvetkov, Luke Zettlemoyer, and Scott Wen-tau Yih. Trusting your evidence: Hallucinate less with context-aware decoding. arXiv preprint arXiv:2305.14739, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 548, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 580 + ], + "type": "text", + "content": "[22] Merrielle Spain and Pietro Perona. Measuring and predicting object importance. Int. J. Comput. Vis., 91(1):59-76, 2011. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 581, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 545, + 635 + ], + "type": "text", + "content": "[23] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 636, + 545, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 691 + ], + "type": "text", + "content": "[24] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[25] Liam van der Poel, Ryan Cotterell, and Clara Meister. Mutual information alleviates hallucinations in abstractive sum" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14311" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 475 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "text", + "content": "marization. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5956-5965, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 2, 3, 4, 5, 6, 11, 14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[26] Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, and Shengyi Huang. Trl: Transformer reinforcement learning. https://github.com/huggingface/trl, 2020.6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[27] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, et al. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 273 + ], + "type": "text", + "content": "[28] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 328 + ], + "type": "text", + "content": "[29] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "text", + "content": "[30] Luca Zancato and Alessandro Chiuso. A novel deep neural network architecture for non-linear system identification. IFAC-PapersOnLine, 54(7):186-191, 2021. 19th IFAC Symposium on System Identification SYSID 2021. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 430 + ], + "type": "text", + "content": "[31] Yiyang Zhou, Chenhang Cui, Jaehong Yoon, Linjun Zhang, Zhun Deng, Chelsea Finn, Mohit Bansal, and Huaxiu Yao. Analyzing and mitigating object hallucination in large vision-language models. arXiv preprint arXiv:2310.00754, 2023. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 431, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 287, + 475 + ], + "type": "text", + "content": "[32] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 7" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14312" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/f9afe157-eaaf-440c-afcf-7604121ce960_content_list.json b/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/f9afe157-eaaf-440c-afcf-7604121ce960_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..27de6a5e7d07e8c950ca02ed397af1a568909ca5 --- /dev/null +++ b/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/f9afe157-eaaf-440c-afcf-7604121ce960_content_list.json @@ -0,0 +1,1601 @@ +[ + { + "type": "text", + "text": "Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering", + "text_level": 1, + "bbox": [ + 86, + 130, + 883, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiawei Yao $^{1}$ Qi Qian $^{2}$ Juhua Hu $^{1*}$", + "bbox": [ + 334, + 179, + 640, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Engineering and Technology, University of Washington, Tacoma, WA 98402, USA \n $^{2}$ Alibaba Group, Bellevue, WA 98004, USA", + "bbox": [ + 109, + 198, + 854, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jwyao, juhuah}@uw.edu, qi.qian@alibaba-inc.com", + "bbox": [ + 274, + 234, + 691, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multiple clustering has gained significant attention in recent years due to its potential to reveal multiple hidden structures of data from different perspectives. The advent of deep multiple clustering techniques has notably advanced the performance by uncovering complex patterns and relationships within large datasets. However, a major challenge arises as users often do not need all the clusterings that algorithms generate, and figuring out the one needed requires a substantial understanding of each clustering result. Traditionally, aligning a user's brief keyword of interest with the corresponding vision components was challenging, but the emergence of multi-modal and large language models (LLMs) has begun to bridge this gap. In response, given unlabeled target visual data, we propose Multi-MaP, a novel method employing a multi-modal proxy learning process. It leverages CLIP encoders to extract coherent text and image embeddings, with GPT-4 integrating users' interests to formulate effective textual contexts. Moreover, reference word constraint and concept-level constraint are designed to learn the optimal text proxy according to the user's interest. Multi-MaP not only adeptly captures a user's interest via a keyword but also facilitates identifying relevant clusterings. Our extensive experiments show that MultiMaP consistently outperforms state-of-the-art methods in all benchmark multi-clustering vision tasks. Our code is available at https://github.com/Alexander-Yao/Multi-MaP.", + "bbox": [ + 76, + 318, + 473, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 729, + 209, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Clustering, which groups data points based on their similarities, has been extensively researched, since huge amount of unlabeled data are becoming more and more available. Traditional methods [2, 16, 19] exploit general-purpose handcrafted features that are not always ideal for specific tasks. Recently, deep clustering algorithms [9, 23, 33] leverage Deep Neural Networks (DNNs) significantly improve performance. However, most algorithms yield a single data", + "bbox": [ + 76, + 755, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2cba1582001304a37fce6343670486d636254288166348832ff61b13105129c.jpg", + "image_caption": [ + "Figure 1. The flow chart of Multi-MaP. Multi-MaP obtains multiple clustering results based on the high-level concepts from users and the reference words from GPT-4." + ], + "image_footnote": [], + "bbox": [ + 501, + 299, + 893, + 462 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "partition, while data can exhibit multiple aspects (e.g., color and species as shown in Fig. 1). Traditional multiple clustering algorithms [1, 12] have been developed to generate different partitions for varying applications, demonstrating the ability to identify multiple distinct clusterings from a dataset. Contemporary advancements in the field reveal a growing inclination among researchers to integrate deep learning methodologies for facilitating multiple clustering outcomes. Predominantly, such techniques capitalize on auto-encoders and data augmentation processes to capture a broad spectrum of feature dimensions, thereby enhancing the performance of multiple clustering [18, 27, 35].", + "bbox": [ + 496, + 520, + 893, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, a common issue arises as users often do not require all the clusterings generated by the algorithm, and identifying the relevant ones necessitates a substantial understanding of each clustering result. Therefore, in this work, we initiate an exploration into a method that is adept at accurately capturing and reflecting a user's interest. Users typically express their interests through concise keywords (e.g., color or species), and aligning these with different visual components precisely is challenging. Fortunately, the advent of multi-modal models like CLIP [25] that aligns images to their corresponding text descriptions, can be helpful to fill this gap. However, unlike methods that employ labeled data to fine-tune pre-trained models [6, 30], multiple", + "bbox": [ + 496, + 704, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14066", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "clustering deals with environments marked by vague or undefined label categories and amounts. Consequently, given only a high-level concept from the user, it is infeasible to fine-tune the pre-trained models to capture a specific aspect of the data, without the detailed labels corresponding to the user's concept.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An intuitive strategy to integrate pre-trained models into clustering is the zero-shot feature extraction, followed by clustering of the resultant embeddings. However, this approach exhibits limitations, particularly in capturing the interests of users within the dataset. Taking the multimodal model CLIP [25] as an example, when feeding image data into CLIP, regardless of what aspects of clustering the user expects, CLIP can only produce the same embeddings. Even considering the scenario that different pretrained models can capture different aspects of the same data as in [9], it is hard to tell which one matches a user's preference. Fortunately, given CLIP's ability to model image-text pairs collaboratively, we can use a user's high-level concept to trigger the corresponding feature extraction from the pre-trained encoders from CLIP. However, no previous work has studied if CLIP has this potential to uncover different aspects of images, which is the focus of this work.", + "bbox": [ + 75, + 185, + 468, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we propose to integrate a user's high-level concept describing the preference using a personalized textual prompt. For example, if a user's focus pertains to the color dimension of fruit, a prospective prompt might be formulated as \"a fruit with the color of *\", wherein the \"*\"\n- placeholder represents the proxy word awaiting determination using the knowledge in CLIP. Thereafter, we can learn the proxy word embedding by maximizing the similarity between the image and text embedding. However, the proxy word embedding is now searched in a continuous space while the original CLIP used discrete tokens, which can downgrade the performance. We prove that the performance can be well guaranteed by selecting the nearest token as the reference, which is unavailable in a clustering task.", + "bbox": [ + 75, + 444, + 468, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fortunately, we can use the user's high-level concept as a reference, which however covers a broad range of tokens under its scope. Therefore, we propose to leverage large language models like GPT-4 to generate candidate tokens using the user's high-level concept, in which the closest can be used as a closer reference token. Furthermore, in some scenarios, users may provide multiple concepts to obtain multiple clusterings simultaneously as shown in Fig. 1, we can also introduce a negative loss with these constrastive concepts to further enhance the learning. Therefore, to capture a user's specific interest and discover personalized clustering structure hidden in the data, we propose a multimodal proxy learning method (Multi-MaP). Multi-MaP incorporates both text prompts and unlabeled images from the clustering task, and leverages CLIP to acquire their respective personalized representations using both reference", + "bbox": [ + 75, + 659, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "word and concept-level constraints. The contributions of this work can be summarized as:", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to explore a deep multiple clustering method that precisely captures a user's interest(s) and generates personalized clustering(s) accordingly.", + "- We propose a novel multi-modal proxy learning method, Multi-MaP, by leveraging the text and image encoders pre-trained by CLIP, where the user's interest can be captured by the personalized text prompts.", + "- Considering the challenge of learning a word proxy in a continuous space while tokens in CLIP were discrete, we theoretically prove that a close reference token can help constrain the search, which motivates the proposed reference word constraint and concept-level constraint.", + "- We conduct extensive experiments on all publicly available visual multiple clustering tasks, which empirically demonstrates the superiority of the proposed Multi-MaP, with a precise capturing of a user's interest.", + "- Finally, to the best of our knowledge, we are the first who demonstrate that CLIP can uncover different semantic aspects of images." + ], + "bbox": [ + 500, + 121, + 890, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 420, + 640, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we briefly review two related directions, that is, multiple clustering and multi-modal models.", + "bbox": [ + 498, + 446, + 890, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multiple Clustering", + "text_level": 1, + "bbox": [ + 500, + 484, + 687, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multiple clustering, as a kind of method that can discover alternative perspectives of data, has attracted considerable attention. Traditional methods for multiple clustering [11] use shallow models to find different ways of grouping data. Some of these methods rely on constraints to produce alternative clusterings. For instance, COALA [1] uses the objects in an existing clustering as constraints for creating a new clustering and Qi et al. [21] formulated multiple clustering as a constrained optimization problem. Other methods exploit different feature subspaces to generate multiple clusterings. For example, Hu et al. [12] found multiple clusterings by maximizing the eigengap in different subspaces. MNMF [34] adds the inner product of similarity matrices as a regularization term to find multiple clusterings. Some methods also use information theory to generate multiple clusterings. Gondek [8] applied conditional information bottleneck and Dang [4] used an expectation maximization framework to optimize mutual information.", + "bbox": [ + 496, + 507, + 890, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, some methods have used DNNs to find multiple clusterings and achieved better results. Wei et al. [31] proposed a deep matrix factorization based method that uses multi-view data to find multiple clusterings. ENRC [18] uses an auto-encoder to learn object features and finds multiple clusterings by optimizing a clustering objective function. iMClusts [27] exploits auto-encoders and multi-head attention to learn features from different perspectives, and", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14067", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "then find multiple clusterings. AugDMC [35] leverages data augmentation to generate diverse aspects of images and learns the representations to discover multiple clusterings. Although existing deep multiple clustering methods have achieved remarkable results, they require users to exert considerable efforts to select the correct clustering they need. In this work, we aim to efficiently and effectively capture users' interests using short keywords and provide clustering results accordingly.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Multi-modal Model", + "text_level": 1, + "bbox": [ + 76, + 233, + 264, + 248 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modal learning refers to the process of learning representations from different types of input modalities, such as image data, text, or speech. As related, we focus on how vision models can benefit from natural language supervision. CLIP [25] is a notable model, which is trained with a dataset containing 400 million text-image pairs from the internet. The objective is to align images to their corresponding text using contrastive learning.", + "bbox": [ + 75, + 253, + 468, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning approaches adapt vision-language models like CLIP to specific downstream image recognition tasks. CoOp [36] and CLIP-Adapter [6] exemplify this, with the latter integrating residual style feature blending to enhance performance on various visual classification tasks. Additionally, insights from TeS [30], further elucidate the effectiveness of fine-tuning strategies in leveraging natural language supervision for enhanced visual understanding. Recognizing the scarcity of labeled data for various tasks, significant research efforts have been dedicated to enhancing zero-shot learning. Some approaches extend beyond CLIP by incorporating other large pre-trained models. For instance, VisDesc [17] harnesses the power of GPT-3 to generate comprehensive contextual descriptions corresponding to given class names, thereby demonstrating superior performance compared to CLIP's basic prompts. UPL [13] and TPT [29] leverages unlabeled data to optimize learnable input text prompts. InMaP [24] recovers the proxy of each class in the vision space with the help from the text proxy. All of these methods aim to improve the performance of vision classification tasks, while clustering is a different scenario that we do not have class names that we can exploit to extract useful information from CLIP. In this work, we consider leveraging the CLIP encoder to extract coherent text and image embeddings, and using GPT-4 to integrate users' interests for the purpose of multiple clustering.", + "bbox": [ + 75, + 375, + 472, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 773, + 289, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first briefly review the training objective in CLIP as follows, and then describe the details of our proposed method based on that.", + "bbox": [ + 75, + 791, + 468, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multi-modal Pre-training", + "text_level": 1, + "bbox": [ + 76, + 847, + 310, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of image-text pairs as $\\{x_{i},t_{i}\\}_{i = 1}^{n}$ , where $x_{i}$ is an image and $t_i$ is the corresponding text description, their", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "vision and text representations can be obtained by two encoders as $\\mathbf{x}_i = f(x_i)$ and $\\mathbf{t}_i = h(t_i)$ . $f(\\cdot)$ and $h(\\cdot)$ are vision and text encoders for optimization, where $\\mathbf{x}_i$ and $\\mathbf{t}_i$ have the unit norm. Then, these two encoders can be learned by minimizing the contrastive loss as", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {f, h} \\sum_ {i} - \\log \\frac {\\exp (\\mathbf {x} _ {i} ^ {\\top} \\mathbf {t} _ {i} / \\tau)}{\\sum_ {j} \\exp (\\mathbf {x} _ {i} ^ {\\top} \\mathbf {t} _ {j}) / \\tau) - \\log \\frac {\\exp (\\mathbf {t} _ {i} ^ {\\top} \\mathbf {x} _ {i} / \\tau)}{\\sum_ {j} \\exp (\\mathbf {t} _ {i} ^ {\\top} \\mathbf {x} _ {j}) / \\tau)}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 176, + 890, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature.", + "bbox": [ + 500, + 224, + 683, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This contrastive loss aims to pull the image and its description together while pushing away the irrelevant text [22], which enables the emerging multi-modal applications, e.g., zero-shot transfer [24, 25], text-to-image generation [28], etc.", + "bbox": [ + 496, + 239, + 890, + 315 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-modal Proxy Learning", + "text_level": 1, + "bbox": [ + 500, + 324, + 758, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the pre-trained vision and text encoders from CLIP, this work takes one step further to investigate if we can extract user-specific information from the alignment between images and text.", + "bbox": [ + 496, + 348, + 890, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concretely, given an image of fruit [12] as shown in Fig. 2, some users may be interested in only one specific property of the object, e.g., color. In this scenario, applying the vision encoder to extract the representation for the whole image can miss the preference of users. To mitigate the problem, we propose to explore the proxy representation from the image with the guidance from the text using users' preference, named Multi-Modal Proxy learning (Multi-MaP).", + "bbox": [ + 496, + 407, + 890, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recall that CLIP is pre-trained by images and text descriptions, where the text prompt is \"a photo of a fruit\" for an image containing \"fruit\". Now given a user's preference (e.g., color), we can rewrite the prompt as \"fruit with the color of *\" denoted by $t_i^*$ for image $x_i$ , where \" * \" is the proxy word and its token embedding is $\\mathbf{w}_i$ that is learnable. Then, we can align image and text prompt representations to obtain the appropriate proxy embedding for a user's interest. Since there is no negative pairs, only the similarity between positive pairs can be optimized as", + "bbox": [ + 496, + 542, + 890, + 695 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\mathbf {w} _ {i}\\right) = - \\langle f \\left(x _ {i}\\right), h \\left(t _ {i} ^ {*}\\right) \\rangle \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 705, + 890, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the vision and text encoders are frozen, and $\\mathbf{w}_i$ is the only variable for learning as the representation of the proxy word. By maximizing the similarity to the image representation, Multi-MaP aims to learn the optimal text proxy according to the user's interest.", + "bbox": [ + 496, + 734, + 890, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, it should be noted that the text encoder was pre-trained with discrete text tokens, while the domain of $\\mathbf{w}_i$ in Eqn. 1 is unconstrained. Therefore, the text representation extracted from the frozen text encoder can be inaccurate for $\\mathbf{w}_i$ that degenerates the performance, which is demonstrated in the following theory.", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14068", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2ba575801ef6635c7c562da723dd4dc6de5af78eadb4d7edac62571b4717dbfa.jpg", + "image_caption": [ + "Figure 2. Multi-MaP framework. In the training process of Multi-MaP, the vision and text encoders are frozen and the proxy word embeddings $\\mathbf{w}_i$ are learnable. Specifically, it first constructs the prompt embeddings based on the reference words provided by GPT-4 using a user's high-level concept, and then selects a reference word $z_i$ for each image according to the similarity between the prompt embeddings $\\mathbf{t}_i$ and the image embeddings $\\mathbf{x}_i$ . Then, it combines the prompt and the reference words to form the new prompt embeddings $\\mathbf{t}_i^*$ and maximizes the similarity to the image representation, so the proxy word embeddings $\\mathbf{w}_i$ can capture the desired image features. In addition, the proxy word embeddings $\\mathbf{w}_i$ should be close to the target concept word $\\mathbf{u}_1$ and the selected reference word $\\mathbf{z}_i$ to construct the concept-level constraint and reference word constraint, which capture the features related to the user's interest." + ], + "image_footnote": [], + "bbox": [ + 117, + 109, + 849, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the sake of simplicity, we assume $h'(t) \\in \\mathbb{R}$ is defined on the whole set but only has the right estimation on a discrete set as $T = \\{t_i\\}$ and the counterpart with the unconstrained set is denoted as $H(w)$ . According to the definition, we have $\\forall t \\in T, h'(t) = H(t)$ . The gap between the estimation from $h'$ and $H$ on unconstrained variable $w$ can be depicted as follows.", + "bbox": [ + 75, + 516, + 470, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. Given $w \\notin T$ and $t \\in T$ , if assuming $h'$ and $H$ are $L_h$ and $L_H$ -Lipschitz continuous, we have", + "bbox": [ + 76, + 628, + 468, + 659 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\| h ^ {\\prime} (w) - H (w) \\| _ {2} \\leq \\left(L _ {h} + L _ {H}\\right) \\| t - w \\| _ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 666, + 415, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. According to the definition, we have", + "bbox": [ + 76, + 690, + 370, + 704 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left\\| h ^ {\\prime} (w) - H (w) \\right\\| _ {2} = \\left\\| h ^ {\\prime} (w) - h ^ {\\prime} (t) + h ^ {\\prime} (t) - H (w) \\right\\| _ {2} \\\\ \\leq \\| h ^ {\\prime} (w) - h ^ {\\prime} (t) \\| _ {2} + \\| h ^ {\\prime} (t) - H (w) \\| _ {2} \\\\ = \\| h ^ {\\prime} (w) - h ^ {\\prime} (t) \\| _ {2} + \\| H (t) - H (w) \\| _ {2} \\\\ \\leq \\left(L _ {h} + L _ {H}\\right) \\| t - w \\| _ {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 710, + 460, + 785 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a3aeb8352ad7a5f0985af472da22dc1d96157a30fa0f311f31584f4132f28b69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 792, + 467, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark Theorem 1 implies that the distance of the estimation $h'(w)$ to the ground-truth result $H(w)$ is bounded by that of $w$ to an arbitrary discrete token $t$ . Therefore, by selecting the nearest token as the reference, the bound can be improved as shown in the following corollary.", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Corollary 2. With the assumptions in Theorem 1 and letting $t' = \\arg \\min_i \\| t_i - w\\| _2$ , we have", + "bbox": [ + 498, + 517, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| h ^ {\\prime} (w) - H (w) \\right\\| _ {2} \\leq \\left(L _ {h} + L _ {H}\\right) \\| t ^ {\\prime} - w \\| _ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 553, + 841, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Concept-level Constraint", + "text_level": 1, + "bbox": [ + 500, + 587, + 728, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to the above analysis, a good reference $t'$ can help guarantee the performance. Fortunately, the input concept (e.g., color) from the user can be leveraged as the reference to constrain the freedom of the proxy word. Therefore, given the target concept word $u$ , we can obtain its token embedding as $\\mathbf{u} = \\phi(u)$ . Then, to learn appropriate representations from the proxy embedding, the original problem can be rewritten with the constraint as", + "bbox": [ + 496, + 609, + 890, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathbf {w} _ {i}) = - \\langle f (x _ {i}), h \\left(t _ {i} ^ {*}\\right) \\rangle \\quad s. t. \\quad \\| \\mathbf {w} _ {i} - \\mathbf {u} \\| _ {2} ^ {2} \\leq \\lambda\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 737, + 862, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The constrained problem is equivalent to", + "bbox": [ + 500, + 761, + 771, + 776 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\mathbf {w} _ {i}\\right) = - \\langle f \\left(x _ {i}\\right), h \\left(t _ {i} ^ {*}\\right) \\rangle + \\alpha \\| \\mathbf {w} _ {i} - \\mathbf {u} \\| _ {2} ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 782, + 890, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "following [3], which can be optimized effectively by gradient descent.", + "bbox": [ + 496, + 806, + 890, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Constrained Optimization with Reference Word", + "text_level": 1, + "bbox": [ + 500, + 845, + 888, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, it is well known that the user concept is often with a large scope covering a broad range of words (e.g., color", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14069", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/023cb8528467ed42b5c8af2d7528723dccaf6a140745fdfbe0012979083f01b6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Datasets# Samples# Clusters
ALOI2882;2
Card8,02913;4
CMUface6404;20;2;4
Fruit1053;3
Fruit3604,8564;4
Stanford Cars1,2004;3
Flowers1,6004;4
", + "bbox": [ + 143, + 88, + 401, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Dataset Statistics.", + "bbox": [ + 192, + 212, + 352, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "covers all including but not limited to 'red', 'blue', 'green', etc.). As suggested by our above theoretical analysis, it is desired if the reference is as close as possible. In a clustering scenario and given only the user's high-level concept, it is challenging to find a closer reference word to further constrain the proxy learning. Fortunately, with the development of large language models (LLMs), we can leverage them (e.g., GPT-4) to provide relevant words according to a user's high-level concept as the candidate set and develop a selection strategy to obtain a closer reference word for each image. While the responses gathered from GPT-4 might not always precisely align with the data's ground truth, they indisputably furnish valuable candidate features, enriching the capabilities of Multi-MaP.", + "bbox": [ + 75, + 234, + 468, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To elucidate, considering the task of clustering a fruit dataset based on the concept of color, we can pose a query to GPT-4 as \"What are the common colors of fruit?\" The response we obtain is \"Common colors of fruit include red, yellow, green, orange, purple, and blue\". The colors enumerated in this response can serve as reference word candidates. It is worth noting that while the color spectrum derived from this method might surpass the actual colors present in the data, we collect all of them into the candidate set of reference words $\\{z_k\\}_k$ , where $z_1$ : \"red\", $z_2$ : \"yellow\", etc. Then, their text representations can be obtained from the prompt $t_k$ as \"fruit with the color of $z_k$ \". Given the image $x_i$ , the closest reference can be observed as", + "bbox": [ + 75, + 446, + 470, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i} = \\arg \\max _ {k} \\left\\langle \\mathbf {x} _ {i}, \\mathbf {t} _ {k} \\right\\rangle\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 656, + 346, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{t}_k = h(t_k)$ . After that, $\\mathbf{w}_i$ can be initialized with the token embedding of $z_i$ as $\\mathbf{z}_i = \\phi(z_i)$ . Moreover, we change the regularization using a closer reference word compared to the high-level concept as", + "bbox": [ + 75, + 689, + 468, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\mathbf {w} _ {i}\\right) = - \\langle f \\left(x _ {i}\\right), h \\left(t _ {i} ^ {*}\\right) \\rangle + \\alpha \\| \\mathbf {w} _ {i} - \\mathbf {z} _ {i} \\| _ {2} ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 762, + 468, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 Contrastive Concepts", + "text_level": 1, + "bbox": [ + 76, + 801, + 281, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In some application scenarios, one user may need more than one clustering and provide high-level concepts as $\\{u_j\\}$ , e.g., $u_1$ : \"color\", $u_2$ : \"species\", etc. For the concept \"color\", the irrelevant concept \"species\" can be leveraged as the negative constraint for the learning of proxy word. Concretely,", + "bbox": [ + 75, + 825, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "let $u_{w}$ denote the target concept word, a contrastive loss can be adopted as regularization", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR (\\mathbf {w} _ {i}) = - \\log \\frac {\\exp (\\mathbf {w} _ {i} ^ {\\top} \\mathbf {u} _ {w})}{\\sum_ {j} \\exp (\\mathbf {w} _ {i} ^ {\\top} \\mathbf {u} _ {j})}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 132, + 807, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the final objective becomes", + "bbox": [ + 500, + 180, + 710, + 194 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} \\left(\\mathbf {w} _ {i}\\right) = \\\\ - \\left\\langle f \\left(x _ {i}\\right), h \\left(t _ {i} ^ {*}\\right) \\right\\rangle + \\alpha \\| \\mathbf {w} _ {i} - \\mathbf {z} _ {i} \\| _ {2} ^ {2} - \\beta \\log \\frac {\\exp \\left(\\mathbf {w} _ {i} ^ {\\top} \\mathbf {u} _ {w}\\right)}{\\sum_ {j} \\exp \\left(\\mathbf {w} _ {i} ^ {\\top} \\mathbf {u} _ {j}\\right)} \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 207, + 890, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the first term is to infer the user-specific feature, while the latter two terms constrain the proxy word to the reference words for the appropriate representation extraction from the pre-trained text encoder. The overall framework of Multi-MaP is illustrated in Fig. 2.", + "bbox": [ + 496, + 290, + 890, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 373, + 633, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate our proposed method, we evaluate MultiMaP on all publicly available image datasets in multiple clustering, including ALOI [7], Stanford Cars [14], Card [35], CMUface [10], Flowers [20], Fruit [12], and Fruit360 [35] as summarized in Table 1.", + "bbox": [ + 496, + 398, + 890, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare Multi-MaP against five state-of-the-art methods: MSC [12] is a traditional multiple clustering method that uses hand-crafted features; MCV [9] leverages multiple feature extractors to represent different \"views\" of the same data and employs a multi-input neural network to enhance clustering outcomes; ENRC [18] is a deep multiple clustering method that integrates auto-encoder and clustering objective to generate different clusterings; iM-Clusts [27] makes use of the expressive representational power of deep autoencoders and multi-head attention to accomplish multiple clusterings; AugDMC [35] leverages augmentations to learn different image representations to achieve multiple clustering.", + "bbox": [ + 496, + 474, + 892, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 680, + 676, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ Adam and set momentum as 0.9 to train the model for 1000 epochs. All hyperparameters are searched according to the loss score of Multi-MaP, where the learning rate is searched in $\\{0.1, 0.05, 0.01, 0.005, 0.001, 0.0005\\}$ , weight decay is in $\\{0.0005, 0.0001, 0.00005, 0.00001, 0\\}$ , $\\alpha, \\beta$ are in $\\{0.0, 0.1, 0.2, \\ldots, 1.0\\}$ , and $\\lambda$ is fixed as 1 for all the experiments. For the non-pre-trained methods, we perform k-means [15] 10 times due to its randomness and evaluate the average clustering performance using two quantitative metrics, that is, Normalized Mutual Information (NMI) [32] and Rand index (RI) [26]. These measures range in $[0, 1]$ , and higher scores imply more accurate", + "bbox": [ + 496, + 704, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14070", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8eee278c969f47dbe4bdb51f69502e190b17b7eabe5a1802911b9b9899b279ae.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetClusteringMSCMCVENRCiMClustersAugDMCMulti-MaP
NMIRINMIRINMIRINMIRINMIRINMIRI
ALOIColor0.15630.34280.69820.74390.98330.98921.00001.00001.00001.00001.00001.0000
Shape0.29680.51990.73590.82610.97320.98610.99630.99891.00001.00001.00001.0000
FruitColor0.68860.80510.62660.76850.71030.85110.73510.86320.85170.91080.86190.9526
Species0.16270.60450.27330.65970.31870.65360.30290.67430.35460.73991.00001.0000
Fruit360Color0.25440.60540.37760.67910.42640.68680.40970.68410.45940.73920.62390.8243
Species0.21840.58050.29850.61760.41420.69840.38610.67320.51390.74300.52840.7582
CardOrder0.08070.78050.07920.71280.12250.73130.11440.76580.14400.82670.36530.8587
Suits0.04970.35870.04300.36380.06760.38010.07160.37150.08730.42280.27340.7039
CMUfaceEmotion0.12840.67360.14330.52680.15920.66300.04220.59320.01610.53670.17860.7105
Glass0.14200.57450.12010.49050.14930.62090.19290.56270.10390.53610.34020.7068
Identity0.38920.73260.46370.62470.56070.76350.51090.82600.58750.83340.66250.9496
Pose0.36870.63220.32540.60280.22900.50290.44370.61140.13200.55170.46930.6624
Stanford CarsColor0.23310.61580.21030.58020.24650.67790.23360.65520.27360.75250.73600.9193
Type0.13250.53360.16500.56340.20630.62170.19630.56430.23640.73560.63550.8399
FlowersColor0.25610.59650.29380.58600.33290.62140.31690.61270.35560.69310.64260.7984
Species0.13260.52730.15610.60650.18940.61950.18870.60770.19960.62270.60130.8103
", + "bbox": [ + 138, + 88, + 826, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "clustering results. The experiments are conducted with GPU NVIDIA GeForce RTX 2080 Ti.", + "bbox": [ + 75, + 349, + 468, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It should also be noted that some data are difficult to obtain corresponding candidate labels from GPT-4 or the labels do not provide semantic features, such as names. For example, for the identity clustering for the CMUface dataset [10], different identities represent different people and the semantic meaning of names should not affect the clustering results. In this case, we randomly extract 10 words from WordNet [5] as reference words, in order to make the candidate labels more distinctive. For instance, we randomly choose \"abstain, candid, function, haphazard, knot, luxury, nonchalance, pension, resilience, taciturn\" for the above scenario. Furthermore, all publicly available multiple clustering datasets provide each ground-truth clustering a high-level concept, e.g., 'shape', 'pose', etc. Therefore, in the experiment, we directly use them as users' preferences for our evaluation purposes.", + "bbox": [ + 75, + 378, + 468, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance Comparison", + "text_level": 1, + "bbox": [ + 76, + 633, + 310, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiments, after we obtain the proxy word embedding of each image for a desired concept, we feed them to k-means to obtain the corresponding clustering. Since k-means is random, we repeat ten times and the average performance is reported in Table 2. The best results are marked by bold numbers.", + "bbox": [ + 75, + 657, + 468, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We can observe that Multi-MaP outperforms the baselines in all the cases, indicating the superiority of the proposed method. This also shows a strong generalization ability of the pre-trained model by CLIP, which can capture the features of data in different aspects.", + "bbox": [ + 75, + 748, + 468, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since our method uses the CLIP encoder and GPT-4 to obtain clustering results, a natural question arises how the performance would be if we directly use them in a zero-short manner. Therefore, we provide two zero-shot variants of CLIP, that is, $\\mathrm{CLIP}_{\\mathrm{GPT}}$ that uses GPT-4 to obtain the can", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b79a813f9965f48a9a79949debd7ca4e45602ae683d2015a1ddd5fb74def29ae.jpg", + "table_caption": [ + "Table 2. Quantitative comparison. The significantly best results with $95\\%$ confidence are in bold." + ], + "table_footnote": [], + "table_body": "
DatasetClusteringCLIPGPTCLIPlabelMulti-MaP
NMIRINMIRINMIRI
ALOIColor0.85810.94071.00001.00001.00001.0000
Shape1.00001.00001.00001.00001.00001.0000
FruitColor0.79120.90750.86290.97800.86190.9526
Species0.97930.99191.00001.00001.00001.0000
Fruit360Color0.56130.73050.57460.76730.62390.8243
Species0.43700.75520.53640.76310.52840.7582
CardOrder0.35180.84580.35180.84580.36530.8587
Suits0.27110.61230.27110.61230.27340.7039
CMUfaceEmotion0.15760.65320.15900.66190.17860.7105
Glass0.29050.68690.46860.75050.34020.7068
Identity0.19980.63880.26770.75450.66250.9496
Pose0.40880.64730.46910.64090.46930.6624
Stanford CarsColor0.65390.82370.68300.86420.73600.9193
Type0.62070.79310.64290.84560.63550.8399
FlowersColor0.56530.76290.58280.78360.64260.7984
Species0.56200.75530.60190.79960.60130.8103
", + "bbox": [ + 503, + 345, + 888, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Variants of CLIP. The significantly best results with $95\\%$ confidence are in bold.", + "bbox": [ + 500, + 561, + 888, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "didate labels and predicts labels through zero-shot classification with all candidate labels as class names, and $\\mathrm{CLIP}_{\\mathrm{label}}$ that performs zero-shot classification with all ground truth labels. It should be noted that $\\mathrm{CLIP}_{\\mathrm{label}}$ uses an unfair setting with a ground-truth label set known in advance, which is expected to provide the best performance using CLIP in a zero-short manner. The results are shown in Table 3.", + "bbox": [ + 496, + 598, + 890, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As expected, $\\mathrm{CLIP}_{\\mathrm{label}}$ achieves better performance than $\\mathrm{CLIP}_{\\mathrm{GPT}}$ in almost all cases, since $\\mathrm{CLIP}_{\\mathrm{label}}$ uses a fixed ground truth label set, while $\\mathrm{CLIP}_{\\mathrm{GPT}}$ uses candidate labels that may not match the ground truth exactly, introducing noise. Note that $\\mathrm{CLIP}_{\\mathrm{GPT}}$ and $\\mathrm{CLIP}_{\\mathrm{label}}$ achieve the same results on Cards, because the candidate labels provided by GPT-4 are exactly the same as the true labels.", + "bbox": [ + 496, + 704, + 890, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides, Multi-MaP performs better than $\\mathrm{CLIP}_{\\mathrm{GPT}}$ in almost all cases, indicating that the proposed method can learn more effective features through its training process. Moreover, although $\\mathrm{CLIP}_{\\mathrm{label}}$ uses the ground truth, which is expected to be the best, Multi-MaP still outperforms $\\mathrm{CLIP}_{\\mathrm{label}}$ in some cases, such as the clustering of color for", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14071", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5da8772cbc53ecd16c3723cff932647c8c5d3b0c4849a0fcdc2f39fd59cb1e3e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Multi-MaPpMulti-MaPcMulti-MaPrMulti-MaPcrMulti-MaP
ModulesProxy Learning✓✓✓✓✓
Concept Word×✓×✓✓
Reference Word××✓✓✓
Contrastive Concepts××××✓
NMI↑RI↑NMI↑RI↑NMI↑RI↑NMI↑RI↑NMI↑RI↑
ALOI [7]Color0.96190.98261.00001.00000.97950.98691.00001.00001.00001.0000
Shape1.00001.00001.00001.00001.00001.00001.00001.00001.00001.0000
Fruit [12]Color0.76420.84390.82150.92830.81360.90730.84840.93080.86190.9526
Species1.00001.00001.00001.00001.00001.00001.00001.00001.00001.0000
Fruit360 [35]Color0.56430.76650.62170.78360.59100.77460.60890.79650.62390.8243
Species0.50770.73680.51370.74360.50940.74250.51990.74280.52840.7582
Card [35]Order0.19320.81520.35680.84720.31130.82290.36160.80940.36530.8587
Suits0.23750.62820.26960.66410.24980.63650.25620.65990.27340.7039
CMUface [10]Emotion0.16900.61700.17140.62290.16970.63600.17130.68430.17860.7105
Glass0.31120.69110.32690.71360.31620.69170.33700.71080.34020.7068
Identity0.56170.82340.62430.83590.58390.82630.63910.89460.66250.9496
Pose0.43610.63860.45500.64990.43810.64290.43870.64890.46930.6624
Stanford cars [14]Color0.59390.78350.68360.86590.67290.86380.71120.91170.73600.9193
Type0.55690.79960.63830.82710.60910.80460.62890.81810.63550.8399
Flowers [20]Color0.57830.77230.58300.78330.59870.78490.62160.79410.64260.7984
Species0.57040.76080.57440.78420.57230.78110.58460.78920.60130.8103
", + "bbox": [ + 140, + 87, + 831, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Components ablation. All of our components boost performance consistently in all benchmark multi-clustering vision tasks.", + "bbox": [ + 91, + 388, + 875, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d379e2a9c14ac905850983fa8ae1b6801bed353d8651054dddb942dca9822a2.jpg", + "image_caption": [ + "(a) Parameter sensitivity of $\\alpha$", + "Figure 3. Parameter analysis of $\\alpha$ and $\\beta$ on Fruit [12]." + ], + "image_footnote": [], + "bbox": [ + 96, + 426, + 274, + 547 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9aa1ae24c98b35eb8f0bd4d28573803976cadca5c5595260233e2adab90e438e.jpg", + "image_caption": [ + "(b) Parameter sensitivity of $\\beta$" + ], + "image_footnote": [], + "bbox": [ + 279, + 426, + 455, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fruit360 dataset. This is because CLIP is more inclined to capture the features of one aspect of the data, while Multi-MaP learns better embedding of different aspects by training with the supervision of users' interests. Furthermore, Multi-MaP also achieves very competitive results as $\\mathrm{CLIP}_{\\mathrm{label}}$ in the remaining cases, which further demonstrates the effectiveness of the proposed method.", + "bbox": [ + 75, + 588, + 468, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 708, + 230, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of the proposed method, we show the gain from four components in Multi-MaP in Table 4. Let \"Multi-MaP $p$ \" denote the proxy learning without concept-level constraint, \"Multi-MaP $c$ \" denote the variant optimized by solely applying concept word to constrain the freedom of the proxy word, \"Multi-MaP $r$ \" denote the variant optimized by reference word provided by GPT-4 to find a closer reference word to further constrain the proxy learning, and \"Multi-MaP $cr$ \" denote the variant leveraged both concept word and reference word.", + "bbox": [ + 75, + 732, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can observe that the proxy learning only with concept", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "word or reference word, i.e., $\\mathrm{Multi - MaP}_c$ and $\\mathrm{Multi - MaP}_r$ , performs better than $\\mathrm{Multi - MaP}_p$ . This shows that the proposed concept-level constraint and constrained optimization with reference words play an important role in the model as demonstrated by our theoretical analysis. Moreover, the model with combined components, i.e., $\\mathrm{Multi - MaP}_{cr}$ , achieves better results than $\\mathrm{Multi - MaP}_r$ and $\\mathrm{Multi - MaP}_c$ . This indicates the effectiveness of the combination of reference words and reference concepts. Finally, our proposal using all including the contrastive concepts can further improve the performance, and thus provide the best results on all cases. This further demonstrates our proposal.", + "bbox": [ + 496, + 430, + 890, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Parameter Analysis", + "text_level": 1, + "bbox": [ + 498, + 619, + 687, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further investigate the effect of the reference word constraint weight $\\alpha$ and concept-level constraint weight $\\beta$ varying from 0.0 to 0.5. The results of Multi-MaP on Fruit datasets are shown in Fig. 3 (a) and Fig. 3 (b), respectively. As $\\alpha$ and $\\beta$ change, the proposed method keeps a species score of 1, since the image encoder can capture very effective species features. The above results also show that the proposed method can effectively capture useful information from images, reference words, and target concepts. As $\\alpha$ increases, the proposed method first increases and then decreases, and reaches the maximum values at $\\alpha = 0.4$ . Similar results can be observed for $\\beta$ that the performance of the proposed method first increases and then decreases as $\\beta$ increases, and reaches the maximum values at $\\beta = 0.3$ . Indicating a suitable value of $\\alpha$ or $\\beta$ is helpful for Multi-MaP to obtain more effective embeddings for multiple clustering tasks. More studies such as efficiency analysis can be found", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14072", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d47eed2ba007e88973ba891e69b4e708cc98869c6042b469f2ffed4e5e3ac151.jpg", + "image_caption": [ + "(a) Color using ground truth" + ], + "image_footnote": [], + "bbox": [ + 116, + 99, + 287, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/68305eda48f13a2802bee29840a20200e88b2c25c145aebc5e3e32b8d7b46396.jpg", + "image_caption": [ + "(b) Color using reference words" + ], + "image_footnote": [], + "bbox": [ + 303, + 101, + 480, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9cc9acdca6050ec7a34c0cff89c072dcce12242d8f3a52229a5da63be06a9538.jpg", + "image_caption": [ + "(c) Color of Multi-MaP", + "(d) Legend of color" + ], + "image_footnote": [ + "red yellow", + "green orange", + "purple blue" + ], + "bbox": [ + 529, + 101, + 671, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4232e29350efc6f472a931a4f2780a4e8977717639f2aec56c9172d58296d812.jpg", + "image_caption": [ + "(e) Species using ground truth" + ], + "image_footnote": [], + "bbox": [ + 137, + 226, + 284, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f62f49f5cb29726cf762be8dc6bd08433e39a964631a3e5850911afe87ab465.jpg", + "image_caption": [ + "(f) Species using reference words" + ], + "image_footnote": [], + "bbox": [ + 318, + 224, + 475, + 313 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d7301845ed65e734b73418bd46dee7d5539e0cdd9d5b4b96946db9fa0bf95ebe.jpg", + "image_caption": [ + "(g) Species of Multi-MaP", + "(h) Legend of species", + "Figure 4. Visualization of feature embeddings and related labels. The points represent the image or pseudo-word embeddings, and the triangles represent the prompt or label embeddings. Different colors represent different labels, which are indicated by the text next to the triangles." + ], + "image_footnote": [ + "- apple\n- orange\n- banana\n- strawberry\n- grapes\n- raspberry\n- blueberry\n- cherry\n- pear\n- plum\n- peach\n- nectarine\n- pineapple\n- kiwi\n- watermelon\n- cantaloupe\n- mango" + ], + "bbox": [ + 526, + 227, + 669, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in the supplementary.", + "bbox": [ + 76, + 398, + 220, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualization", + "text_level": 1, + "bbox": [ + 76, + 422, + 212, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further demonstrate the effectiveness of the proposed method, we visualize the representations obtained in $\\mathrm{CLIP}_{\\mathrm{label}}$ , $\\mathrm{CLIP}_{\\mathrm{GPT}}$ , and Multi-MaP. Specifically, for $\\mathrm{CLIP}_{\\mathrm{label}}$ and $\\mathrm{CLIP}_{\\mathrm{GPT}}$ , we visualize the image representations as well as prompts generated with real labels and candidate labels, respectively. For the Multi-MaP, we visualize the word embedding $w_{*}$ and the candidate labels selected for initialization. The results are shown in Fig. 4. For species clustering, we can see that image embeddings show very clear boundaries and correspond well to the prompt for $\\mathrm{CLIP}_{\\mathrm{label}}$ , which indicates that CLIP can effectively capture the features of the species in the data. $\\mathrm{CLIP}_{\\mathrm{GPT}}$ uses the candidate labels to generate prompts, which introduces more noise, but benefits from the CLIP text encoder, the image embedding can keep a relatively far distance from most of the irrelevant prompts. However, since there are still a few images that are labeled as peaches (i.e., a noisy label), it performs slightly worse than $\\mathrm{CLIP}_{\\mathrm{label}}$ , as shown in Fig. 4(e). Besides, Multi-MaP can capture the image and users' interests in the training process, therefore it compensates for the shortcomings of $\\mathrm{CLIP}_{\\mathrm{GPT}}$ and achieves better results. On the other hand, for color clustering, the prompts are farther away from $\\mathrm{CLIP}_{\\mathrm{label}}$ and $\\mathrm{CLIP}_{\\mathrm{GPT}}$ , that indicates the image embeddings mainly capture the features of species, which have no direct connection with color. $\\mathrm{CLIP}_{\\mathrm{GPT}}$ generates the prompt from the candidate label, which has more noise than the ground truth label, resulting in worse performance than $\\mathrm{CLIP}_{\\mathrm{label}}$ . The proposed method can distinguish different colors more clearly, because it can learn from the user's interest and capture the color-related features. How", + "bbox": [ + 75, + 446, + 472, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ever, some red color embeddings are closer to purple, because some images in the datasets are actually purple, but labeled as red. To sum up, the proposed method can learn more effective embeddings based on the users' interests for multiple clustering tasks.", + "bbox": [ + 498, + 398, + 890, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 481, + 617, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To conclude, our study thoroughly investigates the significant challenges that current advanced deep learning techniques face in multiple clustering. A key issue is that users often do not need every clustering result produced by an algorithm, and selecting the most relevant one requires an in-depth understanding of each outcome. To overcome this challenge, the proposed method introduces a novel multimodal proxy learning process, which effectively aligns a user's brief keyword describing the interest with the corresponding vision components. By integrating a multi-modal model and GPT-4 to precisely capture a user's interest using a keyword, the proposed approach uses both reference word constraint and concept-level constraint to discover personalized clustering result(s), which can also lead to enhanced performance. Experiments on diverse datasets demonstrate the superiority of the proposed method in multiple clustering tasks with a precise capture of a user's interest. The proposed method is limited by data with semantic meaningful labels, although we can use WordNet to help, whose comprehensive study will be our future work.", + "bbox": [ + 496, + 503, + 890, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 814, + 679, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is supported in part by Advata Gift funding. All opinions, findings, conclusions and recommendations in this paper are those of the author and do not necessarily reflect the views of the funding agencies.", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14073", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eric Bae and James Bailey. Coala: A novel approach for the extraction of an alternate clustering of high quality and high dissimilarity. In ICDM, pages 53-62. IEEE, 2006. 1, 2", + "[2] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning. Springer, 2006. 1", + "[3] Stephen P. Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University Press, 2014. 4", + "[4] Xuan Hong Dang and James Bailey. Generation of alternative clusterings using the cami approach. In Proceedings of the 2010 SIAM International Conference on Data Mining, pages 118-129. SIAM, 2010. 2", + "[5] Christiane Fellbaum. Wordnet. In Theory and applications of ontology: computer applications, pages 231-243. Springer, 2010. 6", + "[6] 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. International Journal of Computer Vision, pages 1-15, 2023. 1, 3", + "[7] Jan-Mark Geusebroek, Gertjan J Burghouts, and Arnold WM Smeulders. The amstedam library of object images. International Journal of Computer Vision, 61:103-112, 2005. 5, 7", + "[8] David Gondek and Thomas Hofmann. Conditional information bottleneck clustering. In 3rd IEEE international conference on data mining, workshop on clustering large data sets, pages 36-42, 2003. 2", + "[9] Joris Guérin and Byron Boots. Improving image clustering with multiple pretrained cnn feature extractors. In *British Machine Vision Conference* 2018, BMVC 2018, 2018. 1, 2, 5", + "[10] Stephan Gunnemann, Ines Farber, Matthias Rüdiger, and Thomas Seidl. Smvc: semi-supervised multi-view clustering in subspace projections. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 253-262, 2014. 5, 6, 7", + "[11] Juhua Hu and Jian Pei. Subspace multi-clustering: a review. Knowledge and information systems, 56(2):257-284, 2018. 2", + "[12] Juhua Hu, Qi Qian, Jian Pei, Rong Jin, and Shenghuo Zhu. Finding multiple stable clusterings. Knowledge and Information Systems, 51(3):991-1021, 2017. 1, 2, 3, 5, 7", + "[13] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 3", + "[14] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5, 7", + "[15] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 5", + "[16] James MacQueen et al. Some methods for classification and analysis of multivariate observations. In Proceedings of the fifth Berkeley symposium on mathematical statistics and probability, pages 281-297. Oakland, CA, USA, 1967. 1" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 3", + "[18] Lukas Miklautz, Dominik Mautz, Muzaffer Can Altinigneli, Christian Böhm, and Claudia Plant. Deep embedded nonredundant clustering. In Proceedings of the AAAI conference on artificial intelligence, pages 5174-5181, 2020. 1, 2, 5", + "[19] Andrew Ng, Michael Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 14, 2001. 1", + "[20] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 5, 7", + "[21] ZiJie Qi and Ian Davidson. A principled and flexible framework for finding alternative clusterings. In SIGKDD, pages 717-726, 2009. 2", + "[22] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Tacoma Tacoma, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In ICCV, pages 6449-6457. IEEE, 2019. 3", + "[23] Qi Qian, Yuanhong Xu, Juhua Hu, Hao Li, and Rong Jin. Unsupervised visual representation learning by online constrained k-means. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16640-16649, 2022. 1", + "[24] Qi Qian, Yuanhong Xu, and Juhua Hu. Intra-modal proxy learning for zero-shot visual categorization with clip. In Thirty-seventh Conference on Neural Information Processing Systems, NeurIPS 2023, 2023. 3", + "[25] 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, 2, 3", + "[26] William M Rand. Objective criteria for the evaluation of clustering methods. Journal of the American Statistical Association, 66(336):846-850, 1971. 5", + "[27] Liangrui Ren, Guoxian Yu, Jun Wang, Lei Liu, Carlotta Domeniconi, and Xiangliang Zhang. A diversified attention model for interpretable multiple clusterings. IEEE Transactions on Knowledge and Data Engineering, 2022. 1, 2, 5", + "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685. IEEE, 2022. 3", + "[29] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Advances in Neural Information Processing Systems, 35:14274-14289, 2022. 3", + "[30] Junyang Wang, Yuanhong Xu, Juhua Hu, Ming Yan, Jitao Sang, and Qi Qian. Improved visual fine-tuning with natural language supervision. arXiv preprint arXiv:2304.01489, 2023. 1, 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14074", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Shaowei Wei, Jun Wang, Guoxian Yu, Carlotta Domeni-coni, and Xiangliang Zhang. Multi-view multiple clusterings using deep matrix factorization. In Proceedings of the AAAI conference on artificial intelligence, pages 6348–6355, 2020. 2", + "[32] JV White, Sam Steingold, and CG Fournelle. Performance metrics for group-detection algorithms. Proceedings of Interface, 2004, 2004. 5", + "[33] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487, 2016. 1", + "[34] Sen Yang and Lijun Zhang. Non-redundant multiple clustering by nonnegative matrix factorization. Machine Learning, 106(5):695-712, 2017. 2", + "[35] Jiawei Yao, Enbei Liu, Maham Rashid, and Juhua Hu. Augdmc: Data augmentation guided deep multiple clustering. In INNS DLIA@IJCNN, 2023. 1, 3, 5, 7", + "[36] 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. 3" + ], + "bbox": [ + 78, + 90, + 470, + 388 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14075", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/f9afe157-eaaf-440c-afcf-7604121ce960_model.json b/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/f9afe157-eaaf-440c-afcf-7604121ce960_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9befead1719dff01e99578372d9a1b4d971afe66 --- /dev/null +++ b/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/f9afe157-eaaf-440c-afcf-7604121ce960_model.json @@ -0,0 +1,2145 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.088, + 0.131, + 0.885, + 0.155 + ], + "angle": 0, + "content": "Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering" + }, + { + "type": "text", + "bbox": [ + 0.335, + 0.18, + 0.642, + 0.199 + ], + "angle": 0, + "content": "Jiawei Yao\\(^{1}\\) Qi Qian\\(^{2}\\) Juhua Hu\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.199, + 0.856, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Engineering and Technology, University of Washington, Tacoma, WA 98402, USA \n\\(^{2}\\) Alibaba Group, Bellevue, WA 98004, USA" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.236, + 0.692, + 0.253 + ], + "angle": 0, + "content": "{jwyao, juhuah}@uw.edu, qi.qian@alibaba-inc.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.319, + 0.474, + 0.712 + ], + "angle": 0, + "content": "Multiple clustering has gained significant attention in recent years due to its potential to reveal multiple hidden structures of data from different perspectives. The advent of deep multiple clustering techniques has notably advanced the performance by uncovering complex patterns and relationships within large datasets. However, a major challenge arises as users often do not need all the clusterings that algorithms generate, and figuring out the one needed requires a substantial understanding of each clustering result. Traditionally, aligning a user's brief keyword of interest with the corresponding vision components was challenging, but the emergence of multi-modal and large language models (LLMs) has begun to bridge this gap. In response, given unlabeled target visual data, we propose Multi-MaP, a novel method employing a multi-modal proxy learning process. It leverages CLIP encoders to extract coherent text and image embeddings, with GPT-4 integrating users' interests to formulate effective textual contexts. Moreover, reference word constraint and concept-level constraint are designed to learn the optimal text proxy according to the user's interest. Multi-MaP not only adeptly captures a user's interest via a keyword but also facilitates identifying relevant clusterings. Our extensive experiments show that MultiMaP consistently outperforms state-of-the-art methods in all benchmark multi-clustering vision tasks. Our code is available at https://github.com/Alexander-Yao/Multi-MaP." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.731, + 0.21, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.756, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Clustering, which groups data points based on their similarities, has been extensively researched, since huge amount of unlabeled data are becoming more and more available. Traditional methods [2, 16, 19] exploit general-purpose handcrafted features that are not always ideal for specific tasks. Recently, deep clustering algorithms [9, 23, 33] leverage Deep Neural Networks (DNNs) significantly improve performance. However, most algorithms yield a single data" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.3, + 0.895, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.464, + 0.893, + 0.506 + ], + "angle": 0, + "content": "Figure 1. The flow chart of Multi-MaP. Multi-MaP obtains multiple clustering results based on the high-level concepts from users and the reference words from GPT-4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.895, + 0.703 + ], + "angle": 0, + "content": "partition, while data can exhibit multiple aspects (e.g., color and species as shown in Fig. 1). Traditional multiple clustering algorithms [1, 12] have been developed to generate different partitions for varying applications, demonstrating the ability to identify multiple distinct clusterings from a dataset. Contemporary advancements in the field reveal a growing inclination among researchers to integrate deep learning methodologies for facilitating multiple clustering outcomes. Predominantly, such techniques capitalize on auto-encoders and data augmentation processes to capture a broad spectrum of feature dimensions, thereby enhancing the performance of multiple clustering [18, 27, 35]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.896, + 0.903 + ], + "angle": 0, + "content": "However, a common issue arises as users often do not require all the clusterings generated by the algorithm, and identifying the relevant ones necessitates a substantial understanding of each clustering result. Therefore, in this work, we initiate an exploration into a method that is adept at accurately capturing and reflecting a user's interest. Users typically express their interests through concise keywords (e.g., color or species), and aligning these with different visual components precisely is challenging. Fortunately, the advent of multi-modal models like CLIP [25] that aligns images to their corresponding text descriptions, can be helpful to fill this gap. However, unlike methods that employ labeled data to fine-tune pre-trained models [6, 30], multiple" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14066" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "clustering deals with environments marked by vague or undefined label categories and amounts. Consequently, given only a high-level concept from the user, it is infeasible to fine-tune the pre-trained models to capture a specific aspect of the data, without the detailed labels corresponding to the user's concept." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.186, + 0.47, + 0.442 + ], + "angle": 0, + "content": "An intuitive strategy to integrate pre-trained models into clustering is the zero-shot feature extraction, followed by clustering of the resultant embeddings. However, this approach exhibits limitations, particularly in capturing the interests of users within the dataset. Taking the multimodal model CLIP [25] as an example, when feeding image data into CLIP, regardless of what aspects of clustering the user expects, CLIP can only produce the same embeddings. Even considering the scenario that different pretrained models can capture different aspects of the same data as in [9], it is hard to tell which one matches a user's preference. Fortunately, given CLIP's ability to model image-text pairs collaboratively, we can use a user's high-level concept to trigger the corresponding feature extraction from the pre-trained encoders from CLIP. However, no previous work has studied if CLIP has this potential to uncover different aspects of images, which is the focus of this work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.656 + ], + "angle": 0, + "content": "Specifically, we propose to integrate a user's high-level concept describing the preference using a personalized textual prompt. For example, if a user's focus pertains to the color dimension of fruit, a prospective prompt might be formulated as \"a fruit with the color of *\", wherein the \"*\"\n- placeholder represents the proxy word awaiting determination using the knowledge in CLIP. Thereafter, we can learn the proxy word embedding by maximizing the similarity between the image and text embedding. However, the proxy word embedding is now searched in a continuous space while the original CLIP used discrete tokens, which can downgrade the performance. We prove that the performance can be well guaranteed by selecting the nearest token as the reference, which is unavailable in a clustering task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Fortunately, we can use the user's high-level concept as a reference, which however covers a broad range of tokens under its scope. Therefore, we propose to leverage large language models like GPT-4 to generate candidate tokens using the user's high-level concept, in which the closest can be used as a closer reference token. Furthermore, in some scenarios, users may provide multiple concepts to obtain multiple clusterings simultaneously as shown in Fig. 1, we can also introduce a negative loss with these constrastive concepts to further enhance the learning. Therefore, to capture a user's specific interest and discover personalized clustering structure hidden in the data, we propose a multimodal proxy learning method (Multi-MaP). Multi-MaP incorporates both text prompts and unlabeled images from the clustering task, and leverages CLIP to acquire their respective personalized representations using both reference" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "word and concept-level constraints. The contributions of this work can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.892, + 0.167 + ], + "angle": 0, + "content": "- We are the first to explore a deep multiple clustering method that precisely captures a user's interest(s) and generates personalized clustering(s) accordingly." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.168, + 0.892, + 0.227 + ], + "angle": 0, + "content": "- We propose a novel multi-modal proxy learning method, Multi-MaP, by leveraging the text and image encoders pre-trained by CLIP, where the user's interest can be captured by the personalized text prompts." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.228, + 0.892, + 0.302 + ], + "angle": 0, + "content": "- Considering the challenge of learning a word proxy in a continuous space while tokens in CLIP were discrete, we theoretically prove that a close reference token can help constrain the search, which motivates the proposed reference word constraint and concept-level constraint." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.304, + 0.892, + 0.363 + ], + "angle": 0, + "content": "- We conduct extensive experiments on all publicly available visual multiple clustering tasks, which empirically demonstrates the superiority of the proposed Multi-MaP, with a precise capturing of a user's interest." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.364, + 0.892, + 0.409 + ], + "angle": 0, + "content": "- Finally, to the best of our knowledge, we are the first who demonstrate that CLIP can uncover different semantic aspects of images." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.122, + 0.892, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.421, + 0.642, + 0.437 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.447, + 0.892, + 0.477 + ], + "angle": 0, + "content": "In this section, we briefly review two related directions, that is, multiple clustering and multi-modal models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.485, + 0.689, + 0.502 + ], + "angle": 0, + "content": "2.1. Multiple Clustering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Multiple clustering, as a kind of method that can discover alternative perspectives of data, has attracted considerable attention. Traditional methods for multiple clustering [11] use shallow models to find different ways of grouping data. Some of these methods rely on constraints to produce alternative clusterings. For instance, COALA [1] uses the objects in an existing clustering as constraints for creating a new clustering and Qi et al. [21] formulated multiple clustering as a constrained optimization problem. Other methods exploit different feature subspaces to generate multiple clusterings. For example, Hu et al. [12] found multiple clusterings by maximizing the eigengap in different subspaces. MNMF [34] adds the inner product of similarity matrices as a regularization term to find multiple clusterings. Some methods also use information theory to generate multiple clusterings. Gondek [8] applied conditional information bottleneck and Dang [4] used an expectation maximization framework to optimize mutual information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recently, some methods have used DNNs to find multiple clusterings and achieved better results. Wei et al. [31] proposed a deep matrix factorization based method that uses multi-view data to find multiple clusterings. ENRC [18] uses an auto-encoder to learn object features and finds multiple clusterings by optimizing a clustering objective function. iMClusts [27] exploits auto-encoders and multi-head attention to learn features from different perspectives, and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14067" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "then find multiple clusterings. AugDMC [35] leverages data augmentation to generate diverse aspects of images and learns the representations to discover multiple clusterings. Although existing deep multiple clustering methods have achieved remarkable results, they require users to exert considerable efforts to select the correct clustering they need. In this work, we aim to efficiently and effectively capture users' interests using short keywords and provide clustering results accordingly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.234, + 0.266, + 0.249 + ], + "angle": 0, + "content": "2.2. Multi-modal Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.47, + 0.375 + ], + "angle": 0, + "content": "Multi-modal learning refers to the process of learning representations from different types of input modalities, such as image data, text, or speech. As related, we focus on how vision models can benefit from natural language supervision. CLIP [25] is a notable model, which is trained with a dataset containing 400 million text-image pairs from the internet. The objective is to align images to their corresponding text using contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.473, + 0.768 + ], + "angle": 0, + "content": "Fine-tuning approaches adapt vision-language models like CLIP to specific downstream image recognition tasks. CoOp [36] and CLIP-Adapter [6] exemplify this, with the latter integrating residual style feature blending to enhance performance on various visual classification tasks. Additionally, insights from TeS [30], further elucidate the effectiveness of fine-tuning strategies in leveraging natural language supervision for enhanced visual understanding. Recognizing the scarcity of labeled data for various tasks, significant research efforts have been dedicated to enhancing zero-shot learning. Some approaches extend beyond CLIP by incorporating other large pre-trained models. For instance, VisDesc [17] harnesses the power of GPT-3 to generate comprehensive contextual descriptions corresponding to given class names, thereby demonstrating superior performance compared to CLIP's basic prompts. UPL [13] and TPT [29] leverages unlabeled data to optimize learnable input text prompts. InMaP [24] recovers the proxy of each class in the vision space with the help from the text proxy. All of these methods aim to improve the performance of vision classification tasks, while clustering is a different scenario that we do not have class names that we can exploit to extract useful information from CLIP. In this work, we consider leveraging the CLIP encoder to extract coherent text and image embeddings, and using GPT-4 to integrate users' interests for the purpose of multiple clustering." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.774, + 0.29, + 0.79 + ], + "angle": 0, + "content": "3. The Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.469, + 0.838 + ], + "angle": 0, + "content": "In this section, we first briefly review the training objective in CLIP as follows, and then describe the details of our proposed method based on that." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.312, + 0.865 + ], + "angle": 0, + "content": "3.1. Multi-modal Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Given a set of image-text pairs as \\(\\{x_{i},t_{i}\\}_{i = 1}^{n}\\), where \\(x_{i}\\) is an image and \\(t_i\\) is the corresponding text description, their" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "vision and text representations can be obtained by two encoders as \\(\\mathbf{x}_i = f(x_i)\\) and \\(\\mathbf{t}_i = h(t_i)\\). \\(f(\\cdot)\\) and \\(h(\\cdot)\\) are vision and text encoders for optimization, where \\(\\mathbf{x}_i\\) and \\(\\mathbf{t}_i\\) have the unit norm. Then, these two encoders can be learned by minimizing the contrastive loss as" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.177, + 0.892, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\min _ {f, h} \\sum_ {i} - \\log \\frac {\\exp (\\mathbf {x} _ {i} ^ {\\top} \\mathbf {t} _ {i} / \\tau)}{\\sum_ {j} \\exp (\\mathbf {x} _ {i} ^ {\\top} \\mathbf {t} _ {j}) / \\tau) - \\log \\frac {\\exp (\\mathbf {t} _ {i} ^ {\\top} \\mathbf {x} _ {i} / \\tau)}{\\sum_ {j} \\exp (\\mathbf {t} _ {i} ^ {\\top} \\mathbf {x} _ {j}) / \\tau)}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.226, + 0.684, + 0.24 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.24, + 0.892, + 0.316 + ], + "angle": 0, + "content": "This contrastive loss aims to pull the image and its description together while pushing away the irrelevant text [22], which enables the emerging multi-modal applications, e.g., zero-shot transfer [24, 25], text-to-image generation [28], etc." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.325, + 0.759, + 0.342 + ], + "angle": 0, + "content": "3.2. Multi-modal Proxy Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.408 + ], + "angle": 0, + "content": "Given the pre-trained vision and text encoders from CLIP, this work takes one step further to investigate if we can extract user-specific information from the alignment between images and text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.892, + 0.544 + ], + "angle": 0, + "content": "Concretely, given an image of fruit [12] as shown in Fig. 2, some users may be interested in only one specific property of the object, e.g., color. In this scenario, applying the vision encoder to extract the representation for the whole image can miss the preference of users. To mitigate the problem, we propose to explore the proxy representation from the image with the guidance from the text using users' preference, named Multi-Modal Proxy learning (Multi-MaP)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.892, + 0.696 + ], + "angle": 0, + "content": "Recall that CLIP is pre-trained by images and text descriptions, where the text prompt is \"a photo of a fruit\" for an image containing \"fruit\". Now given a user's preference (e.g., color), we can rewrite the prompt as \"fruit with the color of *\" denoted by \\( t_i^* \\) for image \\( x_i \\), where \" * \" is the proxy word and its token embedding is \\( \\mathbf{w}_i \\) that is learnable. Then, we can align image and text prompt representations to obtain the appropriate proxy embedding for a user's interest. Since there is no negative pairs, only the similarity between positive pairs can be optimized as" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.707, + 0.892, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\mathbf {w} _ {i}\\right) = - \\langle f \\left(x _ {i}\\right), h \\left(t _ {i} ^ {*}\\right) \\rangle \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.809 + ], + "angle": 0, + "content": "where the vision and text encoders are frozen, and \\(\\mathbf{w}_i\\) is the only variable for learning as the representation of the proxy word. By maximizing the similarity to the image representation, Multi-MaP aims to learn the optimal text proxy according to the user's interest." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, it should be noted that the text encoder was pre-trained with discrete text tokens, while the domain of \\(\\mathbf{w}_i\\) in Eqn. 1 is unconstrained. Therefore, the text representation extracted from the frozen text encoder can be inaccurate for \\(\\mathbf{w}_i\\) that degenerates the performance, which is demonstrated in the following theory." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14068" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.11, + 0.85, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.411, + 0.892, + 0.509 + ], + "angle": 0, + "content": "Figure 2. Multi-MaP framework. In the training process of Multi-MaP, the vision and text encoders are frozen and the proxy word embeddings \\(\\mathbf{w}_i\\) are learnable. Specifically, it first constructs the prompt embeddings based on the reference words provided by GPT-4 using a user's high-level concept, and then selects a reference word \\(z_i\\) for each image according to the similarity between the prompt embeddings \\(\\mathbf{t}_i\\) and the image embeddings \\(\\mathbf{x}_i\\). Then, it combines the prompt and the reference words to form the new prompt embeddings \\(\\mathbf{t}_i^*\\) and maximizes the similarity to the image representation, so the proxy word embeddings \\(\\mathbf{w}_i\\) can capture the desired image features. In addition, the proxy word embeddings \\(\\mathbf{w}_i\\) should be close to the target concept word \\(\\mathbf{u}_1\\) and the selected reference word \\(\\mathbf{z}_i\\) to construct the concept-level constraint and reference word constraint, which capture the features related to the user's interest." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.471, + 0.623 + ], + "angle": 0, + "content": "For the sake of simplicity, we assume \\( h'(t) \\in \\mathbb{R} \\) is defined on the whole set but only has the right estimation on a discrete set as \\( T = \\{t_i\\} \\) and the counterpart with the unconstrained set is denoted as \\( H(w) \\). According to the definition, we have \\( \\forall t \\in T, h'(t) = H(t) \\). The gap between the estimation from \\( h' \\) and \\( H \\) on unconstrained variable \\( w \\) can be depicted as follows." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.629, + 0.47, + 0.66 + ], + "angle": 0, + "content": "Theorem 1. Given \\( w \\notin T \\) and \\( t \\in T \\), if assuming \\( h' \\) and \\( H \\) are \\( L_h \\) and \\( L_H \\)-Lipschitz continuous, we have" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.667, + 0.416, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\| h ^ {\\prime} (w) - H (w) \\| _ {2} \\leq \\left(L _ {h} + L _ {H}\\right) \\| t - w \\| _ {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.691, + 0.371, + 0.705 + ], + "angle": 0, + "content": "Proof. According to the definition, we have" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.712, + 0.462, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left\\| h ^ {\\prime} (w) - H (w) \\right\\| _ {2} = \\left\\| h ^ {\\prime} (w) - h ^ {\\prime} (t) + h ^ {\\prime} (t) - H (w) \\right\\| _ {2} \\\\ \\leq \\| h ^ {\\prime} (w) - h ^ {\\prime} (t) \\| _ {2} + \\| h ^ {\\prime} (t) - H (w) \\| _ {2} \\\\ = \\| h ^ {\\prime} (w) - h ^ {\\prime} (t) \\| _ {2} + \\| H (t) - H (w) \\| _ {2} \\\\ \\leq \\left(L _ {h} + L _ {H}\\right) \\| t - w \\| _ {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.794, + 0.468, + 0.805 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Remark Theorem 1 implies that the distance of the estimation \\( h'(w) \\) to the ground-truth result \\( H(w) \\) is bounded by that of \\( w \\) to an arbitrary discrete token \\( t \\). Therefore, by selecting the nearest token as the reference, the bound can be improved as shown in the following corollary." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.518, + 0.892, + 0.548 + ], + "angle": 0, + "content": "Corollary 2. With the assumptions in Theorem 1 and letting \\(t' = \\arg \\min_i \\| t_i - w\\| _2\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.554, + 0.842, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\left\\| h ^ {\\prime} (w) - H (w) \\right\\| _ {2} \\leq \\left(L _ {h} + L _ {H}\\right) \\| t ^ {\\prime} - w \\| _ {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.588, + 0.729, + 0.603 + ], + "angle": 0, + "content": "3.2.1 Concept-level Constraint" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.732 + ], + "angle": 0, + "content": "According to the above analysis, a good reference \\( t' \\) can help guarantee the performance. Fortunately, the input concept (e.g., color) from the user can be leveraged as the reference to constrain the freedom of the proxy word. Therefore, given the target concept word \\( u \\), we can obtain its token embedding as \\( \\mathbf{u} = \\phi(u) \\). Then, to learn appropriate representations from the proxy embedding, the original problem can be rewritten with the constraint as" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.738, + 0.864, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathbf {w} _ {i}) = - \\langle f (x _ {i}), h \\left(t _ {i} ^ {*}\\right) \\rangle \\quad s. t. \\quad \\| \\mathbf {w} _ {i} - \\mathbf {u} \\| _ {2} ^ {2} \\leq \\lambda\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.762, + 0.772, + 0.777 + ], + "angle": 0, + "content": "The constrained problem is equivalent to" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.783, + 0.892, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\mathbf {w} _ {i}\\right) = - \\langle f \\left(x _ {i}\\right), h \\left(t _ {i} ^ {*}\\right) \\rangle + \\alpha \\| \\mathbf {w} _ {i} - \\mathbf {u} \\| _ {2} ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.807, + 0.891, + 0.836 + ], + "angle": 0, + "content": "following [3], which can be optimized effectively by gradient descent." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.889, + 0.862 + ], + "angle": 0, + "content": "3.2.2 Constrained Optimization with Reference Word" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, it is well known that the user concept is often with a large scope covering a broad range of words (e.g., color" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14069" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.145, + 0.089, + 0.403, + 0.212 + ], + "angle": 0, + "content": "
Datasets# Samples# Clusters
ALOI2882;2
Card8,02913;4
CMUface6404;20;2;4
Fruit1053;3
Fruit3604,8564;4
Stanford Cars1,2004;3
Flowers1,6004;4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.193, + 0.213, + 0.353, + 0.224 + ], + "angle": 0, + "content": "Table 1. Dataset Statistics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.47, + 0.446 + ], + "angle": 0, + "content": "covers all including but not limited to 'red', 'blue', 'green', etc.). As suggested by our above theoretical analysis, it is desired if the reference is as close as possible. In a clustering scenario and given only the user's high-level concept, it is challenging to find a closer reference word to further constrain the proxy learning. Fortunately, with the development of large language models (LLMs), we can leverage them (e.g., GPT-4) to provide relevant words according to a user's high-level concept as the candidate set and develop a selection strategy to obtain a closer reference word for each image. While the responses gathered from GPT-4 might not always precisely align with the data's ground truth, they indisputably furnish valuable candidate features, enriching the capabilities of Multi-MaP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.471, + 0.645 + ], + "angle": 0, + "content": "To elucidate, considering the task of clustering a fruit dataset based on the concept of color, we can pose a query to GPT-4 as \"What are the common colors of fruit?\" The response we obtain is \"Common colors of fruit include red, yellow, green, orange, purple, and blue\". The colors enumerated in this response can serve as reference word candidates. It is worth noting that while the color spectrum derived from this method might surpass the actual colors present in the data, we collect all of them into the candidate set of reference words \\(\\{z_k\\}_k\\), where \\(z_1\\): \"red\", \\(z_2\\): \"yellow\", etc. Then, their text representations can be obtained from the prompt \\(t_k\\) as \"fruit with the color of \\(z_k\\)\". Given the image \\(x_i\\), the closest reference can be observed as" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.657, + 0.348, + 0.678 + ], + "angle": 0, + "content": "\\[\nz _ {i} = \\arg \\max _ {k} \\left\\langle \\mathbf {x} _ {i}, \\mathbf {t} _ {k} \\right\\rangle\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.751 + ], + "angle": 0, + "content": "where \\(\\mathbf{t}_k = h(t_k)\\). After that, \\(\\mathbf{w}_i\\) can be initialized with the token embedding of \\(z_i\\) as \\(\\mathbf{z}_i = \\phi(z_i)\\). Moreover, we change the regularization using a closer reference word compared to the high-level concept as" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.763, + 0.469, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\mathbf {w} _ {i}\\right) = - \\langle f \\left(x _ {i}\\right), h \\left(t _ {i} ^ {*}\\right) \\rangle + \\alpha \\| \\mathbf {w} _ {i} - \\mathbf {z} _ {i} \\| _ {2} ^ {2} \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.282, + 0.817 + ], + "angle": 0, + "content": "3.2.3 Contrastive Concepts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.902 + ], + "angle": 0, + "content": "In some application scenarios, one user may need more than one clustering and provide high-level concepts as \\(\\{u_j\\}\\), e.g., \\(u_1\\): \"color\", \\(u_2\\): \"species\", etc. For the concept \"color\", the irrelevant concept \"species\" can be leveraged as the negative constraint for the learning of proxy word. Concretely," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "let \\( u_{w} \\) denote the target concept word, a contrastive loss can be adopted as regularization" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.133, + 0.808, + 0.171 + ], + "angle": 0, + "content": "\\[\nR (\\mathbf {w} _ {i}) = - \\log \\frac {\\exp (\\mathbf {w} _ {i} ^ {\\top} \\mathbf {u} _ {w})}{\\sum_ {j} \\exp (\\mathbf {w} _ {i} ^ {\\top} \\mathbf {u} _ {j})}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.181, + 0.712, + 0.195 + ], + "angle": 0, + "content": "and the final objective becomes" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.208, + 0.892, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} \\left(\\mathbf {w} _ {i}\\right) = \\\\ - \\left\\langle f \\left(x _ {i}\\right), h \\left(t _ {i} ^ {*}\\right) \\right\\rangle + \\alpha \\| \\mathbf {w} _ {i} - \\mathbf {z} _ {i} \\| _ {2} ^ {2} - \\beta \\log \\frac {\\exp \\left(\\mathbf {w} _ {i} ^ {\\top} \\mathbf {u} _ {w}\\right)}{\\sum_ {j} \\exp \\left(\\mathbf {w} _ {i} ^ {\\top} \\mathbf {u} _ {j}\\right)} \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.367 + ], + "angle": 0, + "content": "where the first term is to infer the user-specific feature, while the latter two terms constrain the proxy word to the reference words for the appropriate representation extraction from the pre-trained text encoder. The overall framework of Multi-MaP is illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.374, + 0.634, + 0.391 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.474 + ], + "angle": 0, + "content": "To demonstrate our proposed method, we evaluate MultiMaP on all publicly available image datasets in multiple clustering, including ALOI [7], Stanford Cars [14], Card [35], CMUface [10], Flowers [20], Fruit [12], and Fruit360 [35] as summarized in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.893, + 0.672 + ], + "angle": 0, + "content": "We compare Multi-MaP against five state-of-the-art methods: MSC [12] is a traditional multiple clustering method that uses hand-crafted features; MCV [9] leverages multiple feature extractors to represent different \"views\" of the same data and employs a multi-input neural network to enhance clustering outcomes; ENRC [18] is a deep multiple clustering method that integrates auto-encoder and clustering objective to generate different clusterings; iM-Clusts [27] makes use of the expressive representational power of deep autoencoders and multi-head attention to accomplish multiple clusterings; AugDMC [35] leverages augmentations to learn different image representations to achieve multiple clustering." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.677, + 0.698 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We employ Adam and set momentum as 0.9 to train the model for 1000 epochs. All hyperparameters are searched according to the loss score of Multi-MaP, where the learning rate is searched in \\(\\{0.1, 0.05, 0.01, 0.005, 0.001, 0.0005\\}\\), weight decay is in \\(\\{0.0005, 0.0001, 0.00005, 0.00001, 0\\}\\), \\(\\alpha, \\beta\\) are in \\(\\{0.0, 0.1, 0.2, \\ldots, 1.0\\}\\), and \\(\\lambda\\) is fixed as 1 for all the experiments. For the non-pre-trained methods, we perform k-means [15] 10 times due to its randomness and evaluate the average clustering performance using two quantitative metrics, that is, Normalized Mutual Information (NMI) [32] and Rand index (RI) [26]. These measures range in \\([0, 1]\\), and higher scores imply more accurate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14070" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.14, + 0.089, + 0.827, + 0.319 + ], + "angle": 0, + "content": "
DatasetClusteringMSCMCVENRCiMClustersAugDMCMulti-MaP
NMIRINMIRINMIRINMIRINMIRINMIRI
ALOIColor0.15630.34280.69820.74390.98330.98921.00001.00001.00001.00001.00001.0000
Shape0.29680.51990.73590.82610.97320.98610.99630.99891.00001.00001.00001.0000
FruitColor0.68860.80510.62660.76850.71030.85110.73510.86320.85170.91080.86190.9526
Species0.16270.60450.27330.65970.31870.65360.30290.67430.35460.73991.00001.0000
Fruit360Color0.25440.60540.37760.67910.42640.68680.40970.68410.45940.73920.62390.8243
Species0.21840.58050.29850.61760.41420.69840.38610.67320.51390.74300.52840.7582
CardOrder0.08070.78050.07920.71280.12250.73130.11440.76580.14400.82670.36530.8587
Suits0.04970.35870.04300.36380.06760.38010.07160.37150.08730.42280.27340.7039
CMUfaceEmotion0.12840.67360.14330.52680.15920.66300.04220.59320.01610.53670.17860.7105
Glass0.14200.57450.12010.49050.14930.62090.19290.56270.10390.53610.34020.7068
Identity0.38920.73260.46370.62470.56070.76350.51090.82600.58750.83340.66250.9496
Pose0.36870.63220.32540.60280.22900.50290.44370.61140.13200.55170.46930.6624
Stanford CarsColor0.23310.61580.21030.58020.24650.67790.23360.65520.27360.75250.73600.9193
Type0.13250.53360.16500.56340.20630.62170.19630.56430.23640.73560.63550.8399
FlowersColor0.25610.59650.29380.58600.33290.62140.31690.61270.35560.69310.64260.7984
Species0.13260.52730.15610.60650.18940.61950.18870.60770.19960.62270.60130.8103
" + }, + { + "type": "table_caption", + "bbox": [ + 0.197, + 0.32, + 0.772, + 0.334 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison. The significantly best results with \\(95\\%\\) confidence are in bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.469, + 0.379 + ], + "angle": 0, + "content": "clustering results. The experiments are conducted with GPU NVIDIA GeForce RTX 2080 Ti." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.47, + 0.623 + ], + "angle": 0, + "content": "It should also be noted that some data are difficult to obtain corresponding candidate labels from GPT-4 or the labels do not provide semantic features, such as names. For example, for the identity clustering for the CMUface dataset [10], different identities represent different people and the semantic meaning of names should not affect the clustering results. In this case, we randomly extract 10 words from WordNet [5] as reference words, in order to make the candidate labels more distinctive. For instance, we randomly choose \"abstain, candid, function, haphazard, knot, luxury, nonchalance, pension, resilience, taciturn\" for the above scenario. Furthermore, all publicly available multiple clustering datasets provide each ground-truth clustering a high-level concept, e.g., 'shape', 'pose', etc. Therefore, in the experiment, we directly use them as users' preferences for our evaluation purposes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.312, + 0.65 + ], + "angle": 0, + "content": "4.2. Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.469, + 0.748 + ], + "angle": 0, + "content": "In our experiments, after we obtain the proxy word embedding of each image for a desired concept, we feed them to k-means to obtain the corresponding clustering. Since k-means is random, we repeat ten times and the average performance is reported in Table 2. The best results are marked by bold numbers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.469, + 0.825 + ], + "angle": 0, + "content": "We can observe that Multi-MaP outperforms the baselines in all the cases, indicating the superiority of the proposed method. This also shows a strong generalization ability of the pre-trained model by CLIP, which can capture the features of data in different aspects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Since our method uses the CLIP encoder and GPT-4 to obtain clustering results, a natural question arises how the performance would be if we directly use them in a zero-short manner. Therefore, we provide two zero-shot variants of CLIP, that is, \\(\\mathrm{CLIP}_{\\mathrm{GPT}}\\) that uses GPT-4 to obtain the can" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.347, + 0.89, + 0.562 + ], + "angle": 0, + "content": "
DatasetClusteringCLIPGPTCLIPlabelMulti-MaP
NMIRINMIRINMIRI
ALOIColor0.85810.94071.00001.00001.00001.0000
Shape1.00001.00001.00001.00001.00001.0000
FruitColor0.79120.90750.86290.97800.86190.9526
Species0.97930.99191.00001.00001.00001.0000
Fruit360Color0.56130.73050.57460.76730.62390.8243
Species0.43700.75520.53640.76310.52840.7582
CardOrder0.35180.84580.35180.84580.36530.8587
Suits0.27110.61230.27110.61230.27340.7039
CMUfaceEmotion0.15760.65320.15900.66190.17860.7105
Glass0.29050.68690.46860.75050.34020.7068
Identity0.19980.63880.26770.75450.66250.9496
Pose0.40880.64730.46910.64090.46930.6624
Stanford CarsColor0.65390.82370.68300.86420.73600.9193
Type0.62070.79310.64290.84560.63550.8399
FlowersColor0.56530.76290.58280.78360.64260.7984
Species0.56200.75530.60190.79960.60130.8103
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.563, + 0.89, + 0.589 + ], + "angle": 0, + "content": "Table 3. Variants of CLIP. The significantly best results with \\(95\\%\\) confidence are in bold." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.704 + ], + "angle": 0, + "content": "didate labels and predicts labels through zero-shot classification with all candidate labels as class names, and \\(\\mathrm{CLIP}_{\\mathrm{label}}\\) that performs zero-shot classification with all ground truth labels. It should be noted that \\(\\mathrm{CLIP}_{\\mathrm{label}}\\) uses an unfair setting with a ground-truth label set known in advance, which is expected to provide the best performance using CLIP in a zero-short manner. The results are shown in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.81 + ], + "angle": 0, + "content": "As expected, \\(\\mathrm{CLIP}_{\\mathrm{label}}\\) achieves better performance than \\(\\mathrm{CLIP}_{\\mathrm{GPT}}\\) in almost all cases, since \\(\\mathrm{CLIP}_{\\mathrm{label}}\\) uses a fixed ground truth label set, while \\(\\mathrm{CLIP}_{\\mathrm{GPT}}\\) uses candidate labels that may not match the ground truth exactly, introducing noise. Note that \\(\\mathrm{CLIP}_{\\mathrm{GPT}}\\) and \\(\\mathrm{CLIP}_{\\mathrm{label}}\\) achieve the same results on Cards, because the candidate labels provided by GPT-4 are exactly the same as the true labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Besides, Multi-MaP performs better than \\(\\mathrm{CLIP}_{\\mathrm{GPT}}\\) in almost all cases, indicating that the proposed method can learn more effective features through its training process. Moreover, although \\(\\mathrm{CLIP}_{\\mathrm{label}}\\) uses the ground truth, which is expected to be the best, Multi-MaP still outperforms \\(\\mathrm{CLIP}_{\\mathrm{label}}\\) in some cases, such as the clustering of color for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14071" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.142, + 0.088, + 0.832, + 0.381 + ], + "angle": 0, + "content": "
Multi-MaPpMulti-MaPcMulti-MaPrMulti-MaPcrMulti-MaP
ModulesProxy Learning✓✓✓✓✓
Concept Word×✓×✓✓
Reference Word××✓✓✓
Contrastive Concepts××××✓
NMI↑RI↑NMI↑RI↑NMI↑RI↑NMI↑RI↑NMI↑RI↑
ALOI [7]Color0.96190.98261.00001.00000.97950.98691.00001.00001.00001.0000
Shape1.00001.00001.00001.00001.00001.00001.00001.00001.00001.0000
Fruit [12]Color0.76420.84390.82150.92830.81360.90730.84840.93080.86190.9526
Species1.00001.00001.00001.00001.00001.00001.00001.00001.00001.0000
Fruit360 [35]Color0.56430.76650.62170.78360.59100.77460.60890.79650.62390.8243
Species0.50770.73680.51370.74360.50940.74250.51990.74280.52840.7582
Card [35]Order0.19320.81520.35680.84720.31130.82290.36160.80940.36530.8587
Suits0.23750.62820.26960.66410.24980.63650.25620.65990.27340.7039
CMUface [10]Emotion0.16900.61700.17140.62290.16970.63600.17130.68430.17860.7105
Glass0.31120.69110.32690.71360.31620.69170.33700.71080.34020.7068
Identity0.56170.82340.62430.83590.58390.82630.63910.89460.66250.9496
Pose0.43610.63860.45500.64990.43810.64290.43870.64890.46930.6624
Stanford cars [14]Color0.59390.78350.68360.86590.67290.86380.71120.91170.73600.9193
Type0.55690.79960.63830.82710.60910.80460.62890.81810.63550.8399
Flowers [20]Color0.57830.77230.58300.78330.59870.78490.62160.79410.64260.7984
Species0.57040.76080.57440.78420.57230.78110.58460.78920.60130.8103
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.39, + 0.877, + 0.407 + ], + "angle": 0, + "content": "Table 4. Components ablation. All of our components boost performance consistently in all benchmark multi-clustering vision tasks." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.428, + 0.275, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.552, + 0.262, + 0.564 + ], + "angle": 0, + "content": "(a) Parameter sensitivity of \\(\\alpha\\)" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.428, + 0.457, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.552, + 0.444, + 0.564 + ], + "angle": 0, + "content": "(b) Parameter sensitivity of \\(\\beta\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.566, + 0.434, + 0.58 + ], + "angle": 0, + "content": "Figure 3. Parameter analysis of \\(\\alpha\\) and \\(\\beta\\) on Fruit [12]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.47, + 0.695 + ], + "angle": 0, + "content": "Fruit360 dataset. This is because CLIP is more inclined to capture the features of one aspect of the data, while Multi-MaP learns better embedding of different aspects by training with the supervision of users' interests. Furthermore, Multi-MaP also achieves very competitive results as \\(\\mathrm{CLIP}_{\\mathrm{label}}\\) in the remaining cases, which further demonstrates the effectiveness of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.709, + 0.231, + 0.725 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.884 + ], + "angle": 0, + "content": "To validate the effectiveness of the proposed method, we show the gain from four components in Multi-MaP in Table 4. Let \"Multi-MaP \\( p \\)\" denote the proxy learning without concept-level constraint, \"Multi-MaP \\( c \\)\" denote the variant optimized by solely applying concept word to constrain the freedom of the proxy word, \"Multi-MaP \\( r \\)\" denote the variant optimized by reference word provided by GPT-4 to find a closer reference word to further constrain the proxy learning, and \"Multi-MaP \\( cr \\)\" denote the variant leveraged both concept word and reference word." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We can observe that the proxy learning only with concept" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.613 + ], + "angle": 0, + "content": "word or reference word, i.e., \\(\\mathrm{Multi - MaP}_c\\) and \\(\\mathrm{Multi - MaP}_r\\), performs better than \\(\\mathrm{Multi - MaP}_p\\). This shows that the proposed concept-level constraint and constrained optimization with reference words play an important role in the model as demonstrated by our theoretical analysis. Moreover, the model with combined components, i.e., \\(\\mathrm{Multi - MaP}_{cr}\\), achieves better results than \\(\\mathrm{Multi - MaP}_r\\) and \\(\\mathrm{Multi - MaP}_c\\). This indicates the effectiveness of the combination of reference words and reference concepts. Finally, our proposal using all including the contrastive concepts can further improve the performance, and thus provide the best results on all cases. This further demonstrates our proposal." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.621, + 0.689, + 0.637 + ], + "angle": 0, + "content": "4.4. Parameter Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We further investigate the effect of the reference word constraint weight \\(\\alpha\\) and concept-level constraint weight \\(\\beta\\) varying from 0.0 to 0.5. The results of Multi-MaP on Fruit datasets are shown in Fig. 3 (a) and Fig. 3 (b), respectively. As \\(\\alpha\\) and \\(\\beta\\) change, the proposed method keeps a species score of 1, since the image encoder can capture very effective species features. The above results also show that the proposed method can effectively capture useful information from images, reference words, and target concepts. As \\(\\alpha\\) increases, the proposed method first increases and then decreases, and reaches the maximum values at \\(\\alpha = 0.4\\). Similar results can be observed for \\(\\beta\\) that the performance of the proposed method first increases and then decreases as \\(\\beta\\) increases, and reaches the maximum values at \\(\\beta = 0.3\\). Indicating a suitable value of \\(\\alpha\\) or \\(\\beta\\) is helpful for Multi-MaP to obtain more effective embeddings for multiple clustering tasks. More studies such as efficiency analysis can be found" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14072" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.101, + 0.288, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.2, + 0.29, + 0.214 + ], + "angle": 0, + "content": "(a) Color using ground truth" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.102, + 0.481, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.2, + 0.495, + 0.214 + ], + "angle": 0, + "content": "(b) Color using reference words" + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.102, + 0.673, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.2, + 0.666, + 0.214 + ], + "angle": 0, + "content": "(c) Color of Multi-MaP" + }, + { + "type": "image_footnote", + "bbox": [ + 0.698, + 0.133, + 0.838, + 0.143 + ], + "angle": 0, + "content": "red yellow" + }, + { + "type": "image_footnote", + "bbox": [ + 0.701, + 0.144, + 0.837, + 0.154 + ], + "angle": 0, + "content": "green orange" + }, + { + "type": "image_footnote", + "bbox": [ + 0.701, + 0.155, + 0.825, + 0.165 + ], + "angle": 0, + "content": "purple blue" + }, + { + "type": "list", + "bbox": [ + 0.701, + 0.144, + 0.837, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.227, + 0.285, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.326, + 0.295, + 0.34 + ], + "angle": 0, + "content": "(e) Species using ground truth" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.226, + 0.477, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.326, + 0.499, + 0.34 + ], + "angle": 0, + "content": "(f) Species using reference words" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.228, + 0.671, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.326, + 0.672, + 0.34 + ], + "angle": 0, + "content": "(g) Species of Multi-MaP" + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.2, + 0.834, + 0.214 + ], + "angle": 0, + "content": "(d) Legend of color" + }, + { + "type": "image_footnote", + "bbox": [ + 0.701, + 0.225, + 0.859, + 0.322 + ], + "angle": 0, + "content": "- apple\n- orange\n- banana\n- strawberry\n- grapes\n- raspberry\n- blueberry\n- cherry\n- pear\n- plum\n- peach\n- nectarine\n- pineapple\n- kiwi\n- watermelon\n- cantaloupe\n- mango" + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.326, + 0.841, + 0.34 + ], + "angle": 0, + "content": "(h) Legend of species" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.351, + 0.895, + 0.395 + ], + "angle": 0, + "content": "Figure 4. Visualization of feature embeddings and related labels. The points represent the image or pseudo-word embeddings, and the triangles represent the prompt or label embeddings. Different colors represent different labels, which are indicated by the text next to the triangles." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.4, + 0.221, + 0.415 + ], + "angle": 0, + "content": "in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.424, + 0.214, + 0.439 + ], + "angle": 0, + "content": "4.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.473, + 0.901 + ], + "angle": 0, + "content": "To further demonstrate the effectiveness of the proposed method, we visualize the representations obtained in \\(\\mathrm{CLIP}_{\\mathrm{label}}\\), \\(\\mathrm{CLIP}_{\\mathrm{GPT}}\\), and Multi-MaP. Specifically, for \\(\\mathrm{CLIP}_{\\mathrm{label}}\\) and \\(\\mathrm{CLIP}_{\\mathrm{GPT}}\\), we visualize the image representations as well as prompts generated with real labels and candidate labels, respectively. For the Multi-MaP, we visualize the word embedding \\(w_{*}\\) and the candidate labels selected for initialization. The results are shown in Fig. 4. For species clustering, we can see that image embeddings show very clear boundaries and correspond well to the prompt for \\(\\mathrm{CLIP}_{\\mathrm{label}}\\), which indicates that CLIP can effectively capture the features of the species in the data. \\(\\mathrm{CLIP}_{\\mathrm{GPT}}\\) uses the candidate labels to generate prompts, which introduces more noise, but benefits from the CLIP text encoder, the image embedding can keep a relatively far distance from most of the irrelevant prompts. However, since there are still a few images that are labeled as peaches (i.e., a noisy label), it performs slightly worse than \\(\\mathrm{CLIP}_{\\mathrm{label}}\\), as shown in Fig. 4(e). Besides, Multi-MaP can capture the image and users' interests in the training process, therefore it compensates for the shortcomings of \\(\\mathrm{CLIP}_{\\mathrm{GPT}}\\) and achieves better results. On the other hand, for color clustering, the prompts are farther away from \\(\\mathrm{CLIP}_{\\mathrm{label}}\\) and \\(\\mathrm{CLIP}_{\\mathrm{GPT}}\\), that indicates the image embeddings mainly capture the features of species, which have no direct connection with color. \\(\\mathrm{CLIP}_{\\mathrm{GPT}}\\) generates the prompt from the candidate label, which has more noise than the ground truth label, resulting in worse performance than \\(\\mathrm{CLIP}_{\\mathrm{label}}\\). The proposed method can distinguish different colors more clearly, because it can learn from the user's interest and capture the color-related features. How" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.4, + 0.892, + 0.474 + ], + "angle": 0, + "content": "ever, some red color embeddings are closer to purple, because some images in the datasets are actually purple, but labeled as red. To sum up, the proposed method can learn more effective embeddings based on the users' interests for multiple clustering tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.482, + 0.619, + 0.498 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.807 + ], + "angle": 0, + "content": "To conclude, our study thoroughly investigates the significant challenges that current advanced deep learning techniques face in multiple clustering. A key issue is that users often do not need every clustering result produced by an algorithm, and selecting the most relevant one requires an in-depth understanding of each outcome. To overcome this challenge, the proposed method introduces a novel multimodal proxy learning process, which effectively aligns a user's brief keyword describing the interest with the corresponding vision components. By integrating a multi-modal model and GPT-4 to precisely capture a user's interest using a keyword, the proposed approach uses both reference word constraint and concept-level constraint to discover personalized clustering result(s), which can also lead to enhanced performance. Experiments on diverse datasets demonstrate the superiority of the proposed method in multiple clustering tasks with a precise capture of a user's interest. The proposed method is limited by data with semantic meaningful labels, although we can use WordNet to help, whose comprehensive study will be our future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.681, + 0.832 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This research is supported in part by Advata Gift funding. All opinions, findings, conclusions and recommendations in this paper are those of the author and do not necessarily reflect the views of the funding agencies." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14073" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Eric Bae and James Bailey. Coala: A novel approach for the extraction of an alternate clustering of high quality and high dissimilarity. In ICDM, pages 53-62. IEEE, 2006. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[2] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning. Springer, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.468, + 0.215 + ], + "angle": 0, + "content": "[3] Stephen P. Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University Press, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[4] Xuan Hong Dang and James Bailey. Generation of alternative clusterings using the cami approach. In Proceedings of the 2010 SIAM International Conference on Data Mining, pages 118-129. SIAM, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.275, + 0.471, + 0.315 + ], + "angle": 0, + "content": "[5] Christiane Fellbaum. Wordnet. In Theory and applications of ontology: computer applications, pages 231-243. Springer, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.47, + 0.385 + ], + "angle": 0, + "content": "[6] 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. International Journal of Computer Vision, pages 1-15, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[7] Jan-Mark Geusebroek, Gertjan J Burghouts, and Arnold WM Smeulders. The amstedam library of object images. International Journal of Computer Vision, 61:103-112, 2005. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.468, + 0.5 + ], + "angle": 0, + "content": "[8] David Gondek and Thomas Hofmann. Conditional information bottleneck clustering. In 3rd IEEE international conference on data mining, workshop on clustering large data sets, pages 36-42, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.468, + 0.556 + ], + "angle": 0, + "content": "[9] Joris Guérin and Byron Boots. Improving image clustering with multiple pretrained cnn feature extractors. In *British Machine Vision Conference* 2018, BMVC 2018, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.559, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[10] Stephan Gunnemann, Ines Farber, Matthias Rüdiger, and Thomas Seidl. Smvc: semi-supervised multi-view clustering in subspace projections. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 253-262, 2014. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[11] Juhua Hu and Jian Pei. Subspace multi-clustering: a review. Knowledge and information systems, 56(2):257-284, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[12] Juhua Hu, Qi Qian, Jian Pei, Rong Jin, and Shenghuo Zhu. Finding multiple stable clusterings. Knowledge and Information Systems, 51(3):991-1021, 2017. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[13] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[14] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[15] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[16] James MacQueen et al. Some methods for classification and analysis of multivariate observations. In Proceedings of the fifth Berkeley symposium on mathematical statistics and probability, pages 281-297. Oakland, CA, USA, 1967. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[17] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[18] Lukas Miklautz, Dominik Mautz, Muzaffer Can Altinigneli, Christian Böhm, and Claudia Plant. Deep embedded nonredundant clustering. In Proceedings of the AAAI conference on artificial intelligence, pages 5174-5181, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.195, + 0.892, + 0.236 + ], + "angle": 0, + "content": "[19] Andrew Ng, Michael Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 14, 2001. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.239, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[20] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.297, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[21] ZiJie Qi and Ian Davidson. A principled and flexible framework for finding alternative clusterings. In SIGKDD, pages 717-726, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.341, + 0.892, + 0.395 + ], + "angle": 0, + "content": "[22] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Tacoma Tacoma, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In ICCV, pages 6449-6457. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.399, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[23] Qi Qian, Yuanhong Xu, Juhua Hu, Hao Li, and Rong Jin. Unsupervised visual representation learning by online constrained k-means. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16640-16649, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.526 + ], + "angle": 0, + "content": "[24] Qi Qian, Yuanhong Xu, and Juhua Hu. Intra-modal proxy learning for zero-shot visual categorization with clip. In Thirty-seventh Conference on Neural Information Processing Systems, NeurIPS 2023, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.528, + 0.892, + 0.61 + ], + "angle": 0, + "content": "[25] 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, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.614, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[26] William M Rand. Objective criteria for the evaluation of clustering methods. Journal of the American Statistical Association, 66(336):846-850, 1971. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.658, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[27] Liangrui Ren, Guoxian Yu, Jun Wang, Lei Liu, Carlotta Domeniconi, and Xiangliang Zhang. A diversified attention model for interpretable multiple clusterings. IEEE Transactions on Knowledge and Data Engineering, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[29] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Advances in Neural Information Processing Systems, 35:14274-14289, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[30] Junyang Wang, Yuanhong Xu, Juhua Hu, Ming Yan, Jitao Sang, and Qi Qian. Improved visual fine-tuning with natural language supervision. arXiv preprint arXiv:2304.01489, 2023. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14074" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[31] Shaowei Wei, Jun Wang, Guoxian Yu, Carlotta Domeni-coni, and Xiangliang Zhang. Multi-view multiple clusterings using deep matrix factorization. In Proceedings of the AAAI conference on artificial intelligence, pages 6348–6355, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[32] JV White, Sam Steingold, and CG Fournelle. Performance metrics for group-detection algorithms. Proceedings of Interface, 2004, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[33] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[34] Sen Yang and Lijun Zhang. Non-redundant multiple clustering by nonnegative matrix factorization. Machine Learning, 106(5):695-712, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.333 + ], + "angle": 0, + "content": "[35] Jiawei Yao, Enbei Liu, Maham Rashid, and Juhua Hu. Augdmc: Data augmentation guided deep multiple clustering. In INNS DLIA@IJCNN, 2023. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[36] 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. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14075" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/f9afe157-eaaf-440c-afcf-7604121ce960_origin.pdf b/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/f9afe157-eaaf-440c-afcf-7604121ce960_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7dbbb788d77c3ea13b20e7958a21efc9134a9da0 --- /dev/null +++ b/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/f9afe157-eaaf-440c-afcf-7604121ce960_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2547b5985b89caa94d1d7bbb0612fa5987eb93f50f4fbbbc3d9dab3571505e41 +size 1518290 diff --git a/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/full.md b/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e560681d976971e045b87ac30c90dd2323a88acb --- /dev/null +++ b/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/full.md @@ -0,0 +1,334 @@ +# Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering + +Jiawei Yao $^{1}$ Qi Qian $^{2}$ Juhua Hu $^{1*}$ + +$^{1}$ School of Engineering and Technology, University of Washington, Tacoma, WA 98402, USA + $^{2}$ Alibaba Group, Bellevue, WA 98004, USA + +{jwyao, juhuah}@uw.edu, qi.qian@alibaba-inc.com + +# Abstract + +Multiple clustering has gained significant attention in recent years due to its potential to reveal multiple hidden structures of data from different perspectives. The advent of deep multiple clustering techniques has notably advanced the performance by uncovering complex patterns and relationships within large datasets. However, a major challenge arises as users often do not need all the clusterings that algorithms generate, and figuring out the one needed requires a substantial understanding of each clustering result. Traditionally, aligning a user's brief keyword of interest with the corresponding vision components was challenging, but the emergence of multi-modal and large language models (LLMs) has begun to bridge this gap. In response, given unlabeled target visual data, we propose Multi-MaP, a novel method employing a multi-modal proxy learning process. It leverages CLIP encoders to extract coherent text and image embeddings, with GPT-4 integrating users' interests to formulate effective textual contexts. Moreover, reference word constraint and concept-level constraint are designed to learn the optimal text proxy according to the user's interest. Multi-MaP not only adeptly captures a user's interest via a keyword but also facilitates identifying relevant clusterings. Our extensive experiments show that MultiMaP consistently outperforms state-of-the-art methods in all benchmark multi-clustering vision tasks. Our code is available at https://github.com/Alexander-Yao/Multi-MaP. + +# 1. Introduction + +Clustering, which groups data points based on their similarities, has been extensively researched, since huge amount of unlabeled data are becoming more and more available. Traditional methods [2, 16, 19] exploit general-purpose handcrafted features that are not always ideal for specific tasks. Recently, deep clustering algorithms [9, 23, 33] leverage Deep Neural Networks (DNNs) significantly improve performance. However, most algorithms yield a single data + +![](images/a2cba1582001304a37fce6343670486d636254288166348832ff61b13105129c.jpg) +Figure 1. The flow chart of Multi-MaP. Multi-MaP obtains multiple clustering results based on the high-level concepts from users and the reference words from GPT-4. + +partition, while data can exhibit multiple aspects (e.g., color and species as shown in Fig. 1). Traditional multiple clustering algorithms [1, 12] have been developed to generate different partitions for varying applications, demonstrating the ability to identify multiple distinct clusterings from a dataset. Contemporary advancements in the field reveal a growing inclination among researchers to integrate deep learning methodologies for facilitating multiple clustering outcomes. Predominantly, such techniques capitalize on auto-encoders and data augmentation processes to capture a broad spectrum of feature dimensions, thereby enhancing the performance of multiple clustering [18, 27, 35]. + +However, a common issue arises as users often do not require all the clusterings generated by the algorithm, and identifying the relevant ones necessitates a substantial understanding of each clustering result. Therefore, in this work, we initiate an exploration into a method that is adept at accurately capturing and reflecting a user's interest. Users typically express their interests through concise keywords (e.g., color or species), and aligning these with different visual components precisely is challenging. Fortunately, the advent of multi-modal models like CLIP [25] that aligns images to their corresponding text descriptions, can be helpful to fill this gap. However, unlike methods that employ labeled data to fine-tune pre-trained models [6, 30], multiple + +clustering deals with environments marked by vague or undefined label categories and amounts. Consequently, given only a high-level concept from the user, it is infeasible to fine-tune the pre-trained models to capture a specific aspect of the data, without the detailed labels corresponding to the user's concept. + +An intuitive strategy to integrate pre-trained models into clustering is the zero-shot feature extraction, followed by clustering of the resultant embeddings. However, this approach exhibits limitations, particularly in capturing the interests of users within the dataset. Taking the multimodal model CLIP [25] as an example, when feeding image data into CLIP, regardless of what aspects of clustering the user expects, CLIP can only produce the same embeddings. Even considering the scenario that different pretrained models can capture different aspects of the same data as in [9], it is hard to tell which one matches a user's preference. Fortunately, given CLIP's ability to model image-text pairs collaboratively, we can use a user's high-level concept to trigger the corresponding feature extraction from the pre-trained encoders from CLIP. However, no previous work has studied if CLIP has this potential to uncover different aspects of images, which is the focus of this work. + +Specifically, we propose to integrate a user's high-level concept describing the preference using a personalized textual prompt. For example, if a user's focus pertains to the color dimension of fruit, a prospective prompt might be formulated as "a fruit with the color of *", wherein the "*" +- placeholder represents the proxy word awaiting determination using the knowledge in CLIP. Thereafter, we can learn the proxy word embedding by maximizing the similarity between the image and text embedding. However, the proxy word embedding is now searched in a continuous space while the original CLIP used discrete tokens, which can downgrade the performance. We prove that the performance can be well guaranteed by selecting the nearest token as the reference, which is unavailable in a clustering task. + +Fortunately, we can use the user's high-level concept as a reference, which however covers a broad range of tokens under its scope. Therefore, we propose to leverage large language models like GPT-4 to generate candidate tokens using the user's high-level concept, in which the closest can be used as a closer reference token. Furthermore, in some scenarios, users may provide multiple concepts to obtain multiple clusterings simultaneously as shown in Fig. 1, we can also introduce a negative loss with these constrastive concepts to further enhance the learning. Therefore, to capture a user's specific interest and discover personalized clustering structure hidden in the data, we propose a multimodal proxy learning method (Multi-MaP). Multi-MaP incorporates both text prompts and unlabeled images from the clustering task, and leverages CLIP to acquire their respective personalized representations using both reference + +word and concept-level constraints. The contributions of this work can be summarized as: + +- We are the first to explore a deep multiple clustering method that precisely captures a user's interest(s) and generates personalized clustering(s) accordingly. +- We propose a novel multi-modal proxy learning method, Multi-MaP, by leveraging the text and image encoders pre-trained by CLIP, where the user's interest can be captured by the personalized text prompts. +- Considering the challenge of learning a word proxy in a continuous space while tokens in CLIP were discrete, we theoretically prove that a close reference token can help constrain the search, which motivates the proposed reference word constraint and concept-level constraint. +- We conduct extensive experiments on all publicly available visual multiple clustering tasks, which empirically demonstrates the superiority of the proposed Multi-MaP, with a precise capturing of a user's interest. +- Finally, to the best of our knowledge, we are the first who demonstrate that CLIP can uncover different semantic aspects of images. + +# 2. Related Work + +In this section, we briefly review two related directions, that is, multiple clustering and multi-modal models. + +# 2.1. Multiple Clustering + +Multiple clustering, as a kind of method that can discover alternative perspectives of data, has attracted considerable attention. Traditional methods for multiple clustering [11] use shallow models to find different ways of grouping data. Some of these methods rely on constraints to produce alternative clusterings. For instance, COALA [1] uses the objects in an existing clustering as constraints for creating a new clustering and Qi et al. [21] formulated multiple clustering as a constrained optimization problem. Other methods exploit different feature subspaces to generate multiple clusterings. For example, Hu et al. [12] found multiple clusterings by maximizing the eigengap in different subspaces. MNMF [34] adds the inner product of similarity matrices as a regularization term to find multiple clusterings. Some methods also use information theory to generate multiple clusterings. Gondek [8] applied conditional information bottleneck and Dang [4] used an expectation maximization framework to optimize mutual information. + +Recently, some methods have used DNNs to find multiple clusterings and achieved better results. Wei et al. [31] proposed a deep matrix factorization based method that uses multi-view data to find multiple clusterings. ENRC [18] uses an auto-encoder to learn object features and finds multiple clusterings by optimizing a clustering objective function. iMClusts [27] exploits auto-encoders and multi-head attention to learn features from different perspectives, and + +then find multiple clusterings. AugDMC [35] leverages data augmentation to generate diverse aspects of images and learns the representations to discover multiple clusterings. Although existing deep multiple clustering methods have achieved remarkable results, they require users to exert considerable efforts to select the correct clustering they need. In this work, we aim to efficiently and effectively capture users' interests using short keywords and provide clustering results accordingly. + +# 2.2. Multi-modal Model + +Multi-modal learning refers to the process of learning representations from different types of input modalities, such as image data, text, or speech. As related, we focus on how vision models can benefit from natural language supervision. CLIP [25] is a notable model, which is trained with a dataset containing 400 million text-image pairs from the internet. The objective is to align images to their corresponding text using contrastive learning. + +Fine-tuning approaches adapt vision-language models like CLIP to specific downstream image recognition tasks. CoOp [36] and CLIP-Adapter [6] exemplify this, with the latter integrating residual style feature blending to enhance performance on various visual classification tasks. Additionally, insights from TeS [30], further elucidate the effectiveness of fine-tuning strategies in leveraging natural language supervision for enhanced visual understanding. Recognizing the scarcity of labeled data for various tasks, significant research efforts have been dedicated to enhancing zero-shot learning. Some approaches extend beyond CLIP by incorporating other large pre-trained models. For instance, VisDesc [17] harnesses the power of GPT-3 to generate comprehensive contextual descriptions corresponding to given class names, thereby demonstrating superior performance compared to CLIP's basic prompts. UPL [13] and TPT [29] leverages unlabeled data to optimize learnable input text prompts. InMaP [24] recovers the proxy of each class in the vision space with the help from the text proxy. All of these methods aim to improve the performance of vision classification tasks, while clustering is a different scenario that we do not have class names that we can exploit to extract useful information from CLIP. In this work, we consider leveraging the CLIP encoder to extract coherent text and image embeddings, and using GPT-4 to integrate users' interests for the purpose of multiple clustering. + +# 3. The Proposed Method + +In this section, we first briefly review the training objective in CLIP as follows, and then describe the details of our proposed method based on that. + +# 3.1. Multi-modal Pre-training + +Given a set of image-text pairs as $\{x_{i},t_{i}\}_{i = 1}^{n}$ , where $x_{i}$ is an image and $t_i$ is the corresponding text description, their + +vision and text representations can be obtained by two encoders as $\mathbf{x}_i = f(x_i)$ and $\mathbf{t}_i = h(t_i)$ . $f(\cdot)$ and $h(\cdot)$ are vision and text encoders for optimization, where $\mathbf{x}_i$ and $\mathbf{t}_i$ have the unit norm. Then, these two encoders can be learned by minimizing the contrastive loss as + +$$ +\min _ {f, h} \sum_ {i} - \log \frac {\exp (\mathbf {x} _ {i} ^ {\top} \mathbf {t} _ {i} / \tau)}{\sum_ {j} \exp (\mathbf {x} _ {i} ^ {\top} \mathbf {t} _ {j}) / \tau) - \log \frac {\exp (\mathbf {t} _ {i} ^ {\top} \mathbf {x} _ {i} / \tau)}{\sum_ {j} \exp (\mathbf {t} _ {i} ^ {\top} \mathbf {x} _ {j}) / \tau)} +$$ + +where $\tau$ is the temperature. + +This contrastive loss aims to pull the image and its description together while pushing away the irrelevant text [22], which enables the emerging multi-modal applications, e.g., zero-shot transfer [24, 25], text-to-image generation [28], etc. + +# 3.2. Multi-modal Proxy Learning + +Given the pre-trained vision and text encoders from CLIP, this work takes one step further to investigate if we can extract user-specific information from the alignment between images and text. + +Concretely, given an image of fruit [12] as shown in Fig. 2, some users may be interested in only one specific property of the object, e.g., color. In this scenario, applying the vision encoder to extract the representation for the whole image can miss the preference of users. To mitigate the problem, we propose to explore the proxy representation from the image with the guidance from the text using users' preference, named Multi-Modal Proxy learning (Multi-MaP). + +Recall that CLIP is pre-trained by images and text descriptions, where the text prompt is "a photo of a fruit" for an image containing "fruit". Now given a user's preference (e.g., color), we can rewrite the prompt as "fruit with the color of *" denoted by $t_i^*$ for image $x_i$ , where " * " is the proxy word and its token embedding is $\mathbf{w}_i$ that is learnable. Then, we can align image and text prompt representations to obtain the appropriate proxy embedding for a user's interest. Since there is no negative pairs, only the similarity between positive pairs can be optimized as + +$$ +\mathcal {L} \left(\mathbf {w} _ {i}\right) = - \langle f \left(x _ {i}\right), h \left(t _ {i} ^ {*}\right) \rangle \tag {1} +$$ + +where the vision and text encoders are frozen, and $\mathbf{w}_i$ is the only variable for learning as the representation of the proxy word. By maximizing the similarity to the image representation, Multi-MaP aims to learn the optimal text proxy according to the user's interest. + +However, it should be noted that the text encoder was pre-trained with discrete text tokens, while the domain of $\mathbf{w}_i$ in Eqn. 1 is unconstrained. Therefore, the text representation extracted from the frozen text encoder can be inaccurate for $\mathbf{w}_i$ that degenerates the performance, which is demonstrated in the following theory. + +![](images/2ba575801ef6635c7c562da723dd4dc6de5af78eadb4d7edac62571b4717dbfa.jpg) +Figure 2. Multi-MaP framework. In the training process of Multi-MaP, the vision and text encoders are frozen and the proxy word embeddings $\mathbf{w}_i$ are learnable. Specifically, it first constructs the prompt embeddings based on the reference words provided by GPT-4 using a user's high-level concept, and then selects a reference word $z_i$ for each image according to the similarity between the prompt embeddings $\mathbf{t}_i$ and the image embeddings $\mathbf{x}_i$ . Then, it combines the prompt and the reference words to form the new prompt embeddings $\mathbf{t}_i^*$ and maximizes the similarity to the image representation, so the proxy word embeddings $\mathbf{w}_i$ can capture the desired image features. In addition, the proxy word embeddings $\mathbf{w}_i$ should be close to the target concept word $\mathbf{u}_1$ and the selected reference word $\mathbf{z}_i$ to construct the concept-level constraint and reference word constraint, which capture the features related to the user's interest. + +For the sake of simplicity, we assume $h'(t) \in \mathbb{R}$ is defined on the whole set but only has the right estimation on a discrete set as $T = \{t_i\}$ and the counterpart with the unconstrained set is denoted as $H(w)$ . According to the definition, we have $\forall t \in T, h'(t) = H(t)$ . The gap between the estimation from $h'$ and $H$ on unconstrained variable $w$ can be depicted as follows. + +Theorem 1. Given $w \notin T$ and $t \in T$ , if assuming $h'$ and $H$ are $L_h$ and $L_H$ -Lipschitz continuous, we have + +$$ +\| h ^ {\prime} (w) - H (w) \| _ {2} \leq \left(L _ {h} + L _ {H}\right) \| t - w \| _ {2} +$$ + +Proof. According to the definition, we have + +$$ +\begin{array}{l} \left\| h ^ {\prime} (w) - H (w) \right\| _ {2} = \left\| h ^ {\prime} (w) - h ^ {\prime} (t) + h ^ {\prime} (t) - H (w) \right\| _ {2} \\ \leq \| h ^ {\prime} (w) - h ^ {\prime} (t) \| _ {2} + \| h ^ {\prime} (t) - H (w) \| _ {2} \\ = \| h ^ {\prime} (w) - h ^ {\prime} (t) \| _ {2} + \| H (t) - H (w) \| _ {2} \\ \leq \left(L _ {h} + L _ {H}\right) \| t - w \| _ {2} \\ \end{array} +$$ + +![](images/a3aeb8352ad7a5f0985af472da22dc1d96157a30fa0f311f31584f4132f28b69.jpg) + +Remark Theorem 1 implies that the distance of the estimation $h'(w)$ to the ground-truth result $H(w)$ is bounded by that of $w$ to an arbitrary discrete token $t$ . Therefore, by selecting the nearest token as the reference, the bound can be improved as shown in the following corollary. + +Corollary 2. With the assumptions in Theorem 1 and letting $t' = \arg \min_i \| t_i - w\| _2$ , we have + +$$ +\left\| h ^ {\prime} (w) - H (w) \right\| _ {2} \leq \left(L _ {h} + L _ {H}\right) \| t ^ {\prime} - w \| _ {2} +$$ + +# 3.2.1 Concept-level Constraint + +According to the above analysis, a good reference $t'$ can help guarantee the performance. Fortunately, the input concept (e.g., color) from the user can be leveraged as the reference to constrain the freedom of the proxy word. Therefore, given the target concept word $u$ , we can obtain its token embedding as $\mathbf{u} = \phi(u)$ . Then, to learn appropriate representations from the proxy embedding, the original problem can be rewritten with the constraint as + +$$ +\mathcal {L} (\mathbf {w} _ {i}) = - \langle f (x _ {i}), h \left(t _ {i} ^ {*}\right) \rangle \quad s. t. \quad \| \mathbf {w} _ {i} - \mathbf {u} \| _ {2} ^ {2} \leq \lambda +$$ + +The constrained problem is equivalent to + +$$ +\mathcal {L} \left(\mathbf {w} _ {i}\right) = - \langle f \left(x _ {i}\right), h \left(t _ {i} ^ {*}\right) \rangle + \alpha \| \mathbf {w} _ {i} - \mathbf {u} \| _ {2} ^ {2} \tag {2} +$$ + +following [3], which can be optimized effectively by gradient descent. + +# 3.2.2 Constrained Optimization with Reference Word + +However, it is well known that the user concept is often with a large scope covering a broad range of words (e.g., color + +
Datasets# Samples# Clusters
ALOI2882;2
Card8,02913;4
CMUface6404;20;2;4
Fruit1053;3
Fruit3604,8564;4
Stanford Cars1,2004;3
Flowers1,6004;4
+ +Table 1. Dataset Statistics. + +covers all including but not limited to 'red', 'blue', 'green', etc.). As suggested by our above theoretical analysis, it is desired if the reference is as close as possible. In a clustering scenario and given only the user's high-level concept, it is challenging to find a closer reference word to further constrain the proxy learning. Fortunately, with the development of large language models (LLMs), we can leverage them (e.g., GPT-4) to provide relevant words according to a user's high-level concept as the candidate set and develop a selection strategy to obtain a closer reference word for each image. While the responses gathered from GPT-4 might not always precisely align with the data's ground truth, they indisputably furnish valuable candidate features, enriching the capabilities of Multi-MaP. + +To elucidate, considering the task of clustering a fruit dataset based on the concept of color, we can pose a query to GPT-4 as "What are the common colors of fruit?" The response we obtain is "Common colors of fruit include red, yellow, green, orange, purple, and blue". The colors enumerated in this response can serve as reference word candidates. It is worth noting that while the color spectrum derived from this method might surpass the actual colors present in the data, we collect all of them into the candidate set of reference words $\{z_k\}_k$ , where $z_1$ : "red", $z_2$ : "yellow", etc. Then, their text representations can be obtained from the prompt $t_k$ as "fruit with the color of $z_k$ ". Given the image $x_i$ , the closest reference can be observed as + +$$ +z _ {i} = \arg \max _ {k} \left\langle \mathbf {x} _ {i}, \mathbf {t} _ {k} \right\rangle +$$ + +where $\mathbf{t}_k = h(t_k)$ . After that, $\mathbf{w}_i$ can be initialized with the token embedding of $z_i$ as $\mathbf{z}_i = \phi(z_i)$ . Moreover, we change the regularization using a closer reference word compared to the high-level concept as + +$$ +\mathcal {L} \left(\mathbf {w} _ {i}\right) = - \langle f \left(x _ {i}\right), h \left(t _ {i} ^ {*}\right) \rangle + \alpha \| \mathbf {w} _ {i} - \mathbf {z} _ {i} \| _ {2} ^ {2} \tag {3} +$$ + +# 3.2.3 Contrastive Concepts + +In some application scenarios, one user may need more than one clustering and provide high-level concepts as $\{u_j\}$ , e.g., $u_1$ : "color", $u_2$ : "species", etc. For the concept "color", the irrelevant concept "species" can be leveraged as the negative constraint for the learning of proxy word. Concretely, + +let $u_{w}$ denote the target concept word, a contrastive loss can be adopted as regularization + +$$ +R (\mathbf {w} _ {i}) = - \log \frac {\exp (\mathbf {w} _ {i} ^ {\top} \mathbf {u} _ {w})}{\sum_ {j} \exp (\mathbf {w} _ {i} ^ {\top} \mathbf {u} _ {j})} +$$ + +and the final objective becomes + +$$ +\begin{array}{l} \mathcal {L} \left(\mathbf {w} _ {i}\right) = \\ - \left\langle f \left(x _ {i}\right), h \left(t _ {i} ^ {*}\right) \right\rangle + \alpha \| \mathbf {w} _ {i} - \mathbf {z} _ {i} \| _ {2} ^ {2} - \beta \log \frac {\exp \left(\mathbf {w} _ {i} ^ {\top} \mathbf {u} _ {w}\right)}{\sum_ {j} \exp \left(\mathbf {w} _ {i} ^ {\top} \mathbf {u} _ {j}\right)} \tag {4} \\ \end{array} +$$ + +where the first term is to infer the user-specific feature, while the latter two terms constrain the proxy word to the reference words for the appropriate representation extraction from the pre-trained text encoder. The overall framework of Multi-MaP is illustrated in Fig. 2. + +# 4. Experiments + +To demonstrate our proposed method, we evaluate MultiMaP on all publicly available image datasets in multiple clustering, including ALOI [7], Stanford Cars [14], Card [35], CMUface [10], Flowers [20], Fruit [12], and Fruit360 [35] as summarized in Table 1. + +We compare Multi-MaP against five state-of-the-art methods: MSC [12] is a traditional multiple clustering method that uses hand-crafted features; MCV [9] leverages multiple feature extractors to represent different "views" of the same data and employs a multi-input neural network to enhance clustering outcomes; ENRC [18] is a deep multiple clustering method that integrates auto-encoder and clustering objective to generate different clusterings; iM-Clusts [27] makes use of the expressive representational power of deep autoencoders and multi-head attention to accomplish multiple clusterings; AugDMC [35] leverages augmentations to learn different image representations to achieve multiple clustering. + +# 4.1. Experiment Setup + +We employ Adam and set momentum as 0.9 to train the model for 1000 epochs. All hyperparameters are searched according to the loss score of Multi-MaP, where the learning rate is searched in $\{0.1, 0.05, 0.01, 0.005, 0.001, 0.0005\}$ , weight decay is in $\{0.0005, 0.0001, 0.00005, 0.00001, 0\}$ , $\alpha, \beta$ are in $\{0.0, 0.1, 0.2, \ldots, 1.0\}$ , and $\lambda$ is fixed as 1 for all the experiments. For the non-pre-trained methods, we perform k-means [15] 10 times due to its randomness and evaluate the average clustering performance using two quantitative metrics, that is, Normalized Mutual Information (NMI) [32] and Rand index (RI) [26]. These measures range in $[0, 1]$ , and higher scores imply more accurate + +
DatasetClusteringMSCMCVENRCiMClustersAugDMCMulti-MaP
NMIRINMIRINMIRINMIRINMIRINMIRI
ALOIColor0.15630.34280.69820.74390.98330.98921.00001.00001.00001.00001.00001.0000
Shape0.29680.51990.73590.82610.97320.98610.99630.99891.00001.00001.00001.0000
FruitColor0.68860.80510.62660.76850.71030.85110.73510.86320.85170.91080.86190.9526
Species0.16270.60450.27330.65970.31870.65360.30290.67430.35460.73991.00001.0000
Fruit360Color0.25440.60540.37760.67910.42640.68680.40970.68410.45940.73920.62390.8243
Species0.21840.58050.29850.61760.41420.69840.38610.67320.51390.74300.52840.7582
CardOrder0.08070.78050.07920.71280.12250.73130.11440.76580.14400.82670.36530.8587
Suits0.04970.35870.04300.36380.06760.38010.07160.37150.08730.42280.27340.7039
CMUfaceEmotion0.12840.67360.14330.52680.15920.66300.04220.59320.01610.53670.17860.7105
Glass0.14200.57450.12010.49050.14930.62090.19290.56270.10390.53610.34020.7068
Identity0.38920.73260.46370.62470.56070.76350.51090.82600.58750.83340.66250.9496
Pose0.36870.63220.32540.60280.22900.50290.44370.61140.13200.55170.46930.6624
Stanford CarsColor0.23310.61580.21030.58020.24650.67790.23360.65520.27360.75250.73600.9193
Type0.13250.53360.16500.56340.20630.62170.19630.56430.23640.73560.63550.8399
FlowersColor0.25610.59650.29380.58600.33290.62140.31690.61270.35560.69310.64260.7984
Species0.13260.52730.15610.60650.18940.61950.18870.60770.19960.62270.60130.8103
+ +clustering results. The experiments are conducted with GPU NVIDIA GeForce RTX 2080 Ti. + +It should also be noted that some data are difficult to obtain corresponding candidate labels from GPT-4 or the labels do not provide semantic features, such as names. For example, for the identity clustering for the CMUface dataset [10], different identities represent different people and the semantic meaning of names should not affect the clustering results. In this case, we randomly extract 10 words from WordNet [5] as reference words, in order to make the candidate labels more distinctive. For instance, we randomly choose "abstain, candid, function, haphazard, knot, luxury, nonchalance, pension, resilience, taciturn" for the above scenario. Furthermore, all publicly available multiple clustering datasets provide each ground-truth clustering a high-level concept, e.g., 'shape', 'pose', etc. Therefore, in the experiment, we directly use them as users' preferences for our evaluation purposes. + +# 4.2. Performance Comparison + +In our experiments, after we obtain the proxy word embedding of each image for a desired concept, we feed them to k-means to obtain the corresponding clustering. Since k-means is random, we repeat ten times and the average performance is reported in Table 2. The best results are marked by bold numbers. + +We can observe that Multi-MaP outperforms the baselines in all the cases, indicating the superiority of the proposed method. This also shows a strong generalization ability of the pre-trained model by CLIP, which can capture the features of data in different aspects. + +Since our method uses the CLIP encoder and GPT-4 to obtain clustering results, a natural question arises how the performance would be if we directly use them in a zero-short manner. Therefore, we provide two zero-shot variants of CLIP, that is, $\mathrm{CLIP}_{\mathrm{GPT}}$ that uses GPT-4 to obtain the can + +Table 2. Quantitative comparison. The significantly best results with $95\%$ confidence are in bold. + +
DatasetClusteringCLIPGPTCLIPlabelMulti-MaP
NMIRINMIRINMIRI
ALOIColor0.85810.94071.00001.00001.00001.0000
Shape1.00001.00001.00001.00001.00001.0000
FruitColor0.79120.90750.86290.97800.86190.9526
Species0.97930.99191.00001.00001.00001.0000
Fruit360Color0.56130.73050.57460.76730.62390.8243
Species0.43700.75520.53640.76310.52840.7582
CardOrder0.35180.84580.35180.84580.36530.8587
Suits0.27110.61230.27110.61230.27340.7039
CMUfaceEmotion0.15760.65320.15900.66190.17860.7105
Glass0.29050.68690.46860.75050.34020.7068
Identity0.19980.63880.26770.75450.66250.9496
Pose0.40880.64730.46910.64090.46930.6624
Stanford CarsColor0.65390.82370.68300.86420.73600.9193
Type0.62070.79310.64290.84560.63550.8399
FlowersColor0.56530.76290.58280.78360.64260.7984
Species0.56200.75530.60190.79960.60130.8103
+ +Table 3. Variants of CLIP. The significantly best results with $95\%$ confidence are in bold. + +didate labels and predicts labels through zero-shot classification with all candidate labels as class names, and $\mathrm{CLIP}_{\mathrm{label}}$ that performs zero-shot classification with all ground truth labels. It should be noted that $\mathrm{CLIP}_{\mathrm{label}}$ uses an unfair setting with a ground-truth label set known in advance, which is expected to provide the best performance using CLIP in a zero-short manner. The results are shown in Table 3. + +As expected, $\mathrm{CLIP}_{\mathrm{label}}$ achieves better performance than $\mathrm{CLIP}_{\mathrm{GPT}}$ in almost all cases, since $\mathrm{CLIP}_{\mathrm{label}}$ uses a fixed ground truth label set, while $\mathrm{CLIP}_{\mathrm{GPT}}$ uses candidate labels that may not match the ground truth exactly, introducing noise. Note that $\mathrm{CLIP}_{\mathrm{GPT}}$ and $\mathrm{CLIP}_{\mathrm{label}}$ achieve the same results on Cards, because the candidate labels provided by GPT-4 are exactly the same as the true labels. + +Besides, Multi-MaP performs better than $\mathrm{CLIP}_{\mathrm{GPT}}$ in almost all cases, indicating that the proposed method can learn more effective features through its training process. Moreover, although $\mathrm{CLIP}_{\mathrm{label}}$ uses the ground truth, which is expected to be the best, Multi-MaP still outperforms $\mathrm{CLIP}_{\mathrm{label}}$ in some cases, such as the clustering of color for + +
Multi-MaPpMulti-MaPcMulti-MaPrMulti-MaPcrMulti-MaP
ModulesProxy Learning✓✓✓✓✓
Concept Word×✓×✓✓
Reference Word××✓✓✓
Contrastive Concepts××××✓
NMI↑RI↑NMI↑RI↑NMI↑RI↑NMI↑RI↑NMI↑RI↑
ALOI [7]Color0.96190.98261.00001.00000.97950.98691.00001.00001.00001.0000
Shape1.00001.00001.00001.00001.00001.00001.00001.00001.00001.0000
Fruit [12]Color0.76420.84390.82150.92830.81360.90730.84840.93080.86190.9526
Species1.00001.00001.00001.00001.00001.00001.00001.00001.00001.0000
Fruit360 [35]Color0.56430.76650.62170.78360.59100.77460.60890.79650.62390.8243
Species0.50770.73680.51370.74360.50940.74250.51990.74280.52840.7582
Card [35]Order0.19320.81520.35680.84720.31130.82290.36160.80940.36530.8587
Suits0.23750.62820.26960.66410.24980.63650.25620.65990.27340.7039
CMUface [10]Emotion0.16900.61700.17140.62290.16970.63600.17130.68430.17860.7105
Glass0.31120.69110.32690.71360.31620.69170.33700.71080.34020.7068
Identity0.56170.82340.62430.83590.58390.82630.63910.89460.66250.9496
Pose0.43610.63860.45500.64990.43810.64290.43870.64890.46930.6624
Stanford cars [14]Color0.59390.78350.68360.86590.67290.86380.71120.91170.73600.9193
Type0.55690.79960.63830.82710.60910.80460.62890.81810.63550.8399
Flowers [20]Color0.57830.77230.58300.78330.59870.78490.62160.79410.64260.7984
Species0.57040.76080.57440.78420.57230.78110.58460.78920.60130.8103
+ +Table 4. Components ablation. All of our components boost performance consistently in all benchmark multi-clustering vision tasks. + +![](images/1d379e2a9c14ac905850983fa8ae1b6801bed353d8651054dddb942dca9822a2.jpg) +(a) Parameter sensitivity of $\alpha$ +Figure 3. Parameter analysis of $\alpha$ and $\beta$ on Fruit [12]. + +![](images/9aa1ae24c98b35eb8f0bd4d28573803976cadca5c5595260233e2adab90e438e.jpg) +(b) Parameter sensitivity of $\beta$ + +Fruit360 dataset. This is because CLIP is more inclined to capture the features of one aspect of the data, while Multi-MaP learns better embedding of different aspects by training with the supervision of users' interests. Furthermore, Multi-MaP also achieves very competitive results as $\mathrm{CLIP}_{\mathrm{label}}$ in the remaining cases, which further demonstrates the effectiveness of the proposed method. + +# 4.3. Ablation Study + +To validate the effectiveness of the proposed method, we show the gain from four components in Multi-MaP in Table 4. Let "Multi-MaP $p$ " denote the proxy learning without concept-level constraint, "Multi-MaP $c$ " denote the variant optimized by solely applying concept word to constrain the freedom of the proxy word, "Multi-MaP $r$ " denote the variant optimized by reference word provided by GPT-4 to find a closer reference word to further constrain the proxy learning, and "Multi-MaP $cr$ " denote the variant leveraged both concept word and reference word. + +We can observe that the proxy learning only with concept + +word or reference word, i.e., $\mathrm{Multi - MaP}_c$ and $\mathrm{Multi - MaP}_r$ , performs better than $\mathrm{Multi - MaP}_p$ . This shows that the proposed concept-level constraint and constrained optimization with reference words play an important role in the model as demonstrated by our theoretical analysis. Moreover, the model with combined components, i.e., $\mathrm{Multi - MaP}_{cr}$ , achieves better results than $\mathrm{Multi - MaP}_r$ and $\mathrm{Multi - MaP}_c$ . This indicates the effectiveness of the combination of reference words and reference concepts. Finally, our proposal using all including the contrastive concepts can further improve the performance, and thus provide the best results on all cases. This further demonstrates our proposal. + +# 4.4. Parameter Analysis + +We further investigate the effect of the reference word constraint weight $\alpha$ and concept-level constraint weight $\beta$ varying from 0.0 to 0.5. The results of Multi-MaP on Fruit datasets are shown in Fig. 3 (a) and Fig. 3 (b), respectively. As $\alpha$ and $\beta$ change, the proposed method keeps a species score of 1, since the image encoder can capture very effective species features. The above results also show that the proposed method can effectively capture useful information from images, reference words, and target concepts. As $\alpha$ increases, the proposed method first increases and then decreases, and reaches the maximum values at $\alpha = 0.4$ . Similar results can be observed for $\beta$ that the performance of the proposed method first increases and then decreases as $\beta$ increases, and reaches the maximum values at $\beta = 0.3$ . Indicating a suitable value of $\alpha$ or $\beta$ is helpful for Multi-MaP to obtain more effective embeddings for multiple clustering tasks. More studies such as efficiency analysis can be found + +![](images/d47eed2ba007e88973ba891e69b4e708cc98869c6042b469f2ffed4e5e3ac151.jpg) +(a) Color using ground truth + +![](images/68305eda48f13a2802bee29840a20200e88b2c25c145aebc5e3e32b8d7b46396.jpg) +(b) Color using reference words + +(c) Color of Multi-MaP +(d) Legend of color +![](images/9cc9acdca6050ec7a34c0cff89c072dcce12242d8f3a52229a5da63be06a9538.jpg) +red yellow +green orange +purple blue + +![](images/4232e29350efc6f472a931a4f2780a4e8977717639f2aec56c9172d58296d812.jpg) +(e) Species using ground truth + +![](images/5f62f49f5cb29726cf762be8dc6bd08433e39a964631a3e5850911afe87ab465.jpg) +(f) Species using reference words + +(g) Species of Multi-MaP +(h) Legend of species +Figure 4. Visualization of feature embeddings and related labels. The points represent the image or pseudo-word embeddings, and the triangles represent the prompt or label embeddings. Different colors represent different labels, which are indicated by the text next to the triangles. +![](images/d7301845ed65e734b73418bd46dee7d5539e0cdd9d5b4b96946db9fa0bf95ebe.jpg) +- apple +- orange +- banana +- strawberry +- grapes +- raspberry +- blueberry +- cherry +- pear +- plum +- peach +- nectarine +- pineapple +- kiwi +- watermelon +- cantaloupe +- mango + +in the supplementary. + +# 4.5. Visualization + +To further demonstrate the effectiveness of the proposed method, we visualize the representations obtained in $\mathrm{CLIP}_{\mathrm{label}}$ , $\mathrm{CLIP}_{\mathrm{GPT}}$ , and Multi-MaP. Specifically, for $\mathrm{CLIP}_{\mathrm{label}}$ and $\mathrm{CLIP}_{\mathrm{GPT}}$ , we visualize the image representations as well as prompts generated with real labels and candidate labels, respectively. For the Multi-MaP, we visualize the word embedding $w_{*}$ and the candidate labels selected for initialization. The results are shown in Fig. 4. For species clustering, we can see that image embeddings show very clear boundaries and correspond well to the prompt for $\mathrm{CLIP}_{\mathrm{label}}$ , which indicates that CLIP can effectively capture the features of the species in the data. $\mathrm{CLIP}_{\mathrm{GPT}}$ uses the candidate labels to generate prompts, which introduces more noise, but benefits from the CLIP text encoder, the image embedding can keep a relatively far distance from most of the irrelevant prompts. However, since there are still a few images that are labeled as peaches (i.e., a noisy label), it performs slightly worse than $\mathrm{CLIP}_{\mathrm{label}}$ , as shown in Fig. 4(e). Besides, Multi-MaP can capture the image and users' interests in the training process, therefore it compensates for the shortcomings of $\mathrm{CLIP}_{\mathrm{GPT}}$ and achieves better results. On the other hand, for color clustering, the prompts are farther away from $\mathrm{CLIP}_{\mathrm{label}}$ and $\mathrm{CLIP}_{\mathrm{GPT}}$ , that indicates the image embeddings mainly capture the features of species, which have no direct connection with color. $\mathrm{CLIP}_{\mathrm{GPT}}$ generates the prompt from the candidate label, which has more noise than the ground truth label, resulting in worse performance than $\mathrm{CLIP}_{\mathrm{label}}$ . The proposed method can distinguish different colors more clearly, because it can learn from the user's interest and capture the color-related features. How + +ever, some red color embeddings are closer to purple, because some images in the datasets are actually purple, but labeled as red. To sum up, the proposed method can learn more effective embeddings based on the users' interests for multiple clustering tasks. + +# 5. Conclusion + +To conclude, our study thoroughly investigates the significant challenges that current advanced deep learning techniques face in multiple clustering. A key issue is that users often do not need every clustering result produced by an algorithm, and selecting the most relevant one requires an in-depth understanding of each outcome. To overcome this challenge, the proposed method introduces a novel multimodal proxy learning process, which effectively aligns a user's brief keyword describing the interest with the corresponding vision components. By integrating a multi-modal model and GPT-4 to precisely capture a user's interest using a keyword, the proposed approach uses both reference word constraint and concept-level constraint to discover personalized clustering result(s), which can also lead to enhanced performance. Experiments on diverse datasets demonstrate the superiority of the proposed method in multiple clustering tasks with a precise capture of a user's interest. The proposed method is limited by data with semantic meaningful labels, although we can use WordNet to help, whose comprehensive study will be our future work. + +# 6. Acknowledgement + +This research is supported in part by Advata Gift funding. All opinions, findings, conclusions and recommendations in this paper are those of the author and do not necessarily reflect the views of the funding agencies. + +# References + +[1] Eric Bae and James Bailey. Coala: A novel approach for the extraction of an alternate clustering of high quality and high dissimilarity. In ICDM, pages 53-62. IEEE, 2006. 1, 2 +[2] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning. Springer, 2006. 1 +[3] Stephen P. Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University Press, 2014. 4 +[4] Xuan Hong Dang and James Bailey. Generation of alternative clusterings using the cami approach. In Proceedings of the 2010 SIAM International Conference on Data Mining, pages 118-129. SIAM, 2010. 2 +[5] Christiane Fellbaum. Wordnet. In Theory and applications of ontology: computer applications, pages 231-243. Springer, 2010. 6 +[6] 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. International Journal of Computer Vision, pages 1-15, 2023. 1, 3 +[7] Jan-Mark Geusebroek, Gertjan J Burghouts, and Arnold WM Smeulders. The amstedam library of object images. International Journal of Computer Vision, 61:103-112, 2005. 5, 7 +[8] David Gondek and Thomas Hofmann. Conditional information bottleneck clustering. In 3rd IEEE international conference on data mining, workshop on clustering large data sets, pages 36-42, 2003. 2 +[9] Joris Guérin and Byron Boots. Improving image clustering with multiple pretrained cnn feature extractors. In *British Machine Vision Conference* 2018, BMVC 2018, 2018. 1, 2, 5 +[10] Stephan Gunnemann, Ines Farber, Matthias Rüdiger, and Thomas Seidl. Smvc: semi-supervised multi-view clustering in subspace projections. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 253-262, 2014. 5, 6, 7 +[11] Juhua Hu and Jian Pei. Subspace multi-clustering: a review. Knowledge and information systems, 56(2):257-284, 2018. 2 +[12] Juhua Hu, Qi Qian, Jian Pei, Rong Jin, and Shenghuo Zhu. Finding multiple stable clusterings. Knowledge and Information Systems, 51(3):991-1021, 2017. 1, 2, 3, 5, 7 +[13] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 3 +[14] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5, 7 +[15] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 5 +[16] James MacQueen et al. Some methods for classification and analysis of multivariate observations. In Proceedings of the fifth Berkeley symposium on mathematical statistics and probability, pages 281-297. Oakland, CA, USA, 1967. 1 + +[17] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 3 +[18] Lukas Miklautz, Dominik Mautz, Muzaffer Can Altinigneli, Christian Böhm, and Claudia Plant. Deep embedded nonredundant clustering. In Proceedings of the AAAI conference on artificial intelligence, pages 5174-5181, 2020. 1, 2, 5 +[19] Andrew Ng, Michael Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 14, 2001. 1 +[20] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 5, 7 +[21] ZiJie Qi and Ian Davidson. A principled and flexible framework for finding alternative clusterings. In SIGKDD, pages 717-726, 2009. 2 +[22] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Tacoma Tacoma, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In ICCV, pages 6449-6457. IEEE, 2019. 3 +[23] Qi Qian, Yuanhong Xu, Juhua Hu, Hao Li, and Rong Jin. Unsupervised visual representation learning by online constrained k-means. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16640-16649, 2022. 1 +[24] Qi Qian, Yuanhong Xu, and Juhua Hu. Intra-modal proxy learning for zero-shot visual categorization with clip. In Thirty-seventh Conference on Neural Information Processing Systems, NeurIPS 2023, 2023. 3 +[25] 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, 2, 3 +[26] William M Rand. Objective criteria for the evaluation of clustering methods. Journal of the American Statistical Association, 66(336):846-850, 1971. 5 +[27] Liangrui Ren, Guoxian Yu, Jun Wang, Lei Liu, Carlotta Domeniconi, and Xiangliang Zhang. A diversified attention model for interpretable multiple clusterings. IEEE Transactions on Knowledge and Data Engineering, 2022. 1, 2, 5 +[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685. IEEE, 2022. 3 +[29] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Advances in Neural Information Processing Systems, 35:14274-14289, 2022. 3 +[30] Junyang Wang, Yuanhong Xu, Juhua Hu, Ming Yan, Jitao Sang, and Qi Qian. Improved visual fine-tuning with natural language supervision. arXiv preprint arXiv:2304.01489, 2023. 1, 3 + +[31] Shaowei Wei, Jun Wang, Guoxian Yu, Carlotta Domeni-coni, and Xiangliang Zhang. Multi-view multiple clusterings using deep matrix factorization. In Proceedings of the AAAI conference on artificial intelligence, pages 6348–6355, 2020. 2 +[32] JV White, Sam Steingold, and CG Fournelle. Performance metrics for group-detection algorithms. Proceedings of Interface, 2004, 2004. 5 +[33] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487, 2016. 1 +[34] Sen Yang and Lijun Zhang. Non-redundant multiple clustering by nonnegative matrix factorization. Machine Learning, 106(5):695-712, 2017. 2 +[35] Jiawei Yao, Enbei Liu, Maham Rashid, and Juhua Hu. Augdmc: Data augmentation guided deep multiple clustering. In INNS DLIA@IJCNN, 2023. 1, 3, 5, 7 +[36] 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. 3 \ No newline at end of file diff --git a/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/images.zip b/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5267a0fdac4b70e33da9e768f155607dd65de913 --- /dev/null +++ b/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb9d8374f0885eb16ee94e128d3771d8a14926a72fbb5f5affbc7bd10eb19270 +size 603056 diff --git a/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/layout.json b/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9fe11e2224a4700174b32c98436af2a5d1a2475e --- /dev/null +++ b/2024/Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering/layout.json @@ -0,0 +1,9011 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 103, + 541, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 541, + 122 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 541, + 122 + ], + "type": "text", + "content": "Multi-Modal Proxy Learning Towards Personalized Visual Multiple Clustering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 142, + 392, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 142, + 392, + 157 + ], + "spans": [ + { + "bbox": [ + 205, + 142, + 392, + 157 + ], + "type": "text", + "content": "Jiawei Yao" + }, + { + "bbox": [ + 205, + 142, + 392, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 205, + 142, + 392, + 157 + ], + "type": "text", + "content": " Qi Qian" + }, + { + "bbox": [ + 205, + 142, + 392, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 205, + 142, + 392, + 157 + ], + "type": "text", + "content": " Juhua Hu" + }, + { + "bbox": [ + 205, + 142, + 392, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 157, + 523, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 157, + 523, + 185 + ], + "spans": [ + { + "bbox": [ + 67, + 157, + 523, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 157, + 523, + 185 + ], + "type": "text", + "content": " School of Engineering and Technology, University of Washington, Tacoma, WA 98402, USA \n" + }, + { + "bbox": [ + 67, + 157, + 523, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 157, + 523, + 185 + ], + "type": "text", + "content": " Alibaba Group, Bellevue, WA 98004, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 168, + 186, + 423, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 186, + 423, + 200 + ], + "spans": [ + { + "bbox": [ + 168, + 186, + 423, + 200 + ], + "type": "text", + "content": "{jwyao, juhuah}@uw.edu, qi.qian@alibaba-inc.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 252, + 290, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 290, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 290, + 563 + ], + "type": "text", + "content": "Multiple clustering has gained significant attention in recent years due to its potential to reveal multiple hidden structures of data from different perspectives. The advent of deep multiple clustering techniques has notably advanced the performance by uncovering complex patterns and relationships within large datasets. However, a major challenge arises as users often do not need all the clusterings that algorithms generate, and figuring out the one needed requires a substantial understanding of each clustering result. Traditionally, aligning a user's brief keyword of interest with the corresponding vision components was challenging, but the emergence of multi-modal and large language models (LLMs) has begun to bridge this gap. In response, given unlabeled target visual data, we propose Multi-MaP, a novel method employing a multi-modal proxy learning process. It leverages CLIP encoders to extract coherent text and image embeddings, with GPT-4 integrating users' interests to formulate effective textual contexts. Moreover, reference word constraint and concept-level constraint are designed to learn the optimal text proxy according to the user's interest. Multi-MaP not only adeptly captures a user's interest via a keyword but also facilitates identifying relevant clusterings. Our extensive experiments show that MultiMaP consistently outperforms state-of-the-art methods in all benchmark multi-clustering vision tasks. Our code is available at https://github.com/Alexander-Yao/Multi-MaP." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 598, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 287, + 696 + ], + "type": "text", + "content": "Clustering, which groups data points based on their similarities, has been extensively researched, since huge amount of unlabeled data are becoming more and more available. Traditional methods [2, 16, 19] exploit general-purpose handcrafted features that are not always ideal for specific tasks. Recently, deep clustering algorithms [9, 23, 33] leverage Deep Neural Networks (DNNs) significantly improve performance. However, most algorithms yield a single data" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 237, + 547, + 366 + ], + "blocks": [ + { + "bbox": [ + 307, + 237, + 547, + 366 + ], + "lines": [ + { + "bbox": [ + 307, + 237, + 547, + 366 + ], + "spans": [ + { + "bbox": [ + 307, + 237, + 547, + 366 + ], + "type": "image", + "image_path": "a2cba1582001304a37fce6343670486d636254288166348832ff61b13105129c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 367, + 546, + 400 + ], + "lines": [ + { + "bbox": [ + 305, + 367, + 546, + 400 + ], + "spans": [ + { + "bbox": [ + 305, + 367, + 546, + 400 + ], + "type": "text", + "content": "Figure 1. The flow chart of Multi-MaP. Multi-MaP obtains multiple clustering results based on the high-level concepts from users and the reference words from GPT-4." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 412, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 547, + 556 + ], + "type": "text", + "content": "partition, while data can exhibit multiple aspects (e.g., color and species as shown in Fig. 1). Traditional multiple clustering algorithms [1, 12] have been developed to generate different partitions for varying applications, demonstrating the ability to identify multiple distinct clusterings from a dataset. Contemporary advancements in the field reveal a growing inclination among researchers to integrate deep learning methodologies for facilitating multiple clustering outcomes. Predominantly, such techniques capitalize on auto-encoders and data augmentation processes to capture a broad spectrum of feature dimensions, thereby enhancing the performance of multiple clustering [18, 27, 35]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": "However, a common issue arises as users often do not require all the clusterings generated by the algorithm, and identifying the relevant ones necessitates a substantial understanding of each clustering result. Therefore, in this work, we initiate an exploration into a method that is adept at accurately capturing and reflecting a user's interest. Users typically express their interests through concise keywords (e.g., color or species), and aligning these with different visual components precisely is challenging. Fortunately, the advent of multi-modal models like CLIP [25] that aligns images to their corresponding text descriptions, can be helpful to fill this gap. However, unlike methods that employ labeled data to fine-tune pre-trained models [6, 30], multiple" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14066" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "clustering deals with environments marked by vague or undefined label categories and amounts. Consequently, given only a high-level concept from the user, it is infeasible to fine-tune the pre-trained models to capture a specific aspect of the data, without the detailed labels corresponding to the user's concept." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 147, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 147, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 287, + 350 + ], + "type": "text", + "content": "An intuitive strategy to integrate pre-trained models into clustering is the zero-shot feature extraction, followed by clustering of the resultant embeddings. However, this approach exhibits limitations, particularly in capturing the interests of users within the dataset. Taking the multimodal model CLIP [25] as an example, when feeding image data into CLIP, regardless of what aspects of clustering the user expects, CLIP can only produce the same embeddings. Even considering the scenario that different pretrained models can capture different aspects of the same data as in [9], it is hard to tell which one matches a user's preference. Fortunately, given CLIP's ability to model image-text pairs collaboratively, we can use a user's high-level concept to trigger the corresponding feature extraction from the pre-trained encoders from CLIP. However, no previous work has studied if CLIP has this potential to uncover different aspects of images, which is the focus of this work." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 352, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 519 + ], + "type": "text", + "content": "Specifically, we propose to integrate a user's high-level concept describing the preference using a personalized textual prompt. For example, if a user's focus pertains to the color dimension of fruit, a prospective prompt might be formulated as \"a fruit with the color of *\", wherein the \"*\"\n- placeholder represents the proxy word awaiting determination using the knowledge in CLIP. Thereafter, we can learn the proxy word embedding by maximizing the similarity between the image and text embedding. However, the proxy word embedding is now searched in a continuous space while the original CLIP used discrete tokens, which can downgrade the performance. We prove that the performance can be well guaranteed by selecting the nearest token as the reference, which is unavailable in a clustering task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "Fortunately, we can use the user's high-level concept as a reference, which however covers a broad range of tokens under its scope. Therefore, we propose to leverage large language models like GPT-4 to generate candidate tokens using the user's high-level concept, in which the closest can be used as a closer reference token. Furthermore, in some scenarios, users may provide multiple concepts to obtain multiple clusterings simultaneously as shown in Fig. 1, we can also introduce a negative loss with these constrastive concepts to further enhance the learning. Therefore, to capture a user's specific interest and discover personalized clustering structure hidden in the data, we propose a multimodal proxy learning method (Multi-MaP). Multi-MaP incorporates both text prompts and unlabeled images from the clustering task, and leverages CLIP to acquire their respective personalized representations using both reference" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "word and concept-level constraints. The contributions of this work can be summarized as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 96, + 545, + 323 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 96, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 545, + 132 + ], + "type": "text", + "content": "- We are the first to explore a deep multiple clustering method that precisely captures a user's interest(s) and generates personalized clustering(s) accordingly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 133, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 545, + 179 + ], + "type": "text", + "content": "- We propose a novel multi-modal proxy learning method, Multi-MaP, by leveraging the text and image encoders pre-trained by CLIP, where the user's interest can be captured by the personalized text prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 180, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 545, + 239 + ], + "type": "text", + "content": "- Considering the challenge of learning a word proxy in a continuous space while tokens in CLIP were discrete, we theoretically prove that a close reference token can help constrain the search, which motivates the proposed reference word constraint and concept-level constraint." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 240, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 240, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 306, + 240, + 545, + 287 + ], + "type": "text", + "content": "- We conduct extensive experiments on all publicly available visual multiple clustering tasks, which empirically demonstrates the superiority of the proposed Multi-MaP, with a precise capturing of a user's interest." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 288, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 545, + 323 + ], + "type": "text", + "content": "- Finally, to the best of our knowledge, we are the first who demonstrate that CLIP can uncover different semantic aspects of images." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 333, + 392, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 392, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 392, + 346 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 354, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 377 + ], + "type": "text", + "content": "In this section, we briefly review two related directions, that is, multiple clustering and multi-modal models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 384, + 421, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 421, + 397 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 421, + 397 + ], + "type": "text", + "content": "2.1. Multiple Clustering" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": "Multiple clustering, as a kind of method that can discover alternative perspectives of data, has attracted considerable attention. Traditional methods for multiple clustering [11] use shallow models to find different ways of grouping data. Some of these methods rely on constraints to produce alternative clusterings. For instance, COALA [1] uses the objects in an existing clustering as constraints for creating a new clustering and Qi et al. [21] formulated multiple clustering as a constrained optimization problem. Other methods exploit different feature subspaces to generate multiple clusterings. For example, Hu et al. [12] found multiple clusterings by maximizing the eigengap in different subspaces. MNMF [34] adds the inner product of similarity matrices as a regularization term to find multiple clusterings. Some methods also use information theory to generate multiple clusterings. Gondek [8] applied conditional information bottleneck and Dang [4] used an expectation maximization framework to optimize mutual information." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Recently, some methods have used DNNs to find multiple clusterings and achieved better results. Wei et al. [31] proposed a deep matrix factorization based method that uses multi-view data to find multiple clusterings. ENRC [18] uses an auto-encoder to learn object features and finds multiple clusterings by optimizing a clustering objective function. iMClusts [27] exploits auto-encoders and multi-head attention to learn features from different perspectives, and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14067" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "then find multiple clusterings. AugDMC [35] leverages data augmentation to generate diverse aspects of images and learns the representations to discover multiple clusterings. Although existing deep multiple clustering methods have achieved remarkable results, they require users to exert considerable efforts to select the correct clustering they need. In this work, we aim to efficiently and effectively capture users' interests using short keywords and provide clustering results accordingly." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 185, + 162, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 185, + 162, + 197 + ], + "spans": [ + { + "bbox": [ + 47, + 185, + 162, + 197 + ], + "type": "text", + "content": "2.2. Multi-modal Model" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 201, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 297 + ], + "type": "text", + "content": "Multi-modal learning refers to the process of learning representations from different types of input modalities, such as image data, text, or speech. As related, we focus on how vision models can benefit from natural language supervision. CLIP [25] is a notable model, which is trained with a dataset containing 400 million text-image pairs from the internet. The objective is to align images to their corresponding text using contrastive learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 297, + 289, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 289, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 289, + 608 + ], + "type": "text", + "content": "Fine-tuning approaches adapt vision-language models like CLIP to specific downstream image recognition tasks. CoOp [36] and CLIP-Adapter [6] exemplify this, with the latter integrating residual style feature blending to enhance performance on various visual classification tasks. Additionally, insights from TeS [30], further elucidate the effectiveness of fine-tuning strategies in leveraging natural language supervision for enhanced visual understanding. Recognizing the scarcity of labeled data for various tasks, significant research efforts have been dedicated to enhancing zero-shot learning. Some approaches extend beyond CLIP by incorporating other large pre-trained models. For instance, VisDesc [17] harnesses the power of GPT-3 to generate comprehensive contextual descriptions corresponding to given class names, thereby demonstrating superior performance compared to CLIP's basic prompts. UPL [13] and TPT [29] leverages unlabeled data to optimize learnable input text prompts. InMaP [24] recovers the proxy of each class in the vision space with the help from the text proxy. All of these methods aim to improve the performance of vision classification tasks, while clustering is a different scenario that we do not have class names that we can exploit to extract useful information from CLIP. In this work, we consider leveraging the CLIP encoder to extract coherent text and image embeddings, and using GPT-4 to integrate users' interests for the purpose of multiple clustering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 613, + 177, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 177, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 177, + 625 + ], + "type": "text", + "content": "3. The Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 627, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 287, + 663 + ], + "type": "text", + "content": "In this section, we first briefly review the training objective in CLIP as follows, and then describe the details of our proposed method based on that." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 190, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 190, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 190, + 685 + ], + "type": "text", + "content": "3.1. Multi-modal Pre-training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Given a set of image-text pairs as " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{x_{i},t_{i}\\}_{i = 1}^{n}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " is an image and " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " is the corresponding text description, their" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "vision and text representations can be obtained by two encoders as " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = f(x_i)" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i = h(t_i)" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " are vision and text encoders for optimization, where " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " have the unit norm. Then, these two encoders can be learned by minimizing the contrastive loss as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 140, + 545, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 140, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 306, + 140, + 545, + 171 + ], + "type": "interline_equation", + "content": "\\min _ {f, h} \\sum_ {i} - \\log \\frac {\\exp (\\mathbf {x} _ {i} ^ {\\top} \\mathbf {t} _ {i} / \\tau)}{\\sum_ {j} \\exp (\\mathbf {x} _ {i} ^ {\\top} \\mathbf {t} _ {j}) / \\tau) - \\log \\frac {\\exp (\\mathbf {t} _ {i} ^ {\\top} \\mathbf {x} _ {i} / \\tau)}{\\sum_ {j} \\exp (\\mathbf {t} _ {i} ^ {\\top} \\mathbf {x} _ {j}) / \\tau)}", + "image_path": "bc2fde35cf37d02017a84507ba8bb604ca0a8bf995224123e760f3ecfae87d33.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 178, + 418, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 418, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 418, + 190 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 178, + 418, + 190 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 306, + 178, + 418, + 190 + ], + "type": "text", + "content": " is the temperature." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 190, + 545, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 250 + ], + "type": "text", + "content": "This contrastive loss aims to pull the image and its description together while pushing away the irrelevant text [22], which enables the emerging multi-modal applications, e.g., zero-shot transfer [24, 25], text-to-image generation [28], etc." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 257, + 464, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 257, + 464, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 257, + 464, + 270 + ], + "type": "text", + "content": "3.2. Multi-modal Proxy Learning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 276, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 323 + ], + "type": "text", + "content": "Given the pre-trained vision and text encoders from CLIP, this work takes one step further to investigate if we can extract user-specific information from the alignment between images and text." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 323, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 430 + ], + "type": "text", + "content": "Concretely, given an image of fruit [12] as shown in Fig. 2, some users may be interested in only one specific property of the object, e.g., color. In this scenario, applying the vision encoder to extract the representation for the whole image can miss the preference of users. To mitigate the problem, we propose to explore the proxy representation from the image with the guidance from the text using users' preference, named Multi-Modal Proxy learning (Multi-MaP)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 430, + 545, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 551 + ], + "type": "text", + "content": "Recall that CLIP is pre-trained by images and text descriptions, where the text prompt is \"a photo of a fruit\" for an image containing \"fruit\". Now given a user's preference (e.g., color), we can rewrite the prompt as \"fruit with the color of *\" denoted by " + }, + { + "bbox": [ + 304, + 430, + 545, + 551 + ], + "type": "inline_equation", + "content": "t_i^*" + }, + { + "bbox": [ + 304, + 430, + 545, + 551 + ], + "type": "text", + "content": " for image " + }, + { + "bbox": [ + 304, + 430, + 545, + 551 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 430, + 545, + 551 + ], + "type": "text", + "content": ", where \" * \" is the proxy word and its token embedding is " + }, + { + "bbox": [ + 304, + 430, + 545, + 551 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 304, + 430, + 545, + 551 + ], + "type": "text", + "content": " that is learnable. Then, we can align image and text prompt representations to obtain the appropriate proxy embedding for a user's interest. Since there is no negative pairs, only the similarity between positive pairs can be optimized as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 362, + 559, + 545, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 559, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 362, + 559, + 545, + 573 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\mathbf {w} _ {i}\\right) = - \\langle f \\left(x _ {i}\\right), h \\left(t _ {i} ^ {*}\\right) \\rangle \\tag {1}", + "image_path": "e909913ceea0374096f9893ad5a18f28ad79b40e853fd769e526478a72be1268.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 582, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 640 + ], + "type": "text", + "content": "where the vision and text encoders are frozen, and " + }, + { + "bbox": [ + 304, + 582, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 304, + 582, + 545, + 640 + ], + "type": "text", + "content": " is the only variable for learning as the representation of the proxy word. By maximizing the similarity to the image representation, Multi-MaP aims to learn the optimal text proxy according to the user's interest." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "However, it should be noted that the text encoder was pre-trained with discrete text tokens, while the domain of " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " in Eqn. 1 is unconstrained. Therefore, the text representation extracted from the frozen text encoder can be inaccurate for " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " that degenerates the performance, which is demonstrated in the following theory." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14068" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 87, + 520, + 323 + ], + "blocks": [ + { + "bbox": [ + 72, + 87, + 520, + 323 + ], + "lines": [ + { + "bbox": [ + 72, + 87, + 520, + 323 + ], + "spans": [ + { + "bbox": [ + 72, + 87, + 520, + 323 + ], + "type": "image", + "image_path": "2ba575801ef6635c7c562da723dd4dc6de5af78eadb4d7edac62571b4717dbfa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "text", + "content": "Figure 2. Multi-MaP framework. In the training process of Multi-MaP, the vision and text encoders are frozen and the proxy word embeddings " + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "text", + "content": " are learnable. Specifically, it first constructs the prompt embeddings based on the reference words provided by GPT-4 using a user's high-level concept, and then selects a reference word " + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "text", + "content": " for each image according to the similarity between the prompt embeddings " + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "text", + "content": " and the image embeddings " + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "text", + "content": ". Then, it combines the prompt and the reference words to form the new prompt embeddings " + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i^*" + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "text", + "content": " and maximizes the similarity to the image representation, so the proxy word embeddings " + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "text", + "content": " can capture the desired image features. In addition, the proxy word embeddings " + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "text", + "content": " should be close to the target concept word " + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_1" + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "text", + "content": " and the selected reference word " + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 46, + 325, + 545, + 403 + ], + "type": "text", + "content": " to construct the concept-level constraint and reference word constraint, which capture the features related to the user's interest." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "text", + "content": "For the sake of simplicity, we assume " + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "inline_equation", + "content": "h'(t) \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "text", + "content": " is defined on the whole set but only has the right estimation on a discrete set as " + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "inline_equation", + "content": "T = \\{t_i\\}" + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "text", + "content": " and the counterpart with the unconstrained set is denoted as " + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "inline_equation", + "content": "H(w)" + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "text", + "content": ". According to the definition, we have " + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "inline_equation", + "content": "\\forall t \\in T, h'(t) = H(t)" + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "text", + "content": ". The gap between the estimation from " + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "inline_equation", + "content": "h'" + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "text", + "content": " on unconstrained variable " + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 409, + 288, + 493 + ], + "type": "text", + "content": " can be depicted as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "content": "Theorem 1. Given " + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "inline_equation", + "content": "w \\notin T" + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "inline_equation", + "content": "t \\in T" + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "content": ", if assuming " + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "inline_equation", + "content": "h'" + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "inline_equation", + "content": "L_h" + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "inline_equation", + "content": "L_H" + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "content": "-Lipschitz continuous, we have" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 528, + 254, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 528, + 254, + 541 + ], + "spans": [ + { + "bbox": [ + 80, + 528, + 254, + 541 + ], + "type": "interline_equation", + "content": "\\| h ^ {\\prime} (w) - H (w) \\| _ {2} \\leq \\left(L _ {h} + L _ {H}\\right) \\| t - w \\| _ {2}", + "image_path": "a7525f58ba0fc826388b10c6f37569d9129d265aa3f370b569f14284f5e5f2e6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 547, + 227, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 547, + 227, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 547, + 227, + 558 + ], + "type": "text", + "content": "Proof. According to the definition, we have" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 563, + 282, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 563, + 282, + 622 + ], + "spans": [ + { + "bbox": [ + 52, + 563, + 282, + 622 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left\\| h ^ {\\prime} (w) - H (w) \\right\\| _ {2} = \\left\\| h ^ {\\prime} (w) - h ^ {\\prime} (t) + h ^ {\\prime} (t) - H (w) \\right\\| _ {2} \\\\ \\leq \\| h ^ {\\prime} (w) - h ^ {\\prime} (t) \\| _ {2} + \\| h ^ {\\prime} (t) - H (w) \\| _ {2} \\\\ = \\| h ^ {\\prime} (w) - h ^ {\\prime} (t) \\| _ {2} + \\| H (t) - H (w) \\| _ {2} \\\\ \\leq \\left(L _ {h} + L _ {H}\\right) \\| t - w \\| _ {2} \\\\ \\end{array}", + "image_path": "b0c9b2b8d4c2b734e76c59034c9711e3c69764f692bb391025ab37d8d0489aee.jpg" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 277, + 628, + 286, + 637 + ], + "blocks": [ + { + "bbox": [ + 277, + 628, + 286, + 637 + ], + "lines": [ + { + "bbox": [ + 277, + 628, + 286, + 637 + ], + "spans": [ + { + "bbox": [ + 277, + 628, + 286, + 637 + ], + "type": "image", + "image_path": "a3aeb8352ad7a5f0985af472da22dc1d96157a30fa0f311f31584f4132f28b69.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Remark Theorem 1 implies that the distance of the estimation " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "h'(w)" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " to the ground-truth result " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "H(w)" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " is bounded by that of " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " to an arbitrary discrete token " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": ". Therefore, by selecting the nearest token as the reference, the bound can be improved as shown in the following corollary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 410, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 410, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 545, + 434 + ], + "type": "text", + "content": "Corollary 2. With the assumptions in Theorem 1 and letting " + }, + { + "bbox": [ + 305, + 410, + 545, + 434 + ], + "type": "inline_equation", + "content": "t' = \\arg \\min_i \\| t_i - w\\| _2" + }, + { + "bbox": [ + 305, + 410, + 545, + 434 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 337, + 438, + 515, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 438, + 515, + 452 + ], + "spans": [ + { + "bbox": [ + 337, + 438, + 515, + 452 + ], + "type": "interline_equation", + "content": "\\left\\| h ^ {\\prime} (w) - H (w) \\right\\| _ {2} \\leq \\left(L _ {h} + L _ {H}\\right) \\| t ^ {\\prime} - w \\| _ {2}", + "image_path": "36557b9e841611f1ecf296121a3b67cae0c3806793143edfae4ab2e2bfb294f5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 465, + 446, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 446, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 446, + 477 + ], + "type": "text", + "content": "3.2.1 Concept-level Constraint" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "type": "text", + "content": "According to the above analysis, a good reference " + }, + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "type": "text", + "content": " can help guarantee the performance. Fortunately, the input concept (e.g., color) from the user can be leveraged as the reference to constrain the freedom of the proxy word. Therefore, given the target concept word " + }, + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "type": "text", + "content": ", we can obtain its token embedding as " + }, + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = \\phi(u)" + }, + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "type": "text", + "content": ". Then, to learn appropriate representations from the proxy embedding, the original problem can be rewritten with the constraint as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 323, + 584, + 528, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 584, + 528, + 598 + ], + "spans": [ + { + "bbox": [ + 323, + 584, + 528, + 598 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathbf {w} _ {i}) = - \\langle f (x _ {i}), h \\left(t _ {i} ^ {*}\\right) \\rangle \\quad s. t. \\quad \\| \\mathbf {w} _ {i} - \\mathbf {u} \\| _ {2} ^ {2} \\leq \\lambda", + "image_path": "3010f316c2b9b317a3b454d0a130c6ee11456346e9019e6a73ce76ee760b8670.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 603, + 472, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 603, + 472, + 615 + ], + "spans": [ + { + "bbox": [ + 306, + 603, + 472, + 615 + ], + "type": "text", + "content": "The constrained problem is equivalent to" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 331, + 620, + 545, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 620, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 331, + 620, + 545, + 633 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\mathbf {w} _ {i}\\right) = - \\langle f \\left(x _ {i}\\right), h \\left(t _ {i} ^ {*}\\right) \\rangle + \\alpha \\| \\mathbf {w} _ {i} - \\mathbf {u} \\| _ {2} ^ {2} \\tag {2}", + "image_path": "462a6fc13574da82ab83f237086501ac446ab5f47304ff9f1977601835b5a285.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 639, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 662 + ], + "type": "text", + "content": "following [3], which can be optimized effectively by gradient descent." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 670, + 544, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 544, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 544, + 682 + ], + "type": "text", + "content": "3.2.2 Constrained Optimization with Reference Word" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "However, it is well known that the user concept is often with a large scope covering a broad range of words (e.g., color" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14069" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 70, + 246, + 167 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 246, + 167 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 246, + 167 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 246, + 167 + ], + "type": "table", + "html": "
Datasets# Samples# Clusters
ALOI2882;2
Card8,02913;4
CMUface6404;20;2;4
Fruit1053;3
Fruit3604,8564;4
Stanford Cars1,2004;3
Flowers1,6004;4
", + "image_path": "023cb8528467ed42b5c8af2d7528723dccaf6a140745fdfbe0012979083f01b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 168, + 216, + 177 + ], + "lines": [ + { + "bbox": [ + 118, + 168, + 216, + 177 + ], + "spans": [ + { + "bbox": [ + 118, + 168, + 216, + 177 + ], + "type": "text", + "content": "Table 1. Dataset Statistics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 186, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 353 + ], + "type": "text", + "content": "covers all including but not limited to 'red', 'blue', 'green', etc.). As suggested by our above theoretical analysis, it is desired if the reference is as close as possible. In a clustering scenario and given only the user's high-level concept, it is challenging to find a closer reference word to further constrain the proxy learning. Fortunately, with the development of large language models (LLMs), we can leverage them (e.g., GPT-4) to provide relevant words according to a user's high-level concept as the candidate set and develop a selection strategy to obtain a closer reference word for each image. While the responses gathered from GPT-4 might not always precisely align with the data's ground truth, they indisputably furnish valuable candidate features, enriching the capabilities of Multi-MaP." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "type": "text", + "content": "To elucidate, considering the task of clustering a fruit dataset based on the concept of color, we can pose a query to GPT-4 as \"What are the common colors of fruit?\" The response we obtain is \"Common colors of fruit include red, yellow, green, orange, purple, and blue\". The colors enumerated in this response can serve as reference word candidates. It is worth noting that while the color spectrum derived from this method might surpass the actual colors present in the data, we collect all of them into the candidate set of reference words " + }, + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\{z_k\\}_k" + }, + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "type": "inline_equation", + "content": "z_1" + }, + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "type": "text", + "content": ": \"red\", " + }, + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "type": "inline_equation", + "content": "z_2" + }, + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "type": "text", + "content": ": \"yellow\", etc. Then, their text representations can be obtained from the prompt " + }, + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "type": "text", + "content": " as \"fruit with the color of " + }, + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "type": "inline_equation", + "content": "z_k" + }, + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "type": "text", + "content": "\". Given the image " + }, + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 354, + 288, + 510 + ], + "type": "text", + "content": ", the closest reference can be observed as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 122, + 520, + 212, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 520, + 212, + 536 + ], + "spans": [ + { + "bbox": [ + 122, + 520, + 212, + 536 + ], + "type": "interline_equation", + "content": "z _ {i} = \\arg \\max _ {k} \\left\\langle \\mathbf {x} _ {i}, \\mathbf {t} _ {k} \\right\\rangle", + "image_path": "c55ddbe7cb6fa0aff174e50afd97edf74d93218b6d86e1d0376c5dfbb80e5601.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 546, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 546, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_k = h(t_k)" + }, + { + "bbox": [ + 46, + 546, + 287, + 594 + ], + "type": "text", + "content": ". After that, " + }, + { + "bbox": [ + 46, + 546, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 46, + 546, + 287, + 594 + ], + "type": "text", + "content": " can be initialized with the token embedding of " + }, + { + "bbox": [ + 46, + 546, + 287, + 594 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 46, + 546, + 287, + 594 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 546, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i = \\phi(z_i)" + }, + { + "bbox": [ + 46, + 546, + 287, + 594 + ], + "type": "text", + "content": ". Moreover, we change the regularization using a closer reference word compared to the high-level concept as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 604, + 287, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 604, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 80, + 604, + 287, + 618 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\mathbf {w} _ {i}\\right) = - \\langle f \\left(x _ {i}\\right), h \\left(t _ {i} ^ {*}\\right) \\rangle + \\alpha \\| \\mathbf {w} _ {i} - \\mathbf {z} _ {i} \\| _ {2} ^ {2} \\tag {3}", + "image_path": "ae7f401c036c275ad32d143fe48de8dc69b3a993f76e6ecb45b229b306580885.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 635, + 172, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 172, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 172, + 647 + ], + "type": "text", + "content": "3.2.3 Contrastive Concepts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": "In some application scenarios, one user may need more than one clustering and provide high-level concepts as " + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{u_j\\}" + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": ", e.g., " + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "u_1" + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": ": \"color\", " + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "u_2" + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": ": \"species\", etc. For the concept \"color\", the irrelevant concept \"species\" can be leveraged as the negative constraint for the learning of proxy word. Concretely," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "let " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "u_{w}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " denote the target concept word, a contrastive loss can be adopted as regularization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 358, + 105, + 494, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 105, + 494, + 135 + ], + "spans": [ + { + "bbox": [ + 358, + 105, + 494, + 135 + ], + "type": "interline_equation", + "content": "R (\\mathbf {w} _ {i}) = - \\log \\frac {\\exp (\\mathbf {w} _ {i} ^ {\\top} \\mathbf {u} _ {w})}{\\sum_ {j} \\exp (\\mathbf {w} _ {i} ^ {\\top} \\mathbf {u} _ {j})}", + "image_path": "3b5fff37c747a8cd163c058219d46cc48b3282494a2750c35fbf96330155bff8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 143, + 435, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 143, + 435, + 154 + ], + "spans": [ + { + "bbox": [ + 306, + 143, + 435, + 154 + ], + "type": "text", + "content": "and the final objective becomes" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 164, + 545, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 164, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 164, + 545, + 220 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} \\left(\\mathbf {w} _ {i}\\right) = \\\\ - \\left\\langle f \\left(x _ {i}\\right), h \\left(t _ {i} ^ {*}\\right) \\right\\rangle + \\alpha \\| \\mathbf {w} _ {i} - \\mathbf {z} _ {i} \\| _ {2} ^ {2} - \\beta \\log \\frac {\\exp \\left(\\mathbf {w} _ {i} ^ {\\top} \\mathbf {u} _ {w}\\right)}{\\sum_ {j} \\exp \\left(\\mathbf {w} _ {i} ^ {\\top} \\mathbf {u} _ {j}\\right)} \\tag {4} \\\\ \\end{array}", + "image_path": "3b13ad3b96ed99578cfc1ea207ea4efb9ced1e9413bf75888a4915849a4353a5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 230, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 290 + ], + "type": "text", + "content": "where the first term is to infer the user-specific feature, while the latter two terms constrain the proxy word to the reference words for the appropriate representation extraction from the pre-trained text encoder. The overall framework of Multi-MaP is illustrated in Fig. 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 296, + 388, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 296, + 388, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 296, + 388, + 309 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 316, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 375 + ], + "type": "text", + "content": "To demonstrate our proposed method, we evaluate MultiMaP on all publicly available image datasets in multiple clustering, including ALOI [7], Stanford Cars [14], Card [35], CMUface [10], Flowers [20], Fruit [12], and Fruit360 [35] as summarized in Table 1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 376, + 546, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 532 + ], + "type": "text", + "content": "We compare Multi-MaP against five state-of-the-art methods: MSC [12] is a traditional multiple clustering method that uses hand-crafted features; MCV [9] leverages multiple feature extractors to represent different \"views\" of the same data and employs a multi-input neural network to enhance clustering outcomes; ENRC [18] is a deep multiple clustering method that integrates auto-encoder and clustering objective to generate different clusterings; iM-Clusts [27] makes use of the expressive representational power of deep autoencoders and multi-head attention to accomplish multiple clusterings; AugDMC [35] leverages augmentations to learn different image representations to achieve multiple clustering." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 539, + 414, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 414, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 414, + 552 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": "We employ Adam and set momentum as 0.9 to train the model for 1000 epochs. All hyperparameters are searched according to the loss score of Multi-MaP, where the learning rate is searched in " + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\{0.1, 0.05, 0.01, 0.005, 0.001, 0.0005\\}" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": ", weight decay is in " + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\{0.0005, 0.0001, 0.00005, 0.00001, 0\\}" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": " are in " + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\{0.0, 0.1, 0.2, \\ldots, 1.0\\}" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": " is fixed as 1 for all the experiments. For the non-pre-trained methods, we perform k-means [15] 10 times due to its randomness and evaluate the average clustering performance using two quantitative metrics, that is, Normalized Mutual Information (NMI) [32] and Rand index (RI) [26]. These measures range in " + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "inline_equation", + "content": "[0, 1]" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": ", and higher scores imply more accurate" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14070" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 70, + 506, + 252 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 506, + 252 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 506, + 252 + ], + "type": "table", + "html": "
DatasetClusteringMSCMCVENRCiMClustersAugDMCMulti-MaP
NMIRINMIRINMIRINMIRINMIRINMIRI
ALOIColor0.15630.34280.69820.74390.98330.98921.00001.00001.00001.00001.00001.0000
Shape0.29680.51990.73590.82610.97320.98610.99630.99891.00001.00001.00001.0000
FruitColor0.68860.80510.62660.76850.71030.85110.73510.86320.85170.91080.86190.9526
Species0.16270.60450.27330.65970.31870.65360.30290.67430.35460.73991.00001.0000
Fruit360Color0.25440.60540.37760.67910.42640.68680.40970.68410.45940.73920.62390.8243
Species0.21840.58050.29850.61760.41420.69840.38610.67320.51390.74300.52840.7582
CardOrder0.08070.78050.07920.71280.12250.73130.11440.76580.14400.82670.36530.8587
Suits0.04970.35870.04300.36380.06760.38010.07160.37150.08730.42280.27340.7039
CMUfaceEmotion0.12840.67360.14330.52680.15920.66300.04220.59320.01610.53670.17860.7105
Glass0.14200.57450.12010.49050.14930.62090.19290.56270.10390.53610.34020.7068
Identity0.38920.73260.46370.62470.56070.76350.51090.82600.58750.83340.66250.9496
Pose0.36870.63220.32540.60280.22900.50290.44370.61140.13200.55170.46930.6624
Stanford CarsColor0.23310.61580.21030.58020.24650.67790.23360.65520.27360.75250.73600.9193
Type0.13250.53360.16500.56340.20630.62170.19630.56430.23640.73560.63550.8399
FlowersColor0.25610.59650.29380.58600.33290.62140.31690.61270.35560.69310.64260.7984
Species0.13260.52730.15610.60650.18940.61950.18870.60770.19960.62270.60130.8103
", + "image_path": "8eee278c969f47dbe4bdb51f69502e190b17b7eabe5a1802911b9b9899b279ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 300 + ], + "type": "text", + "content": "clustering results. The experiments are conducted with GPU NVIDIA GeForce RTX 2080 Ti." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 493 + ], + "type": "text", + "content": "It should also be noted that some data are difficult to obtain corresponding candidate labels from GPT-4 or the labels do not provide semantic features, such as names. For example, for the identity clustering for the CMUface dataset [10], different identities represent different people and the semantic meaning of names should not affect the clustering results. In this case, we randomly extract 10 words from WordNet [5] as reference words, in order to make the candidate labels more distinctive. For instance, we randomly choose \"abstain, candid, function, haphazard, knot, luxury, nonchalance, pension, resilience, taciturn\" for the above scenario. Furthermore, all publicly available multiple clustering datasets provide each ground-truth clustering a high-level concept, e.g., 'shape', 'pose', etc. Therefore, in the experiment, we directly use them as users' preferences for our evaluation purposes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 502, + 190, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 190, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 190, + 514 + ], + "type": "text", + "content": "4.2. Performance Comparison" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 592 + ], + "type": "text", + "content": "In our experiments, after we obtain the proxy word embedding of each image for a desired concept, we feed them to k-means to obtain the corresponding clustering. Since k-means is random, we repeat ten times and the average performance is reported in Table 2. The best results are marked by bold numbers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 593, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 653 + ], + "type": "text", + "content": "We can observe that Multi-MaP outperforms the baselines in all the cases, indicating the superiority of the proposed method. This also shows a strong generalization ability of the pre-trained model by CLIP, which can capture the features of data in different aspects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Since our method uses the CLIP encoder and GPT-4 to obtain clustering results, a natural question arises how the performance would be if we directly use them in a zero-short manner. Therefore, we provide two zero-shot variants of CLIP, that is, " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{GPT}}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " that uses GPT-4 to obtain the can" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 274, + 544, + 445 + ], + "blocks": [ + { + "bbox": [ + 120, + 253, + 472, + 264 + ], + "lines": [ + { + "bbox": [ + 120, + 253, + 472, + 264 + ], + "spans": [ + { + "bbox": [ + 120, + 253, + 472, + 264 + ], + "type": "text", + "content": "Table 2. Quantitative comparison. The significantly best results with " + }, + { + "bbox": [ + 120, + 253, + 472, + 264 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 120, + 253, + 472, + 264 + ], + "type": "text", + "content": " confidence are in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 274, + 544, + 445 + ], + "lines": [ + { + "bbox": [ + 308, + 274, + 544, + 445 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 544, + 445 + ], + "type": "table", + "html": "
DatasetClusteringCLIPGPTCLIPlabelMulti-MaP
NMIRINMIRINMIRI
ALOIColor0.85810.94071.00001.00001.00001.0000
Shape1.00001.00001.00001.00001.00001.0000
FruitColor0.79120.90750.86290.97800.86190.9526
Species0.97930.99191.00001.00001.00001.0000
Fruit360Color0.56130.73050.57460.76730.62390.8243
Species0.43700.75520.53640.76310.52840.7582
CardOrder0.35180.84580.35180.84580.36530.8587
Suits0.27110.61230.27110.61230.27340.7039
CMUfaceEmotion0.15760.65320.15900.66190.17860.7105
Glass0.29050.68690.46860.75050.34020.7068
Identity0.19980.63880.26770.75450.66250.9496
Pose0.40880.64730.46910.64090.46930.6624
Stanford CarsColor0.65390.82370.68300.86420.73600.9193
Type0.62070.79310.64290.84560.63550.8399
FlowersColor0.56530.76290.58280.78360.64260.7984
Species0.56200.75530.60190.79960.60130.8103
", + "image_path": "b79a813f9965f48a9a79949debd7ca4e45602ae683d2015a1ddd5fb74def29ae.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 445, + 544, + 466 + ], + "lines": [ + { + "bbox": [ + 306, + 445, + 544, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 544, + 466 + ], + "type": "text", + "content": "Table 3. Variants of CLIP. The significantly best results with " + }, + { + "bbox": [ + 306, + 445, + 544, + 466 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 306, + 445, + 544, + 466 + ], + "type": "text", + "content": " confidence are in bold." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "content": "didate labels and predicts labels through zero-shot classification with all candidate labels as class names, and " + }, + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{label}}" + }, + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "content": " that performs zero-shot classification with all ground truth labels. It should be noted that " + }, + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{label}}" + }, + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "content": " uses an unfair setting with a ground-truth label set known in advance, which is expected to provide the best performance using CLIP in a zero-short manner. The results are shown in Table 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "text", + "content": "As expected, " + }, + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{label}}" + }, + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "text", + "content": " achieves better performance than " + }, + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{GPT}}" + }, + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "text", + "content": " in almost all cases, since " + }, + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{label}}" + }, + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "text", + "content": " uses a fixed ground truth label set, while " + }, + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{GPT}}" + }, + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "text", + "content": " uses candidate labels that may not match the ground truth exactly, introducing noise. Note that " + }, + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{GPT}}" + }, + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{label}}" + }, + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "text", + "content": " achieve the same results on Cards, because the candidate labels provided by GPT-4 are exactly the same as the true labels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Besides, Multi-MaP performs better than " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{GPT}}" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " in almost all cases, indicating that the proposed method can learn more effective features through its training process. Moreover, although " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{label}}" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " uses the ground truth, which is expected to be the best, Multi-MaP still outperforms " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{label}}" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " in some cases, such as the clustering of color for" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14071" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 69, + 509, + 301 + ], + "blocks": [ + { + "bbox": [ + 86, + 69, + 509, + 301 + ], + "lines": [ + { + "bbox": [ + 86, + 69, + 509, + 301 + ], + "spans": [ + { + "bbox": [ + 86, + 69, + 509, + 301 + ], + "type": "table", + "html": "
Multi-MaPpMulti-MaPcMulti-MaPrMulti-MaPcrMulti-MaP
ModulesProxy Learning✓✓✓✓✓
Concept Word×✓×✓✓
Reference Word××✓✓✓
Contrastive Concepts××××✓
NMI↑RI↑NMI↑RI↑NMI↑RI↑NMI↑RI↑NMI↑RI↑
ALOI [7]Color0.96190.98261.00001.00000.97950.98691.00001.00001.00001.0000
Shape1.00001.00001.00001.00001.00001.00001.00001.00001.00001.0000
Fruit [12]Color0.76420.84390.82150.92830.81360.90730.84840.93080.86190.9526
Species1.00001.00001.00001.00001.00001.00001.00001.00001.00001.0000
Fruit360 [35]Color0.56430.76650.62170.78360.59100.77460.60890.79650.62390.8243
Species0.50770.73680.51370.74360.50940.74250.51990.74280.52840.7582
Card [35]Order0.19320.81520.35680.84720.31130.82290.36160.80940.36530.8587
Suits0.23750.62820.26960.66410.24980.63650.25620.65990.27340.7039
CMUface [10]Emotion0.16900.61700.17140.62290.16970.63600.17130.68430.17860.7105
Glass0.31120.69110.32690.71360.31620.69170.33700.71080.34020.7068
Identity0.56170.82340.62430.83590.58390.82630.63910.89460.66250.9496
Pose0.43610.63860.45500.64990.43810.64290.43870.64890.46930.6624
Stanford cars [14]Color0.59390.78350.68360.86590.67290.86380.71120.91170.73600.9193
Type0.55690.79960.63830.82710.60910.80460.62890.81810.63550.8399
Flowers [20]Color0.57830.77230.58300.78330.59870.78490.62160.79410.64260.7984
Species0.57040.76080.57440.78420.57230.78110.58460.78920.60130.8103
", + "image_path": "5da8772cbc53ecd16c3723cff932647c8c5d3b0c4849a0fcdc2f39fd59cb1e3e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 308, + 536, + 322 + ], + "lines": [ + { + "bbox": [ + 56, + 308, + 536, + 322 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 536, + 322 + ], + "type": "text", + "content": "Table 4. Components ablation. All of our components boost performance consistently in all benchmark multi-clustering vision tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 338, + 168, + 434 + ], + "blocks": [ + { + "bbox": [ + 59, + 338, + 168, + 434 + ], + "lines": [ + { + "bbox": [ + 59, + 338, + 168, + 434 + ], + "spans": [ + { + "bbox": [ + 59, + 338, + 168, + 434 + ], + "type": "image", + "image_path": "1d379e2a9c14ac905850983fa8ae1b6801bed353d8651054dddb942dca9822a2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 437, + 160, + 446 + ], + "lines": [ + { + "bbox": [ + 63, + 437, + 160, + 446 + ], + "spans": [ + { + "bbox": [ + 63, + 437, + 160, + 446 + ], + "type": "text", + "content": "(a) Parameter sensitivity of " + }, + { + "bbox": [ + 63, + 437, + 160, + 446 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 448, + 265, + 459 + ], + "lines": [ + { + "bbox": [ + 67, + 448, + 265, + 459 + ], + "spans": [ + { + "bbox": [ + 67, + 448, + 265, + 459 + ], + "type": "text", + "content": "Figure 3. Parameter analysis of " + }, + { + "bbox": [ + 67, + 448, + 265, + 459 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 67, + 448, + 265, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 448, + 265, + 459 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 67, + 448, + 265, + 459 + ], + "type": "text", + "content": " on Fruit [12]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 338, + 279, + 434 + ], + "blocks": [ + { + "bbox": [ + 171, + 338, + 279, + 434 + ], + "lines": [ + { + "bbox": [ + 171, + 338, + 279, + 434 + ], + "spans": [ + { + "bbox": [ + 171, + 338, + 279, + 434 + ], + "type": "image", + "image_path": "9aa1ae24c98b35eb8f0bd4d28573803976cadca5c5595260233e2adab90e438e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 437, + 271, + 446 + ], + "lines": [ + { + "bbox": [ + 175, + 437, + 271, + 446 + ], + "spans": [ + { + "bbox": [ + 175, + 437, + 271, + 446 + ], + "type": "text", + "content": "(b) Parameter sensitivity of " + }, + { + "bbox": [ + 175, + 437, + 271, + 446 + ], + "type": "inline_equation", + "content": "\\beta" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "text", + "content": "Fruit360 dataset. This is because CLIP is more inclined to capture the features of one aspect of the data, while Multi-MaP learns better embedding of different aspects by training with the supervision of users' interests. Furthermore, Multi-MaP also achieves very competitive results as " + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{label}}" + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "text", + "content": " in the remaining cases, which further demonstrates the effectiveness of the proposed method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 561, + 141, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 141, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 141, + 574 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "text", + "content": "To validate the effectiveness of the proposed method, we show the gain from four components in Multi-MaP in Table 4. Let \"Multi-MaP " + }, + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "text", + "content": "\" denote the proxy learning without concept-level constraint, \"Multi-MaP " + }, + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "text", + "content": "\" denote the variant optimized by solely applying concept word to constrain the freedom of the proxy word, \"Multi-MaP " + }, + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "text", + "content": "\" denote the variant optimized by reference word provided by GPT-4 to find a closer reference word to further constrain the proxy learning, and \"Multi-MaP " + }, + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "inline_equation", + "content": "cr" + }, + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "text", + "content": "\" denote the variant leveraged both concept word and reference word." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "We can observe that the proxy learning only with concept" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "type": "text", + "content": "word or reference word, i.e., " + }, + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{Multi - MaP}_c" + }, + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{Multi - MaP}_r" + }, + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "type": "text", + "content": ", performs better than " + }, + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{Multi - MaP}_p" + }, + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "type": "text", + "content": ". This shows that the proposed concept-level constraint and constrained optimization with reference words play an important role in the model as demonstrated by our theoretical analysis. Moreover, the model with combined components, i.e., " + }, + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{Multi - MaP}_{cr}" + }, + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "type": "text", + "content": ", achieves better results than " + }, + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{Multi - MaP}_r" + }, + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{Multi - MaP}_c" + }, + { + "bbox": [ + 304, + 341, + 545, + 485 + ], + "type": "text", + "content": ". This indicates the effectiveness of the combination of reference words and reference concepts. Finally, our proposal using all including the contrastive concepts can further improve the performance, and thus provide the best results on all cases. This further demonstrates our proposal." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 491, + 421, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 421, + 504 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 421, + 504 + ], + "type": "text", + "content": "4.4. Parameter Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "We further investigate the effect of the reference word constraint weight " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " and concept-level constraint weight " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " varying from 0.0 to 0.5. The results of Multi-MaP on Fruit datasets are shown in Fig. 3 (a) and Fig. 3 (b), respectively. As " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " change, the proposed method keeps a species score of 1, since the image encoder can capture very effective species features. The above results also show that the proposed method can effectively capture useful information from images, reference words, and target concepts. As " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " increases, the proposed method first increases and then decreases, and reaches the maximum values at " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 0.4" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ". Similar results can be observed for " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " that the performance of the proposed method first increases and then decreases as " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " increases, and reaches the maximum values at " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta = 0.3" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ". Indicating a suitable value of " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " is helpful for Multi-MaP to obtain more effective embeddings for multiple clustering tasks. More studies such as efficiency analysis can be found" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14072" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 79, + 176, + 149 + ], + "blocks": [ + { + "bbox": [ + 71, + 79, + 176, + 149 + ], + "lines": [ + { + "bbox": [ + 71, + 79, + 176, + 149 + ], + "spans": [ + { + "bbox": [ + 71, + 79, + 176, + 149 + ], + "type": "image", + "image_path": "d47eed2ba007e88973ba891e69b4e708cc98869c6042b469f2ffed4e5e3ac151.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 158, + 177, + 169 + ], + "lines": [ + { + "bbox": [ + 75, + 158, + 177, + 169 + ], + "spans": [ + { + "bbox": [ + 75, + 158, + 177, + 169 + ], + "type": "text", + "content": "(a) Color using ground truth" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 80, + 294, + 146 + ], + "blocks": [ + { + "bbox": [ + 186, + 80, + 294, + 146 + ], + "lines": [ + { + "bbox": [ + 186, + 80, + 294, + 146 + ], + "spans": [ + { + "bbox": [ + 186, + 80, + 294, + 146 + ], + "type": "image", + "image_path": "68305eda48f13a2802bee29840a20200e88b2c25c145aebc5e3e32b8d7b46396.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 158, + 302, + 169 + ], + "lines": [ + { + "bbox": [ + 187, + 158, + 302, + 169 + ], + "spans": [ + { + "bbox": [ + 187, + 158, + 302, + 169 + ], + "type": "text", + "content": "(b) Color using reference words" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 324, + 80, + 411, + 148 + ], + "blocks": [ + { + "bbox": [ + 324, + 80, + 411, + 148 + ], + "lines": [ + { + "bbox": [ + 324, + 80, + 411, + 148 + ], + "spans": [ + { + "bbox": [ + 324, + 80, + 411, + 148 + ], + "type": "image", + "image_path": "9cc9acdca6050ec7a34c0cff89c072dcce12242d8f3a52229a5da63be06a9538.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 158, + 407, + 169 + ], + "lines": [ + { + "bbox": [ + 321, + 158, + 407, + 169 + ], + "spans": [ + { + "bbox": [ + 321, + 158, + 407, + 169 + ], + "type": "text", + "content": "(c) Color of Multi-MaP" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 105, + 512, + 113 + ], + "lines": [ + { + "bbox": [ + 427, + 105, + 512, + 113 + ], + "spans": [ + { + "bbox": [ + 427, + 105, + 512, + 113 + ], + "type": "text", + "content": "red yellow" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 429, + 114, + 512, + 121 + ], + "lines": [ + { + "bbox": [ + 429, + 114, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 429, + 114, + 512, + 121 + ], + "type": "text", + "content": "green orange" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 429, + 122, + 504, + 130 + ], + "lines": [ + { + "bbox": [ + 429, + 122, + 504, + 130 + ], + "spans": [ + { + "bbox": [ + 429, + 122, + 504, + 130 + ], + "type": "text", + "content": "purple blue" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 440, + 158, + 510, + 169 + ], + "lines": [ + { + "bbox": [ + 440, + 158, + 510, + 169 + ], + "spans": [ + { + "bbox": [ + 440, + 158, + 510, + 169 + ], + "type": "text", + "content": "(d) Legend of color" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 84, + 179, + 174, + 247 + ], + "blocks": [ + { + "bbox": [ + 84, + 179, + 174, + 247 + ], + "lines": [ + { + "bbox": [ + 84, + 179, + 174, + 247 + ], + "spans": [ + { + "bbox": [ + 84, + 179, + 174, + 247 + ], + "type": "image", + "image_path": "4232e29350efc6f472a931a4f2780a4e8977717639f2aec56c9172d58296d812.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 258, + 180, + 269 + ], + "lines": [ + { + "bbox": [ + 71, + 258, + 180, + 269 + ], + "spans": [ + { + "bbox": [ + 71, + 258, + 180, + 269 + ], + "type": "text", + "content": "(e) Species using ground truth" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 195, + 178, + 291, + 248 + ], + "blocks": [ + { + "bbox": [ + 195, + 178, + 291, + 248 + ], + "lines": [ + { + "bbox": [ + 195, + 178, + 291, + 248 + ], + "spans": [ + { + "bbox": [ + 195, + 178, + 291, + 248 + ], + "type": "image", + "image_path": "5f62f49f5cb29726cf762be8dc6bd08433e39a964631a3e5850911afe87ab465.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 258, + 305, + 269 + ], + "lines": [ + { + "bbox": [ + 184, + 258, + 305, + 269 + ], + "spans": [ + { + "bbox": [ + 184, + 258, + 305, + 269 + ], + "type": "text", + "content": "(f) Species using reference words" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 322, + 180, + 410, + 247 + ], + "blocks": [ + { + "bbox": [ + 322, + 180, + 410, + 247 + ], + "lines": [ + { + "bbox": [ + 322, + 180, + 410, + 247 + ], + "spans": [ + { + "bbox": [ + 322, + 180, + 410, + 247 + ], + "type": "image", + "image_path": "d7301845ed65e734b73418bd46dee7d5539e0cdd9d5b4b96946db9fa0bf95ebe.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 258, + 411, + 269 + ], + "lines": [ + { + "bbox": [ + 317, + 258, + 411, + 269 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 411, + 269 + ], + "type": "text", + "content": "(g) Species of Multi-MaP" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 178, + 525, + 255 + ], + "lines": [ + { + "bbox": [ + 429, + 178, + 525, + 255 + ], + "spans": [ + { + "bbox": [ + 429, + 178, + 525, + 255 + ], + "type": "text", + "content": "- apple\n- orange\n- banana\n- strawberry\n- grapes\n- raspberry\n- blueberry\n- cherry\n- pear\n- plum\n- peach\n- nectarine\n- pineapple\n- kiwi\n- watermelon\n- cantaloupe\n- mango" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 436, + 258, + 514, + 269 + ], + "lines": [ + { + "bbox": [ + 436, + 258, + 514, + 269 + ], + "spans": [ + { + "bbox": [ + 436, + 258, + 514, + 269 + ], + "type": "text", + "content": "(h) Legend of species" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 277, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 45, + 277, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 45, + 277, + 547, + 312 + ], + "type": "text", + "content": "Figure 4. Visualization of feature embeddings and related labels. The points represent the image or pseudo-word embeddings, and the triangles represent the prompt or label embeddings. Different colors represent different labels, which are indicated by the text next to the triangles." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 316, + 135, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 135, + 328 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 135, + 328 + ], + "type": "text", + "content": "in the supplementary." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 335, + 130, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 130, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 130, + 347 + ], + "type": "text", + "content": "4.5. Visualization" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": "To further demonstrate the effectiveness of the proposed method, we visualize the representations obtained in " + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{label}}" + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{GPT}}" + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": ", and Multi-MaP. Specifically, for " + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{label}}" + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{GPT}}" + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": ", we visualize the image representations as well as prompts generated with real labels and candidate labels, respectively. For the Multi-MaP, we visualize the word embedding " + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "inline_equation", + "content": "w_{*}" + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": " and the candidate labels selected for initialization. The results are shown in Fig. 4. For species clustering, we can see that image embeddings show very clear boundaries and correspond well to the prompt for " + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{label}}" + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": ", which indicates that CLIP can effectively capture the features of the species in the data. " + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{GPT}}" + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": " uses the candidate labels to generate prompts, which introduces more noise, but benefits from the CLIP text encoder, the image embedding can keep a relatively far distance from most of the irrelevant prompts. However, since there are still a few images that are labeled as peaches (i.e., a noisy label), it performs slightly worse than " + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{label}}" + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": ", as shown in Fig. 4(e). Besides, Multi-MaP can capture the image and users' interests in the training process, therefore it compensates for the shortcomings of " + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{GPT}}" + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": " and achieves better results. On the other hand, for color clustering, the prompts are farther away from " + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{label}}" + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{GPT}}" + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": ", that indicates the image embeddings mainly capture the features of species, which have no direct connection with color. " + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{GPT}}" + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": " generates the prompt from the candidate label, which has more noise than the ground truth label, resulting in worse performance than " + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{label}}" + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": ". The proposed method can distinguish different colors more clearly, because it can learn from the user's interest and capture the color-related features. How" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 316, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 375 + ], + "type": "text", + "content": "ever, some red color embeddings are closer to purple, because some images in the datasets are actually purple, but labeled as red. To sum up, the proposed method can learn more effective embeddings based on the users' interests for multiple clustering tasks." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 381, + 378, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 378, + 394 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 378, + 394 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 399, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 639 + ], + "type": "text", + "content": "To conclude, our study thoroughly investigates the significant challenges that current advanced deep learning techniques face in multiple clustering. A key issue is that users often do not need every clustering result produced by an algorithm, and selecting the most relevant one requires an in-depth understanding of each outcome. To overcome this challenge, the proposed method introduces a novel multimodal proxy learning process, which effectively aligns a user's brief keyword describing the interest with the corresponding vision components. By integrating a multi-modal model and GPT-4 to precisely capture a user's interest using a keyword, the proposed approach uses both reference word constraint and concept-level constraint to discover personalized clustering result(s), which can also lead to enhanced performance. Experiments on diverse datasets demonstrate the superiority of the proposed method in multiple clustering tasks with a precise capture of a user's interest. The proposed method is limited by data with semantic meaningful labels, although we can use WordNet to help, whose comprehensive study will be our future work." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 645, + 416, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 416, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 416, + 658 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "This research is supported in part by Advata Gift funding. All opinions, findings, conclusions and recommendations in this paper are those of the author and do not necessarily reflect the views of the funding agencies." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14073" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Eric Bae and James Bailey. Coala: A novel approach for the extraction of an alternate clustering of high quality and high dissimilarity. In ICDM, pages 53-62. IEEE, 2006. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 147 + ], + "type": "text", + "content": "[2] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning. Springer, 2006. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 286, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 286, + 170 + ], + "type": "text", + "content": "[3] Stephen P. Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University Press, 2014. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 171, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 216 + ], + "type": "text", + "content": "[4] Xuan Hong Dang and James Bailey. Generation of alternative clusterings using the cami approach. In Proceedings of the 2010 SIAM International Conference on Data Mining, pages 118-129. SIAM, 2010. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 217, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 249 + ], + "type": "text", + "content": "[5] Christiane Fellbaum. Wordnet. In Theory and applications of ontology: computer applications, pages 231-243. Springer, 2010. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 251, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 304 + ], + "type": "text", + "content": "[6] 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. International Journal of Computer Vision, pages 1-15, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 307, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 287, + 350 + ], + "type": "text", + "content": "[7] Jan-Mark Geusebroek, Gertjan J Burghouts, and Arnold WM Smeulders. The amstedam library of object images. International Journal of Computer Vision, 61:103-112, 2005. 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 352, + 286, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 286, + 396 + ], + "type": "text", + "content": "[8] David Gondek and Thomas Hofmann. Conditional information bottleneck clustering. In 3rd IEEE international conference on data mining, workshop on clustering large data sets, pages 36-42, 2003. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 397, + 286, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 286, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 286, + 440 + ], + "type": "text", + "content": "[9] Joris Guérin and Byron Boots. Improving image clustering with multiple pretrained cnn feature extractors. In *British Machine Vision Conference* 2018, BMVC 2018, 2018. 1, 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 442, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 442, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 442, + 287, + 497 + ], + "type": "text", + "content": "[10] Stephan Gunnemann, Ines Farber, Matthias Rüdiger, and Thomas Seidl. Smvc: semi-supervised multi-view clustering in subspace projections. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 253-262, 2014. 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 498, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 529 + ], + "type": "text", + "content": "[11] Juhua Hu and Jian Pei. Subspace multi-clustering: a review. Knowledge and information systems, 56(2):257-284, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 533, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 565 + ], + "type": "text", + "content": "[12] Juhua Hu, Qi Qian, Jian Pei, Rong Jin, and Shenghuo Zhu. Finding multiple stable clusterings. Knowledge and Information Systems, 51(3):991-1021, 2017. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "text", + "content": "[13] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[14] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 667 + ], + "type": "text", + "content": "[15] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[16] James MacQueen et al. Some methods for classification and analysis of multivariate observations. In Proceedings of the fifth Berkeley symposium on mathematical statistics and probability, pages 281-297. Oakland, CA, USA, 1967. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[17] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "type": "text", + "content": "[18] Lukas Miklautz, Dominik Mautz, Muzaffer Can Altinigneli, Christian Böhm, and Claudia Plant. Deep embedded nonredundant clustering. In Proceedings of the AAAI conference on artificial intelligence, pages 5174-5181, 2020. 1, 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 154, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 154, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 154, + 545, + 186 + ], + "type": "text", + "content": "[19] Andrew Ng, Michael Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 14, 2001. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 189, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 189, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 189, + 545, + 232 + ], + "type": "text", + "content": "[20] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 235, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 235, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 235, + 545, + 266 + ], + "type": "text", + "content": "[21] ZiJie Qi and Ian Davidson. A principled and flexible framework for finding alternative clusterings. In SIGKDD, pages 717-726, 2009. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 270, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 270, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 270, + 545, + 312 + ], + "type": "text", + "content": "[22] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Tacoma Tacoma, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In ICCV, pages 6449-6457. IEEE, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 316, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 369 + ], + "type": "text", + "content": "[23] Qi Qian, Yuanhong Xu, Juhua Hu, Hao Li, and Rong Jin. Unsupervised visual representation learning by online constrained k-means. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16640-16649, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 372, + 545, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 416 + ], + "type": "text", + "content": "[24] Qi Qian, Yuanhong Xu, and Juhua Hu. Intra-modal proxy learning for zero-shot visual categorization with clip. In Thirty-seventh Conference on Neural Information Processing Systems, NeurIPS 2023, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 418, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 545, + 483 + ], + "type": "text", + "content": "[25] 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, 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 486, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 486, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 486, + 545, + 518 + ], + "type": "text", + "content": "[26] William M Rand. Objective criteria for the evaluation of clustering methods. Journal of the American Statistical Association, 66(336):846-850, 1971. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 521, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 521, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 545, + 564 + ], + "type": "text", + "content": "[27] Liangrui Ren, Guoxian Yu, Jun Wang, Lei Liu, Carlotta Domeniconi, and Xiangliang Zhang. A diversified attention model for interpretable multiple clusterings. IEEE Transactions on Knowledge and Data Engineering, 2022. 1, 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "type": "text", + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685. IEEE, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "text", + "content": "[29] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Advances in Neural Information Processing Systems, 35:14274-14289, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[30] Junyang Wang, Yuanhong Xu, Juhua Hu, Ming Yan, Jitao Sang, and Qi Qian. Improved visual fine-tuning with natural language supervision. arXiv preprint arXiv:2304.01489, 2023. 1, 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14074" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 308 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[31] Shaowei Wei, Jun Wang, Guoxian Yu, Carlotta Domeni-coni, and Xiangliang Zhang. Multi-view multiple clusterings using deep matrix factorization. In Proceedings of the AAAI conference on artificial intelligence, pages 6348–6355, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "type": "text", + "content": "[32] JV White, Sam Steingold, and CG Fournelle. Performance metrics for group-detection algorithms. Proceedings of Interface, 2004, 2004. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "type": "text", + "content": "[33] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487, 2016. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "type": "text", + "content": "[34] Sen Yang and Lijun Zhang. Non-redundant multiple clustering by nonnegative matrix factorization. Machine Learning, 106(5):695-712, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 230, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 263 + ], + "type": "text", + "content": "[35] Jiawei Yao, Enbei Liu, Maham Rashid, and Juhua Hu. Augdmc: Data augmentation guided deep multiple clustering. In INNS DLIA@IJCNN, 2023. 1, 3, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[36] 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. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14075" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-Object Tracking in the Dark/4c20e9a2-1106-45cf-a36a-0161982e6015_content_list.json b/2024/Multi-Object Tracking in the Dark/4c20e9a2-1106-45cf-a36a-0161982e6015_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8fdd8515b56b452d2d1c66a0dc0ccb4f1a73c06a --- /dev/null +++ b/2024/Multi-Object Tracking in the Dark/4c20e9a2-1106-45cf-a36a-0161982e6015_content_list.json @@ -0,0 +1,1636 @@ +[ + { + "type": "text", + "text": "Multi-Object Tracking in the Dark", + "text_level": 1, + "bbox": [ + 307, + 130, + 661, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinzhe Wang Kang Ma Qiankun Liu Yunhao Zou Ying Fu * Beijing Institute of Technology", + "bbox": [ + 210, + 180, + 764, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wangxinzhe, makang, liuqk3, zouyunhao, fuying}@bit.edu.cn", + "bbox": [ + 228, + 219, + 741, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Low-light scenes are prevalent in real-world applications (e.g. autonomous driving and surveillance at night). Recently, multi-object tracking in various practical use cases have received much attention, but multi-object tracking in dark scenes is rarely considered. In this paper, we focus on multi-object tracking in dark scenes. To address the lack of datasets, we first build a Low-light Multi-Object Tracking (LMOT) dataset. LMOT provides well-aligned low-light video pairs captured by our dual-camera system, and high-quality multi-object tracking annotations for all videos. Then, we propose a low-light multi-object tracking method, termed as LTrack. We introduce the adaptive low-pass downsample module to enhance low-frequency components of images outside the sensor noises. The degradation suppression learning strategy enables the model to learn invariant information under noise disturbance and image quality degradation. These components improve the robustness of multi-object tracking in dark scenes. We conducted a comprehensive analysis of our LMOT dataset and proposed LTrack. Experimental results demonstrate the superiority of the proposed method and its competitiveness in real night low-light scenes. Dataset and Code: https://github.com/ying-fu/LMOT", + "bbox": [ + 75, + 301, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 680, + 209, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-object tracking (MOT) aims to locate and associate multiple objects in video sequences. It is widely used in many downstream applications, such as video recognition [7, 32], autonomous driving [18], and surveillance [11]. Recently, multi-object tracking in various practical use cases has garnered much attention [8, 9, 18, 34, 42], greatly advancing the development of MOT. However, these works are primarily tailored for high-quality inputs and overlook the prevalent low-light scenes in real-world scenes. Motivated by this, we study multi-object tracking in dark scenes.", + "bbox": [ + 75, + 705, + 470, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to the physical limitations of existing cameras, ac", + "bbox": [ + 96, + 859, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quiring high-quality videos under low-light conditions is difficult. One inherent difficulty in capturing consecutive video frames under such conditions is avoiding motion blur. Current camera technology typically requires short exposure times (usually just a few tens of milliseconds), but in low-light scenarios, the sensor struggles to capture an adequate number of photons within this limited duration. This limitation inevitably leads to degradation in image quality accompanied by higher noise levels. This presents two main challenges for MOT under low-light conditions. The first challenge is for collecting a low-light multi-object tracking dataset. Collecting and annotating a low-light MOT dataset is difficult and expensive. MOT requires dynamic object videos, but videos captured in low-light scenes have extremely low brightness, making it hard to recognize and annotate objects in the videos. The second challenge revolves around low-light multi-object tracking. The popular tracking-by-detection paradigm [3, 10, 29, 30, 62] generally consists of detector, motion-based association module, and appearance-based association module. These modules typically require high-quality input images. The poor quality of low-light images leads to severe performance degradation for both detectors and appearance-based correlation modules. A simple approach is to cascade low-light enhance modules [2, 23, 24, 47], but this introduces additional computational costs. Furthermore, images optimized for visual quality may be suboptimal for downstream tasks [6, 21, 27].", + "bbox": [ + 496, + 268, + 893, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we build a low-light multi-object tracking dataset (LMOT), specifically designed to address the challenges of multi-object tracking in dark scenes. To this end, we develop a dual-camera system that simultaneously captures well-lit and low-light video frames. The video pairs are highly aligned in both spatial and temporal dimensions, offering two key benefits. First, it enables us to annotate on the well-lit videos, resulting in high-quality annotations. Second, the well-lit videos can provide additional supervision information during the training phase, and strongly enhance the performance in the dark scenes. After careful annotation, we collect 32 video sequences $(2.3\\times \\mathrm{MOT17})$ over $35\\mathrm{K}$ frames $(3.1\\times \\mathrm{MOT17})$ ,and over $815\\mathrm{K}$ bounding boxes $(2.8\\times \\mathrm{MOT17})$ .The RAW data is the output of", + "bbox": [ + 496, + 689, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 94, + 886, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "382", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the image sensor and is the input data of the image signal processor (ISP). It saves all information from the image sensor, which is crucial for capturing object information in dark scenes [6, 52, 67]. Therefore, we collect RAW videos for LMOT.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, we propose a low-light multi-object tracking method, termed as LTrack. The low-light video is characterized by substantial sensor noise and poor image quality, which significantly degrades both shallow and deep feature representations, leading to reduced tracking performance. We observe that the sensor noise in low-light images exhibits a similarity to adversarial attacks [35, 43]. To address this issue, our main idea is to learn the invariant semantic information under noise disturbance quality degradation. We present the adaptive low-pass downsample module (ALD). It employs spatial low-pass convolution to extract low-frequency components from images, excluding noises, and adaptively enhance the feature maps. We also present the degradation suppression learning strategy (DSL), which utilizes paired low-light videos to help the model suppress the noise disturbance and encourage image content response in the feature domain. We conduct a comprehensive analysis of our LMOT dataset and validate the superiority of our LTrack in real-world night scenes.", + "bbox": [ + 75, + 167, + 472, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 96, + 455, + 434, + 470 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We build the first low-light multi-object tracking dataset using a carefully constructed dual-camera system. It provides well-aligned low-light videos in RAW format, and high-quality MOT annotations for all videos.", + "- We propose a low-light multi-object tracking method. It utilizes the adaptive low-pass downsample module and the degradation suppression learning strategy to learn to extract invariant features from low-light videos.", + "- We conduct a comprehensive analysis of our dataset and the proposed method. Experimental results demonstrate the superiority of the proposed method and its competitiveness on real night scenes." + ], + "bbox": [ + 76, + 474, + 468, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 671, + 218, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first review the current research status of low-light enhancement and low-light datasets. Then, we summarized the present research for multi-object tracking and tracking in the dark scenes.", + "bbox": [ + 75, + 696, + 468, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Low-light enhancement. Traditional methods for lowlight enhancement are primarily based on histogram equalization and Retinex theory [20, 22, 25]. Recently, deep learning has been explored for many low-level tasks [14, 44, 45, 51, 60, 61, 65, 66], and achieved superior results on low-light enhancement [2, 4, 13, 15, 19, 24, 39, 53]. While these methods are capable of recovering images with high visual quality, they often require heavy computation and may not consider downstream tasks, resulting in suboptimal", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/02e02ea9df4d388f77874d5066d6b2a59c03d1a20f1308b2772c31d5a2c07a46.jpg", + "image_caption": [ + "Figure 1. Our dual-camera system. It consists of two cameras, a beam splitter, and an ND-filter. Two cameras of identical models are meticulously engineered to achieve pixel-by-pixel alignment in the captured video data." + ], + "image_footnote": [], + "bbox": [ + 503, + 89, + 890, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performance. In contrast, our approach focuses on directly learning multi-object tracking from low-light images, thus bypassing the low-light enhancement.", + "bbox": [ + 498, + 354, + 890, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Low-light datasets. The long-short exposure is a widely used method to collect paired low-light images, but can only be used to collect low-light images for static scenes. [4, 5, 49]. To capture dynamic scenes, some works designed the mechatronic system. They obtain paired low-light data by repeating the motion twice [13, 47]. However, these mechatronic systems cannot be used to collect dynamic object videos in the wild. Jiang et al. [23] designed a dual-camera system that simultaneously captures paired well-lit and low-light videos, making it possible to capture dynamic scenes and dynamic object videos for multi-object tracking. Zou et al. [57] setup an optical system to collect paired videos and event streams. These works explore various ways to construct low-light datasets and inspire research on high-level vision tasks in dark scenes, such as LOD [21] for object detection, LIS [6] for instance segmentation, and ExPose [27] for human pose estimation. These datasets provide paired low-light data only for image tasks, and cannot be expanded for multi-object tracking in dark scenes.", + "bbox": [ + 498, + 409, + 892, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-object tracking datasets. MOT15 [26] is the first large-scale benchmark for multi-object tracking. MOT17 [34] stands as one of the most widely applied MOT benchmarks. MOT20 [9] focuses on very crowded scenes. These three datasets are for pedestrians. KITTI [18] and BDD100K [55] are for autonomous driving scenarios. DanceTrack [42] focuses on dancing scenes and is characterized by similar appearance and diverse motions. Recently, SportsMOT [8] aims to track athletes and encourage algorithms to promote both appearance and motion association. These datasets explore multi-object tracking in various practical use cases, but none of them consider multi-object tracking in dark scenes.", + "bbox": [ + 498, + 703, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "383", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f14b87b761c53707cda4a4222aa69192accfca9747e49a4a8a5fec6a6e51d76d.jpg", + "image_caption": [ + "Figure 2. Two example videos from our LMOT dataset. It provides well-aligned low-light video pairs and MOT annotations for all videos. The time interval between adjacent frames is 1s. The first row is the low-light video, the second row is the scaled low-light video and the last row is the well-lit video. Our LMOT dataset is collected from city outdoor scenes." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 488, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/046c4b39280bea23ad40023227baac67a1b0c29921c77ad36239dcbb902d9f77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 88, + 890, + 255 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/75f0c68c13df678050980d1ed59493e89d533c0d802e3dd8ebd401300af0c8c5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetFormatVideosFramesLength (s)BboxTracks
MOT17 [34]sRGB1411,235463292,7331,342
MOT20 [9]sRGB813,4105351,652,0403,456
DanceTrack [42]sRGB100105,8555,292-990
SportMOT [8]sRGB240150,3796,0151,629,4903,401
KITTI [18]sRGB218,000-47,000917
BDD100K [55]sRGB1,600318,000-3,300,000131,000
SWIR [36]SWIR-7,309-57,221-
LMOTRAW3235,1201,756815,5504,090
", + "bbox": [ + 81, + 309, + 465, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object tracking in the dark. To track in low-light scenes, some methods explore to use multi-modal information for single-object tracking, such as event camera [58, 64], depth [40, 54] and thermal [46, 59] devices. Park et al. [36] proposed to use Short-Wave Infrared (SWIR) images for multi-object tracking, since its advantages in terms of robustness in low-light conditions. The common drawback of these methods is that they require additional hardware equipment and cannot be applied to the most widely used CMOS imaging systems.", + "bbox": [ + 75, + 478, + 468, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SORT [1] uses the Kalman Filter motion model and employs IoU for association. ByteTrack [62] enhances tracking performance by considering low-confidence bounding boxes. OS-SORT [3] enhances SORT by restoring lost targets. Recently, Transformer has been explored for MOT [16, 33, 41, 56, 63]. These methods have achieved high performance in many practical scenarios, but they do not consider dealing with low-light conditions. We focus on multi-object tracking under low-light conditions. Based on RAW videos, our method is highly practical with excellent performance and does not require additional hardware.", + "bbox": [ + 75, + 630, + 468, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Low-light Multi-object Tracking Dataset", + "text_level": 1, + "bbox": [ + 76, + 813, + 441, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce our dual camera system and the details of collecting and annotating our low-light multi-object tracking (LMOT) dataset. Then, we analyzed the statistical characteristics of our LMOT dataset.", + "bbox": [ + 75, + 839, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/1a40ca74c89a1ea3f9aa582941c77d145e1a9b36811c8dd1bb7e031822bbd6c6.jpg", + "table_caption": [ + "Table 1. Comparison of statistics between existing MOT datasets and our LMOT dataset" + ], + "table_footnote": [], + "table_body": "
DatasetSplitVideosBboxTracksPaired Well-lit
LMOT-dualtrain11309,4661533✓
val4131,781626✓
test11312,7421644✓
LMOT-realreal661,561287
", + "bbox": [ + 504, + 309, + 888, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2. Detailed statistics and data splits for LMOT.", + "bbox": [ + 535, + 390, + 852, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dataset Construction", + "text_level": 1, + "bbox": [ + 500, + 421, + 697, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-object tracking requires dynamic scenes and object video. To collect low-light videos for multi-object tracking, we build a dual-camera system [23], which is illustrated in Fig. 1. It can simultaneously capture paired lowlight and well-lit video pairs. Its main components include a beam splitter, a neutral density (ND) filter, and two FLIR Grasshopper3 GS3-U3-23S6C cameras. The beam splitter divides the incoming light into two separate paths. This arrangement allows one camera to capture well-lit images directly, while the other camera records low-light images, with the ND-filter attenuating the light intensity. To ensure temporal synchronization of the video frames, we employ the hardware interface to trigger the camera exposure events. Moreover, to avoid frame loss, our dual-camera system uses two independent hardware interfaces for data transmission and is equipped with a high-speed solid-state drive. Thanks to precise calibration, our dual-camera system can capture paired low-light and well-lit videos in real time for dynamic scenes and objects. More details about our dual-camera system are given in supplementary materials.", + "bbox": [ + 496, + 446, + 890, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We save the video frames in RAW format before the images are processed by the camera image signal processor (ISP). In terms of camera settings, we set the exposure time for both cameras to $10ms$ , and the frame rate is fixed at 20. This setting is feasible to avoid motion blur. We adjust the gain level for well-lit cameras to achieve optimal image quality. The gain for low-light cameras is consistently set to the maximum value, to simulate low-light capturing setup in real scenarios. We also collect a real low-light MOT dataset (LMOT-real) to evaluate performance in real-world", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "384", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4c7dae4c6446524a305bb96a7b1e8b7a9ac9a898214ea69671b1262768a5cec9.jpg", + "image_caption": [ + "(a) Number of instances per category" + ], + "image_footnote": [], + "bbox": [ + 86, + 90, + 370, + 193 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1d3709a5a938d9b2e9f9b3d3e8c00ff8410f251cd6fd2bb651690ed41815680a.jpg", + "image_caption": [ + "(b) IoU on adjacent frames", + "Figure 3. (a) Number of instances per category. LMOT consists of 6 categories, most of the instances are the person and car. (b) IoU on adjacent frames. Compared to MOT17, KITTI, and DanceTrack, LMOT has a roughly average score. This indicates that LMOT has a relatively normal movement speed. (c) Cosine distance of appearance features. The cosine distance is smaller under low-light conditions, indicating that the appearance distinguishability is decreased under low-light conditions." + ], + "image_footnote": [], + "bbox": [ + 380, + 90, + 607, + 193 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/10f3feb2bb204853dce35c9fa173111cd49802ff040fddc688a9499a8c5eb335.jpg", + "image_caption": [ + "(c) Cosine distance of appearance features" + ], + "image_footnote": [], + "bbox": [ + 616, + 90, + 883, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dark scenes. These videos are captured using a single camera with the same camera settings.", + "bbox": [ + 75, + 282, + 467, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our LMOT dataset contains a variety of city outdoor scenes, including roads, overpasses, pedestrians, and intersection. The overpass scenes take an overhead shot of objects, while all other scenes are captured from the perspective of pedestrians. To account for the impact of camera motion, we introduce arbitrary horizontal rotations and vertical random movements to the camera. Fig. 2 shows two sampled video sequences from our LMOT dataset.", + "bbox": [ + 75, + 315, + 467, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We annotate six types of moving objects, including car, person, bicycle, motorcycle, bus, and truck. The annotated labels include bounding boxes, identifications, and visibility status. For partly occluded objects, a full box is annotated. For a fully occluded object, an estimated box is annotated. Each object has a unique ID throughout the entire video. Thanks to well-aligned low-light and well-lit videos, we can annotate well-lit videos to simultaneously obtain labels for low-light videos. This greatly reduces the annotation difficulty and enhances quality. Lastly, we carefully review all the annotation results.", + "bbox": [ + 75, + 438, + 467, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dataset Statistic", + "text_level": 1, + "bbox": [ + 76, + 617, + 236, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LMOT is a large-scale dataset that focuses on multi-object tracking in dark scenes. We compare the statistics of LMOT with existing MOT datasets in Tab. 1. It can be seen that LMOT is approximately three times larger than MOT17 [34]. Compared to large-scale datasets such as DanceTrack [42], SportsMOT [25], KITTI [18] and BDD100K [55], the scale of LMOT is still considerable. It should be emphasized that these datasets are not for multi-object tracking in dark scenes and only provide sRGB images. Compared to SWIR, our LMOT dataset has approximately $5 \\times$ frames and $14 \\times$ bounding boxes. The detailed statistics and data splits for LMOT are shown in Tab. 2.", + "bbox": [ + 75, + 642, + 467, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present the number of instances for each category in Fig. 3 (a). The majority of instances are persons and cars. As shown in Fig. 3 (b), the average IoU on adjacent frames of LMOT is lower than MOT17 and DanceTrack, but higher than KITTI. This indicates that the motions in LMOT are", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fast but within normal range. Following [42], we use cosine distance of appearance features ${}^{1}$ to evaluate the appearance similarity. From Fig. 3 (c), we can see that the cosine distance of appearance features under low-light conditions is smaller than that under well-lit conditions. In other words, the appearance of objects will deteriorate under low-light conditions, making them harder to distinguish.", + "bbox": [ + 496, + 282, + 890, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Low-light Multi-object Tracking", + "text_level": 1, + "bbox": [ + 498, + 405, + 795, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we peresnt our low-light multi-object tracking method (LTrack). Our main idea is to learn the invariant semantic information under noise disturbance quality degradation. The overall framework can be seen in Fig. 4.", + "bbox": [ + 496, + 431, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Formulation and Motivation", + "text_level": 1, + "bbox": [ + 498, + 503, + 754, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In low-light scenes, the camera can capture only a small number of photons in a single exposure. Thus, the potential sensor noise is highlighted, resulting in significant degradation to the image quality [50]. We observed that directly feeding low-light images to the network without any special design leads to the feature map degradation and significantly reduces the performance of the model (see in Sec. 5.2). A direct solution is to apply low-light enhancement techniques, which focus on learning a mapping function from low-light images to clean well-lit images. Since it is a highly ill-posed problem, learning such a mapping function requires considerable computing and storage overhead. Although DNN-based methods have achieved excellent performance in low-light enhancement, images enhanced for visual quality may be suboptimal for downstream tasks.", + "bbox": [ + 496, + 527, + 890, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we perform multi-object tracking from low-light images, bypassing the low-light enhancement. Leveraging RAW videos, the network obtains more original scene information compared to sRGB. To enhance the performance and robustness of the multi-object tracking model, our main idea is to earn the invariant semantic information under noise disturbance quality degradation. Thus, we present the adaptive low-pass downsampling (ALD) module", + "bbox": [ + 496, + 753, + 890, + 875 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1We use UniTrack [12] to extract appearance features.", + "bbox": [ + 517, + 886, + 803, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "385", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a123ddd19e490d6e5311f513268160ef5309ce2988220fbd0c730ce7949900fd.jpg", + "image_caption": [ + "Figure 4. The overall framework of the proposed low-light multi-object tracking method, termed as LTrack. It employs adaptive low-pass downsample module and degradation suppression learning strategy, enabling the model to learn invariant features from low-light videos." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 893, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to enhance the low-frequency components of images and filter out high-frequency noise. We also propose the degradation suppression learning strategy (DSL), which utilizes paired low-light videos to help the model suppress image noise disturbance and encourage image content response in the feature domain.", + "bbox": [ + 75, + 342, + 470, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Adaptive Low-pass Downsampling", + "text_level": 1, + "bbox": [ + 76, + 441, + 379, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The downsampling operation reduces the feature size while preserving the most important information. The noise in low-light images introduces high-frequency disturbance to the feature maps, which can mislead the preservation of object information. To weaken the impact of high-frequency noises and enhance the low-frequency part of the feature map, we introduce spatial low-pass convolution (SConv) to extract low-frequency features from the noised feature maps. The softmax function is used to constrain the original convolution kernel to be low-pass as", + "bbox": [ + 75, + 464, + 468, + 614 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {K}} _ {i, j} = \\frac {\\exp \\left(\\mathcal {K} _ {i , j}\\right)}{\\sum_ {p , q} \\exp \\left(\\mathcal {K} _ {p , q}\\right)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 618, + 468, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{K}$ and $\\hat{\\mathcal{K}}$ are the origin convolution kernel and low-pass convolution kernel. We initialize the spatial low-pass convolution kernel using a standard Gaussian kernel. The kernel size is set to 5 to obtain more spatial information. Then, the obtained low-frequency features are adaptively weighted and fused into the original features. We use global average pooling to obtain channel descriptors for both origin features and low-frequency features. These descriptors are then fed into a fully connected layer to compute the weight values.", + "bbox": [ + 75, + 656, + 468, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Degradation Suppression Learning", + "text_level": 1, + "bbox": [ + 76, + 816, + 382, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given that a low-light image and well-lit image pair share the same content, the model should exhibit identical feature responses to them. However, the low-light image results in shallow features full of noise, and the deep feature exhibits", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lower responses to objects (as shown in Fig. 5). To address this, our idea is to suppress image noise in shallow features and use well-lit images to help model learning disturbance invariant information from low-light images. The degradation suppression loss can be expressed as", + "bbox": [ + 496, + 342, + 890, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D S} = \\sum_ {l} \\left\\| \\mathbf {F} _ {l} ^ {\\text {w e l l}} - \\mathbf {F} _ {l} ^ {\\text {l o w}} \\right\\| _ {2} ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 419, + 890, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_l^{well}$ and $\\mathbf{F}_l^{low}$ denotes $l$ -th feature map corresponding to well-lit and low-light image, respectively. To further suppress the noise in shallow features, we also introduce the Total Variation (TV) loss [38] to features as", + "bbox": [ + 498, + 454, + 890, + 515 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T V} = \\sum_ {l} | | \\mathbf {G} ^ {r o w} \\mathbf {F} _ {l} ^ {l o w} | | _ {2} ^ {2} + | | \\mathbf {G} ^ {c o l} \\mathbf {F} _ {l} ^ {l o w} | | _ {2} ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 518, + 890, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{G}^{row}$ and $\\mathbf{G}^{col}$ are the first derivative matrix to role and column. The TV loss adds spatial smoothing constraints to features, which helps model learning to extract noise invariant features. Both the well-lit and low-light images are used to train the model by a common detection loss $L_{det}(\\cdot)$ . The total loss is $\\mathcal{L}_{Total}$ , i.e.,", + "bbox": [ + 498, + 551, + 890, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {T o t a l}} = \\mathcal {L} _ {\\text {d e t}} ^ {\\text {w e l l}} + \\alpha \\mathcal {L} _ {\\text {d e t}} ^ {\\text {l o w}} + \\beta \\mathcal {L} _ {D S} + \\gamma \\mathcal {L} _ {T V} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 646, + 890, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where is the $\\mathcal{L}_{det}^{well}$ and $\\mathcal{L}_{det}^{low}$ are the detection loss for welllit and low-light images, $\\alpha$ and $\\beta$ are loss weights. We set $\\alpha, \\beta$ and $\\gamma$ to 1, 1 and 0.01, respectively.", + "bbox": [ + 498, + 667, + 890, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 729, + 633, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 753, + 674, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Split. In structuring our dataset, we randomly split the videos into training, validation, and testing sets, consisting of 11, 4, and 11 videos respectively. We also provide LMOT-real dataset that is captured in real-world night scenes with 6 videos. Detailed dataset split and statistical information are shown in Tab. 2.", + "bbox": [ + 496, + 777, + 890, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following [8, 42], we recommend using HOTA [31] as the main evaluation metric to simulta", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "386", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c80a3cc4d391a972885812c8a56519f4f260a47438bad78853ea7bc5a552d15f.jpg", + "image_caption": [ + "Figure 5. Visualization of shallow and deep features for well-lit and low-light images. It can be seen that, under low-light conditions, the shallow feature is full of noise, and the deep feature exhibits lower responses to objects." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 215, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eeff969e8471e78601b0b94d4a2a6d161420522a88f3632d691baefaa324ce82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 88, + 341, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9b1169c4dcaf850e15687b2fc7dc336e016615e47bc6e0bb2c8a26590438848d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 88, + 467, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "neously evaluate the performance of detection and association. We also employ AssA and IDF1 [37] to evaluate association performance, MOTA, and Deta to evaluate detection performance. There are two ways to combine metrics for all classes into a single score. One is by averaging metrics over the class values, and the other is by over the detection values. To avoid possible result bias caused by some categories with fewer samples (such as Bus and Truck), we combine scores by averaging over the detection values.", + "bbox": [ + 75, + 289, + 467, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Following ByteTrack [62] and OC-SORT [3], we use YOLOX [17] as our detector. The trackers are pre-trained on the COCO [28] dataset and then trained on our LMOT dataset for 24 epochs. We apply data augmentation strategies including random flip, scale jitter of resizing, and Mosaic. We also use the physical-based noise model [50] for RAW image augment. We use SGD optimizer with weight decay $10^{-4}$ and cosine learning rate schedule, the initial learning rate is $10^{-4}$ and gradually reduces to $10^{-5}$ . We apply linear interpolation as post-processing to all trackers, with the maximum gap set to 20.", + "bbox": [ + 75, + 431, + 467, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Analysis under low-light conditions", + "text_level": 1, + "bbox": [ + 76, + 612, + 382, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We analyze the impact of low-light conditions on multi-object tracking using the LMOT validation set. It should be emphasized that the low-light and well-lit video pairs are perfectly aligned in LMOT.", + "bbox": [ + 75, + 636, + 467, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact to detectors. We first analyze the impact of lighting conditions on the detector. We select YOLOX as the detector since it is widely used in MOT areas [3, 8, 62]. We train the detector using well-lit images (WL), low-light images (LL), and all the images (AL). Then, test them on both well-lit and low-light images. The results are shown in Tab. 3. From the table, we can see that the model trained by welllight images achieves the best result on well-light images, but its performance significantly decreased on low-light images. Further, we visualized the feature maps under these two lighting conditions in Fig. 5. It can be seen that both the shallow feature and deep features of the low-light image are significantly degraded due to the sensor noises. We also", + "bbox": [ + 75, + 704, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cb21001530618e5eda2f6f6ef0632ef1dbb260a7d4fc9996fa2cc1772ed22bba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Training dataWell-litLow-light
WLLLmAPmARAP50mAPmARAP50
✓37.045.065.53.14.86.6
✓23.030.840.816.923.730.3
✓✓28.735.249.117.924.132.2
", + "bbox": [ + 504, + 88, + 888, + 172 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fb02d3c1adbdfabd0627c62d60832df407ccb22a5a6df07b9075074f222b8019.jpg", + "table_caption": [ + "Table 3. Analysis on LMOT validation set for detector (YOLOX [17]). WL and LL indicate the well-lit and low-light, respectively. It shows that it is hard to detect objects under low-light conditions." + ], + "table_footnote": [], + "table_body": "
Cond.Mot.App.HOTAAssAIDF1MOTADetA
-✓86.179.572.982.693.4
WL✓87.377.182.598.898.4
WL✓✓83.778.085.696.890.0
LL✓53.832.140.067.390.2
LL✓✓83.477.585.096.089.8
", + "bbox": [ + 504, + 224, + 888, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Analysis on LMOT validation set for different association models. Cond, Mot, and App indicate the light condition, motion, and appearance, respectively. The detection boxes are ground-truth boxes. It indicates that appearance information is effective for association under well-lit conditions, but dose not as effective under low-light conditions.", + "bbox": [ + 498, + 315, + 890, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "observed that using low-light images can largely improve the performance of the decoder under low-light conditions, and use all images to achieve the best performance. But this result is much lower than that under well-lit conditions.", + "bbox": [ + 496, + 412, + 890, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact to association modules. We analyze the impact of low-light conditions on the object association modules. Motion and appearance are important for object association. Both of them rely on detection boxes to locate the objects. To separate the impact of the detector, we use the ground-truth boxes as the detection boxes. The results are shown in Tab. 4. It can be seen that using only appearance matching under well-lit conditions achieves the best result while using only appearance matching under low-lit conditions results in very poor performance. This indicates that the objects in LMOT have obvious visual distinguishability, but the distinguishability is significantly reduced under low-light conditions. In Fig. 6, we visualize the appearance feature of objects in the LMOT dataset under both well-lit and low-light conditions. We can observe that under well-lit conditions, LMOT is very distinguishable in the feature space. However, under low-light conditions, this discrimination of LMOT decreased significantly.", + "bbox": [ + 496, + 474, + 890, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Results on LMOT dataset", + "text_level": 1, + "bbox": [ + 500, + 756, + 733, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the proposed method with potential low-light multi-object tracking methods. We first train three baseline models using low-light videos, well-lit videos, and both of them, without any modifications to baseline trackers. They are denoted by Base-low, Base-well, and Base-all. The second type of method is tracking after low-light enhancement. These methods use low-light enhancement methods as preprocessing module. Both the enhanced videos and well-lit", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "387", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c620a1e765ee6985ef90d7e4af78f313efd0f995359ef5962b009b4672ec793c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodByteTrack[62]OC-SORT[3]DetectionParams (M)FLOPS (G)
HOTAAssAIDF1HOTAAssAIDF1mAPAP50AP75
Base-low28.042.910.928.840.334.318.832.418.999.00793.29
Base-well13.536.832.913.831.712.05.59.75.599.00793.29
Base-all28.142.632.928.939.834.619.032.519.199.00793.29
LLFlow [48]28.842.934.029.441.035.022.438.322.6116.428755.91
SDSD [47]29.141.135.529.336.936.819.836.219.3103.301136.39
DNF [24]27.640.933.028.136.435.319.135.618.2101.83907.33
SMOID [23]28.039.733.928.736.735.619.135.518.2120.887286.17
RAOD [52]26.142.929.526.142.929.518.331.318.499.07897.04
LTrack (Ours)29.443.235.229.839.036.723.440.623.5100.43800.92
", + "bbox": [ + 80, + 88, + 893, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4dee198d58fca91a3849bc51373e3dd58ceff2ea2c8ab5ee04299a1be024fec7.jpg", + "image_caption": [ + "(a) Well-lit" + ], + "image_footnote": [], + "bbox": [ + 84, + 344, + 271, + 452 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/440d710812ffcd28fa8131bfd71b0b2b86f5a0395b63072b7108f7acd384b1b7.jpg", + "image_caption": [ + "(b) Low-light", + "Figure 6. Visualization of appearance features of LMOT dataset using t-SNE. The same object is coded by the same color. It indicates that the appearance of objects under well-lit is distinguishable, but significantly reduced under low-light conditions" + ], + "image_footnote": [], + "bbox": [ + 274, + 345, + 464, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "videos are used for training. We select four different low-light enhancement techniques for comparison. LLFlow [48] and SDSD [47] are low-light enhancement methods based on RGB images, while DNF [24] and SMOID [23] are based on RAW low-light enhancement methods. LLFlow and DNF take images as input, SDSD and SMOID take videos as input. The RAW object detection method is also trained together on well-lit and low-light videos. Its output bounding boxes are directly fed to the tracker. We test all the potential methods on two state-of-the-art trackers ByteTrack [62] and OC-SORT [3]. We also show mAP, AP50, and AP75 to highlight the detector performance.", + "bbox": [ + 75, + 549, + 468, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Tab. 5 we can see that, the proposed LTrack achieves the best HOTA and is highly competitive on all metrics. For example, the proposed method improves HOTA 1.3 with almost the same parameters and computation as base methods. This strongly proves the effectiveness of the proposed method. Compared with tracking after low-light enhancement, these methods have a large amount of additional parameters and computation but still perform worse than the proposed method. For example, LLFlow delivers 8 times FLOPS but still performs worse than our LTrack. In addition, We observed that there is", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f154aecf2c961420824509d77900cf16efd1f0329eb992449edf0796cf163908.jpg", + "table_caption": [ + "Table 5. Experimental results on LMOT dataset. Base-low, Base-well and Base-all indicate the baseline model trained on low-light, well-lit and all of them, respectively. The number of parameters and prediction latency of each method are reported along with the accuracy" + ], + "table_footnote": [], + "table_body": "
MethodHOTAAssAIDF1MOTADet
Base-low32.533.439.342.427.1
Base-well28.030.434.136.923.3
Base-all34.638.140.044.830.2
LLFlow[48]33.638.038.443.129.7
SDSD[47]31.634.236.738.827.5
DNF[24]33.239.036.944.430.1
SMOID[23]31.234.136.641.226.9
RAOD[52]32.233.539.440.926.5
LTrack (Ours)35.138.940.445.230.7
", + "bbox": [ + 500, + 342, + 913, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Experimental results on LMOT-real dataset with OC-SORT[3]. The best result are shown in boldface", + "bbox": [ + 498, + 536, + 890, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "no significant difference in results between RAW-based image enhancement methods and RGB-based image enhancement methods. This is not consistent with what our method has observed. The reason may be that low light enhancement focuses on image quality restoration and may mislead downstream tasks. As for RAOD [52], it has almost the same number of parameters and computational load as our method, but its performance is much lower than the proposed LTrack. Despite addressing HDR scenes through a preprocessing module for RAW input, it does not perform well under low-light conditions.", + "bbox": [ + 496, + 578, + 893, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Results on Real World", + "text_level": 1, + "bbox": [ + 500, + 756, + 709, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the performance of our method in real-world low-light scenes at night, we evaluate all methods on LMOT-real dataset, using OC-SORT [3]. As shown in Tab. 6, the proposed LTrack performs much better than all comparison methods. Both the methods of tracking after low-light enhancement and the RAW detection method encounter generalizability problems and are not even better than Base-all. This strongly demonstrates the robustness of", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "388", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b74a884963070c68961e8df7f353dd8b70221aebbf367c05876b1cc7c6320461.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DSLALDHOTAAssAIDF1MOTADetA
✓28.142.632.923.818.7
29.042.435.025.820.1
✓28.342.433.124.619.3
✓✓29.443.235.226.120.3
", + "bbox": [ + 81, + 88, + 464, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ddf763f8a6fe037b02cb0c968e65e010dc684f265ce23984a6bb3e0c77eb229e.jpg", + "table_caption": [ + "Table 7. Ablation on adaptive low-pass downsampling (ALD) and degradation suppression learning (DSL)." + ], + "table_footnote": [], + "table_body": "
Data TypeHOTAAssAIDF1MOTADetA
sRGB29.242.734.926.420.5
RAW 8-bit29.042.334.526.120.2
RAW 10-bit29.342.934.826.520.4
RAW 12-bit29.443.235.226.120.3
", + "bbox": [ + 81, + 215, + 464, + 303 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/58ec47597c7db8169e4366b0602d64d7709c2af876bd5cabb2297406d1de92aa.jpg", + "table_caption": [ + "Table 8. Results of different image format and dynamic range on LMOT test set." + ], + "table_footnote": [], + "table_body": "
MethodsHOTAAssAIDF1MOTADetA
Gaussian-Possion29.943.135.726.621.1
Physical-based [50]30.444.136.026.621.2
LMOT (Ours)35.138.940.445.230.7
", + "bbox": [ + 81, + 340, + 464, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our LTrack in real-world dark scenes.", + "bbox": [ + 76, + 455, + 323, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Exploration and Discussion", + "text_level": 1, + "bbox": [ + 76, + 479, + 323, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct extensive analysis and discussion on LMOT datasets and the proposed method.", + "bbox": [ + 75, + 503, + 467, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study. We conduct ablation experiments to validate the effectiveness of our improvements, the results are shown in Tab. 7. It can be seen that all improvements contribute effectively to enhanced performance. Among them, the degradation suppression learning strategy demonstrates the most significant effect, resulting in an approximate 1-point enhancement in HOTA. The best results are achieved when employing all strategies simultaneously. This demonstrates the effectiveness of all our contributions.", + "bbox": [ + 75, + 540, + 467, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "RAW vs. sRGB. We analyze the impact of input data formats, the results are shown in Tab. 8. The 12-bit RAW format achieves significantly better results than sRGB. Because the RAW format saves much more potential information and is helpful for MOT in low-light scenes. We also observed that higher bitwidth is beneficial for performance, which has also been observed in other vision tasks [21, 52].", + "bbox": [ + 75, + 681, + 467, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LMOT vs. Synthetic data. We compare our LMOT dataset with the synthetic low-light data to further demonstrate the value of LMOT. We apply the Gaussian-Possian based and Physical-based [50] low-light data synthesis method to synthesize low-light videos from well-lit videos. We train our LTrack on these types of data and evaluate their performance in real low-light scenes using LMOT-real. As", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/395ebfc15ee0607d5716107e9a2d6011de9c708a55e154985a4cdcaaeff6186c.jpg", + "table_caption": [ + "Table 9. Results with low-light synthesis methods and LMOT dataset on LMOT-real set." + ], + "table_footnote": [], + "table_body": "
CategoriesHOTAAssAIDF1MOTADetA
Person24.230.829.925.119.2
Bicycle16.940.315.68.97.1
Car37.553.347.233.726.5
Motorcycle19.835.022.615.211.4
Bus44.359.951.836.732.9
Truck6.915.66.43.23.1
", + "bbox": [ + 501, + 89, + 893, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10. Results of the proposed method for different categories on LMOT test set.", + "bbox": [ + 500, + 210, + 890, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shown in Tab. 9, the tracker trained on our LMOT dataset has much better performance in real night scenes, which strongly demonstrates the value of our LMOT dataset.", + "bbox": [ + 498, + 256, + 890, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on different category. We also analyze the performance for different categories. From Tab. 10, we can see that cars and buses achieved the best performance because they have regular shapes and larger areas. Trucks exhibited the poorest performance, because they have the fewest instances, making it challenging for the model to learn accurate identification. The person achieves relatively average scores. Bicycles and motorcycles have close scores since they have similar appearance and motion patterns.", + "bbox": [ + 498, + 309, + 890, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 464, + 617, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we investigate the multi-object tracking in the dark scenes. We build a new low-light multi-object tracking (LMOT) dataset, which provides well-aligned low-light video pairs and high-quality multi-object tracking annotations. We observed that low-light images are significantly degraded by the sensor noises, which also degrades the feature maps and significantly deteriorates the model performance. To learn the invariant semantic formation under noise disturbance and quality degradation, we present the adaptive low-pass downsample module and degradation suppression learning. These improvements greatly enhance the robustness of our method in real-world low-light scenes.", + "bbox": [ + 498, + 491, + 890, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. We focus on multi-object tracking in the dark scenes. But we do not consider other degradation environments in the real world, such as rainy and foggy days. In our future work, we will consider exploring multi-object tracking in more real-world scenarios, promoting the development of MOT in real-world applications.", + "bbox": [ + 498, + 672, + 890, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 782, + 658, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Key R&D Program of China (2022YFC3300704), the National Natural Science Foundation of China (62331006,62171038, and 62088101), the R&D Program of Beijing Municipal Education Commission (KZ202211417048), and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 498, + 809, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "389", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, pages 3464-3468, 2016. 3", + "[2] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. arXiv preprint arXiv:2303.06705, 2023. 1, 2", + "[3] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. Observation-centric sort: Rethinking sort for robust multi-object tracking. In CVPR, pages 9686-9696, 2023. 1, 3, 6, 7", + "[4] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In CVPR, pages 3291-3300, 2018. 2", + "[5] Chen Chen, Qifeng Chen, Minh N Do, and Vladlen Koltun. Seeing motion in the dark. In ICCV, pages 3185-3194, 2019.", + "[6] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. IJCV, pages 1-21, 2023. 1, 2", + "[7] Wongun Choi and Silvio Savarese. A unified framework for multi-target tracking and collective activity recognition. In ECCV, pages 215-230, 2012. 1", + "[8] Yutao Cui, Chenkai Zeng, Xiaoyu Zhao, Yichun Yang, Gangshan Wu, and Limin Wang. Sportsmot: A large multi-object tracking dataset in multiple sports scenes. arXiv preprint arXiv:2304.05170, 2023. 1, 2, 3, 5, 6", + "[9] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1, 2, 3", + "[10] Yunhao Du, Zhicheng Zhao, Yang Song, Yanyun Zhao, Fei Su, Tao Gong, and Hongying Meng. *Strongsort: Make deep-sort great again*. IEEE Transactions on Multimedia, 2023. 1", + "[11] Anna Ellis and James Ferryman. Pets2010 and pets2009 evaluation of results using individual ground truthed single views. In 2010 7th IEEE international conference on advanced video and signal based surveillance, pages 135-142, 2010. 1", + "[12] Wang Zhongdao et al. Do different tracking tasks require different appearance models? 2021. 4", + "[13] Huiyuan Fu, Wenkai Zheng, Xicong Wang, Jiaxuan Wang, Heng Zhang, and Huadong Ma. Dancing in the dark: A benchmark towards general low-light video enhancement. In ICCV, pages 12877-12886, 2023. 2", + "[14] Ying Fu, Jian Chen, Tao Zhang, and Yonggang Lin. Residual scale attention network for arbitrary scale image superresolution. Neurocomputing, 427:201-211, 2021. 2", + "[15] Ying Fu, Yang Hong, Linwei Chen, and Shaodi You. Le-gan: Unsupervised low-light image enhancement network using attention module and identity invariant loss. Knowledge-Based Systems, 240:108010, 2022. 2", + "[16] Ruopeng Gao and Limin Wang. Memotr: Long-term memory-augmented transformer for multi-object tracking. In ICCV, pages 9901-9910, 2023. 3" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6", + "[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361, 2012. 1, 2, 3, 4", + "[19] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In CVPR, pages 1780-1789, 2020. 2", + "[20] Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE TIP, 26(2):982-993, 2016. 2", + "[21] Yang Hong, Kaixuan Wei, Linwei Chen, and Ying Fu. Crafting object detection in very low light. In BMVC, page 3, 2021. 1, 2, 8", + "[22] Haidi Ibrahim and Nicholas Sia Pik Kong. Brightness preserving dynamic histogram equalization for image contrast enhancement. IEEE Transactions on Consumer Electronics, 53(4):1752-1758, 2007. 2", + "[23] Haiyang Jiang and Yinqiang Zheng. Learning to see moving objects in the dark. In ICCV, pages 7324-7333, 2019. 1, 2, 3, 7", + "[24] Xin Jin, Ling-Hao Han, Zhen Li, Chun-Le Guo, Zhi Chai, and Chongyi Li. Dnf: Decouple and feedback network for seeing in the dark. In CVPR, pages 18135-18144, 2023. 1, 2, 7", + "[25] Daniel J Jobson, Zia-ur Rahman, and Glenn A Woodell. Properties and performance of a center/surround retina. IEEE TIP, 6(3):451-462, 1997. 2, 4", + "[26] Laura Leal-Taixe, Anton Milan, Ian Reid, Stefan Roth, and Konrad Schindler. Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv preprint arXiv:1504.01942, 2015. 2", + "[27] Sohyun Lee, Jaesung Rim, Boseung Jeong, Geonu Kim, Byungju Woo, Haechan Lee, Sunghyun Cho, and Suha Kwak. Human pose estimation in extremely low-light conditions. In CVPR, pages 704-714, 2023. 1, 2", + "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6", + "[29] Qiankun Liu, Qi Chu, Bin Liu, and Nenghai Yu. Gsm: Graph similarity model for multi-object tracking. In IJCAI, pages 530-536, 2020. 1", + "[30] Qiankun Liu, Dongdong Chen, Qi Chu, Lu Yuan, Bin Liu, Lei Zhang, and Nenghai Yu. Online multi-object tracking with unsupervised re-identification learning and occlusion estimation. Neurocomputing, 483:333–347, 2022. 1", + "[31] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. IJCV, 129:548-578, 2021. 5", + "[32] Kang Ma, Ying Fu, Dezhi Zheng, Chunshui Cao, Xuecai Hu, and Yongzhen Huang. Dynamic aggregated network for gait recognition. In CVPR, pages 22076-22085, 2023. 1" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "390", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, pages 8844-8854, 2022. 3", + "[34] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831, 2016. 1, 2, 3, 4", + "[35] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In CVPR, pages 1765–1773, 2017. 2", + "[36] Jihun Park, Jinseok Hong, Woolil Shim, and Dae-Jin Jung. Multi-object tracking on swir images for city surveillance in an edge-computing environment. Sensors, 23(14):6373, 2023. 3", + "[37] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In ECCV, pages 17-35, 2016. 6", + "[38] Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena, 60(1-4):259-268, 1992. 5", + "[39] Aashish Sharma and Robby T Tan. Nighttime visibility enhancement by increasing the dynamic range and suppression of light effects. In CVPR, pages 11977-11986, 2021. 2", + "[40] Shuran Song and Jianxiong Xiao. Tracking revisited using rgbd camera: Unified benchmark and baselines. In ICCV, pages 233-240, 2013. 3", + "[41] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 3", + "[42] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In CVPR, pages 20993-21002, 2022. 1, 2, 3, 4, 5", + "[43] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 2", + "[44] Ye Tian, Ying Fu, and Jun Zhang. Plug-and-play algorithm for under-sampling fourier single-pixel imaging. Science China. Information Sciences, 65(10):209303, 2022. 2", + "[45] Ye Tian, Ying Fu, and Jun Zhang. Local-enhanced transformer for single-pixel imaging. Optics Letters, 48(10): 2635-2638, 2023. 2", + "[46] Chaoqun Wang, Chunyan Xu, Zhen Cui, Ling Zhou, Tong Zhang, Xiaoya Zhang, and Jian Yang. Cross-modal patternpropagation for rgb-t tracking. In CVPR, pages 7064-7073, 2020. 3", + "[47] Ruixing Wang, Xiaogang Xu, Chi-Wing Fu, Jiangbo Lu, Bei Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In ICCV, pages 9700-9709, 2021. 1, 2, 7", + "[48] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In AAAI, pages 2604-2612, 2022. 7" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 2", + "[50] Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. A physics-based noise formation model for extreme low-light raw denoising. In CVPR, pages 2758-2767, 2020. 4, 6, 8", + "[51] Kaixuan Wei, Angelica Aviles-Rivero, Jingwei Liang, Ying Fu, Hua Huang, and Carola-Bibiane Schonlieb. Tfnpn: Tuning-free plug-and-play proximal algorithms with applications to inverse imaging problems. Journal of Machine Learning Research, 23(16):1-48, 2022. 2", + "[52] Ruikang Xu, Chang Chen, Jingyang Peng, Cheng Li, Yibin Huang, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Toward raw object detection: A new benchmark and a new model. In CVPR, pages 13384-13393, 2023. 2, 7, 8", + "[53] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. In CVPR, pages 17714-17724, 2022. 2", + "[54] Jinyu Yang, Shang Gao, Zhe Li, Feng Zheng, and Ales Leonardis. Resource-efficient rgbd aerial tracking. In CVPR, pages 13374-13383, 2023. 3", + "[55] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2636-2645, 2020. 2, 3, 4", + "[56] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In ECCV, pages 659–675, 2022. 3", + "[57] Fan Zhang, Yu Li, Shaodi You, and Ying Fu. Learning temporal consistency for low light video enhancement from single images. In CVPR, pages 4967-4976, 2021. 2", + "[58] Jiqing Zhang, Xin Yang, Yingkai Fu, Xiaopeng Wei, Baocai Yin, and Bo Dong. Object tracking by jointly exploiting frame and event domain. In ICCV, pages 13043-13052, 2021. 3", + "[59] Pengyu Zhang, Jie Zhao, Dong Wang, Huchuan Lu, and Xiang Ruan. Visible-thermal uav tracking: A large-scale benchmark and new baseline. In CVPR, pages 8886–8895, 2022. 3", + "[60] Tao Zhang, Ying Fu, and Cheng Li. Hyperspectral image denoising with realistic data. In ICCV, pages 2248-2257, 2021. 2", + "[61] Tao Zhang, Ying Fu, and Cheng Li. Deep spatial adaptive network for real image demosaicing. In AAAI, pages 3326-3334, 2022. 2", + "[62] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, pages 1-21, 2022. 1, 3, 6, 7", + "[63] Yang Zhang, Tiancai Wang, and Xiangyu Zhang. Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. In CVPR, pages 22056-22065, 2023. 3", + "[64] Zhiyu Zhu, Junhui Hou, and Dapeng Oliver Wu. Cross-modal orthogonal high-rank augmentation for rgb-event" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "391", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transformer-trackers. In ICCV, pages 22045-22055, 2023. 3", + "[65] Yunhao Zou and Ying Fu. Estimating fine-grained noise model via contrastive learning. In CVPR, pages 12682-12691, 2022. 2", + "[66] Yunhao Zou, Yinqiang Zheng, Tsuyoshi Takatani, and Ying Fu. Learning to reconstruct high speed and high dynamic range videos from events. In CVPR, pages 2024-2033, 2021. 2", + "[67] Yunhao Zou, Chenggang Yan, and Ying Fu. Rawhdr: High dynamic range image reconstruction from a single raw image. In ICCV, pages 12334-12344, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 261 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "392", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Multi-Object Tracking in the Dark/4c20e9a2-1106-45cf-a36a-0161982e6015_model.json b/2024/Multi-Object Tracking in the Dark/4c20e9a2-1106-45cf-a36a-0161982e6015_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3d7c73c324e88f1585c55694371caa568d12a7b8 --- /dev/null +++ b/2024/Multi-Object Tracking in the Dark/4c20e9a2-1106-45cf-a36a-0161982e6015_model.json @@ -0,0 +1,2433 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.308, + 0.131, + 0.662, + 0.153 + ], + "angle": 0, + "content": "Multi-Object Tracking in the Dark" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.181, + 0.766, + 0.218 + ], + "angle": 0, + "content": "Xinzhe Wang Kang Ma Qiankun Liu Yunhao Zou Ying Fu * Beijing Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.22, + 0.742, + 0.235 + ], + "angle": 0, + "content": "{wangxinzhe, makang, liuqk3, zouyunhao, fuying}@bit.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.651 + ], + "angle": 0, + "content": "Low-light scenes are prevalent in real-world applications (e.g. autonomous driving and surveillance at night). Recently, multi-object tracking in various practical use cases have received much attention, but multi-object tracking in dark scenes is rarely considered. In this paper, we focus on multi-object tracking in dark scenes. To address the lack of datasets, we first build a Low-light Multi-Object Tracking (LMOT) dataset. LMOT provides well-aligned low-light video pairs captured by our dual-camera system, and high-quality multi-object tracking annotations for all videos. Then, we propose a low-light multi-object tracking method, termed as LTrack. We introduce the adaptive low-pass downsample module to enhance low-frequency components of images outside the sensor noises. The degradation suppression learning strategy enables the model to learn invariant information under noise disturbance and image quality degradation. These components improve the robustness of multi-object tracking in dark scenes. We conducted a comprehensive analysis of our LMOT dataset and proposed LTrack. Experimental results demonstrate the superiority of the proposed method and its competitiveness in real night low-light scenes. Dataset and Code: https://github.com/ying-fu/LMOT" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.681, + 0.21, + 0.697 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.471, + 0.858 + ], + "angle": 0, + "content": "Multi-object tracking (MOT) aims to locate and associate multiple objects in video sequences. It is widely used in many downstream applications, such as video recognition [7, 32], autonomous driving [18], and surveillance [11]. Recently, multi-object tracking in various practical use cases has garnered much attention [8, 9, 18, 34, 42], greatly advancing the development of MOT. However, these works are primarily tailored for high-quality inputs and overlook the prevalent low-light scenes in real-world scenes. Motivated by this, we study multi-object tracking in dark scenes." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.86, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Due to the physical limitations of existing cameras, ac" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.68 + ], + "angle": 0, + "content": "quiring high-quality videos under low-light conditions is difficult. One inherent difficulty in capturing consecutive video frames under such conditions is avoiding motion blur. Current camera technology typically requires short exposure times (usually just a few tens of milliseconds), but in low-light scenarios, the sensor struggles to capture an adequate number of photons within this limited duration. This limitation inevitably leads to degradation in image quality accompanied by higher noise levels. This presents two main challenges for MOT under low-light conditions. The first challenge is for collecting a low-light multi-object tracking dataset. Collecting and annotating a low-light MOT dataset is difficult and expensive. MOT requires dynamic object videos, but videos captured in low-light scenes have extremely low brightness, making it hard to recognize and annotate objects in the videos. The second challenge revolves around low-light multi-object tracking. The popular tracking-by-detection paradigm [3, 10, 29, 30, 62] generally consists of detector, motion-based association module, and appearance-based association module. These modules typically require high-quality input images. The poor quality of low-light images leads to severe performance degradation for both detectors and appearance-based correlation modules. A simple approach is to cascade low-light enhance modules [2, 23, 24, 47], but this introduces additional computational costs. Furthermore, images optimized for visual quality may be suboptimal for downstream tasks [6, 21, 27]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we build a low-light multi-object tracking dataset (LMOT), specifically designed to address the challenges of multi-object tracking in dark scenes. To this end, we develop a dual-camera system that simultaneously captures well-lit and low-light video frames. The video pairs are highly aligned in both spatial and temporal dimensions, offering two key benefits. First, it enables us to annotate on the well-lit videos, resulting in high-quality annotations. Second, the well-lit videos can provide additional supervision information during the training phase, and strongly enhance the performance in the dark scenes. After careful annotation, we collect 32 video sequences \\((2.3\\times \\mathrm{MOT17})\\) over \\(35\\mathrm{K}\\) frames \\((3.1\\times \\mathrm{MOT17})\\) ,and over \\(815\\mathrm{K}\\) bounding boxes \\((2.8\\times \\mathrm{MOT17})\\) .The RAW data is the output of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.225, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "382" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "the image sensor and is the input data of the image signal processor (ISP). It saves all information from the image sensor, which is crucial for capturing object information in dark scenes [6, 52, 67]. Therefore, we collect RAW videos for LMOT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.473, + 0.455 + ], + "angle": 0, + "content": "Additionally, we propose a low-light multi-object tracking method, termed as LTrack. The low-light video is characterized by substantial sensor noise and poor image quality, which significantly degrades both shallow and deep feature representations, leading to reduced tracking performance. We observe that the sensor noise in low-light images exhibits a similarity to adversarial attacks [35, 43]. To address this issue, our main idea is to learn the invariant semantic information under noise disturbance quality degradation. We present the adaptive low-pass downsample module (ALD). It employs spatial low-pass convolution to extract low-frequency components from images, excluding noises, and adaptively enhance the feature maps. We also present the degradation suppression learning strategy (DSL), which utilizes paired low-light videos to help the model suppress the noise disturbance and encourage image content response in the feature domain. We conduct a comprehensive analysis of our LMOT dataset and validate the superiority of our LTrack in real-world night scenes." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.456, + 0.435, + 0.471 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.475, + 0.47, + 0.535 + ], + "angle": 0, + "content": "- We build the first low-light multi-object tracking dataset using a carefully constructed dual-camera system. It provides well-aligned low-light videos in RAW format, and high-quality MOT annotations for all videos." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.536, + 0.47, + 0.595 + ], + "angle": 0, + "content": "- We propose a low-light multi-object tracking method. It utilizes the adaptive low-pass downsample module and the degradation suppression learning strategy to learn to extract invariant features from low-light videos." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.596, + 0.47, + 0.656 + ], + "angle": 0, + "content": "- We conduct a comprehensive analysis of our dataset and the proposed method. Experimental results demonstrate the superiority of the proposed method and its competitiveness on real night scenes." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.475, + 0.47, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.672, + 0.22, + 0.687 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.47, + 0.759 + ], + "angle": 0, + "content": "In this section, we first review the current research status of low-light enhancement and low-light datasets. Then, we summarized the present research for multi-object tracking and tracking in the dark scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Low-light enhancement. Traditional methods for lowlight enhancement are primarily based on histogram equalization and Retinex theory [20, 22, 25]. Recently, deep learning has been explored for many low-level tasks [14, 44, 45, 51, 60, 61, 65, 66], and achieved superior results on low-light enhancement [2, 4, 13, 15, 19, 24, 39, 53]. While these methods are capable of recovering images with high visual quality, they often require heavy computation and may not consider downstream tasks, resulting in suboptimal" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.09, + 0.891, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.282, + 0.893, + 0.338 + ], + "angle": 0, + "content": "Figure 1. Our dual-camera system. It consists of two cameras, a beam splitter, and an ND-filter. Two cameras of identical models are meticulously engineered to achieve pixel-by-pixel alignment in the captured video data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.356, + 0.892, + 0.403 + ], + "angle": 0, + "content": "performance. In contrast, our approach focuses on directly learning multi-object tracking from low-light images, thus bypassing the low-light enhancement." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.893, + 0.697 + ], + "angle": 0, + "content": "Low-light datasets. The long-short exposure is a widely used method to collect paired low-light images, but can only be used to collect low-light images for static scenes. [4, 5, 49]. To capture dynamic scenes, some works designed the mechatronic system. They obtain paired low-light data by repeating the motion twice [13, 47]. However, these mechatronic systems cannot be used to collect dynamic object videos in the wild. Jiang et al. [23] designed a dual-camera system that simultaneously captures paired well-lit and low-light videos, making it possible to capture dynamic scenes and dynamic object videos for multi-object tracking. Zou et al. [57] setup an optical system to collect paired videos and event streams. These works explore various ways to construct low-light datasets and inspire research on high-level vision tasks in dark scenes, such as LOD [21] for object detection, LIS [6] for instance segmentation, and ExPose [27] for human pose estimation. These datasets provide paired low-light data only for image tasks, and cannot be expanded for multi-object tracking in dark scenes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Multi-object tracking datasets. MOT15 [26] is the first large-scale benchmark for multi-object tracking. MOT17 [34] stands as one of the most widely applied MOT benchmarks. MOT20 [9] focuses on very crowded scenes. These three datasets are for pedestrians. KITTI [18] and BDD100K [55] are for autonomous driving scenarios. DanceTrack [42] focuses on dancing scenes and is characterized by similar appearance and diverse motions. Recently, SportsMOT [8] aims to track athletes and encourage algorithms to promote both appearance and motion association. These datasets explore multi-object tracking in various practical use cases, but none of them consider multi-object tracking in dark scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "383" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.49, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.089, + 0.891, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.893, + 0.298 + ], + "angle": 0, + "content": "Figure 2. Two example videos from our LMOT dataset. It provides well-aligned low-light video pairs and MOT annotations for all videos. The time interval between adjacent frames is 1s. The first row is the low-light video, the second row is the scaled low-light video and the last row is the well-lit video. Our LMOT dataset is collected from city outdoor scenes." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.31, + 0.466, + 0.433 + ], + "angle": 0, + "content": "
DatasetFormatVideosFramesLength (s)BboxTracks
MOT17 [34]sRGB1411,235463292,7331,342
MOT20 [9]sRGB813,4105351,652,0403,456
DanceTrack [42]sRGB100105,8555,292-990
SportMOT [8]sRGB240150,3796,0151,629,4903,401
KITTI [18]sRGB218,000-47,000917
BDD100K [55]sRGB1,600318,000-3,300,000131,000
SWIR [36]SWIR-7,309-57,221-
LMOTRAW3235,1201,756815,5504,090
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.434, + 0.469, + 0.461 + ], + "angle": 0, + "content": "Table 1. Comparison of statistics between existing MOT datasets and our LMOT dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.469, + 0.629 + ], + "angle": 0, + "content": "Object tracking in the dark. To track in low-light scenes, some methods explore to use multi-modal information for single-object tracking, such as event camera [58, 64], depth [40, 54] and thermal [46, 59] devices. Park et al. [36] proposed to use Short-Wave Infrared (SWIR) images for multi-object tracking, since its advantages in terms of robustness in low-light conditions. The common drawback of these methods is that they require additional hardware equipment and cannot be applied to the most widely used CMOS imaging systems." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.469, + 0.797 + ], + "angle": 0, + "content": "SORT [1] uses the Kalman Filter motion model and employs IoU for association. ByteTrack [62] enhances tracking performance by considering low-confidence bounding boxes. OS-SORT [3] enhances SORT by restoring lost targets. Recently, Transformer has been explored for MOT [16, 33, 41, 56, 63]. These methods have achieved high performance in many practical scenarios, but they do not consider dealing with low-light conditions. We focus on multi-object tracking under low-light conditions. Based on RAW videos, our method is highly practical with excellent performance and does not require additional hardware." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.814, + 0.442, + 0.831 + ], + "angle": 0, + "content": "3. Low-light Multi-object Tracking Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.9 + ], + "angle": 0, + "content": "In this section, we first introduce our dual camera system and the details of collecting and annotating our low-light multi-object tracking (LMOT) dataset. Then, we analyzed the statistical characteristics of our LMOT dataset." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.31, + 0.889, + 0.39 + ], + "angle": 0, + "content": "
DatasetSplitVideosBboxTracksPaired Well-lit
LMOT-dualtrain11309,4661533✓
val4131,781626✓
test11312,7421644✓
LMOT-realreal661,561287
" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.391, + 0.853, + 0.404 + ], + "angle": 0, + "content": "Table 2. Detailed statistics and data splits for LMOT." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.422, + 0.699, + 0.437 + ], + "angle": 0, + "content": "3.1. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Multi-object tracking requires dynamic scenes and object video. To collect low-light videos for multi-object tracking, we build a dual-camera system [23], which is illustrated in Fig. 1. It can simultaneously capture paired lowlight and well-lit video pairs. Its main components include a beam splitter, a neutral density (ND) filter, and two FLIR Grasshopper3 GS3-U3-23S6C cameras. The beam splitter divides the incoming light into two separate paths. This arrangement allows one camera to capture well-lit images directly, while the other camera records low-light images, with the ND-filter attenuating the light intensity. To ensure temporal synchronization of the video frames, we employ the hardware interface to trigger the camera exposure events. Moreover, to avoid frame loss, our dual-camera system uses two independent hardware interfaces for data transmission and is equipped with a high-speed solid-state drive. Thanks to precise calibration, our dual-camera system can capture paired low-light and well-lit videos in real time for dynamic scenes and objects. More details about our dual-camera system are given in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We save the video frames in RAW format before the images are processed by the camera image signal processor (ISP). In terms of camera settings, we set the exposure time for both cameras to \\(10ms\\), and the frame rate is fixed at 20. This setting is feasible to avoid motion blur. We adjust the gain level for well-lit cameras to achieve optimal image quality. The gain for low-light cameras is consistently set to the maximum value, to simulate low-light capturing setup in real scenarios. We also collect a real low-light MOT dataset (LMOT-real) to evaluate performance in real-world" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "384" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.092, + 0.372, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.2, + 0.331, + 0.213 + ], + "angle": 0, + "content": "(a) Number of instances per category" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.091, + 0.609, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.2, + 0.568, + 0.213 + ], + "angle": 0, + "content": "(b) IoU on adjacent frames" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.092, + 0.885, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.2, + 0.862, + 0.213 + ], + "angle": 0, + "content": "(c) Cosine distance of appearance features" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.215, + 0.894, + 0.271 + ], + "angle": 0, + "content": "Figure 3. (a) Number of instances per category. LMOT consists of 6 categories, most of the instances are the person and car. (b) IoU on adjacent frames. Compared to MOT17, KITTI, and DanceTrack, LMOT has a roughly average score. This indicates that LMOT has a relatively normal movement speed. (c) Cosine distance of appearance features. The cosine distance is smaller under low-light conditions, indicating that the appearance distinguishability is decreased under low-light conditions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.284, + 0.468, + 0.315 + ], + "angle": 0, + "content": "dark scenes. These videos are captured using a single camera with the same camera settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.468, + 0.436 + ], + "angle": 0, + "content": "Our LMOT dataset contains a variety of city outdoor scenes, including roads, overpasses, pedestrians, and intersection. The overpass scenes take an overhead shot of objects, while all other scenes are captured from the perspective of pedestrians. To account for the impact of camera motion, we introduce arbitrary horizontal rotations and vertical random movements to the camera. Fig. 2 shows two sampled video sequences from our LMOT dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.468, + 0.603 + ], + "angle": 0, + "content": "We annotate six types of moving objects, including car, person, bicycle, motorcycle, bus, and truck. The annotated labels include bounding boxes, identifications, and visibility status. For partly occluded objects, a full box is annotated. For a fully occluded object, an estimated box is annotated. Each object has a unique ID throughout the entire video. Thanks to well-aligned low-light and well-lit videos, we can annotate well-lit videos to simultaneously obtain labels for low-light videos. This greatly reduces the annotation difficulty and enhances quality. Lastly, we carefully review all the annotation results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.618, + 0.238, + 0.632 + ], + "angle": 0, + "content": "3.2. Dataset Statistic" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.468, + 0.822 + ], + "angle": 0, + "content": "LMOT is a large-scale dataset that focuses on multi-object tracking in dark scenes. We compare the statistics of LMOT with existing MOT datasets in Tab. 1. It can be seen that LMOT is approximately three times larger than MOT17 [34]. Compared to large-scale datasets such as DanceTrack [42], SportsMOT [25], KITTI [18] and BDD100K [55], the scale of LMOT is still considerable. It should be emphasized that these datasets are not for multi-object tracking in dark scenes and only provide sRGB images. Compared to SWIR, our LMOT dataset has approximately \\(5 \\times\\) frames and \\(14 \\times\\) bounding boxes. The detailed statistics and data splits for LMOT are shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.899 + ], + "angle": 0, + "content": "We present the number of instances for each category in Fig. 3 (a). The majority of instances are persons and cars. As shown in Fig. 3 (b), the average IoU on adjacent frames of LMOT is lower than MOT17 and DanceTrack, but higher than KITTI. This indicates that the motions in LMOT are" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.284, + 0.892, + 0.39 + ], + "angle": 0, + "content": "fast but within normal range. Following [42], we use cosine distance of appearance features \\( {}^{1} \\) to evaluate the appearance similarity. From Fig. 3 (c), we can see that the cosine distance of appearance features under low-light conditions is smaller than that under well-lit conditions. In other words, the appearance of objects will deteriorate under low-light conditions, making them harder to distinguish." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.406, + 0.796, + 0.423 + ], + "angle": 0, + "content": "4. Low-light Multi-object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.492 + ], + "angle": 0, + "content": "In this section, we peresnt our low-light multi-object tracking method (LTrack). Our main idea is to learn the invariant semantic information under noise disturbance quality degradation. The overall framework can be seen in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.504, + 0.755, + 0.518 + ], + "angle": 0, + "content": "4.1. Formulation and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.753 + ], + "angle": 0, + "content": "In low-light scenes, the camera can capture only a small number of photons in a single exposure. Thus, the potential sensor noise is highlighted, resulting in significant degradation to the image quality [50]. We observed that directly feeding low-light images to the network without any special design leads to the feature map degradation and significantly reduces the performance of the model (see in Sec. 5.2). A direct solution is to apply low-light enhancement techniques, which focus on learning a mapping function from low-light images to clean well-lit images. Since it is a highly ill-posed problem, learning such a mapping function requires considerable computing and storage overhead. Although DNN-based methods have achieved excellent performance in low-light enhancement, images enhanced for visual quality may be suboptimal for downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.755, + 0.892, + 0.876 + ], + "angle": 0, + "content": "In this work, we perform multi-object tracking from low-light images, bypassing the low-light enhancement. Leveraging RAW videos, the network obtains more original scene information compared to sRGB. To enhance the performance and robustness of the multi-object tracking model, our main idea is to earn the invariant semantic information under noise disturbance quality degradation. Thus, we present the adaptive low-pass downsampling (ALD) module" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.805, + 0.9 + ], + "angle": 0, + "content": "1We use UniTrack [12] to extract appearance features." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "385" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.301, + 0.893, + 0.33 + ], + "angle": 0, + "content": "Figure 4. The overall framework of the proposed low-light multi-object tracking method, termed as LTrack. It employs adaptive low-pass downsample module and degradation suppression learning strategy, enabling the model to learn invariant features from low-light videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.471, + 0.432 + ], + "angle": 0, + "content": "to enhance the low-frequency components of images and filter out high-frequency noise. We also propose the degradation suppression learning strategy (DSL), which utilizes paired low-light videos to help the model suppress image noise disturbance and encourage image content response in the feature domain." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.442, + 0.38, + 0.459 + ], + "angle": 0, + "content": "4.2. Adaptive Low-pass Downsampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.47, + 0.616 + ], + "angle": 0, + "content": "The downsampling operation reduces the feature size while preserving the most important information. The noise in low-light images introduces high-frequency disturbance to the feature maps, which can mislead the preservation of object information. To weaken the impact of high-frequency noises and enhance the low-frequency part of the feature map, we introduce spatial low-pass convolution (SConv) to extract low-frequency features from the noised feature maps. The softmax function is used to constrain the original convolution kernel to be low-pass as" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.619, + 0.47, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {K}} _ {i, j} = \\frac {\\exp \\left(\\mathcal {K} _ {i , j}\\right)}{\\sum_ {p , q} \\exp \\left(\\mathcal {K} _ {p , q}\\right)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.809 + ], + "angle": 0, + "content": "where \\(\\mathcal{K}\\) and \\(\\hat{\\mathcal{K}}\\) are the origin convolution kernel and low-pass convolution kernel. We initialize the spatial low-pass convolution kernel using a standard Gaussian kernel. The kernel size is set to 5 to obtain more spatial information. Then, the obtained low-frequency features are adaptively weighted and fused into the original features. We use global average pooling to obtain channel descriptors for both origin features and low-frequency features. These descriptors are then fed into a fully connected layer to compute the weight values." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.383, + 0.835 + ], + "angle": 0, + "content": "4.3. Degradation Suppression Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Given that a low-light image and well-lit image pair share the same content, the model should exhibit identical feature responses to them. However, the low-light image results in shallow features full of noise, and the deep feature exhibits" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.892, + 0.418 + ], + "angle": 0, + "content": "lower responses to objects (as shown in Fig. 5). To address this, our idea is to suppress image noise in shallow features and use well-lit images to help model learning disturbance invariant information from low-light images. The degradation suppression loss can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.42, + 0.892, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D S} = \\sum_ {l} \\left\\| \\mathbf {F} _ {l} ^ {\\text {w e l l}} - \\mathbf {F} _ {l} ^ {\\text {l o w}} \\right\\| _ {2} ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.455, + 0.892, + 0.516 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_l^{well}\\) and \\(\\mathbf{F}_l^{low}\\) denotes \\(l\\)-th feature map corresponding to well-lit and low-light image, respectively. To further suppress the noise in shallow features, we also introduce the Total Variation (TV) loss [38] to features as" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.519, + 0.892, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T V} = \\sum_ {l} | | \\mathbf {G} ^ {r o w} \\mathbf {F} _ {l} ^ {l o w} | | _ {2} ^ {2} + | | \\mathbf {G} ^ {c o l} \\mathbf {F} _ {l} ^ {l o w} | | _ {2} ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.553, + 0.892, + 0.645 + ], + "angle": 0, + "content": "where \\(\\mathbf{G}^{row}\\) and \\(\\mathbf{G}^{col}\\) are the first derivative matrix to role and column. The TV loss adds spatial smoothing constraints to features, which helps model learning to extract noise invariant features. Both the well-lit and low-light images are used to train the model by a common detection loss \\(L_{det}(\\cdot)\\). The total loss is \\(\\mathcal{L}_{Total}\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.647, + 0.892, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {T o t a l}} = \\mathcal {L} _ {\\text {d e t}} ^ {\\text {w e l l}} + \\alpha \\mathcal {L} _ {\\text {d e t}} ^ {\\text {l o w}} + \\beta \\mathcal {L} _ {D S} + \\gamma \\mathcal {L} _ {T V} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.668, + 0.892, + 0.714 + ], + "angle": 0, + "content": "where is the \\( \\mathcal{L}_{det}^{well} \\) and \\( \\mathcal{L}_{det}^{low} \\) are the detection loss for welllit and low-light images, \\( \\alpha \\) and \\( \\beta \\) are loss weights. We set \\( \\alpha, \\beta \\) and \\( \\gamma \\) to 1, 1 and 0.01, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.73, + 0.634, + 0.747 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.676, + 0.772 + ], + "angle": 0, + "content": "5.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Dataset Split. In structuring our dataset, we randomly split the videos into training, validation, and testing sets, consisting of 11, 4, and 11 videos respectively. We also provide LMOT-real dataset that is captured in real-world night scenes with 6 videos. Detailed dataset split and statistical information are shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. Following [8, 42], we recommend using HOTA [31] as the main evaluation metric to simulta" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "386" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.216, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.089, + 0.342, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.089, + 0.468, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.217, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Figure 5. Visualization of shallow and deep features for well-lit and low-light images. It can be seen that, under low-light conditions, the shallow feature is full of noise, and the deep feature exhibits lower responses to objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.468, + 0.425 + ], + "angle": 0, + "content": "neously evaluate the performance of detection and association. We also employ AssA and IDF1 [37] to evaluate association performance, MOTA, and Deta to evaluate detection performance. There are two ways to combine metrics for all classes into a single score. One is by averaging metrics over the class values, and the other is by over the detection values. To avoid possible result bias caused by some categories with fewer samples (such as Bus and Truck), we combine scores by averaging over the detection values." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.468, + 0.6 + ], + "angle": 0, + "content": "Implementation Details. Following ByteTrack [62] and OC-SORT [3], we use YOLOX [17] as our detector. The trackers are pre-trained on the COCO [28] dataset and then trained on our LMOT dataset for 24 epochs. We apply data augmentation strategies including random flip, scale jitter of resizing, and Mosaic. We also use the physical-based noise model [50] for RAW image augment. We use SGD optimizer with weight decay \\(10^{-4}\\) and cosine learning rate schedule, the initial learning rate is \\(10^{-4}\\) and gradually reduces to \\(10^{-5}\\). We apply linear interpolation as post-processing to all trackers, with the maximum gap set to 20." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.613, + 0.383, + 0.629 + ], + "angle": 0, + "content": "5.2. Analysis under low-light conditions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.468, + 0.698 + ], + "angle": 0, + "content": "We analyze the impact of low-light conditions on multi-object tracking using the LMOT validation set. It should be emphasized that the low-light and well-lit video pairs are perfectly aligned in LMOT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Impact to detectors. We first analyze the impact of lighting conditions on the detector. We select YOLOX as the detector since it is widely used in MOT areas [3, 8, 62]. We train the detector using well-lit images (WL), low-light images (LL), and all the images (AL). Then, test them on both well-lit and low-light images. The results are shown in Tab. 3. From the table, we can see that the model trained by welllight images achieves the best result on well-light images, but its performance significantly decreased on low-light images. Further, we visualized the feature maps under these two lighting conditions in Fig. 5. It can be seen that both the shallow feature and deep features of the low-light image are significantly degraded due to the sensor noises. We also" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.889, + 0.174 + ], + "angle": 0, + "content": "
Training dataWell-litLow-light
WLLLmAPmARAP50mAPmARAP50
✓37.045.065.53.14.86.6
✓23.030.840.816.923.730.3
✓✓28.735.249.117.924.132.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.175, + 0.892, + 0.218 + ], + "angle": 0, + "content": "Table 3. Analysis on LMOT validation set for detector (YOLOX [17]). WL and LL indicate the well-lit and low-light, respectively. It shows that it is hard to detect objects under low-light conditions." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.226, + 0.889, + 0.315 + ], + "angle": 0, + "content": "
Cond.Mot.App.HOTAAssAIDF1MOTADetA
-✓86.179.572.982.693.4
WL✓87.377.182.598.898.4
WL✓✓83.778.085.696.890.0
LL✓53.832.140.067.390.2
LL✓✓83.477.585.096.089.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.316, + 0.892, + 0.4 + ], + "angle": 0, + "content": "Table 4. Analysis on LMOT validation set for different association models. Cond, Mot, and App indicate the light condition, motion, and appearance, respectively. The detection boxes are ground-truth boxes. It indicates that appearance information is effective for association under well-lit conditions, but dose not as effective under low-light conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.473 + ], + "angle": 0, + "content": "observed that using low-light images can largely improve the performance of the decoder under low-light conditions, and use all images to achieve the best performance. But this result is much lower than that under well-lit conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.892, + 0.747 + ], + "angle": 0, + "content": "Impact to association modules. We analyze the impact of low-light conditions on the object association modules. Motion and appearance are important for object association. Both of them rely on detection boxes to locate the objects. To separate the impact of the detector, we use the ground-truth boxes as the detection boxes. The results are shown in Tab. 4. It can be seen that using only appearance matching under well-lit conditions achieves the best result while using only appearance matching under low-lit conditions results in very poor performance. This indicates that the objects in LMOT have obvious visual distinguishability, but the distinguishability is significantly reduced under low-light conditions. In Fig. 6, we visualize the appearance feature of objects in the LMOT dataset under both well-lit and low-light conditions. We can observe that under well-lit conditions, LMOT is very distinguishable in the feature space. However, under low-light conditions, this discrimination of LMOT decreased significantly." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.734, + 0.771 + ], + "angle": 0, + "content": "5.3. Results on LMOT dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We compare the proposed method with potential low-light multi-object tracking methods. We first train three baseline models using low-light videos, well-lit videos, and both of them, without any modifications to baseline trackers. They are denoted by Base-low, Base-well, and Base-all. The second type of method is tracking after low-light enhancement. These methods use low-light enhancement methods as preprocessing module. Both the enhanced videos and well-lit" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "387" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.894, + 0.302 + ], + "angle": 0, + "content": "
MethodByteTrack[62]OC-SORT[3]DetectionParams (M)FLOPS (G)
HOTAAssAIDF1HOTAAssAIDF1mAPAP50AP75
Base-low28.042.910.928.840.334.318.832.418.999.00793.29
Base-well13.536.832.913.831.712.05.59.75.599.00793.29
Base-all28.142.632.928.939.834.619.032.519.199.00793.29
LLFlow [48]28.842.934.029.441.035.022.438.322.6116.428755.91
SDSD [47]29.141.135.529.336.936.819.836.219.3103.301136.39
DNF [24]27.640.933.028.136.435.319.135.618.2101.83907.33
SMOID [23]28.039.733.928.736.735.619.135.518.2120.887286.17
RAOD [52]26.142.929.526.142.929.518.331.318.499.07897.04
LTrack (Ours)29.443.235.229.839.036.723.440.623.5100.43800.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.303, + 0.893, + 0.334 + ], + "angle": 0, + "content": "Table 5. Experimental results on LMOT dataset. Base-low, Base-well and Base-all indicate the baseline model trained on low-light, well-lit and all of them, respectively. The number of parameters and prediction latency of each method are reported along with the accuracy" + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.345, + 0.272, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.458, + 0.208, + 0.47 + ], + "angle": 0, + "content": "(a) Well-lit" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.346, + 0.465, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.458, + 0.408, + 0.47 + ], + "angle": 0, + "content": "(b) Low-light" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.529 + ], + "angle": 0, + "content": "Figure 6. Visualization of appearance features of LMOT dataset using t-SNE. The same object is coded by the same color. It indicates that the appearance of objects under well-lit is distinguishable, but significantly reduced under low-light conditions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.732 + ], + "angle": 0, + "content": "videos are used for training. We select four different low-light enhancement techniques for comparison. LLFlow [48] and SDSD [47] are low-light enhancement methods based on RGB images, while DNF [24] and SMOID [23] are based on RAW low-light enhancement methods. LLFlow and DNF take images as input, SDSD and SMOID take videos as input. The RAW object detection method is also trained together on well-lit and low-light videos. Its output bounding boxes are directly fed to the tracker. We test all the potential methods on two state-of-the-art trackers ByteTrack [62] and OC-SORT [3]. We also show mAP, AP50, and AP75 to highlight the detector performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "From Tab. 5 we can see that, the proposed LTrack achieves the best HOTA and is highly competitive on all metrics. For example, the proposed method improves HOTA 1.3 with almost the same parameters and computation as base methods. This strongly proves the effectiveness of the proposed method. Compared with tracking after low-light enhancement, these methods have a large amount of additional parameters and computation but still perform worse than the proposed method. For example, LLFlow delivers 8 times FLOPS but still performs worse than our LTrack. In addition, We observed that there is" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.343, + 0.915, + 0.536 + ], + "angle": 0, + "content": "
MethodHOTAAssAIDF1MOTADet
Base-low32.533.439.342.427.1
Base-well28.030.434.136.923.3
Base-all34.638.140.044.830.2
LLFlow[48]33.638.038.443.129.7
SDSD[47]31.634.236.738.827.5
DNF[24]33.239.036.944.430.1
SMOID[23]31.234.136.641.226.9
RAOD[52]32.233.539.440.926.5
LTrack (Ours)35.138.940.445.230.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.537, + 0.892, + 0.565 + ], + "angle": 0, + "content": "Table 6. Experimental results on LMOT-real dataset with OC-SORT[3]. The best result are shown in boldface" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.894, + 0.746 + ], + "angle": 0, + "content": "no significant difference in results between RAW-based image enhancement methods and RGB-based image enhancement methods. This is not consistent with what our method has observed. The reason may be that low light enhancement focuses on image quality restoration and may mislead downstream tasks. As for RAOD [52], it has almost the same number of parameters and computational load as our method, but its performance is much lower than the proposed LTrack. Despite addressing HDR scenes through a preprocessing module for RAW input, it does not perform well under low-light conditions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.71, + 0.771 + ], + "angle": 0, + "content": "5.4. Results on Real World" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To validate the performance of our method in real-world low-light scenes at night, we evaluate all methods on LMOT-real dataset, using OC-SORT [3]. As shown in Tab. 6, the proposed LTrack performs much better than all comparison methods. Both the methods of tracking after low-light enhancement and the RAW detection method encounter generalizability problems and are not even better than Base-all. This strongly demonstrates the robustness of" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "388" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.465, + 0.178 + ], + "angle": 0, + "content": "
DSLALDHOTAAssAIDF1MOTADetA
✓28.142.632.923.818.7
29.042.435.025.820.1
✓28.342.433.124.619.3
✓✓29.443.235.226.120.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.181, + 0.468, + 0.209 + ], + "angle": 0, + "content": "Table 7. Ablation on adaptive low-pass downsampling (ALD) and degradation suppression learning (DSL)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.216, + 0.465, + 0.304 + ], + "angle": 0, + "content": "
Data TypeHOTAAssAIDF1MOTADetA
sRGB29.242.734.926.420.5
RAW 8-bit29.042.334.526.120.2
RAW 10-bit29.342.934.826.520.4
RAW 12-bit29.443.235.226.120.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.306, + 0.468, + 0.332 + ], + "angle": 0, + "content": "Table 8. Results of different image format and dynamic range on LMOT test set." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.342, + 0.465, + 0.415 + ], + "angle": 0, + "content": "
MethodsHOTAAssAIDF1MOTADetA
Gaussian-Possion29.943.135.726.621.1
Physical-based [50]30.444.136.026.621.2
LMOT (Ours)35.138.940.445.230.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.416, + 0.468, + 0.443 + ], + "angle": 0, + "content": "Table 9. Results with low-light synthesis methods and LMOT dataset on LMOT-real set." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.457, + 0.325, + 0.471 + ], + "angle": 0, + "content": "our LTrack in real-world dark scenes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.481, + 0.325, + 0.497 + ], + "angle": 0, + "content": "5.5. Exploration and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.468, + 0.536 + ], + "angle": 0, + "content": "In this section, we conduct extensive analysis and discussion on LMOT datasets and the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.541, + 0.468, + 0.676 + ], + "angle": 0, + "content": "Ablation Study. We conduct ablation experiments to validate the effectiveness of our improvements, the results are shown in Tab. 7. It can be seen that all improvements contribute effectively to enhanced performance. Among them, the degradation suppression learning strategy demonstrates the most significant effect, resulting in an approximate 1-point enhancement in HOTA. The best results are achieved when employing all strategies simultaneously. This demonstrates the effectiveness of all our contributions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.468, + 0.789 + ], + "angle": 0, + "content": "RAW vs. sRGB. We analyze the impact of input data formats, the results are shown in Tab. 8. The 12-bit RAW format achieves significantly better results than sRGB. Because the RAW format saves much more potential information and is helpful for MOT in low-light scenes. We also observed that higher bitwidth is beneficial for performance, which has also been observed in other vision tasks [21, 52]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "LMOT vs. Synthetic data. We compare our LMOT dataset with the synthetic low-light data to further demonstrate the value of LMOT. We apply the Gaussian-Possian based and Physical-based [50] low-light data synthesis method to synthesize low-light videos from well-lit videos. We train our LTrack on these types of data and evaluate their performance in real low-light scenes using LMOT-real. As" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.09, + 0.894, + 0.212 + ], + "angle": 0, + "content": "
CategoriesHOTAAssAIDF1MOTADetA
Person24.230.829.925.119.2
Bicycle16.940.315.68.97.1
Car37.553.347.233.726.5
Motorcycle19.835.022.615.211.4
Bus44.359.951.836.732.9
Truck6.915.66.43.23.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.212, + 0.892, + 0.239 + ], + "angle": 0, + "content": "Table 10. Results of the proposed method for different categories on LMOT test set." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.257, + 0.892, + 0.303 + ], + "angle": 0, + "content": "shown in Tab. 9, the tracker trained on our LMOT dataset has much better performance in real night scenes, which strongly demonstrates the value of our LMOT dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.31, + 0.892, + 0.447 + ], + "angle": 0, + "content": "Analysis on different category. We also analyze the performance for different categories. From Tab. 10, we can see that cars and buses achieved the best performance because they have regular shapes and larger areas. Trucks exhibited the poorest performance, because they have the fewest instances, making it challenging for the model to learn accurate identification. The person achieves relatively average scores. Bicycles and motorcycles have close scores since they have similar appearance and motion patterns." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.465, + 0.619, + 0.481 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.492, + 0.892, + 0.673 + ], + "angle": 0, + "content": "In this work, we investigate the multi-object tracking in the dark scenes. We build a new low-light multi-object tracking (LMOT) dataset, which provides well-aligned low-light video pairs and high-quality multi-object tracking annotations. We observed that low-light images are significantly degraded by the sensor noises, which also degrades the feature maps and significantly deteriorates the model performance. To learn the invariant semantic formation under noise disturbance and quality degradation, we present the adaptive low-pass downsample module and degradation suppression learning. These improvements greatly enhance the robustness of our method in real-world low-light scenes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.674, + 0.892, + 0.766 + ], + "angle": 0, + "content": "Limitations. We focus on multi-object tracking in the dark scenes. But we do not consider other degradation environments in the real world, such as rainy and foggy days. In our future work, we will consider exploring multi-object tracking in more real-world scenarios, promoting the development of MOT in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.784, + 0.659, + 0.801 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported by the National Key R&D Program of China (2022YFC3300704), the National Natural Science Foundation of China (62331006,62171038, and 62088101), the R&D Program of Beijing Municipal Education Commission (KZ202211417048), and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "389" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.156 + ], + "angle": 0, + "content": "[1] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, pages 3464-3468, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.469, + 0.213 + ], + "angle": 0, + "content": "[2] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. arXiv preprint arXiv:2303.06705, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.469, + 0.268 + ], + "angle": 0, + "content": "[3] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. Observation-centric sort: Rethinking sort for robust multi-object tracking. In CVPR, pages 9686-9696, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.469, + 0.31 + ], + "angle": 0, + "content": "[4] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In CVPR, pages 3291-3300, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.468, + 0.352 + ], + "angle": 0, + "content": "[5] Chen Chen, Qifeng Chen, Minh N Do, and Vladlen Koltun. Seeing motion in the dark. In ICCV, pages 3185-3194, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.469, + 0.395 + ], + "angle": 0, + "content": "[6] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. IJCV, pages 1-21, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[7] Wongun Choi and Silvio Savarese. A unified framework for multi-target tracking and collective activity recognition. In ECCV, pages 215-230, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.469, + 0.494 + ], + "angle": 0, + "content": "[8] Yutao Cui, Chenkai Zeng, Xiaoyu Zhao, Yichun Yang, Gangshan Wu, and Limin Wang. Sportsmot: A large multi-object tracking dataset in multiple sports scenes. arXiv preprint arXiv:2304.05170, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.469, + 0.563 + ], + "angle": 0, + "content": "[9] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.469, + 0.606 + ], + "angle": 0, + "content": "[10] Yunhao Du, Zhicheng Zhao, Yang Song, Yanyun Zhao, Fei Su, Tao Gong, and Hongying Meng. *Strongsort: Make deep-sort great again*. IEEE Transactions on Multimedia, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.469, + 0.674 + ], + "angle": 0, + "content": "[11] Anna Ellis and James Ferryman. Pets2010 and pets2009 evaluation of results using individual ground truthed single views. In 2010 7th IEEE international conference on advanced video and signal based surveillance, pages 135-142, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[12] Wang Zhongdao et al. Do different tracking tasks require different appearance models? 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[13] Huiyuan Fu, Wenkai Zheng, Xicong Wang, Jiaxuan Wang, Heng Zhang, and Huadong Ma. Dancing in the dark: A benchmark towards general low-light video enhancement. In ICCV, pages 12877-12886, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[14] Ying Fu, Jian Chen, Tao Zhang, and Yonggang Lin. Residual scale attention network for arbitrary scale image superresolution. Neurocomputing, 427:201-211, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[15] Ying Fu, Yang Hong, Linwei Chen, and Shaodi You. Le-gan: Unsupervised low-light image enhancement network using attention module and identity invariant loss. Knowledge-Based Systems, 240:108010, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Ruopeng Gao and Limin Wang. Memotr: Long-term memory-augmented transformer for multi-object tracking. In ICCV, pages 9901-9910, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[17] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361, 2012. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[19] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In CVPR, pages 1780-1789, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[20] Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE TIP, 26(2):982-993, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[21] Yang Hong, Kaixuan Wei, Linwei Chen, and Ying Fu. Crafting object detection in very low light. In BMVC, page 3, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.324, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[22] Haidi Ibrahim and Nicholas Sia Pik Kong. Brightness preserving dynamic histogram equalization for image contrast enhancement. IEEE Transactions on Consumer Electronics, 53(4):1752-1758, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.382, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[23] Haiyang Jiang and Yinqiang Zheng. Learning to see moving objects in the dark. In ICCV, pages 7324-7333, 2019. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[24] Xin Jin, Ling-Hao Han, Zhen Li, Chun-Le Guo, Zhi Chai, and Chongyi Li. Dnf: Decouple and feedback network for seeing in the dark. In CVPR, pages 18135-18144, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[25] Daniel J Jobson, Zia-ur Rahman, and Glenn A Woodell. Properties and performance of a center/surround retina. IEEE TIP, 6(3):451-462, 1997. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.527, + 0.892, + 0.582 + ], + "angle": 0, + "content": "[26] Laura Leal-Taixe, Anton Milan, Ian Reid, Stefan Roth, and Konrad Schindler. Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv preprint arXiv:1504.01942, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.584, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[27] Sohyun Lee, Jaesung Rim, Boseung Jeong, Geonu Kim, Byungju Woo, Haechan Lee, Sunghyun Cho, and Suha Kwak. Human pose estimation in extremely low-light conditions. In CVPR, pages 704-714, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.642, + 0.892, + 0.698 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[29] Qiankun Liu, Qi Chu, Bin Liu, and Nenghai Yu. Gsm: Graph similarity model for multi-object tracking. In IJCAI, pages 530-536, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[30] Qiankun Liu, Dongdong Chen, Qi Chu, Lu Yuan, Bin Liu, Lei Zhang, and Nenghai Yu. Online multi-object tracking with unsupervised re-identification learning and occlusion estimation. Neurocomputing, 483:333–347, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[31] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. IJCV, 129:548-578, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Kang Ma, Ying Fu, Dezhi Zheng, Chunshui Cao, Xuecai Hu, and Yongzhen Huang. Dynamic aggregated network for gait recognition. In CVPR, pages 22076-22085, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "390" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[33] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, pages 8844-8854, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[34] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831, 2016. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.235 + ], + "angle": 0, + "content": "[35] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In CVPR, pages 1765–1773, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.238, + 0.469, + 0.292 + ], + "angle": 0, + "content": "[36] Jihun Park, Jinseok Hong, Woolil Shim, and Dae-Jin Jung. Multi-object tracking on swir images for city surveillance in an edge-computing environment. Sensors, 23(14):6373, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.469, + 0.35 + ], + "angle": 0, + "content": "[37] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In ECCV, pages 17-35, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.469, + 0.394 + ], + "angle": 0, + "content": "[38] Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena, 60(1-4):259-268, 1992. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.469, + 0.438 + ], + "angle": 0, + "content": "[39] Aashish Sharma and Robby T Tan. Nighttime visibility enhancement by increasing the dynamic range and suppression of light effects. In CVPR, pages 11977-11986, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[40] Shuran Song and Jianxiong Xiao. Tracking revisited using rgbd camera: Unified benchmark and baselines. In ICCV, pages 233-240, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.539 + ], + "angle": 0, + "content": "[41] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.541, + 0.469, + 0.598 + ], + "angle": 0, + "content": "[42] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In CVPR, pages 20993-21002, 2022. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.6, + 0.469, + 0.653 + ], + "angle": 0, + "content": "[43] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.698 + ], + "angle": 0, + "content": "[44] Ye Tian, Ying Fu, and Jun Zhang. Plug-and-play algorithm for under-sampling fourier single-pixel imaging. Science China. Information Sciences, 65(10):209303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.469, + 0.741 + ], + "angle": 0, + "content": "[45] Ye Tian, Ying Fu, and Jun Zhang. Local-enhanced transformer for single-pixel imaging. Optics Letters, 48(10): 2635-2638, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.798 + ], + "angle": 0, + "content": "[46] Chaoqun Wang, Chunyan Xu, Zhen Cui, Ling Zhou, Tong Zhang, Xiaoya Zhang, and Jian Yang. Cross-modal patternpropagation for rgb-t tracking. In CVPR, pages 7064-7073, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[47] Ruixing Wang, Xiaogang Xu, Chi-Wing Fu, Jiangbo Lu, Bei Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In ICCV, pages 9700-9709, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[48] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In AAAI, pages 2604-2612, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[49] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[50] Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. A physics-based noise formation model for extreme low-light raw denoising. In CVPR, pages 2758-2767, 2020. 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[51] Kaixuan Wei, Angelica Aviles-Rivero, Jingwei Liang, Ying Fu, Hua Huang, and Carola-Bibiane Schonlieb. Tfnpn: Tuning-free plug-and-play proximal algorithms with applications to inverse imaging problems. Journal of Machine Learning Research, 23(16):1-48, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[52] Ruikang Xu, Chang Chen, Jingyang Peng, Cheng Li, Yibin Huang, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Toward raw object detection: A new benchmark and a new model. In CVPR, pages 13384-13393, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[53] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. In CVPR, pages 17714-17724, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[54] Jinyu Yang, Shang Gao, Zhe Li, Feng Zheng, and Ales Leonardis. Resource-efficient rgbd aerial tracking. In CVPR, pages 13374-13383, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[55] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2636-2645, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[56] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In ECCV, pages 659–675, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[57] Fan Zhang, Yu Li, Shaodi You, and Ying Fu. Learning temporal consistency for low light video enhancement from single images. In CVPR, pages 4967-4976, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[58] Jiqing Zhang, Xin Yang, Yingkai Fu, Xiaopeng Wei, Baocai Yin, and Bo Dong. Object tracking by jointly exploiting frame and event domain. In ICCV, pages 13043-13052, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[59] Pengyu Zhang, Jie Zhao, Dong Wang, Huchuan Lu, and Xiang Ruan. Visible-thermal uav tracking: A large-scale benchmark and new baseline. In CVPR, pages 8886–8895, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[60] Tao Zhang, Ying Fu, and Cheng Li. Hyperspectral image denoising with realistic data. In ICCV, pages 2248-2257, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[61] Tao Zhang, Ying Fu, and Cheng Li. Deep spatial adaptive network for real image demosaicing. In AAAI, pages 3326-3334, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[62] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, pages 1-21, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[63] Yang Zhang, Tiancai Wang, and Xiangyu Zhang. Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. In CVPR, pages 22056-22065, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[64] Zhiyu Zhu, Junhui Hou, and Dapeng Oliver Wu. Cross-modal orthogonal high-rank augmentation for rgb-event" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "391" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "transformer-trackers. In ICCV, pages 22045-22055, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[65] Yunhao Zou and Ying Fu. Estimating fine-grained noise model via contrastive learning. In CVPR, pages 12682-12691, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[66] Yunhao Zou, Yinqiang Zheng, Tsuyoshi Takatani, and Ying Fu. Learning to reconstruct high speed and high dynamic range videos from events. In CVPR, pages 2024-2033, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[67] Yunhao Zou, Chenggang Yan, and Ying Fu. Rawhdr: High dynamic range image reconstruction from a single raw image. In ICCV, pages 12334-12344, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "392" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-Object Tracking in the Dark/4c20e9a2-1106-45cf-a36a-0161982e6015_origin.pdf b/2024/Multi-Object Tracking in the Dark/4c20e9a2-1106-45cf-a36a-0161982e6015_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2923651865ad21345dc8116bbe0e4722b623c880 --- /dev/null +++ b/2024/Multi-Object Tracking in the Dark/4c20e9a2-1106-45cf-a36a-0161982e6015_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20e9d2002c00375eaeb91418f85909d444fe60cadc6f9f584b7c78e668f7e85e +size 3719765 diff --git a/2024/Multi-Object Tracking in the Dark/full.md b/2024/Multi-Object Tracking in the Dark/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b11e73cd1e8538fba10d6f0fdd6e28095a2e0178 --- /dev/null +++ b/2024/Multi-Object Tracking in the Dark/full.md @@ -0,0 +1,327 @@ +# Multi-Object Tracking in the Dark + +Xinzhe Wang Kang Ma Qiankun Liu Yunhao Zou Ying Fu * Beijing Institute of Technology + +{wangxinzhe, makang, liuqk3, zouyunhao, fuying}@bit.edu.cn + +# Abstract + +Low-light scenes are prevalent in real-world applications (e.g. autonomous driving and surveillance at night). Recently, multi-object tracking in various practical use cases have received much attention, but multi-object tracking in dark scenes is rarely considered. In this paper, we focus on multi-object tracking in dark scenes. To address the lack of datasets, we first build a Low-light Multi-Object Tracking (LMOT) dataset. LMOT provides well-aligned low-light video pairs captured by our dual-camera system, and high-quality multi-object tracking annotations for all videos. Then, we propose a low-light multi-object tracking method, termed as LTrack. We introduce the adaptive low-pass downsample module to enhance low-frequency components of images outside the sensor noises. The degradation suppression learning strategy enables the model to learn invariant information under noise disturbance and image quality degradation. These components improve the robustness of multi-object tracking in dark scenes. We conducted a comprehensive analysis of our LMOT dataset and proposed LTrack. Experimental results demonstrate the superiority of the proposed method and its competitiveness in real night low-light scenes. Dataset and Code: https://github.com/ying-fu/LMOT + +# 1. Introduction + +Multi-object tracking (MOT) aims to locate and associate multiple objects in video sequences. It is widely used in many downstream applications, such as video recognition [7, 32], autonomous driving [18], and surveillance [11]. Recently, multi-object tracking in various practical use cases has garnered much attention [8, 9, 18, 34, 42], greatly advancing the development of MOT. However, these works are primarily tailored for high-quality inputs and overlook the prevalent low-light scenes in real-world scenes. Motivated by this, we study multi-object tracking in dark scenes. + +Due to the physical limitations of existing cameras, ac + +quiring high-quality videos under low-light conditions is difficult. One inherent difficulty in capturing consecutive video frames under such conditions is avoiding motion blur. Current camera technology typically requires short exposure times (usually just a few tens of milliseconds), but in low-light scenarios, the sensor struggles to capture an adequate number of photons within this limited duration. This limitation inevitably leads to degradation in image quality accompanied by higher noise levels. This presents two main challenges for MOT under low-light conditions. The first challenge is for collecting a low-light multi-object tracking dataset. Collecting and annotating a low-light MOT dataset is difficult and expensive. MOT requires dynamic object videos, but videos captured in low-light scenes have extremely low brightness, making it hard to recognize and annotate objects in the videos. The second challenge revolves around low-light multi-object tracking. The popular tracking-by-detection paradigm [3, 10, 29, 30, 62] generally consists of detector, motion-based association module, and appearance-based association module. These modules typically require high-quality input images. The poor quality of low-light images leads to severe performance degradation for both detectors and appearance-based correlation modules. A simple approach is to cascade low-light enhance modules [2, 23, 24, 47], but this introduces additional computational costs. Furthermore, images optimized for visual quality may be suboptimal for downstream tasks [6, 21, 27]. + +In this paper, we build a low-light multi-object tracking dataset (LMOT), specifically designed to address the challenges of multi-object tracking in dark scenes. To this end, we develop a dual-camera system that simultaneously captures well-lit and low-light video frames. The video pairs are highly aligned in both spatial and temporal dimensions, offering two key benefits. First, it enables us to annotate on the well-lit videos, resulting in high-quality annotations. Second, the well-lit videos can provide additional supervision information during the training phase, and strongly enhance the performance in the dark scenes. After careful annotation, we collect 32 video sequences $(2.3\times \mathrm{MOT17})$ over $35\mathrm{K}$ frames $(3.1\times \mathrm{MOT17})$ ,and over $815\mathrm{K}$ bounding boxes $(2.8\times \mathrm{MOT17})$ .The RAW data is the output of + +the image sensor and is the input data of the image signal processor (ISP). It saves all information from the image sensor, which is crucial for capturing object information in dark scenes [6, 52, 67]. Therefore, we collect RAW videos for LMOT. + +Additionally, we propose a low-light multi-object tracking method, termed as LTrack. The low-light video is characterized by substantial sensor noise and poor image quality, which significantly degrades both shallow and deep feature representations, leading to reduced tracking performance. We observe that the sensor noise in low-light images exhibits a similarity to adversarial attacks [35, 43]. To address this issue, our main idea is to learn the invariant semantic information under noise disturbance quality degradation. We present the adaptive low-pass downsample module (ALD). It employs spatial low-pass convolution to extract low-frequency components from images, excluding noises, and adaptively enhance the feature maps. We also present the degradation suppression learning strategy (DSL), which utilizes paired low-light videos to help the model suppress the noise disturbance and encourage image content response in the feature domain. We conduct a comprehensive analysis of our LMOT dataset and validate the superiority of our LTrack in real-world night scenes. + +In summary, our main contributions are as follows: + +- We build the first low-light multi-object tracking dataset using a carefully constructed dual-camera system. It provides well-aligned low-light videos in RAW format, and high-quality MOT annotations for all videos. +- We propose a low-light multi-object tracking method. It utilizes the adaptive low-pass downsample module and the degradation suppression learning strategy to learn to extract invariant features from low-light videos. +- We conduct a comprehensive analysis of our dataset and the proposed method. Experimental results demonstrate the superiority of the proposed method and its competitiveness on real night scenes. + +# 2. Related Work + +In this section, we first review the current research status of low-light enhancement and low-light datasets. Then, we summarized the present research for multi-object tracking and tracking in the dark scenes. + +Low-light enhancement. Traditional methods for lowlight enhancement are primarily based on histogram equalization and Retinex theory [20, 22, 25]. Recently, deep learning has been explored for many low-level tasks [14, 44, 45, 51, 60, 61, 65, 66], and achieved superior results on low-light enhancement [2, 4, 13, 15, 19, 24, 39, 53]. While these methods are capable of recovering images with high visual quality, they often require heavy computation and may not consider downstream tasks, resulting in suboptimal + +![](images/02e02ea9df4d388f77874d5066d6b2a59c03d1a20f1308b2772c31d5a2c07a46.jpg) +Figure 1. Our dual-camera system. It consists of two cameras, a beam splitter, and an ND-filter. Two cameras of identical models are meticulously engineered to achieve pixel-by-pixel alignment in the captured video data. + +performance. In contrast, our approach focuses on directly learning multi-object tracking from low-light images, thus bypassing the low-light enhancement. + +Low-light datasets. The long-short exposure is a widely used method to collect paired low-light images, but can only be used to collect low-light images for static scenes. [4, 5, 49]. To capture dynamic scenes, some works designed the mechatronic system. They obtain paired low-light data by repeating the motion twice [13, 47]. However, these mechatronic systems cannot be used to collect dynamic object videos in the wild. Jiang et al. [23] designed a dual-camera system that simultaneously captures paired well-lit and low-light videos, making it possible to capture dynamic scenes and dynamic object videos for multi-object tracking. Zou et al. [57] setup an optical system to collect paired videos and event streams. These works explore various ways to construct low-light datasets and inspire research on high-level vision tasks in dark scenes, such as LOD [21] for object detection, LIS [6] for instance segmentation, and ExPose [27] for human pose estimation. These datasets provide paired low-light data only for image tasks, and cannot be expanded for multi-object tracking in dark scenes. + +Multi-object tracking datasets. MOT15 [26] is the first large-scale benchmark for multi-object tracking. MOT17 [34] stands as one of the most widely applied MOT benchmarks. MOT20 [9] focuses on very crowded scenes. These three datasets are for pedestrians. KITTI [18] and BDD100K [55] are for autonomous driving scenarios. DanceTrack [42] focuses on dancing scenes and is characterized by similar appearance and diverse motions. Recently, SportsMOT [8] aims to track athletes and encourage algorithms to promote both appearance and motion association. These datasets explore multi-object tracking in various practical use cases, but none of them consider multi-object tracking in dark scenes. + +![](images/f14b87b761c53707cda4a4222aa69192accfca9747e49a4a8a5fec6a6e51d76d.jpg) +Figure 2. Two example videos from our LMOT dataset. It provides well-aligned low-light video pairs and MOT annotations for all videos. The time interval between adjacent frames is 1s. The first row is the low-light video, the second row is the scaled low-light video and the last row is the well-lit video. Our LMOT dataset is collected from city outdoor scenes. + +![](images/046c4b39280bea23ad40023227baac67a1b0c29921c77ad36239dcbb902d9f77.jpg) + +
DatasetFormatVideosFramesLength (s)BboxTracks
MOT17 [34]sRGB1411,235463292,7331,342
MOT20 [9]sRGB813,4105351,652,0403,456
DanceTrack [42]sRGB100105,8555,292-990
SportMOT [8]sRGB240150,3796,0151,629,4903,401
KITTI [18]sRGB218,000-47,000917
BDD100K [55]sRGB1,600318,000-3,300,000131,000
SWIR [36]SWIR-7,309-57,221-
LMOTRAW3235,1201,756815,5504,090
+ +Object tracking in the dark. To track in low-light scenes, some methods explore to use multi-modal information for single-object tracking, such as event camera [58, 64], depth [40, 54] and thermal [46, 59] devices. Park et al. [36] proposed to use Short-Wave Infrared (SWIR) images for multi-object tracking, since its advantages in terms of robustness in low-light conditions. The common drawback of these methods is that they require additional hardware equipment and cannot be applied to the most widely used CMOS imaging systems. + +SORT [1] uses the Kalman Filter motion model and employs IoU for association. ByteTrack [62] enhances tracking performance by considering low-confidence bounding boxes. OS-SORT [3] enhances SORT by restoring lost targets. Recently, Transformer has been explored for MOT [16, 33, 41, 56, 63]. These methods have achieved high performance in many practical scenarios, but they do not consider dealing with low-light conditions. We focus on multi-object tracking under low-light conditions. Based on RAW videos, our method is highly practical with excellent performance and does not require additional hardware. + +# 3. Low-light Multi-object Tracking Dataset + +In this section, we first introduce our dual camera system and the details of collecting and annotating our low-light multi-object tracking (LMOT) dataset. Then, we analyzed the statistical characteristics of our LMOT dataset. + +Table 1. Comparison of statistics between existing MOT datasets and our LMOT dataset + +
DatasetSplitVideosBboxTracksPaired Well-lit
LMOT-dualtrain11309,4661533✓
val4131,781626✓
test11312,7421644✓
LMOT-realreal661,561287
+ +Table 2. Detailed statistics and data splits for LMOT. + +# 3.1. Dataset Construction + +Multi-object tracking requires dynamic scenes and object video. To collect low-light videos for multi-object tracking, we build a dual-camera system [23], which is illustrated in Fig. 1. It can simultaneously capture paired lowlight and well-lit video pairs. Its main components include a beam splitter, a neutral density (ND) filter, and two FLIR Grasshopper3 GS3-U3-23S6C cameras. The beam splitter divides the incoming light into two separate paths. This arrangement allows one camera to capture well-lit images directly, while the other camera records low-light images, with the ND-filter attenuating the light intensity. To ensure temporal synchronization of the video frames, we employ the hardware interface to trigger the camera exposure events. Moreover, to avoid frame loss, our dual-camera system uses two independent hardware interfaces for data transmission and is equipped with a high-speed solid-state drive. Thanks to precise calibration, our dual-camera system can capture paired low-light and well-lit videos in real time for dynamic scenes and objects. More details about our dual-camera system are given in supplementary materials. + +We save the video frames in RAW format before the images are processed by the camera image signal processor (ISP). In terms of camera settings, we set the exposure time for both cameras to $10ms$ , and the frame rate is fixed at 20. This setting is feasible to avoid motion blur. We adjust the gain level for well-lit cameras to achieve optimal image quality. The gain for low-light cameras is consistently set to the maximum value, to simulate low-light capturing setup in real scenarios. We also collect a real low-light MOT dataset (LMOT-real) to evaluate performance in real-world + +![](images/4c7dae4c6446524a305bb96a7b1e8b7a9ac9a898214ea69671b1262768a5cec9.jpg) +(a) Number of instances per category + +![](images/1d3709a5a938d9b2e9f9b3d3e8c00ff8410f251cd6fd2bb651690ed41815680a.jpg) +(b) IoU on adjacent frames +Figure 3. (a) Number of instances per category. LMOT consists of 6 categories, most of the instances are the person and car. (b) IoU on adjacent frames. Compared to MOT17, KITTI, and DanceTrack, LMOT has a roughly average score. This indicates that LMOT has a relatively normal movement speed. (c) Cosine distance of appearance features. The cosine distance is smaller under low-light conditions, indicating that the appearance distinguishability is decreased under low-light conditions. + +![](images/10f3feb2bb204853dce35c9fa173111cd49802ff040fddc688a9499a8c5eb335.jpg) +(c) Cosine distance of appearance features + +dark scenes. These videos are captured using a single camera with the same camera settings. + +Our LMOT dataset contains a variety of city outdoor scenes, including roads, overpasses, pedestrians, and intersection. The overpass scenes take an overhead shot of objects, while all other scenes are captured from the perspective of pedestrians. To account for the impact of camera motion, we introduce arbitrary horizontal rotations and vertical random movements to the camera. Fig. 2 shows two sampled video sequences from our LMOT dataset. + +We annotate six types of moving objects, including car, person, bicycle, motorcycle, bus, and truck. The annotated labels include bounding boxes, identifications, and visibility status. For partly occluded objects, a full box is annotated. For a fully occluded object, an estimated box is annotated. Each object has a unique ID throughout the entire video. Thanks to well-aligned low-light and well-lit videos, we can annotate well-lit videos to simultaneously obtain labels for low-light videos. This greatly reduces the annotation difficulty and enhances quality. Lastly, we carefully review all the annotation results. + +# 3.2. Dataset Statistic + +LMOT is a large-scale dataset that focuses on multi-object tracking in dark scenes. We compare the statistics of LMOT with existing MOT datasets in Tab. 1. It can be seen that LMOT is approximately three times larger than MOT17 [34]. Compared to large-scale datasets such as DanceTrack [42], SportsMOT [25], KITTI [18] and BDD100K [55], the scale of LMOT is still considerable. It should be emphasized that these datasets are not for multi-object tracking in dark scenes and only provide sRGB images. Compared to SWIR, our LMOT dataset has approximately $5 \times$ frames and $14 \times$ bounding boxes. The detailed statistics and data splits for LMOT are shown in Tab. 2. + +We present the number of instances for each category in Fig. 3 (a). The majority of instances are persons and cars. As shown in Fig. 3 (b), the average IoU on adjacent frames of LMOT is lower than MOT17 and DanceTrack, but higher than KITTI. This indicates that the motions in LMOT are + +fast but within normal range. Following [42], we use cosine distance of appearance features ${}^{1}$ to evaluate the appearance similarity. From Fig. 3 (c), we can see that the cosine distance of appearance features under low-light conditions is smaller than that under well-lit conditions. In other words, the appearance of objects will deteriorate under low-light conditions, making them harder to distinguish. + +# 4. Low-light Multi-object Tracking + +In this section, we peresnt our low-light multi-object tracking method (LTrack). Our main idea is to learn the invariant semantic information under noise disturbance quality degradation. The overall framework can be seen in Fig. 4. + +# 4.1. Formulation and Motivation + +In low-light scenes, the camera can capture only a small number of photons in a single exposure. Thus, the potential sensor noise is highlighted, resulting in significant degradation to the image quality [50]. We observed that directly feeding low-light images to the network without any special design leads to the feature map degradation and significantly reduces the performance of the model (see in Sec. 5.2). A direct solution is to apply low-light enhancement techniques, which focus on learning a mapping function from low-light images to clean well-lit images. Since it is a highly ill-posed problem, learning such a mapping function requires considerable computing and storage overhead. Although DNN-based methods have achieved excellent performance in low-light enhancement, images enhanced for visual quality may be suboptimal for downstream tasks. + +In this work, we perform multi-object tracking from low-light images, bypassing the low-light enhancement. Leveraging RAW videos, the network obtains more original scene information compared to sRGB. To enhance the performance and robustness of the multi-object tracking model, our main idea is to earn the invariant semantic information under noise disturbance quality degradation. Thus, we present the adaptive low-pass downsampling (ALD) module + +![](images/a123ddd19e490d6e5311f513268160ef5309ce2988220fbd0c730ce7949900fd.jpg) +Figure 4. The overall framework of the proposed low-light multi-object tracking method, termed as LTrack. It employs adaptive low-pass downsample module and degradation suppression learning strategy, enabling the model to learn invariant features from low-light videos. + +to enhance the low-frequency components of images and filter out high-frequency noise. We also propose the degradation suppression learning strategy (DSL), which utilizes paired low-light videos to help the model suppress image noise disturbance and encourage image content response in the feature domain. + +# 4.2. Adaptive Low-pass Downsampling + +The downsampling operation reduces the feature size while preserving the most important information. The noise in low-light images introduces high-frequency disturbance to the feature maps, which can mislead the preservation of object information. To weaken the impact of high-frequency noises and enhance the low-frequency part of the feature map, we introduce spatial low-pass convolution (SConv) to extract low-frequency features from the noised feature maps. The softmax function is used to constrain the original convolution kernel to be low-pass as + +$$ +\hat {\mathcal {K}} _ {i, j} = \frac {\exp \left(\mathcal {K} _ {i , j}\right)}{\sum_ {p , q} \exp \left(\mathcal {K} _ {p , q}\right)} \tag {1} +$$ + +where $\mathcal{K}$ and $\hat{\mathcal{K}}$ are the origin convolution kernel and low-pass convolution kernel. We initialize the spatial low-pass convolution kernel using a standard Gaussian kernel. The kernel size is set to 5 to obtain more spatial information. Then, the obtained low-frequency features are adaptively weighted and fused into the original features. We use global average pooling to obtain channel descriptors for both origin features and low-frequency features. These descriptors are then fed into a fully connected layer to compute the weight values. + +# 4.3. Degradation Suppression Learning + +Given that a low-light image and well-lit image pair share the same content, the model should exhibit identical feature responses to them. However, the low-light image results in shallow features full of noise, and the deep feature exhibits + +lower responses to objects (as shown in Fig. 5). To address this, our idea is to suppress image noise in shallow features and use well-lit images to help model learning disturbance invariant information from low-light images. The degradation suppression loss can be expressed as + +$$ +\mathcal {L} _ {D S} = \sum_ {l} \left\| \mathbf {F} _ {l} ^ {\text {w e l l}} - \mathbf {F} _ {l} ^ {\text {l o w}} \right\| _ {2} ^ {2} \tag {2} +$$ + +where $\mathbf{F}_l^{well}$ and $\mathbf{F}_l^{low}$ denotes $l$ -th feature map corresponding to well-lit and low-light image, respectively. To further suppress the noise in shallow features, we also introduce the Total Variation (TV) loss [38] to features as + +$$ +\mathcal {L} _ {T V} = \sum_ {l} | | \mathbf {G} ^ {r o w} \mathbf {F} _ {l} ^ {l o w} | | _ {2} ^ {2} + | | \mathbf {G} ^ {c o l} \mathbf {F} _ {l} ^ {l o w} | | _ {2} ^ {2} \tag {3} +$$ + +where $\mathbf{G}^{row}$ and $\mathbf{G}^{col}$ are the first derivative matrix to role and column. The TV loss adds spatial smoothing constraints to features, which helps model learning to extract noise invariant features. Both the well-lit and low-light images are used to train the model by a common detection loss $L_{det}(\cdot)$ . The total loss is $\mathcal{L}_{Total}$ , i.e., + +$$ +\mathcal {L} _ {\text {T o t a l}} = \mathcal {L} _ {\text {d e t}} ^ {\text {w e l l}} + \alpha \mathcal {L} _ {\text {d e t}} ^ {\text {l o w}} + \beta \mathcal {L} _ {D S} + \gamma \mathcal {L} _ {T V} \tag {4} +$$ + +where is the $\mathcal{L}_{det}^{well}$ and $\mathcal{L}_{det}^{low}$ are the detection loss for welllit and low-light images, $\alpha$ and $\beta$ are loss weights. We set $\alpha, \beta$ and $\gamma$ to 1, 1 and 0.01, respectively. + +# 5. Experiments + +# 5.1. Experiment Setup + +Dataset Split. In structuring our dataset, we randomly split the videos into training, validation, and testing sets, consisting of 11, 4, and 11 videos respectively. We also provide LMOT-real dataset that is captured in real-world night scenes with 6 videos. Detailed dataset split and statistical information are shown in Tab. 2. + +Evaluation Metrics. Following [8, 42], we recommend using HOTA [31] as the main evaluation metric to simulta + +![](images/c80a3cc4d391a972885812c8a56519f4f260a47438bad78853ea7bc5a552d15f.jpg) +Figure 5. Visualization of shallow and deep features for well-lit and low-light images. It can be seen that, under low-light conditions, the shallow feature is full of noise, and the deep feature exhibits lower responses to objects. + +![](images/eeff969e8471e78601b0b94d4a2a6d161420522a88f3632d691baefaa324ce82.jpg) + +![](images/9b1169c4dcaf850e15687b2fc7dc336e016615e47bc6e0bb2c8a26590438848d.jpg) + +neously evaluate the performance of detection and association. We also employ AssA and IDF1 [37] to evaluate association performance, MOTA, and Deta to evaluate detection performance. There are two ways to combine metrics for all classes into a single score. One is by averaging metrics over the class values, and the other is by over the detection values. To avoid possible result bias caused by some categories with fewer samples (such as Bus and Truck), we combine scores by averaging over the detection values. + +Implementation Details. Following ByteTrack [62] and OC-SORT [3], we use YOLOX [17] as our detector. The trackers are pre-trained on the COCO [28] dataset and then trained on our LMOT dataset for 24 epochs. We apply data augmentation strategies including random flip, scale jitter of resizing, and Mosaic. We also use the physical-based noise model [50] for RAW image augment. We use SGD optimizer with weight decay $10^{-4}$ and cosine learning rate schedule, the initial learning rate is $10^{-4}$ and gradually reduces to $10^{-5}$ . We apply linear interpolation as post-processing to all trackers, with the maximum gap set to 20. + +# 5.2. Analysis under low-light conditions + +We analyze the impact of low-light conditions on multi-object tracking using the LMOT validation set. It should be emphasized that the low-light and well-lit video pairs are perfectly aligned in LMOT. + +Impact to detectors. We first analyze the impact of lighting conditions on the detector. We select YOLOX as the detector since it is widely used in MOT areas [3, 8, 62]. We train the detector using well-lit images (WL), low-light images (LL), and all the images (AL). Then, test them on both well-lit and low-light images. The results are shown in Tab. 3. From the table, we can see that the model trained by welllight images achieves the best result on well-light images, but its performance significantly decreased on low-light images. Further, we visualized the feature maps under these two lighting conditions in Fig. 5. It can be seen that both the shallow feature and deep features of the low-light image are significantly degraded due to the sensor noises. We also + +
Training dataWell-litLow-light
WLLLmAPmARAP50mAPmARAP50
✓37.045.065.53.14.86.6
✓23.030.840.816.923.730.3
✓✓28.735.249.117.924.132.2
+ +Table 3. Analysis on LMOT validation set for detector (YOLOX [17]). WL and LL indicate the well-lit and low-light, respectively. It shows that it is hard to detect objects under low-light conditions. + +
Cond.Mot.App.HOTAAssAIDF1MOTADetA
-✓86.179.572.982.693.4
WL✓87.377.182.598.898.4
WL✓✓83.778.085.696.890.0
LL✓53.832.140.067.390.2
LL✓✓83.477.585.096.089.8
+ +Table 4. Analysis on LMOT validation set for different association models. Cond, Mot, and App indicate the light condition, motion, and appearance, respectively. The detection boxes are ground-truth boxes. It indicates that appearance information is effective for association under well-lit conditions, but dose not as effective under low-light conditions. + +observed that using low-light images can largely improve the performance of the decoder under low-light conditions, and use all images to achieve the best performance. But this result is much lower than that under well-lit conditions. + +Impact to association modules. We analyze the impact of low-light conditions on the object association modules. Motion and appearance are important for object association. Both of them rely on detection boxes to locate the objects. To separate the impact of the detector, we use the ground-truth boxes as the detection boxes. The results are shown in Tab. 4. It can be seen that using only appearance matching under well-lit conditions achieves the best result while using only appearance matching under low-lit conditions results in very poor performance. This indicates that the objects in LMOT have obvious visual distinguishability, but the distinguishability is significantly reduced under low-light conditions. In Fig. 6, we visualize the appearance feature of objects in the LMOT dataset under both well-lit and low-light conditions. We can observe that under well-lit conditions, LMOT is very distinguishable in the feature space. However, under low-light conditions, this discrimination of LMOT decreased significantly. + +# 5.3. Results on LMOT dataset + +We compare the proposed method with potential low-light multi-object tracking methods. We first train three baseline models using low-light videos, well-lit videos, and both of them, without any modifications to baseline trackers. They are denoted by Base-low, Base-well, and Base-all. The second type of method is tracking after low-light enhancement. These methods use low-light enhancement methods as preprocessing module. Both the enhanced videos and well-lit + +
MethodByteTrack[62]OC-SORT[3]DetectionParams (M)FLOPS (G)
HOTAAssAIDF1HOTAAssAIDF1mAPAP50AP75
Base-low28.042.910.928.840.334.318.832.418.999.00793.29
Base-well13.536.832.913.831.712.05.59.75.599.00793.29
Base-all28.142.632.928.939.834.619.032.519.199.00793.29
LLFlow [48]28.842.934.029.441.035.022.438.322.6116.428755.91
SDSD [47]29.141.135.529.336.936.819.836.219.3103.301136.39
DNF [24]27.640.933.028.136.435.319.135.618.2101.83907.33
SMOID [23]28.039.733.928.736.735.619.135.518.2120.887286.17
RAOD [52]26.142.929.526.142.929.518.331.318.499.07897.04
LTrack (Ours)29.443.235.229.839.036.723.440.623.5100.43800.92
+ +![](images/4dee198d58fca91a3849bc51373e3dd58ceff2ea2c8ab5ee04299a1be024fec7.jpg) +(a) Well-lit + +![](images/440d710812ffcd28fa8131bfd71b0b2b86f5a0395b63072b7108f7acd384b1b7.jpg) +(b) Low-light +Figure 6. Visualization of appearance features of LMOT dataset using t-SNE. The same object is coded by the same color. It indicates that the appearance of objects under well-lit is distinguishable, but significantly reduced under low-light conditions + +videos are used for training. We select four different low-light enhancement techniques for comparison. LLFlow [48] and SDSD [47] are low-light enhancement methods based on RGB images, while DNF [24] and SMOID [23] are based on RAW low-light enhancement methods. LLFlow and DNF take images as input, SDSD and SMOID take videos as input. The RAW object detection method is also trained together on well-lit and low-light videos. Its output bounding boxes are directly fed to the tracker. We test all the potential methods on two state-of-the-art trackers ByteTrack [62] and OC-SORT [3]. We also show mAP, AP50, and AP75 to highlight the detector performance. + +From Tab. 5 we can see that, the proposed LTrack achieves the best HOTA and is highly competitive on all metrics. For example, the proposed method improves HOTA 1.3 with almost the same parameters and computation as base methods. This strongly proves the effectiveness of the proposed method. Compared with tracking after low-light enhancement, these methods have a large amount of additional parameters and computation but still perform worse than the proposed method. For example, LLFlow delivers 8 times FLOPS but still performs worse than our LTrack. In addition, We observed that there is + +Table 5. Experimental results on LMOT dataset. Base-low, Base-well and Base-all indicate the baseline model trained on low-light, well-lit and all of them, respectively. The number of parameters and prediction latency of each method are reported along with the accuracy + +
MethodHOTAAssAIDF1MOTADet
Base-low32.533.439.342.427.1
Base-well28.030.434.136.923.3
Base-all34.638.140.044.830.2
LLFlow[48]33.638.038.443.129.7
SDSD[47]31.634.236.738.827.5
DNF[24]33.239.036.944.430.1
SMOID[23]31.234.136.641.226.9
RAOD[52]32.233.539.440.926.5
LTrack (Ours)35.138.940.445.230.7
+ +Table 6. Experimental results on LMOT-real dataset with OC-SORT[3]. The best result are shown in boldface + +no significant difference in results between RAW-based image enhancement methods and RGB-based image enhancement methods. This is not consistent with what our method has observed. The reason may be that low light enhancement focuses on image quality restoration and may mislead downstream tasks. As for RAOD [52], it has almost the same number of parameters and computational load as our method, but its performance is much lower than the proposed LTrack. Despite addressing HDR scenes through a preprocessing module for RAW input, it does not perform well under low-light conditions. + +# 5.4. Results on Real World + +To validate the performance of our method in real-world low-light scenes at night, we evaluate all methods on LMOT-real dataset, using OC-SORT [3]. As shown in Tab. 6, the proposed LTrack performs much better than all comparison methods. Both the methods of tracking after low-light enhancement and the RAW detection method encounter generalizability problems and are not even better than Base-all. This strongly demonstrates the robustness of + +
DSLALDHOTAAssAIDF1MOTADetA
✓28.142.632.923.818.7
29.042.435.025.820.1
✓28.342.433.124.619.3
✓✓29.443.235.226.120.3
+ +Table 7. Ablation on adaptive low-pass downsampling (ALD) and degradation suppression learning (DSL). + +
Data TypeHOTAAssAIDF1MOTADetA
sRGB29.242.734.926.420.5
RAW 8-bit29.042.334.526.120.2
RAW 10-bit29.342.934.826.520.4
RAW 12-bit29.443.235.226.120.3
+ +Table 8. Results of different image format and dynamic range on LMOT test set. + +
MethodsHOTAAssAIDF1MOTADetA
Gaussian-Possion29.943.135.726.621.1
Physical-based [50]30.444.136.026.621.2
LMOT (Ours)35.138.940.445.230.7
+ +our LTrack in real-world dark scenes. + +# 5.5. Exploration and Discussion + +In this section, we conduct extensive analysis and discussion on LMOT datasets and the proposed method. + +Ablation Study. We conduct ablation experiments to validate the effectiveness of our improvements, the results are shown in Tab. 7. It can be seen that all improvements contribute effectively to enhanced performance. Among them, the degradation suppression learning strategy demonstrates the most significant effect, resulting in an approximate 1-point enhancement in HOTA. The best results are achieved when employing all strategies simultaneously. This demonstrates the effectiveness of all our contributions. + +RAW vs. sRGB. We analyze the impact of input data formats, the results are shown in Tab. 8. The 12-bit RAW format achieves significantly better results than sRGB. Because the RAW format saves much more potential information and is helpful for MOT in low-light scenes. We also observed that higher bitwidth is beneficial for performance, which has also been observed in other vision tasks [21, 52]. + +LMOT vs. Synthetic data. We compare our LMOT dataset with the synthetic low-light data to further demonstrate the value of LMOT. We apply the Gaussian-Possian based and Physical-based [50] low-light data synthesis method to synthesize low-light videos from well-lit videos. We train our LTrack on these types of data and evaluate their performance in real low-light scenes using LMOT-real. As + +Table 9. Results with low-light synthesis methods and LMOT dataset on LMOT-real set. + +
CategoriesHOTAAssAIDF1MOTADetA
Person24.230.829.925.119.2
Bicycle16.940.315.68.97.1
Car37.553.347.233.726.5
Motorcycle19.835.022.615.211.4
Bus44.359.951.836.732.9
Truck6.915.66.43.23.1
+ +Table 10. Results of the proposed method for different categories on LMOT test set. + +shown in Tab. 9, the tracker trained on our LMOT dataset has much better performance in real night scenes, which strongly demonstrates the value of our LMOT dataset. + +Analysis on different category. We also analyze the performance for different categories. From Tab. 10, we can see that cars and buses achieved the best performance because they have regular shapes and larger areas. Trucks exhibited the poorest performance, because they have the fewest instances, making it challenging for the model to learn accurate identification. The person achieves relatively average scores. Bicycles and motorcycles have close scores since they have similar appearance and motion patterns. + +# 6. Conclusion + +In this work, we investigate the multi-object tracking in the dark scenes. We build a new low-light multi-object tracking (LMOT) dataset, which provides well-aligned low-light video pairs and high-quality multi-object tracking annotations. We observed that low-light images are significantly degraded by the sensor noises, which also degrades the feature maps and significantly deteriorates the model performance. To learn the invariant semantic formation under noise disturbance and quality degradation, we present the adaptive low-pass downsample module and degradation suppression learning. These improvements greatly enhance the robustness of our method in real-world low-light scenes. + +Limitations. We focus on multi-object tracking in the dark scenes. But we do not consider other degradation environments in the real world, such as rainy and foggy days. In our future work, we will consider exploring multi-object tracking in more real-world scenarios, promoting the development of MOT in real-world applications. + +# Acknowledgments + +This work was supported by the National Key R&D Program of China (2022YFC3300704), the National Natural Science Foundation of China (62331006,62171038, and 62088101), the R&D Program of Beijing Municipal Education Commission (KZ202211417048), and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, pages 3464-3468, 2016. 3 +[2] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. arXiv preprint arXiv:2303.06705, 2023. 1, 2 +[3] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. Observation-centric sort: Rethinking sort for robust multi-object tracking. In CVPR, pages 9686-9696, 2023. 1, 3, 6, 7 +[4] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In CVPR, pages 3291-3300, 2018. 2 +[5] Chen Chen, Qifeng Chen, Minh N Do, and Vladlen Koltun. Seeing motion in the dark. In ICCV, pages 3185-3194, 2019. +[6] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. IJCV, pages 1-21, 2023. 1, 2 +[7] Wongun Choi and Silvio Savarese. A unified framework for multi-target tracking and collective activity recognition. In ECCV, pages 215-230, 2012. 1 +[8] Yutao Cui, Chenkai Zeng, Xiaoyu Zhao, Yichun Yang, Gangshan Wu, and Limin Wang. Sportsmot: A large multi-object tracking dataset in multiple sports scenes. arXiv preprint arXiv:2304.05170, 2023. 1, 2, 3, 5, 6 +[9] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1, 2, 3 +[10] Yunhao Du, Zhicheng Zhao, Yang Song, Yanyun Zhao, Fei Su, Tao Gong, and Hongying Meng. *Strongsort: Make deep-sort great again*. IEEE Transactions on Multimedia, 2023. 1 +[11] Anna Ellis and James Ferryman. Pets2010 and pets2009 evaluation of results using individual ground truthed single views. In 2010 7th IEEE international conference on advanced video and signal based surveillance, pages 135-142, 2010. 1 +[12] Wang Zhongdao et al. Do different tracking tasks require different appearance models? 2021. 4 +[13] Huiyuan Fu, Wenkai Zheng, Xicong Wang, Jiaxuan Wang, Heng Zhang, and Huadong Ma. Dancing in the dark: A benchmark towards general low-light video enhancement. In ICCV, pages 12877-12886, 2023. 2 +[14] Ying Fu, Jian Chen, Tao Zhang, and Yonggang Lin. Residual scale attention network for arbitrary scale image superresolution. Neurocomputing, 427:201-211, 2021. 2 +[15] Ying Fu, Yang Hong, Linwei Chen, and Shaodi You. Le-gan: Unsupervised low-light image enhancement network using attention module and identity invariant loss. Knowledge-Based Systems, 240:108010, 2022. 2 +[16] Ruopeng Gao and Limin Wang. Memotr: Long-term memory-augmented transformer for multi-object tracking. In ICCV, pages 9901-9910, 2023. 3 + +[17] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6 +[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361, 2012. 1, 2, 3, 4 +[19] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In CVPR, pages 1780-1789, 2020. 2 +[20] Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE TIP, 26(2):982-993, 2016. 2 +[21] Yang Hong, Kaixuan Wei, Linwei Chen, and Ying Fu. Crafting object detection in very low light. In BMVC, page 3, 2021. 1, 2, 8 +[22] Haidi Ibrahim and Nicholas Sia Pik Kong. Brightness preserving dynamic histogram equalization for image contrast enhancement. IEEE Transactions on Consumer Electronics, 53(4):1752-1758, 2007. 2 +[23] Haiyang Jiang and Yinqiang Zheng. Learning to see moving objects in the dark. In ICCV, pages 7324-7333, 2019. 1, 2, 3, 7 +[24] Xin Jin, Ling-Hao Han, Zhen Li, Chun-Le Guo, Zhi Chai, and Chongyi Li. Dnf: Decouple and feedback network for seeing in the dark. In CVPR, pages 18135-18144, 2023. 1, 2, 7 +[25] Daniel J Jobson, Zia-ur Rahman, and Glenn A Woodell. Properties and performance of a center/surround retina. IEEE TIP, 6(3):451-462, 1997. 2, 4 +[26] Laura Leal-Taixe, Anton Milan, Ian Reid, Stefan Roth, and Konrad Schindler. Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv preprint arXiv:1504.01942, 2015. 2 +[27] Sohyun Lee, Jaesung Rim, Boseung Jeong, Geonu Kim, Byungju Woo, Haechan Lee, Sunghyun Cho, and Suha Kwak. Human pose estimation in extremely low-light conditions. In CVPR, pages 704-714, 2023. 1, 2 +[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6 +[29] Qiankun Liu, Qi Chu, Bin Liu, and Nenghai Yu. Gsm: Graph similarity model for multi-object tracking. In IJCAI, pages 530-536, 2020. 1 +[30] Qiankun Liu, Dongdong Chen, Qi Chu, Lu Yuan, Bin Liu, Lei Zhang, and Nenghai Yu. Online multi-object tracking with unsupervised re-identification learning and occlusion estimation. Neurocomputing, 483:333–347, 2022. 1 +[31] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. IJCV, 129:548-578, 2021. 5 +[32] Kang Ma, Ying Fu, Dezhi Zheng, Chunshui Cao, Xuecai Hu, and Yongzhen Huang. Dynamic aggregated network for gait recognition. In CVPR, pages 22076-22085, 2023. 1 + +[33] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, pages 8844-8854, 2022. 3 +[34] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831, 2016. 1, 2, 3, 4 +[35] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In CVPR, pages 1765–1773, 2017. 2 +[36] Jihun Park, Jinseok Hong, Woolil Shim, and Dae-Jin Jung. Multi-object tracking on swir images for city surveillance in an edge-computing environment. Sensors, 23(14):6373, 2023. 3 +[37] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In ECCV, pages 17-35, 2016. 6 +[38] Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena, 60(1-4):259-268, 1992. 5 +[39] Aashish Sharma and Robby T Tan. Nighttime visibility enhancement by increasing the dynamic range and suppression of light effects. In CVPR, pages 11977-11986, 2021. 2 +[40] Shuran Song and Jianxiong Xiao. Tracking revisited using rgbd camera: Unified benchmark and baselines. In ICCV, pages 233-240, 2013. 3 +[41] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 3 +[42] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In CVPR, pages 20993-21002, 2022. 1, 2, 3, 4, 5 +[43] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 2 +[44] Ye Tian, Ying Fu, and Jun Zhang. Plug-and-play algorithm for under-sampling fourier single-pixel imaging. Science China. Information Sciences, 65(10):209303, 2022. 2 +[45] Ye Tian, Ying Fu, and Jun Zhang. Local-enhanced transformer for single-pixel imaging. Optics Letters, 48(10): 2635-2638, 2023. 2 +[46] Chaoqun Wang, Chunyan Xu, Zhen Cui, Ling Zhou, Tong Zhang, Xiaoya Zhang, and Jian Yang. Cross-modal patternpropagation for rgb-t tracking. In CVPR, pages 7064-7073, 2020. 3 +[47] Ruixing Wang, Xiaogang Xu, Chi-Wing Fu, Jiangbo Lu, Bei Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In ICCV, pages 9700-9709, 2021. 1, 2, 7 +[48] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In AAAI, pages 2604-2612, 2022. 7 + +[49] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 2 +[50] Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. A physics-based noise formation model for extreme low-light raw denoising. In CVPR, pages 2758-2767, 2020. 4, 6, 8 +[51] Kaixuan Wei, Angelica Aviles-Rivero, Jingwei Liang, Ying Fu, Hua Huang, and Carola-Bibiane Schonlieb. Tfnpn: Tuning-free plug-and-play proximal algorithms with applications to inverse imaging problems. Journal of Machine Learning Research, 23(16):1-48, 2022. 2 +[52] Ruikang Xu, Chang Chen, Jingyang Peng, Cheng Li, Yibin Huang, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Toward raw object detection: A new benchmark and a new model. In CVPR, pages 13384-13393, 2023. 2, 7, 8 +[53] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. In CVPR, pages 17714-17724, 2022. 2 +[54] Jinyu Yang, Shang Gao, Zhe Li, Feng Zheng, and Ales Leonardis. Resource-efficient rgbd aerial tracking. In CVPR, pages 13374-13383, 2023. 3 +[55] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2636-2645, 2020. 2, 3, 4 +[56] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In ECCV, pages 659–675, 2022. 3 +[57] Fan Zhang, Yu Li, Shaodi You, and Ying Fu. Learning temporal consistency for low light video enhancement from single images. In CVPR, pages 4967-4976, 2021. 2 +[58] Jiqing Zhang, Xin Yang, Yingkai Fu, Xiaopeng Wei, Baocai Yin, and Bo Dong. Object tracking by jointly exploiting frame and event domain. In ICCV, pages 13043-13052, 2021. 3 +[59] Pengyu Zhang, Jie Zhao, Dong Wang, Huchuan Lu, and Xiang Ruan. Visible-thermal uav tracking: A large-scale benchmark and new baseline. In CVPR, pages 8886–8895, 2022. 3 +[60] Tao Zhang, Ying Fu, and Cheng Li. Hyperspectral image denoising with realistic data. In ICCV, pages 2248-2257, 2021. 2 +[61] Tao Zhang, Ying Fu, and Cheng Li. Deep spatial adaptive network for real image demosaicing. In AAAI, pages 3326-3334, 2022. 2 +[62] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, pages 1-21, 2022. 1, 3, 6, 7 +[63] Yang Zhang, Tiancai Wang, and Xiangyu Zhang. Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. In CVPR, pages 22056-22065, 2023. 3 +[64] Zhiyu Zhu, Junhui Hou, and Dapeng Oliver Wu. Cross-modal orthogonal high-rank augmentation for rgb-event + +transformer-trackers. In ICCV, pages 22045-22055, 2023. 3 +[65] Yunhao Zou and Ying Fu. Estimating fine-grained noise model via contrastive learning. In CVPR, pages 12682-12691, 2022. 2 +[66] Yunhao Zou, Yinqiang Zheng, Tsuyoshi Takatani, and Ying Fu. Learning to reconstruct high speed and high dynamic range videos from events. In CVPR, pages 2024-2033, 2021. 2 +[67] Yunhao Zou, Chenggang Yan, and Ying Fu. Rawhdr: High dynamic range image reconstruction from a single raw image. In ICCV, pages 12334-12344, 2023. 2 \ No newline at end of file diff --git a/2024/Multi-Object Tracking in the Dark/images.zip b/2024/Multi-Object Tracking in the Dark/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ccb4ba2fda683817ec68c9571a852c80e4fc94fa --- /dev/null +++ b/2024/Multi-Object Tracking in the Dark/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d32e4d8c90ce20d186df124e4ecba7695a908ab45013c4bac4e29b41e74370e0 +size 694800 diff --git a/2024/Multi-Object Tracking in the Dark/layout.json b/2024/Multi-Object Tracking in the Dark/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..67e946a7af2785096af0d6af3158b5d48a44ca27 --- /dev/null +++ b/2024/Multi-Object Tracking in the Dark/layout.json @@ -0,0 +1,8058 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 188, + 103, + 405, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 103, + 405, + 121 + ], + "spans": [ + { + "bbox": [ + 188, + 103, + 405, + 121 + ], + "type": "text", + "content": "Multi-Object Tracking in the Dark" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 143, + 468, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 143, + 468, + 172 + ], + "spans": [ + { + "bbox": [ + 129, + 143, + 468, + 172 + ], + "type": "text", + "content": "Xinzhe Wang Kang Ma Qiankun Liu Yunhao Zou Ying Fu * Beijing Institute of Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 174, + 454, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 174, + 454, + 186 + ], + "spans": [ + { + "bbox": [ + 140, + 174, + 454, + 186 + ], + "type": "text", + "content": "{wangxinzhe, makang, liuqk3, zouyunhao, fuying}@bit.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "type": "text", + "content": "Low-light scenes are prevalent in real-world applications (e.g. autonomous driving and surveillance at night). Recently, multi-object tracking in various practical use cases have received much attention, but multi-object tracking in dark scenes is rarely considered. In this paper, we focus on multi-object tracking in dark scenes. To address the lack of datasets, we first build a Low-light Multi-Object Tracking (LMOT) dataset. LMOT provides well-aligned low-light video pairs captured by our dual-camera system, and high-quality multi-object tracking annotations for all videos. Then, we propose a low-light multi-object tracking method, termed as LTrack. We introduce the adaptive low-pass downsample module to enhance low-frequency components of images outside the sensor noises. The degradation suppression learning strategy enables the model to learn invariant information under noise disturbance and image quality degradation. These components improve the robustness of multi-object tracking in dark scenes. We conducted a comprehensive analysis of our LMOT dataset and proposed LTrack. Experimental results demonstrate the superiority of the proposed method and its competitiveness in real night low-light scenes. Dataset and Code: https://github.com/ying-fu/LMOT" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 559, + 288, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 288, + 679 + ], + "type": "text", + "content": "Multi-object tracking (MOT) aims to locate and associate multiple objects in video sequences. It is widely used in many downstream applications, such as video recognition [7, 32], autonomous driving [18], and surveillance [11]. Recently, multi-object tracking in various practical use cases has garnered much attention [8, 9, 18, 34, 42], greatly advancing the development of MOT. However, these works are primarily tailored for high-quality inputs and overlook the prevalent low-light scenes in real-world scenes. Motivated by this, we study multi-object tracking in dark scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 681, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 681, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 59, + 681, + 287, + 693 + ], + "type": "text", + "content": "Due to the physical limitations of existing cameras, ac" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 213, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 538 + ], + "type": "text", + "content": "quiring high-quality videos under low-light conditions is difficult. One inherent difficulty in capturing consecutive video frames under such conditions is avoiding motion blur. Current camera technology typically requires short exposure times (usually just a few tens of milliseconds), but in low-light scenarios, the sensor struggles to capture an adequate number of photons within this limited duration. This limitation inevitably leads to degradation in image quality accompanied by higher noise levels. This presents two main challenges for MOT under low-light conditions. The first challenge is for collecting a low-light multi-object tracking dataset. Collecting and annotating a low-light MOT dataset is difficult and expensive. MOT requires dynamic object videos, but videos captured in low-light scenes have extremely low brightness, making it hard to recognize and annotate objects in the videos. The second challenge revolves around low-light multi-object tracking. The popular tracking-by-detection paradigm [3, 10, 29, 30, 62] generally consists of detector, motion-based association module, and appearance-based association module. These modules typically require high-quality input images. The poor quality of low-light images leads to severe performance degradation for both detectors and appearance-based correlation modules. A simple approach is to cascade low-light enhance modules [2, 23, 24, 47], but this introduces additional computational costs. Furthermore, images optimized for visual quality may be suboptimal for downstream tasks [6, 21, 27]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we build a low-light multi-object tracking dataset (LMOT), specifically designed to address the challenges of multi-object tracking in dark scenes. To this end, we develop a dual-camera system that simultaneously captures well-lit and low-light video frames. The video pairs are highly aligned in both spatial and temporal dimensions, offering two key benefits. First, it enables us to annotate on the well-lit videos, resulting in high-quality annotations. Second, the well-lit videos can provide additional supervision information during the training phase, and strongly enhance the performance in the dark scenes. After careful annotation, we collect 32 video sequences " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "(2.3\\times \\mathrm{MOT17})" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "35\\mathrm{K}" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "(3.1\\times \\mathrm{MOT17})" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": " ,and over " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "815\\mathrm{K}" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": " bounding boxes " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "(2.8\\times \\mathrm{MOT17})" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": " .The RAW data is the output of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "382" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "the image sensor and is the input data of the image signal processor (ISP). It saves all information from the image sensor, which is crucial for capturing object information in dark scenes [6, 52, 67]. Therefore, we collect RAW videos for LMOT." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 289, + 360 + ], + "type": "text", + "content": "Additionally, we propose a low-light multi-object tracking method, termed as LTrack. The low-light video is characterized by substantial sensor noise and poor image quality, which significantly degrades both shallow and deep feature representations, leading to reduced tracking performance. We observe that the sensor noise in low-light images exhibits a similarity to adversarial attacks [35, 43]. To address this issue, our main idea is to learn the invariant semantic information under noise disturbance quality degradation. We present the adaptive low-pass downsample module (ALD). It employs spatial low-pass convolution to extract low-frequency components from images, excluding noises, and adaptively enhance the feature maps. We also present the degradation suppression learning strategy (DSL), which utilizes paired low-light videos to help the model suppress the noise disturbance and encourage image content response in the feature domain. We conduct a comprehensive analysis of our LMOT dataset and validate the superiority of our LTrack in real-world night scenes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 361, + 266, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 361, + 266, + 373 + ], + "spans": [ + { + "bbox": [ + 59, + 361, + 266, + 373 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 376, + 287, + 519 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 376, + 287, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 423 + ], + "type": "text", + "content": "- We build the first low-light multi-object tracking dataset using a carefully constructed dual-camera system. It provides well-aligned low-light videos in RAW format, and high-quality MOT annotations for all videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 424, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 424, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 287, + 471 + ], + "type": "text", + "content": "- We propose a low-light multi-object tracking method. It utilizes the adaptive low-pass downsample module and the degradation suppression learning strategy to learn to extract invariant features from low-light videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 472, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 287, + 519 + ], + "type": "text", + "content": "- We conduct a comprehensive analysis of our dataset and the proposed method. Experimental results demonstrate the superiority of the proposed method and its competitiveness on real night scenes." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 532, + 134, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 134, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 134, + 544 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 552, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 601 + ], + "type": "text", + "content": "In this section, we first review the current research status of low-light enhancement and low-light datasets. Then, we summarized the present research for multi-object tracking and tracking in the dark scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Low-light enhancement. Traditional methods for lowlight enhancement are primarily based on histogram equalization and Retinex theory [20, 22, 25]. Recently, deep learning has been explored for many low-level tasks [14, 44, 45, 51, 60, 61, 65, 66], and achieved superior results on low-light enhancement [2, 4, 13, 15, 19, 24, 39, 53]. While these methods are capable of recovering images with high visual quality, they often require heavy computation and may not consider downstream tasks, resulting in suboptimal" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 71, + 545, + 222 + ], + "blocks": [ + { + "bbox": [ + 308, + 71, + 545, + 222 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 545, + 222 + ], + "type": "image", + "image_path": "02e02ea9df4d388f77874d5066d6b2a59c03d1a20f1308b2772c31d5a2c07a46.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 223, + 546, + 267 + ], + "lines": [ + { + "bbox": [ + 305, + 223, + 546, + 267 + ], + "spans": [ + { + "bbox": [ + 305, + 223, + 546, + 267 + ], + "type": "text", + "content": "Figure 1. Our dual-camera system. It consists of two cameras, a beam splitter, and an ND-filter. Two cameras of identical models are meticulously engineered to achieve pixel-by-pixel alignment in the captured video data." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 281, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 319 + ], + "type": "text", + "content": "performance. In contrast, our approach focuses on directly learning multi-object tracking from low-light images, thus bypassing the low-light enhancement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 324, + 546, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 546, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 546, + 552 + ], + "type": "text", + "content": "Low-light datasets. The long-short exposure is a widely used method to collect paired low-light images, but can only be used to collect low-light images for static scenes. [4, 5, 49]. To capture dynamic scenes, some works designed the mechatronic system. They obtain paired low-light data by repeating the motion twice [13, 47]. However, these mechatronic systems cannot be used to collect dynamic object videos in the wild. Jiang et al. [23] designed a dual-camera system that simultaneously captures paired well-lit and low-light videos, making it possible to capture dynamic scenes and dynamic object videos for multi-object tracking. Zou et al. [57] setup an optical system to collect paired videos and event streams. These works explore various ways to construct low-light datasets and inspire research on high-level vision tasks in dark scenes, such as LOD [21] for object detection, LIS [6] for instance segmentation, and ExPose [27] for human pose estimation. These datasets provide paired low-light data only for image tasks, and cannot be expanded for multi-object tracking in dark scenes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 557, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 547, + 713 + ], + "type": "text", + "content": "Multi-object tracking datasets. MOT15 [26] is the first large-scale benchmark for multi-object tracking. MOT17 [34] stands as one of the most widely applied MOT benchmarks. MOT20 [9] focuses on very crowded scenes. These three datasets are for pedestrians. KITTI [18] and BDD100K [55] are for autonomous driving scenarios. DanceTrack [42] focuses on dancing scenes and is characterized by similar appearance and diverse motions. Recently, SportsMOT [8] aims to track athletes and encourage algorithms to promote both appearance and motion association. These datasets explore multi-object tracking in various practical use cases, but none of them consider multi-object tracking in dark scenes." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "383" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 299, + 202 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 299, + 202 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 299, + 202 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 299, + 202 + ], + "type": "image", + "image_path": "f14b87b761c53707cda4a4222aa69192accfca9747e49a4a8a5fec6a6e51d76d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 546, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 546, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 546, + 236 + ], + "type": "text", + "content": "Figure 2. Two example videos from our LMOT dataset. It provides well-aligned low-light video pairs and MOT annotations for all videos. The time interval between adjacent frames is 1s. The first row is the low-light video, the second row is the scaled low-light video and the last row is the well-lit video. Our LMOT dataset is collected from city outdoor scenes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 70, + 545, + 202 + ], + "blocks": [ + { + "bbox": [ + 301, + 70, + 545, + 202 + ], + "lines": [ + { + "bbox": [ + 301, + 70, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 301, + 70, + 545, + 202 + ], + "type": "image", + "image_path": "046c4b39280bea23ad40023227baac67a1b0c29921c77ad36239dcbb902d9f77.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 245, + 285, + 342 + ], + "blocks": [ + { + "bbox": [ + 50, + 245, + 285, + 342 + ], + "lines": [ + { + "bbox": [ + 50, + 245, + 285, + 342 + ], + "spans": [ + { + "bbox": [ + 50, + 245, + 285, + 342 + ], + "type": "table", + "html": "
DatasetFormatVideosFramesLength (s)BboxTracks
MOT17 [34]sRGB1411,235463292,7331,342
MOT20 [9]sRGB813,4105351,652,0403,456
DanceTrack [42]sRGB100105,8555,292-990
SportMOT [8]sRGB240150,3796,0151,629,4903,401
KITTI [18]sRGB218,000-47,000917
BDD100K [55]sRGB1,600318,000-3,300,000131,000
SWIR [36]SWIR-7,309-57,221-
LMOTRAW3235,1201,756815,5504,090
", + "image_path": "75f0c68c13df678050980d1ed59493e89d533c0d802e3dd8ebd401300af0c8c5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 379, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 498 + ], + "type": "text", + "content": "Object tracking in the dark. To track in low-light scenes, some methods explore to use multi-modal information for single-object tracking, such as event camera [58, 64], depth [40, 54] and thermal [46, 59] devices. Park et al. [36] proposed to use Short-Wave Infrared (SWIR) images for multi-object tracking, since its advantages in terms of robustness in low-light conditions. The common drawback of these methods is that they require additional hardware equipment and cannot be applied to the most widely used CMOS imaging systems." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 499, + 287, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 631 + ], + "type": "text", + "content": "SORT [1] uses the Kalman Filter motion model and employs IoU for association. ByteTrack [62] enhances tracking performance by considering low-confidence bounding boxes. OS-SORT [3] enhances SORT by restoring lost targets. Recently, Transformer has been explored for MOT [16, 33, 41, 56, 63]. These methods have achieved high performance in many practical scenarios, but they do not consider dealing with low-light conditions. We focus on multi-object tracking under low-light conditions. Based on RAW videos, our method is highly practical with excellent performance and does not require additional hardware." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 644, + 270, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 644, + 270, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 644, + 270, + 658 + ], + "type": "text", + "content": "3. Low-light Multi-object Tracking Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": "In this section, we first introduce our dual camera system and the details of collecting and annotating our low-light multi-object tracking (LMOT) dataset. Then, we analyzed the statistical characteristics of our LMOT dataset." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 245, + 544, + 308 + ], + "blocks": [ + { + "bbox": [ + 47, + 343, + 287, + 365 + ], + "lines": [ + { + "bbox": [ + 47, + 343, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 287, + 365 + ], + "type": "text", + "content": "Table 1. Comparison of statistics between existing MOT datasets and our LMOT dataset" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 245, + 544, + 308 + ], + "lines": [ + { + "bbox": [ + 309, + 245, + 544, + 308 + ], + "spans": [ + { + "bbox": [ + 309, + 245, + 544, + 308 + ], + "type": "table", + "html": "
DatasetSplitVideosBboxTracksPaired Well-lit
LMOT-dualtrain11309,4661533✓
val4131,781626✓
test11312,7421644✓
LMOT-realreal661,561287
", + "image_path": "1a40ca74c89a1ea3f9aa582941c77d145e1a9b36811c8dd1bb7e031822bbd6c6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 328, + 309, + 522, + 319 + ], + "lines": [ + { + "bbox": [ + 328, + 309, + 522, + 319 + ], + "spans": [ + { + "bbox": [ + 328, + 309, + 522, + 319 + ], + "type": "text", + "content": "Table 2. Detailed statistics and data splits for LMOT." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 334, + 427, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 334, + 427, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 427, + 346 + ], + "type": "text", + "content": "3.1. Dataset Construction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 354, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 593 + ], + "type": "text", + "content": "Multi-object tracking requires dynamic scenes and object video. To collect low-light videos for multi-object tracking, we build a dual-camera system [23], which is illustrated in Fig. 1. It can simultaneously capture paired lowlight and well-lit video pairs. Its main components include a beam splitter, a neutral density (ND) filter, and two FLIR Grasshopper3 GS3-U3-23S6C cameras. The beam splitter divides the incoming light into two separate paths. This arrangement allows one camera to capture well-lit images directly, while the other camera records low-light images, with the ND-filter attenuating the light intensity. To ensure temporal synchronization of the video frames, we employ the hardware interface to trigger the camera exposure events. Moreover, to avoid frame loss, our dual-camera system uses two independent hardware interfaces for data transmission and is equipped with a high-speed solid-state drive. Thanks to precise calibration, our dual-camera system can capture paired low-light and well-lit videos in real time for dynamic scenes and objects. More details about our dual-camera system are given in supplementary materials." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "We save the video frames in RAW format before the images are processed by the camera image signal processor (ISP). In terms of camera settings, we set the exposure time for both cameras to " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "10ms" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ", and the frame rate is fixed at 20. This setting is feasible to avoid motion blur. We adjust the gain level for well-lit cameras to achieve optimal image quality. The gain for low-light cameras is consistently set to the maximum value, to simulate low-light capturing setup in real scenarios. We also collect a real low-light MOT dataset (LMOT-real) to evaluate performance in real-world" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "384" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 72, + 227, + 153 + ], + "blocks": [ + { + "bbox": [ + 53, + 72, + 227, + 153 + ], + "lines": [ + { + "bbox": [ + 53, + 72, + 227, + 153 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 227, + 153 + ], + "type": "image", + "image_path": "4c7dae4c6446524a305bb96a7b1e8b7a9ac9a898214ea69671b1262768a5cec9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 158, + 202, + 168 + ], + "lines": [ + { + "bbox": [ + 80, + 158, + 202, + 168 + ], + "spans": [ + { + "bbox": [ + 80, + 158, + 202, + 168 + ], + "type": "text", + "content": "(a) Number of instances per category" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 233, + 72, + 372, + 153 + ], + "blocks": [ + { + "bbox": [ + 233, + 72, + 372, + 153 + ], + "lines": [ + { + "bbox": [ + 233, + 72, + 372, + 153 + ], + "spans": [ + { + "bbox": [ + 233, + 72, + 372, + 153 + ], + "type": "image", + "image_path": "1d3709a5a938d9b2e9f9b3d3e8c00ff8410f251cd6fd2bb651690ed41815680a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 158, + 347, + 168 + ], + "lines": [ + { + "bbox": [ + 258, + 158, + 347, + 168 + ], + "spans": [ + { + "bbox": [ + 258, + 158, + 347, + 168 + ], + "type": "text", + "content": "(b) IoU on adjacent frames" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 170, + 547, + 214 + ], + "lines": [ + { + "bbox": [ + 46, + 170, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 547, + 214 + ], + "type": "text", + "content": "Figure 3. (a) Number of instances per category. LMOT consists of 6 categories, most of the instances are the person and car. (b) IoU on adjacent frames. Compared to MOT17, KITTI, and DanceTrack, LMOT has a roughly average score. This indicates that LMOT has a relatively normal movement speed. (c) Cosine distance of appearance features. The cosine distance is smaller under low-light conditions, indicating that the appearance distinguishability is decreased under low-light conditions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 377, + 72, + 541, + 153 + ], + "blocks": [ + { + "bbox": [ + 377, + 72, + 541, + 153 + ], + "lines": [ + { + "bbox": [ + 377, + 72, + 541, + 153 + ], + "spans": [ + { + "bbox": [ + 377, + 72, + 541, + 153 + ], + "type": "image", + "image_path": "10f3feb2bb204853dce35c9fa173111cd49802ff040fddc688a9499a8c5eb335.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 158, + 527, + 168 + ], + "lines": [ + { + "bbox": [ + 389, + 158, + 527, + 168 + ], + "spans": [ + { + "bbox": [ + 389, + 158, + 527, + 168 + ], + "type": "text", + "content": "(c) Cosine distance of appearance features" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 224, + 286, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 286, + 249 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 286, + 249 + ], + "type": "text", + "content": "dark scenes. These videos are captured using a single camera with the same camera settings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 250, + 286, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 286, + 345 + ], + "type": "text", + "content": "Our LMOT dataset contains a variety of city outdoor scenes, including roads, overpasses, pedestrians, and intersection. The overpass scenes take an overhead shot of objects, while all other scenes are captured from the perspective of pedestrians. To account for the impact of camera motion, we introduce arbitrary horizontal rotations and vertical random movements to the camera. Fig. 2 shows two sampled video sequences from our LMOT dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 347, + 286, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 286, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 286, + 477 + ], + "type": "text", + "content": "We annotate six types of moving objects, including car, person, bicycle, motorcycle, bus, and truck. The annotated labels include bounding boxes, identifications, and visibility status. For partly occluded objects, a full box is annotated. For a fully occluded object, an estimated box is annotated. Each object has a unique ID throughout the entire video. Thanks to well-aligned low-light and well-lit videos, we can annotate well-lit videos to simultaneously obtain labels for low-light videos. This greatly reduces the annotation difficulty and enhances quality. Lastly, we carefully review all the annotation results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 489, + 145, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 145, + 500 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 145, + 500 + ], + "type": "text", + "content": "3.2. Dataset Statistic" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 509, + 286, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 286, + 651 + ], + "type": "text", + "content": "LMOT is a large-scale dataset that focuses on multi-object tracking in dark scenes. We compare the statistics of LMOT with existing MOT datasets in Tab. 1. It can be seen that LMOT is approximately three times larger than MOT17 [34]. Compared to large-scale datasets such as DanceTrack [42], SportsMOT [25], KITTI [18] and BDD100K [55], the scale of LMOT is still considerable. It should be emphasized that these datasets are not for multi-object tracking in dark scenes and only provide sRGB images. Compared to SWIR, our LMOT dataset has approximately " + }, + { + "bbox": [ + 46, + 509, + 286, + 651 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 46, + 509, + 286, + 651 + ], + "type": "text", + "content": " frames and " + }, + { + "bbox": [ + 46, + 509, + 286, + 651 + ], + "type": "inline_equation", + "content": "14 \\times" + }, + { + "bbox": [ + 46, + 509, + 286, + 651 + ], + "type": "text", + "content": " bounding boxes. The detailed statistics and data splits for LMOT are shown in Tab. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "We present the number of instances for each category in Fig. 3 (a). The majority of instances are persons and cars. As shown in Fig. 3 (b), the average IoU on adjacent frames of LMOT is lower than MOT17 and DanceTrack, but higher than KITTI. This indicates that the motions in LMOT are" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 224, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 308 + ], + "type": "text", + "content": "fast but within normal range. Following [42], we use cosine distance of appearance features " + }, + { + "bbox": [ + 304, + 224, + 545, + 308 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 304, + 224, + 545, + 308 + ], + "type": "text", + "content": " to evaluate the appearance similarity. From Fig. 3 (c), we can see that the cosine distance of appearance features under low-light conditions is smaller than that under well-lit conditions. In other words, the appearance of objects will deteriorate under low-light conditions, making them harder to distinguish." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 321, + 487, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 321, + 487, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 321, + 487, + 335 + ], + "type": "text", + "content": "4. Low-light Multi-object Tracking" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 342, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 389 + ], + "type": "text", + "content": "In this section, we peresnt our low-light multi-object tracking method (LTrack). Our main idea is to learn the invariant semantic information under noise disturbance quality degradation. The overall framework can be seen in Fig. 4." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 399, + 462, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 462, + 410 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 462, + 410 + ], + "type": "text", + "content": "4.1. Formulation and Motivation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 418, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 596 + ], + "type": "text", + "content": "In low-light scenes, the camera can capture only a small number of photons in a single exposure. Thus, the potential sensor noise is highlighted, resulting in significant degradation to the image quality [50]. We observed that directly feeding low-light images to the network without any special design leads to the feature map degradation and significantly reduces the performance of the model (see in Sec. 5.2). A direct solution is to apply low-light enhancement techniques, which focus on learning a mapping function from low-light images to clean well-lit images. Since it is a highly ill-posed problem, learning such a mapping function requires considerable computing and storage overhead. Although DNN-based methods have achieved excellent performance in low-light enhancement, images enhanced for visual quality may be suboptimal for downstream tasks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 597, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 545, + 693 + ], + "type": "text", + "content": "In this work, we perform multi-object tracking from low-light images, bypassing the low-light enhancement. Leveraging RAW videos, the network obtains more original scene information compared to sRGB. To enhance the performance and robustness of the multi-object tracking model, our main idea is to earn the invariant semantic information under noise disturbance quality degradation. Thus, we present the adaptive low-pass downsampling (ALD) module" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 492, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 492, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 492, + 712 + ], + "type": "text", + "content": "1We use UniTrack [12] to extract appearance features." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "385" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 547, + 235 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 235 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 235 + ], + "type": "image", + "image_path": "a123ddd19e490d6e5311f513268160ef5309ce2988220fbd0c730ce7949900fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 238, + 546, + 261 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 546, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 546, + 261 + ], + "type": "text", + "content": "Figure 4. The overall framework of the proposed low-light multi-object tracking method, termed as LTrack. It employs adaptive low-pass downsample module and degradation suppression learning strategy, enabling the model to learn invariant features from low-light videos." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 271, + 288, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 288, + 342 + ], + "type": "text", + "content": "to enhance the low-frequency components of images and filter out high-frequency noise. We also propose the degradation suppression learning strategy (DSL), which utilizes paired low-light videos to help the model suppress image noise disturbance and encourage image content response in the feature domain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 350, + 232, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 232, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 232, + 363 + ], + "type": "text", + "content": "4.2. Adaptive Low-pass Downsampling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 368, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 287, + 487 + ], + "type": "text", + "content": "The downsampling operation reduces the feature size while preserving the most important information. The noise in low-light images introduces high-frequency disturbance to the feature maps, which can mislead the preservation of object information. To weaken the impact of high-frequency noises and enhance the low-frequency part of the feature map, we introduce spatial low-pass convolution (SConv) to extract low-frequency features from the noised feature maps. The softmax function is used to constrain the original convolution kernel to be low-pass as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 490, + 287, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 490, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 116, + 490, + 287, + 517 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {K}} _ {i, j} = \\frac {\\exp \\left(\\mathcal {K} _ {i , j}\\right)}{\\sum_ {p , q} \\exp \\left(\\mathcal {K} _ {p , q}\\right)} \\tag {1}", + "image_path": "647b8b4b8e7a48bd38e37584cd941697fbb3736901893031b9c32ca85dde2091.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 520, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 520, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 46, + 520, + 287, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 520, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{K}}" + }, + { + "bbox": [ + 46, + 520, + 287, + 640 + ], + "type": "text", + "content": " are the origin convolution kernel and low-pass convolution kernel. We initialize the spatial low-pass convolution kernel using a standard Gaussian kernel. The kernel size is set to 5 to obtain more spatial information. Then, the obtained low-frequency features are adaptively weighted and fused into the original features. We use global average pooling to obtain channel descriptors for both origin features and low-frequency features. These descriptors are then fed into a fully connected layer to compute the weight values." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 234, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 234, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 234, + 661 + ], + "type": "text", + "content": "4.3. Degradation Suppression Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "Given that a low-light image and well-lit image pair share the same content, the model should exhibit identical feature responses to them. However, the low-light image results in shallow features full of noise, and the deep feature exhibits" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 271, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 545, + 331 + ], + "type": "text", + "content": "lower responses to objects (as shown in Fig. 5). To address this, our idea is to suppress image noise in shallow features and use well-lit images to help model learning disturbance invariant information from low-light images. The degradation suppression loss can be expressed as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 364, + 332, + 545, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 332, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 364, + 332, + 545, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D S} = \\sum_ {l} \\left\\| \\mathbf {F} _ {l} ^ {\\text {w e l l}} - \\mathbf {F} _ {l} ^ {\\text {l o w}} \\right\\| _ {2} ^ {2} \\tag {2}", + "image_path": "643832ee85e113d08706fd033277b39937f8675184be6ff77e0b26d57c305b71.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 360, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 360, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 305, + 360, + 545, + 408 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 360, + 545, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l^{well}" + }, + { + "bbox": [ + 305, + 360, + 545, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 360, + 545, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l^{low}" + }, + { + "bbox": [ + 305, + 360, + 545, + 408 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 305, + 360, + 545, + 408 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 360, + 545, + 408 + ], + "type": "text", + "content": "-th feature map corresponding to well-lit and low-light image, respectively. To further suppress the noise in shallow features, we also introduce the Total Variation (TV) loss [38] to features as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 411, + 545, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 411, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 335, + 411, + 545, + 435 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T V} = \\sum_ {l} | | \\mathbf {G} ^ {r o w} \\mathbf {F} _ {l} ^ {l o w} | | _ {2} ^ {2} + | | \\mathbf {G} ^ {c o l} \\mathbf {F} _ {l} ^ {l o w} | | _ {2} ^ {2} \\tag {3}", + "image_path": "1e41c5678fdbee70168d3788a8069241fb8f72e0e4356bfa703902bf81530d0b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 437, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 437, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^{row}" + }, + { + "bbox": [ + 305, + 437, + 545, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 437, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^{col}" + }, + { + "bbox": [ + 305, + 437, + 545, + 510 + ], + "type": "text", + "content": " are the first derivative matrix to role and column. The TV loss adds spatial smoothing constraints to features, which helps model learning to extract noise invariant features. Both the well-lit and low-light images are used to train the model by a common detection loss " + }, + { + "bbox": [ + 305, + 437, + 545, + 510 + ], + "type": "inline_equation", + "content": "L_{det}(\\cdot)" + }, + { + "bbox": [ + 305, + 437, + 545, + 510 + ], + "type": "text", + "content": ". The total loss is " + }, + { + "bbox": [ + 305, + 437, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Total}" + }, + { + "bbox": [ + 305, + 437, + 545, + 510 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 336, + 512, + 545, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 512, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 336, + 512, + 545, + 526 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {T o t a l}} = \\mathcal {L} _ {\\text {d e t}} ^ {\\text {w e l l}} + \\alpha \\mathcal {L} _ {\\text {d e t}} ^ {\\text {l o w}} + \\beta \\mathcal {L} _ {D S} + \\gamma \\mathcal {L} _ {T V} \\tag {4}", + "image_path": "3f300549cb7eaef08ad7e26831b911f001593b01d9f56446f4eee59d52474eca.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "type": "text", + "content": "where is the " + }, + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{det}^{well}" + }, + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{det}^{low}" + }, + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "type": "text", + "content": " are the detection loss for welllit and low-light images, " + }, + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "type": "text", + "content": " are loss weights. We set " + }, + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 529, + 545, + 565 + ], + "type": "text", + "content": " to 1, 1 and 0.01, respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 578, + 388, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 578, + 388, + 591 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 388, + 591 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 597, + 413, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 413, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 413, + 611 + ], + "type": "text", + "content": "5.1. Experiment Setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "type": "text", + "content": "Dataset Split. In structuring our dataset, we randomly split the videos into training, validation, and testing sets, consisting of 11, 4, and 11 videos respectively. We also provide LMOT-real dataset that is captured in real-world night scenes with 6 videos. Detailed dataset split and statistical information are shown in Tab. 2." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. Following [8, 42], we recommend using HOTA [31] as the main evaluation metric to simulta" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "386" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 132, + 171 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 132, + 171 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 132, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 132, + 171 + ], + "type": "image", + "image_path": "c80a3cc4d391a972885812c8a56519f4f260a47438bad78853ea7bc5a552d15f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "text", + "content": "Figure 5. Visualization of shallow and deep features for well-lit and low-light images. It can be seen that, under low-light conditions, the shallow feature is full of noise, and the deep feature exhibits lower responses to objects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 133, + 70, + 209, + 171 + ], + "blocks": [ + { + "bbox": [ + 133, + 70, + 209, + 171 + ], + "lines": [ + { + "bbox": [ + 133, + 70, + 209, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 70, + 209, + 171 + ], + "type": "image", + "image_path": "eeff969e8471e78601b0b94d4a2a6d161420522a88f3632d691baefaa324ce82.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 211, + 70, + 286, + 171 + ], + "blocks": [ + { + "bbox": [ + 211, + 70, + 286, + 171 + ], + "lines": [ + { + "bbox": [ + 211, + 70, + 286, + 171 + ], + "spans": [ + { + "bbox": [ + 211, + 70, + 286, + 171 + ], + "type": "image", + "image_path": "9b1169c4dcaf850e15687b2fc7dc336e016615e47bc6e0bb2c8a26590438848d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "text", + "content": "neously evaluate the performance of detection and association. We also employ AssA and IDF1 [37] to evaluate association performance, MOTA, and Deta to evaluate detection performance. There are two ways to combine metrics for all classes into a single score. One is by averaging metrics over the class values, and the other is by over the detection values. To avoid possible result bias caused by some categories with fewer samples (such as Bus and Truck), we combine scores by averaging over the detection values." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 342, + 286, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 286, + 475 + ], + "type": "text", + "content": "Implementation Details. Following ByteTrack [62] and OC-SORT [3], we use YOLOX [17] as our detector. The trackers are pre-trained on the COCO [28] dataset and then trained on our LMOT dataset for 24 epochs. We apply data augmentation strategies including random flip, scale jitter of resizing, and Mosaic. We also use the physical-based noise model [50] for RAW image augment. We use SGD optimizer with weight decay " + }, + { + "bbox": [ + 46, + 342, + 286, + 475 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 342, + 286, + 475 + ], + "type": "text", + "content": " and cosine learning rate schedule, the initial learning rate is " + }, + { + "bbox": [ + 46, + 342, + 286, + 475 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 342, + 286, + 475 + ], + "type": "text", + "content": " and gradually reduces to " + }, + { + "bbox": [ + 46, + 342, + 286, + 475 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 46, + 342, + 286, + 475 + ], + "type": "text", + "content": ". We apply linear interpolation as post-processing to all trackers, with the maximum gap set to 20." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 485, + 234, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 485, + 234, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 485, + 234, + 498 + ], + "type": "text", + "content": "5.2. Analysis under low-light conditions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 504, + 286, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 286, + 552 + ], + "type": "text", + "content": "We analyze the impact of low-light conditions on multi-object tracking using the LMOT validation set. It should be emphasized that the low-light and well-lit video pairs are perfectly aligned in LMOT." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "content": "Impact to detectors. We first analyze the impact of lighting conditions on the detector. We select YOLOX as the detector since it is widely used in MOT areas [3, 8, 62]. We train the detector using well-lit images (WL), low-light images (LL), and all the images (AL). Then, test them on both well-lit and low-light images. The results are shown in Tab. 3. From the table, we can see that the model trained by welllight images achieves the best result on well-light images, but its performance significantly decreased on low-light images. Further, we visualized the feature maps under these two lighting conditions in Fig. 5. It can be seen that both the shallow feature and deep features of the low-light image are significantly degraded due to the sensor noises. We also" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 544, + 137 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 544, + 137 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 137 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 137 + ], + "type": "table", + "html": "
Training dataWell-litLow-light
WLLLmAPmARAP50mAPmARAP50
✓37.045.065.53.14.86.6
✓23.030.840.816.923.730.3
✓✓28.735.249.117.924.132.2
", + "image_path": "cb21001530618e5eda2f6f6ef0632ef1dbb260a7d4fc9996fa2cc1772ed22bba.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 178, + 544, + 249 + ], + "blocks": [ + { + "bbox": [ + 306, + 138, + 545, + 172 + ], + "lines": [ + { + "bbox": [ + 306, + 138, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 306, + 138, + 545, + 172 + ], + "type": "text", + "content": "Table 3. Analysis on LMOT validation set for detector (YOLOX [17]). WL and LL indicate the well-lit and low-light, respectively. It shows that it is hard to detect objects under low-light conditions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 178, + 544, + 249 + ], + "lines": [ + { + "bbox": [ + 309, + 178, + 544, + 249 + ], + "spans": [ + { + "bbox": [ + 309, + 178, + 544, + 249 + ], + "type": "table", + "html": "
Cond.Mot.App.HOTAAssAIDF1MOTADetA
-✓86.179.572.982.693.4
WL✓87.377.182.598.898.4
WL✓✓83.778.085.696.890.0
LL✓53.832.140.067.390.2
LL✓✓83.477.585.096.089.8
", + "image_path": "fb02d3c1adbdfabd0627c62d60832df407ccb22a5a6df07b9075074f222b8019.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 250, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 305, + 250, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 545, + 316 + ], + "type": "text", + "content": "Table 4. Analysis on LMOT validation set for different association models. Cond, Mot, and App indicate the light condition, motion, and appearance, respectively. The detection boxes are ground-truth boxes. It indicates that appearance information is effective for association under well-lit conditions, but dose not as effective under low-light conditions." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 327, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 374 + ], + "type": "text", + "content": "observed that using low-light images can largely improve the performance of the decoder under low-light conditions, and use all images to achieve the best performance. But this result is much lower than that under well-lit conditions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 376, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 591 + ], + "type": "text", + "content": "Impact to association modules. We analyze the impact of low-light conditions on the object association modules. Motion and appearance are important for object association. Both of them rely on detection boxes to locate the objects. To separate the impact of the detector, we use the ground-truth boxes as the detection boxes. The results are shown in Tab. 4. It can be seen that using only appearance matching under well-lit conditions achieves the best result while using only appearance matching under low-lit conditions results in very poor performance. This indicates that the objects in LMOT have obvious visual distinguishability, but the distinguishability is significantly reduced under low-light conditions. In Fig. 6, we visualize the appearance feature of objects in the LMOT dataset under both well-lit and low-light conditions. We can observe that under well-lit conditions, LMOT is very distinguishable in the feature space. However, under low-light conditions, this discrimination of LMOT decreased significantly." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 599, + 449, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 449, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 449, + 610 + ], + "type": "text", + "content": "5.3. Results on LMOT dataset" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We compare the proposed method with potential low-light multi-object tracking methods. We first train three baseline models using low-light videos, well-lit videos, and both of them, without any modifications to baseline trackers. They are denoted by Base-low, Base-well, and Base-all. The second type of method is tracking after low-light enhancement. These methods use low-light enhancement methods as preprocessing module. Both the enhanced videos and well-lit" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "387" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 547, + 239 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 239 + ], + "type": "table", + "html": "
MethodByteTrack[62]OC-SORT[3]DetectionParams (M)FLOPS (G)
HOTAAssAIDF1HOTAAssAIDF1mAPAP50AP75
Base-low28.042.910.928.840.334.318.832.418.999.00793.29
Base-well13.536.832.913.831.712.05.59.75.599.00793.29
Base-all28.142.632.928.939.834.619.032.519.199.00793.29
LLFlow [48]28.842.934.029.441.035.022.438.322.6116.428755.91
SDSD [47]29.141.135.529.336.936.819.836.219.3103.301136.39
DNF [24]27.640.933.028.136.435.319.135.618.2101.83907.33
SMOID [23]28.039.733.928.736.735.619.135.518.2120.887286.17
RAOD [52]26.142.929.526.142.929.518.331.318.499.07897.04
LTrack (Ours)29.443.235.229.839.036.723.440.623.5100.43800.92
", + "image_path": "c620a1e765ee6985ef90d7e4af78f313efd0f995359ef5962b009b4672ec793c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 273, + 166, + 358 + ], + "blocks": [ + { + "bbox": [ + 52, + 273, + 166, + 358 + ], + "lines": [ + { + "bbox": [ + 52, + 273, + 166, + 358 + ], + "spans": [ + { + "bbox": [ + 52, + 273, + 166, + 358 + ], + "type": "image", + "image_path": "4dee198d58fca91a3849bc51373e3dd58ceff2ea2c8ab5ee04299a1be024fec7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 362, + 127, + 372 + ], + "lines": [ + { + "bbox": [ + 89, + 362, + 127, + 372 + ], + "spans": [ + { + "bbox": [ + 89, + 362, + 127, + 372 + ], + "type": "text", + "content": "(a) Well-lit" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 274, + 284, + 358 + ], + "blocks": [ + { + "bbox": [ + 168, + 274, + 284, + 358 + ], + "lines": [ + { + "bbox": [ + 168, + 274, + 284, + 358 + ], + "spans": [ + { + "bbox": [ + 168, + 274, + 284, + 358 + ], + "type": "image", + "image_path": "440d710812ffcd28fa8131bfd71b0b2b86f5a0395b63072b7108f7acd384b1b7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 362, + 249, + 372 + ], + "lines": [ + { + "bbox": [ + 204, + 362, + 249, + 372 + ], + "spans": [ + { + "bbox": [ + 204, + 362, + 249, + 372 + ], + "type": "text", + "content": "(b) Low-light" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 373, + 287, + 418 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 418 + ], + "type": "text", + "content": "Figure 6. Visualization of appearance features of LMOT dataset using t-SNE. The same object is coded by the same color. It indicates that the appearance of objects under well-lit is distinguishable, but significantly reduced under low-light conditions" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 435, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 579 + ], + "type": "text", + "content": "videos are used for training. We select four different low-light enhancement techniques for comparison. LLFlow [48] and SDSD [47] are low-light enhancement methods based on RGB images, while DNF [24] and SMOID [23] are based on RAW low-light enhancement methods. LLFlow and DNF take images as input, SDSD and SMOID take videos as input. The RAW object detection method is also trained together on well-lit and low-light videos. Its output bounding boxes are directly fed to the tracker. We test all the potential methods on two state-of-the-art trackers ByteTrack [62] and OC-SORT [3]. We also show mAP, AP50, and AP75 to highlight the detector performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "From Tab. 5 we can see that, the proposed LTrack achieves the best HOTA and is highly competitive on all metrics. For example, the proposed method improves HOTA 1.3 with almost the same parameters and computation as base methods. This strongly proves the effectiveness of the proposed method. Compared with tracking after low-light enhancement, these methods have a large amount of additional parameters and computation but still perform worse than the proposed method. For example, LLFlow delivers 8 times FLOPS but still performs worse than our LTrack. In addition, We observed that there is" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 271, + 559, + 424 + ], + "blocks": [ + { + "bbox": [ + 47, + 239, + 546, + 264 + ], + "lines": [ + { + "bbox": [ + 47, + 239, + 546, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 546, + 264 + ], + "type": "text", + "content": "Table 5. Experimental results on LMOT dataset. Base-low, Base-well and Base-all indicate the baseline model trained on low-light, well-lit and all of them, respectively. The number of parameters and prediction latency of each method are reported along with the accuracy" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 271, + 559, + 424 + ], + "lines": [ + { + "bbox": [ + 306, + 271, + 559, + 424 + ], + "spans": [ + { + "bbox": [ + 306, + 271, + 559, + 424 + ], + "type": "table", + "html": "
MethodHOTAAssAIDF1MOTADet
Base-low32.533.439.342.427.1
Base-well28.030.434.136.923.3
Base-all34.638.140.044.830.2
LLFlow[48]33.638.038.443.129.7
SDSD[47]31.634.236.738.827.5
DNF[24]33.239.036.944.430.1
SMOID[23]31.234.136.641.226.9
RAOD[52]32.233.539.440.926.5
LTrack (Ours)35.138.940.445.230.7
", + "image_path": "f154aecf2c961420824509d77900cf16efd1f0329eb992449edf0796cf163908.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 425, + 545, + 447 + ], + "lines": [ + { + "bbox": [ + 305, + 425, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 545, + 447 + ], + "type": "text", + "content": "Table 6. Experimental results on LMOT-real dataset with OC-SORT[3]. The best result are shown in boldface" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 458, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 547, + 590 + ], + "type": "text", + "content": "no significant difference in results between RAW-based image enhancement methods and RGB-based image enhancement methods. This is not consistent with what our method has observed. The reason may be that low light enhancement focuses on image quality restoration and may mislead downstream tasks. As for RAOD [52], it has almost the same number of parameters and computational load as our method, but its performance is much lower than the proposed LTrack. Despite addressing HDR scenes through a preprocessing module for RAW input, it does not perform well under low-light conditions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 599, + 434, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 434, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 434, + 610 + ], + "type": "text", + "content": "5.4. Results on Real World" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "To validate the performance of our method in real-world low-light scenes at night, we evaluate all methods on LMOT-real dataset, using OC-SORT [3]. As shown in Tab. 6, the proposed LTrack performs much better than all comparison methods. Both the methods of tracking after low-light enhancement and the RAW detection method encounter generalizability problems and are not even better than Base-all. This strongly demonstrates the robustness of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "388" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 284, + 140 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 284, + 140 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 284, + 140 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 284, + 140 + ], + "type": "table", + "html": "
DSLALDHOTAAssAIDF1MOTADetA
✓28.142.632.923.818.7
29.042.435.025.820.1
✓28.342.433.124.619.3
✓✓29.443.235.226.120.3
", + "image_path": "b74a884963070c68961e8df7f353dd8b70221aebbf367c05876b1cc7c6320461.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 171, + 284, + 240 + ], + "blocks": [ + { + "bbox": [ + 47, + 143, + 286, + 165 + ], + "lines": [ + { + "bbox": [ + 47, + 143, + 286, + 165 + ], + "spans": [ + { + "bbox": [ + 47, + 143, + 286, + 165 + ], + "type": "text", + "content": "Table 7. Ablation on adaptive low-pass downsampling (ALD) and degradation suppression learning (DSL)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 171, + 284, + 240 + ], + "lines": [ + { + "bbox": [ + 50, + 171, + 284, + 240 + ], + "spans": [ + { + "bbox": [ + 50, + 171, + 284, + 240 + ], + "type": "table", + "html": "
Data TypeHOTAAssAIDF1MOTADetA
sRGB29.242.734.926.420.5
RAW 8-bit29.042.334.526.120.2
RAW 10-bit29.342.934.826.520.4
RAW 12-bit29.443.235.226.120.3
", + "image_path": "ddf763f8a6fe037b02cb0c968e65e010dc684f265ce23984a6bb3e0c77eb229e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 270, + 284, + 328 + ], + "blocks": [ + { + "bbox": [ + 47, + 242, + 286, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 242, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 242, + 286, + 262 + ], + "type": "text", + "content": "Table 8. Results of different image format and dynamic range on LMOT test set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 270, + 284, + 328 + ], + "lines": [ + { + "bbox": [ + 50, + 270, + 284, + 328 + ], + "spans": [ + { + "bbox": [ + 50, + 270, + 284, + 328 + ], + "type": "table", + "html": "
MethodsHOTAAssAIDF1MOTADetA
Gaussian-Possion29.943.135.726.621.1
Physical-based [50]30.444.136.026.621.2
LMOT (Ours)35.138.940.445.230.7
", + "image_path": "58ec47597c7db8169e4366b0602d64d7709c2af876bd5cabb2297406d1de92aa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 361, + 198, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 198, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 198, + 373 + ], + "type": "text", + "content": "our LTrack in real-world dark scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 380, + 198, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 198, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 198, + 393 + ], + "type": "text", + "content": "5.5. Exploration and Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 399, + 286, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 286, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 286, + 424 + ], + "type": "text", + "content": "In this section, we conduct extensive analysis and discussion on LMOT datasets and the proposed method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 428, + 286, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 428, + 286, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 428, + 286, + 535 + ], + "type": "text", + "content": "Ablation Study. We conduct ablation experiments to validate the effectiveness of our improvements, the results are shown in Tab. 7. It can be seen that all improvements contribute effectively to enhanced performance. Among them, the degradation suppression learning strategy demonstrates the most significant effect, resulting in an approximate 1-point enhancement in HOTA. The best results are achieved when employing all strategies simultaneously. This demonstrates the effectiveness of all our contributions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 540, + 286, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 286, + 624 + ], + "type": "text", + "content": "RAW vs. sRGB. We analyze the impact of input data formats, the results are shown in Tab. 8. The 12-bit RAW format achieves significantly better results than sRGB. Because the RAW format saves much more potential information and is helpful for MOT in low-light scenes. We also observed that higher bitwidth is beneficial for performance, which has also been observed in other vision tasks [21, 52]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "LMOT vs. Synthetic data. We compare our LMOT dataset with the synthetic low-light data to further demonstrate the value of LMOT. We apply the Gaussian-Possian based and Physical-based [50] low-light data synthesis method to synthesize low-light videos from well-lit videos. We train our LTrack on these types of data and evaluate their performance in real low-light scenes using LMOT-real. As" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 547, + 167 + ], + "blocks": [ + { + "bbox": [ + 47, + 329, + 286, + 350 + ], + "lines": [ + { + "bbox": [ + 47, + 329, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 329, + 286, + 350 + ], + "type": "text", + "content": "Table 9. Results with low-light synthesis methods and LMOT dataset on LMOT-real set." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 71, + 547, + 167 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 547, + 167 + ], + "type": "table", + "html": "
CategoriesHOTAAssAIDF1MOTADetA
Person24.230.829.925.119.2
Bicycle16.940.315.68.97.1
Car37.553.347.233.726.5
Motorcycle19.835.022.615.211.4
Bus44.359.951.836.732.9
Truck6.915.66.43.23.1
", + "image_path": "395ebfc15ee0607d5716107e9a2d6011de9c708a55e154985a4cdcaaeff6186c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 167, + 545, + 189 + ], + "lines": [ + { + "bbox": [ + 306, + 167, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 545, + 189 + ], + "type": "text", + "content": "Table 10. Results of the proposed method for different categories on LMOT test set." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 203, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 203, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 203, + 545, + 239 + ], + "type": "text", + "content": "shown in Tab. 9, the tracker trained on our LMOT dataset has much better performance in real night scenes, which strongly demonstrates the value of our LMOT dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 245, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 245, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 245, + 545, + 354 + ], + "type": "text", + "content": "Analysis on different category. We also analyze the performance for different categories. From Tab. 10, we can see that cars and buses achieved the best performance because they have regular shapes and larger areas. Trucks exhibited the poorest performance, because they have the fewest instances, making it challenging for the model to learn accurate identification. The person achieves relatively average scores. Bicycles and motorcycles have close scores since they have similar appearance and motion patterns." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 368, + 378, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 368, + 378, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 368, + 378, + 380 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 389, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 389, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 545, + 533 + ], + "type": "text", + "content": "In this work, we investigate the multi-object tracking in the dark scenes. We build a new low-light multi-object tracking (LMOT) dataset, which provides well-aligned low-light video pairs and high-quality multi-object tracking annotations. We observed that low-light images are significantly degraded by the sensor noises, which also degrades the feature maps and significantly deteriorates the model performance. To learn the invariant semantic formation under noise disturbance and quality degradation, we present the adaptive low-pass downsample module and degradation suppression learning. These improvements greatly enhance the robustness of our method in real-world low-light scenes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 533, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 533, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 545, + 606 + ], + "type": "text", + "content": "Limitations. We focus on multi-object tracking in the dark scenes. But we do not consider other degradation environments in the real world, such as rainy and foggy days. In our future work, we will consider exploring multi-object tracking in more real-world scenarios, promoting the development of MOT in real-world applications." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 620, + 403, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 403, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 403, + 634 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "type": "text", + "content": "This work was supported by the National Key R&D Program of China (2022YFC3300704), the National Natural Science Foundation of China (62331006,62171038, and 62088101), the R&D Program of Beijing Municipal Education Commission (KZ202211417048), and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "389" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, pages 3464-3468, 2016. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 168 + ], + "type": "text", + "content": "[2] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. arXiv preprint arXiv:2303.06705, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 287, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 287, + 212 + ], + "type": "text", + "content": "[3] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. Observation-centric sort: Rethinking sort for robust multi-object tracking. In CVPR, pages 9686-9696, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 245 + ], + "type": "text", + "content": "[4] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In CVPR, pages 3291-3300, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 246, + 286, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 286, + 278 + ], + "type": "text", + "content": "[5] Chen Chen, Qifeng Chen, Minh N Do, and Vladlen Koltun. Seeing motion in the dark. In ICCV, pages 3185-3194, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 312 + ], + "type": "text", + "content": "[6] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. IJCV, pages 1-21, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 313, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 287, + 346 + ], + "type": "text", + "content": "[7] Wongun Choi and Silvio Savarese. A unified framework for multi-target tracking and collective activity recognition. In ECCV, pages 215-230, 2012. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 346, + 287, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 391 + ], + "type": "text", + "content": "[8] Yutao Cui, Chenkai Zeng, Xiaoyu Zhao, Yichun Yang, Gangshan Wu, and Limin Wang. Sportsmot: A large multi-object tracking dataset in multiple sports scenes. arXiv preprint arXiv:2304.05170, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 392, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 287, + 445 + ], + "type": "text", + "content": "[9] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "type": "text", + "content": "[10] Yunhao Du, Zhicheng Zhao, Yang Song, Yanyun Zhao, Fei Su, Tao Gong, and Hongying Meng. *Strongsort: Make deep-sort great again*. IEEE Transactions on Multimedia, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 480, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 533 + ], + "type": "text", + "content": "[11] Anna Ellis and James Ferryman. Pets2010 and pets2009 evaluation of results using individual ground truthed single views. In 2010 7th IEEE international conference on advanced video and signal based surveillance, pages 135-142, 2010. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 535, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 557 + ], + "type": "text", + "content": "[12] Wang Zhongdao et al. Do different tracking tasks require different appearance models? 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "text", + "content": "[13] Huiyuan Fu, Wenkai Zheng, Xicong Wang, Jiaxuan Wang, Heng Zhang, and Huadong Ma. Dancing in the dark: A benchmark towards general low-light video enhancement. In ICCV, pages 12877-12886, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 601, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 635 + ], + "type": "text", + "content": "[14] Ying Fu, Jian Chen, Tao Zhang, and Yonggang Lin. Residual scale attention network for arbitrary scale image superresolution. Neurocomputing, 427:201-211, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "text", + "content": "[15] Ying Fu, Yang Hong, Linwei Chen, and Shaodi You. Le-gan: Unsupervised low-light image enhancement network using attention module and identity invariant loss. Knowledge-Based Systems, 240:108010, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[16] Ruopeng Gao and Limin Wang. Memotr: Long-term memory-augmented transformer for multi-object tracking. In ICCV, pages 9901-9910, 2023. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361, 2012. 1, 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 186 + ], + "type": "text", + "content": "[19] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In CVPR, pages 1780-1789, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 220 + ], + "type": "text", + "content": "[20] Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE TIP, 26(2):982-993, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "type": "text", + "content": "[21] Yang Hong, Kaixuan Wei, Linwei Chen, and Ying Fu. Crafting object detection in very low light. In BMVC, page 3, 2021. 1, 2, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 256, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 545, + 300 + ], + "type": "text", + "content": "[22] Haidi Ibrahim and Nicholas Sia Pik Kong. Brightness preserving dynamic histogram equalization for image contrast enhancement. IEEE Transactions on Consumer Electronics, 53(4):1752-1758, 2007. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 302, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 302, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 302, + 545, + 335 + ], + "type": "text", + "content": "[23] Haiyang Jiang and Yinqiang Zheng. Learning to see moving objects in the dark. In ICCV, pages 7324-7333, 2019. 1, 2, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 337, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 545, + 380 + ], + "type": "text", + "content": "[24] Xin Jin, Ling-Hao Han, Zhen Li, Chun-Le Guo, Zhi Chai, and Chongyi Li. Dnf: Decouple and feedback network for seeing in the dark. In CVPR, pages 18135-18144, 2023. 1, 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "type": "text", + "content": "[25] Daniel J Jobson, Zia-ur Rahman, and Glenn A Woodell. Properties and performance of a center/surround retina. IEEE TIP, 6(3):451-462, 1997. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "type": "text", + "content": "[26] Laura Leal-Taixe, Anton Milan, Ian Reid, Stefan Roth, and Konrad Schindler. Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv preprint arXiv:1504.01942, 2015. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 462, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 462, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 307, + 462, + 545, + 507 + ], + "type": "text", + "content": "[27] Sohyun Lee, Jaesung Rim, Boseung Jeong, Geonu Kim, Byungju Woo, Haechan Lee, Sunghyun Cho, and Suha Kwak. Human pose estimation in extremely low-light conditions. In CVPR, pages 704-714, 2023. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 508, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 508, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 508, + 545, + 552 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "type": "text", + "content": "[29] Qiankun Liu, Qi Chu, Bin Liu, and Nenghai Yu. Gsm: Graph similarity model for multi-object tracking. In IJCAI, pages 530-536, 2020. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 588, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 633 + ], + "type": "text", + "content": "[30] Qiankun Liu, Dongdong Chen, Qi Chu, Lu Yuan, Bin Liu, Lei Zhang, and Nenghai Yu. Online multi-object tracking with unsupervised re-identification learning and occlusion estimation. Neurocomputing, 483:333–347, 2022. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 634, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 545, + 678 + ], + "type": "text", + "content": "[31] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. IJCV, 129:548-578, 2021. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 679, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 545, + 713 + ], + "type": "text", + "content": "[32] Kang Ma, Ying Fu, Dezhi Zheng, Chunshui Cao, Xuecai Hu, and Yongzhen Huang. Dynamic aggregated network for gait recognition. In CVPR, pages 22076-22085, 2023. 1" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "390" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[33] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, pages 8844-8854, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 152 + ], + "type": "text", + "content": "[34] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831, 2016. 1, 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "type": "text", + "content": "[35] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In CVPR, pages 1765–1773, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 188, + 287, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 188, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 48, + 188, + 287, + 231 + ], + "type": "text", + "content": "[36] Jihun Park, Jinseok Hong, Woolil Shim, and Dae-Jin Jung. Multi-object tracking on swir images for city surveillance in an edge-computing environment. Sensors, 23(14):6373, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 233, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 287, + 277 + ], + "type": "text", + "content": "[37] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In ECCV, pages 17-35, 2016. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 279, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 287, + 312 + ], + "type": "text", + "content": "[38] Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena, 60(1-4):259-268, 1992. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 314, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 314, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 287, + 346 + ], + "type": "text", + "content": "[39] Aashish Sharma and Robby T Tan. Nighttime visibility enhancement by increasing the dynamic range and suppression of light effects. In CVPR, pages 11977-11986, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 349, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 380 + ], + "type": "text", + "content": "[40] Shuran Song and Jianxiong Xiao. Tracking revisited using rgbd camera: Unified benchmark and baselines. In ICCV, pages 233-240, 2013. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "text", + "content": "[41] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 428, + 287, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 428, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 48, + 428, + 287, + 473 + ], + "type": "text", + "content": "[42] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In CVPR, pages 20993-21002, 2022. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 475, + 287, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 517 + ], + "type": "text", + "content": "[43] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 520, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 552 + ], + "type": "text", + "content": "[44] Ye Tian, Ying Fu, and Jun Zhang. Plug-and-play algorithm for under-sampling fourier single-pixel imaging. Science China. Information Sciences, 65(10):209303, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 586 + ], + "type": "text", + "content": "[45] Ye Tian, Ying Fu, and Jun Zhang. Local-enhanced transformer for single-pixel imaging. Optics Letters, 48(10): 2635-2638, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "text", + "content": "[46] Chaoqun Wang, Chunyan Xu, Zhen Cui, Ling Zhou, Tong Zhang, Xiaoya Zhang, and Jian Yang. Cross-modal patternpropagation for rgb-t tracking. In CVPR, pages 7064-7073, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[47] Ruixing Wang, Xiaogang Xu, Chi-Wing Fu, Jiangbo Lu, Bei Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In ICCV, pages 9700-9709, 2021. 1, 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[48] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In AAAI, pages 2604-2612, 2022. 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[49] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[50] Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. A physics-based noise formation model for extreme low-light raw denoising. In CVPR, pages 2758-2767, 2020. 4, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[51] Kaixuan Wei, Angelica Aviles-Rivero, Jingwei Liang, Ying Fu, Hua Huang, and Carola-Bibiane Schonlieb. Tfnpn: Tuning-free plug-and-play proximal algorithms with applications to inverse imaging problems. Journal of Machine Learning Research, 23(16):1-48, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[52] Ruikang Xu, Chang Chen, Jingyang Peng, Cheng Li, Yibin Huang, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Toward raw object detection: A new benchmark and a new model. In CVPR, pages 13384-13393, 2023. 2, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[53] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. In CVPR, pages 17714-17724, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[54] Jinyu Yang, Shang Gao, Zhe Li, Feng Zheng, and Ales Leonardis. Resource-efficient rgbd aerial tracking. In CVPR, pages 13374-13383, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "type": "text", + "content": "[55] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2636-2645, 2020. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "type": "text", + "content": "[56] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In ECCV, pages 659–675, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "type": "text", + "content": "[57] Fan Zhang, Yu Li, Shaodi You, and Ying Fu. Learning temporal consistency for low light video enhancement from single images. In CVPR, pages 4967-4976, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "type": "text", + "content": "[58] Jiqing Zhang, Xin Yang, Yingkai Fu, Xiaopeng Wei, Baocai Yin, and Bo Dong. Object tracking by jointly exploiting frame and event domain. In ICCV, pages 13043-13052, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "type": "text", + "content": "[59] Pengyu Zhang, Jie Zhao, Dong Wang, Huchuan Lu, and Xiang Ruan. Visible-thermal uav tracking: A large-scale benchmark and new baseline. In CVPR, pages 8886–8895, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "type": "text", + "content": "[60] Tao Zhang, Ying Fu, and Cheng Li. Hyperspectral image denoising with realistic data. In ICCV, pages 2248-2257, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "type": "text", + "content": "[61] Tao Zhang, Ying Fu, and Cheng Li. Deep spatial adaptive network for real image demosaicing. In AAAI, pages 3326-3334, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "text", + "content": "[62] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, pages 1-21, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[63] Yang Zhang, Tiancai Wang, and Xiangyu Zhang. Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. In CVPR, pages 22056-22065, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[64] Zhiyu Zhu, Junhui Hou, and Dapeng Oliver Wu. Cross-modal orthogonal high-rank augmentation for rgb-event" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "391" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 207 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "text", + "content": "transformer-trackers. In ICCV, pages 22045-22055, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "text", + "content": "[65] Yunhao Zou and Ying Fu. Estimating fine-grained noise model via contrastive learning. In CVPR, pages 12682-12691, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 173 + ], + "type": "text", + "content": "[66] Yunhao Zou, Yinqiang Zheng, Tsuyoshi Takatani, and Ying Fu. Learning to reconstruct high speed and high dynamic range videos from events. In CVPR, pages 2024-2033, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[67] Yunhao Zou, Chenggang Yan, and Ying Fu. Rawhdr: High dynamic range image reconstruction from a single raw image. In ICCV, pages 12334-12344, 2023. 2" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "392" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/20977810-85e5-40f1-a4c6-cf52c533cece_content_list.json b/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/20977810-85e5-40f1-a4c6-cf52c533cece_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..64bb73b0369a6e40c4600c1e0cf6f9dd93a88c6b --- /dev/null +++ b/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/20977810-85e5-40f1-a4c6-cf52c533cece_content_list.json @@ -0,0 +1,1382 @@ +[ + { + "type": "text", + "text": "Multi-Scale 3D Gaussian Splating for Anti-Aliased Rendering", + "text_level": 1, + "bbox": [ + 169, + 130, + 802, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiwen Yan", + "bbox": [ + 223, + 181, + 323, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weng Fei Low", + "bbox": [ + 364, + 181, + 483, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Chen", + "bbox": [ + 524, + 181, + 594, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gim Hee Lee", + "bbox": [ + 635, + 181, + 743, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science, National University of Singapore", + "bbox": [ + 217, + 199, + 751, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yan.zhiwen@u.nus.edu {wengfei.low, chenyu}@comp.nus.edu.sg gimhee.lee@nus.edu.sg", + "bbox": [ + 112, + 219, + 849, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/22407358080dc2e589a148aa68a17808d0b136374d0c8104810b69f9dbc561a9.jpg", + "image_caption": [ + "Figure 1. The rendering quality and speed of the original 3D Gaussian splatting[12] deteriorate severely at low resolutions or from distant cameras due to aliasing. Conversely, our multi-scale 3D Gaussians representation utilizes selective rendering to achieve faster $(160\\% - 2400\\%)$ at $4\\times -128\\times$ resolution) and more accurate rendering at lower resolutions." + ], + "image_footnote": [], + "bbox": [ + 86, + 250, + 893, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 590, + 313, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussians have recently emerged as a highly efficient representation for 3D reconstruction and rendering. Despite its high rendering quality and speed at high resolutions, they both deteriorate drastically when rendered at lower resolutions or from far away camera position. During low resolution or far away rendering, the pixel size of the image can fall below the Nyquist frequency compared to the screen size of each splatted 3D Gaussian and leads to aliasing effect. The rendering is also drastically slowed down by the sequential alpha blending of more splatted Gaussians per pixel. To address these issues, we propose a multi-scale 3D Gaussian splatting algorithm, which maintains Gaussians at different scales to represent the same scene. Higher-resolution images are rendered with more small Gaussians, and lower-resolution images are rendered with fewer larger Gaussians. With similar training time, our algorithm can achieve $13\\% -66\\%$ PSNR and $160\\% -2400\\%$ rendering speed improvement at $4\\times -128\\times$ scale rendering", + "bbox": [ + 75, + 628, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on Mip-NeRF360 dataset compared to the single scale 3D Gaussian splatting. More results and code are released on our project page.", + "bbox": [ + 498, + 593, + 890, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 660, + 632, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian Splatting[12] has recently emerged as a highly efficient representation for novel view synthesis. Compared to the time-consuming ray marching used in most neural radiance fields (NeRF) [2, 15, 16], a high-resolution image can be rendered in real-time by rasterizing the splatted 3D Gaussians. However, this rasterization algorithm is subjected to severe aliasing effect and speed deterioration when rendering the same scene at low resolution or from distant positions as shown in Fig. 1. This limitation significantly constrains the application of the 3D Gaussian splatting algorithm in reconstructing and rendering large-scale scenes.", + "bbox": [ + 496, + 686, + 890, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aliasing effect is a consequence of inadequate sampling frequency failing to capture the continuous signal accurately. In the context of rendering, image pixels are sampled", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20923", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "with an interval of one-pixel size. The signal can be considered as the 3D scene represented implicitly as in NeRF or explicitly as in 3D Gaussians. When part of the 3D scene is represented with high details but rendered with low resolution or from distant positions, the disparity between the low sampling and high signal frequencies culminates in aliasing artifacts. A naive solution is to render at high resolution and subsequently down-scale the rendered image to a lower resolution. However, this solution is not viable for scenes containing both near and far regions which are very common. Due to the inability of 3D Gaussian splatting algorithm to accommodate varying resolutions within a single image, rendering the entire image with a even higher resolution for the sake of far away regions is neither time nor memory efficient.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We postulate that the pronounced aliasing artifacts observed when rendering with 3D Gaussians, as opposed to other techniques such as NeRF, are primarily attributable to the splatting of small Gaussians. 3D regions with intricate details are represented with large amount of small Gaussians. When rendering these regions with low resolution or from a distant view, many splatted small Gaussians are cramped in one pixel and therefore the pixel color of this region is dominated by the front-most Gaussian, even if this Gaussian is much smaller than others and not at the center. This problem is further aggravated by the low pass filter in [12, 19] applied to each individual Gaussian with the intention to mitigate aliasing on edges at high resolutions. This problem is explained in more detail in Sec. 3.2.", + "bbox": [ + 75, + 318, + 472, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to the aliasing artifacts, the rendering speed of 3D Gaussians is also affected at low resolution. The number of 3D Gaussians that need to be rendered remains constant at lower resolutions, but they are more concentrated to fewer pixels. The Gaussians that are splatted to the same pixel cannot be rendered in parallel. This means that the image rendering is even slower at lower resolution in comparison with NeRF rendering time that reduces linearly with decreasing resolution. Hence, although aliasing is not a problem exclusive to 3D Gaussian splattering, it is more prominent and more difficult to tackle.", + "bbox": [ + 75, + 530, + 472, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions To mitigate the aliasing problem for 3D Gaussian splatting, we propose a novel multi-scale 3D Gaussians to represent the scene at different levels of detail (LOD) as shown in Fig. 2. This is inspired by the mipmap and LOD algorithms widely used in computer graphics, which pre-computes textures and polygons at different scales to be rendered under different resolutions and distances. Similarly, we add larger, coarser Gaussians for lower resolutions by aggregating the smaller and finer Gaussians from higher resolutions. Depending on the pixel coverage of the splatted Gaussians during rendering, only a subset of the Gaussians is used. A simplified explanation", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for this is that the coarse Gaussians are used to render low-resolution images and the fine Gaussians are used to render high-resolution images. With fewer than $5\\%$ number of Gaussians added and a similar training time, our method can achieve $13\\% - 66\\%$ PSNR and $160\\% - 2400\\%$ rendering speed improvements at $4 \\times -128 \\times$ scale rendering on MipNeRF360 dataset[2], while maintaining a comparable rendering quality and speed at $1 \\times$ scale.", + "bbox": [ + 496, + 90, + 893, + 213 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 498, + 227, + 650, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Anti-Aliasing in Computer Graphics", + "text_level": 1, + "bbox": [ + 498, + 252, + 818, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ailiasing is a long-standing problem for computer graphics when rendering a scene to a discrete image. Traditional anti-aliasing techniques primarily target mesh representations. Supersampling Anti-Aliasing (SSAA) [5] renders the scene at a higher resolution before downscaling, leading to significantly more time and memory demain, and therefore is less used in real-time applications. The Multisample Anti-Aliasing (MSAA) [1, 5] algorithm selectively supersamples pixels on the edges, reducing resource and time consumption. This technique is not very suitable for 3D Gaussian splatting because of its requirement for regular grids and lack of support for variable sampling resolution at different pixels. The more recent Fast Approximate Anti-Aliasing (FXAA) [11, 14] is a post processing algorithm that smooths the jagged edges after the image is rendered. Unfortunately, this technique is also not suitable for Gaussian representation as the front-most Gaussian dominates the pixel color and produces chunky instead of jagged artifacts in mesh rendering.", + "bbox": [ + 496, + 276, + 893, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to the supersampling methods mentioned above, our method takes the inspiration from hierarchical mipmap [18] and level of details (LOD) [7, 9] algorithms to address the aliasing for 3D Gaussians. Mipmap uses multi-scale textures for the rendering at different resolution or from different distances. LOD algorithm represents the models in a scene with different complexity to be rendered at different distances. Both techniques not only mitigate the aliasing effect by reducing the complexity of the scene representation, but also enhances rendering speed, particularly for large-scale scenes.", + "bbox": [ + 496, + 564, + 895, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Anti-Aliasing in Neural Representation", + "text_level": 1, + "bbox": [ + 498, + 739, + 836, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recent success of neural representations especially Neural Radiance Fields (NeRF) [6, 15, 16] has also inspired some works to develop algorithms against aliasing effect on neural representations beyond the traditional mesh representation. Mip-NeRF [2, 3] employ low pass filters on the positional encoding of the input spatial coordinates to reduce the scene signal frequency. Building on the hash grid representation used by InstantNGP [16] with no position encoding, Zip-NeRF [4] proposes a multi-sampling", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20924", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/49c1372bcd42994df6ad48a0c39401d577612edebdd55ad6e09fcfe3214ccdd2.jpg", + "image_caption": [ + "Figure 2. Overall pipeline of our algorithm. At the early stage of training (left), small Gaussians below certain size threshold in each voxel are aggregated, enlarged and inserted into the scene at different resolution scale. During rendering (right), the multi-scale Gaussians of the appropriate \"pixel coverage\" at the current render resolution are selected for rendering. If the rendering resolution scale equals to the scale of the Gaussians, the expected \"pixel coverage\" range of the Gaussians are updated accordingly." + ], + "image_footnote": [], + "bbox": [ + 119, + 88, + 429, + 282 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/528e9aadd169d7d48229070c53928db14ac276d72d9fede1bc9d2acdf1ffda30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 88, + 849, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "strategy in the conical frustum instead of the camera ray, at the cost of $6 \\times$ rendering time. Similar to the mipmap algorithm in mesh texture rendering, Tri-MipNeRF [10] and MipGrid [17] proposes to use multi-scale feature grids for rendering at different resolution or distance.", + "bbox": [ + 75, + 367, + 468, + 443 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conversely, 3D Gaussian splatting [12] presents unique anti-aliasing challenges due to its distinct scene representation. It does not have any positional encoding or feature grid, and its requirement for regular grids conflicts the more flexible multi-sampling strategies. The concentration of small Gaussians in detail-rich regions exacerbates aliasing and speed issues, more so than in NeRF representations. To the best of our knowledge, we are the first to propose an anti-aliasing algorithm for scene reconstruction using 3D Gaussian splatting.", + "bbox": [ + 75, + 443, + 470, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 606, + 217, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 76, + 631, + 285, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian splatting is first proposed in EWA Splatting [19], and later used by [12] for scene reconstruction and novel view synthesis. The scene is represented by a set of $\\mathbf{K}$ 3D Gaussians $\\{\\mathcal{G}_{\\hat{V}^k,\\hat{\\mu}^k},\\sigma_k,\\mathbf{c}_k\\mid k\\in [1,\\mathbf{K}]\\}$ with variance $\\hat{V}^k$ , center $\\hat{\\mu}^k$ , density $\\sigma_{k}$ and color $\\mathbf{c}^k$ . During rendering, the 3D Gaussians are splatted to the 2D screen to by the perspective transformation to form 2D Gaussians $\\mathcal{G}_{V^k,\\mu^k}$ . The image is then divided into $16\\times 16$ regular tiles and all 2D Gaussians touching each tile are sorted based on their original depth. The color of each pixel in the tile is then rasterized from the sequential alpha blending the 2D Gaussians from front to back.", + "bbox": [ + 75, + 654, + 468, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Cause of Aliasing in 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 76, + 847, + 444, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Aliasing can occur when sampling a continuous signal $g(x)$ with a discrete sampling function $\\delta_s(x, \\Delta x) =$", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\sum_{n = -\\infty}^{\\infty}\\delta (x - n\\cdot \\Delta x)$ , where $\\delta$ is a impulse function. The result of the sample in the spatial domain is:", + "bbox": [ + 498, + 367, + 890, + 398 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng _ {s} (x) = \\delta_ {s} (x, \\Delta x) \\cdot g (x). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 407, + 890, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This sampled function converted into the frequency domain using Fourier transform operator $\\mathcal{F}$ becomes:", + "bbox": [ + 498, + 433, + 890, + 463 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {F} \\left[ g _ {s} (u) \\right] = \\frac {1}{\\Delta x} \\sum_ {\\substack {k = - \\infty \\\\ \\infty}} ^ {\\infty} \\delta \\left(u - \\frac {k}{\\Delta x}\\right) * \\mathcal {F} [ g (x) ] \\tag{2} \\\\ = \\frac {1}{\\Delta x} \\sum_ {k = - \\infty} ^ {\\infty} \\mathbf {G} (u - \\frac {k}{\\Delta x}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 472, + 890, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When the highest frequency component $f_{max}$ of the signal is greater than half of the sampling frequency $f_{s} = \\frac{1}{\\Delta x}$ , $\\mathbf{G}(u - \\frac{k}{\\Delta x})$ in the summation sequence would overlap with each other and causes the sampled signal to diverge from the actual signal. This phenomenon is the aliasing effect and the minimum sampling frequency needed to avoid aliasing is $f_{Ny} = 2 \\cdot f_{max}$ , known as the Nyquist frequency.", + "bbox": [ + 498, + 563, + 890, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The EWA splatting [19] used by 3D Gaussian splatting [12] also tries to mitigate the aliasing problem by applying a low pass filter to the rendered color. To approximate this efficiently, it applies a Gaussian kernel $h(\\mathbf{x})$ as the low pass filter on each splatted 2D signal $g_{c}(\\mathbf{x})$ independently to produce a band limited signal:", + "bbox": [ + 496, + 671, + 890, + 762 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} g _ {c} ^ {\\prime} (\\mathbf {x}) = g _ {c} (\\mathbf {x}) * h (\\mathbf {x}) \\\\ \\approx \\sum_ {k} \\sigma_ {k} \\mathbf {c} _ {k} T _ {k} \\int_ {\\mathcal {R} ^ {2}} q _ {k} (\\eta) h (\\mathbf {x} - \\eta) d \\eta \\tag {3} \\\\ = \\sum_ {k} \\sigma_ {k} \\mathbf {c} _ {k} T _ {k} \\cdot (q _ {k} * h) (\\mathbf {x}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 768, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{R}^2$ is the range of one pixel, $q_{k}(\\mathbf{x})$ is the 2D integrated Gaussian kernel, and $\\sigma_k,\\mathbf{c}_k,T_k$ are the opacity,", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20925", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "color, and transmittance at each Gaussian, respectively. By combining the reconstruction Gaussian kernel $\\mathcal{G}_{V^k}$ and low pass Gaussian kernel $\\mathcal{G}_{V^h}$ of covariance matrix $V^{k}$ and $V^{h}$ , the band limit function becomes:", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} g _ {c} ^ {\\prime} (\\mathbf {x}) = \\sum_ {k} \\alpha_ {k} \\cdot (\\mathcal {G} _ {V ^ {k}} * \\mathcal {G} _ {V ^ {h}}) (\\mathbf {x}) \\\\ = \\sum_ {k} \\alpha_ {k} \\cdot \\mathcal {G} _ {V ^ {k} + V ^ {h}} (\\mathbf {x}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 160, + 385, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{k}$ represents all coefficients invariant of $\\mathbf{x}$ at each Gaussian and $V^{h}$ is determined by the screen pixel size. A simple understanding of this is that the covariance of each 3D Gaussian is increased based on the screen pixel size.", + "bbox": [ + 75, + 238, + 468, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This method of applying a low pass filter to each 3D Gaussian independently helps to smooth the edges of the Gaussians when the Gaussians are not too small compared to the pixel size. However, it also gives rise to two substantial issues at low resolutions:", + "bbox": [ + 75, + 299, + 468, + 373 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. $V^{h}$ added to the original covariance $V^{k}$ effectively increases the extent of each Gaussian, especially when $V^{h}$ is large compared to $V^{k}$ at low resolutions. Small Gaussians in the front dominate the color of the pixel and cause severe artifacts shown in Fig. 7.", + "2. The number of Gaussians involved in the sequential $\\sum_{k}$ for each pixel scales increases with decreasing image resolution. Due to the incremental calculation of the transmittance $T_{k}$ , the rendering even slower at lower resolutions." + ], + "bbox": [ + 76, + 375, + 468, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Our Method", + "text_level": 1, + "bbox": [ + 76, + 540, + 207, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Multi-Scale Gaussians Based on Pixel Coverage", + "text_level": 1, + "bbox": [ + 76, + 565, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate the aliasing artifacts of 3D Gaussians [12] while avoiding the two problems of the EWA splatting [19], we introduce multi-scale 3D Gaussians (cf. Fig. 2) that tackle the problem on the scene-level instead of on each individual Gaussian. The 3D scene is represented with Gaussians from 4 levels of detail, corresponding to the $1 \\times$ , $4 \\times$ , $16 \\times$ , and $64 \\times$ downsampled resolution. Small finer-level Gaussians are aggregated to create larger Gaussians for coarser levels during training. Each 3D Gaussian $\\mathcal{G}_k^l$ belongs to one of the levels $l$ and is included or excluded independently during the rendering based on its \"pixel coverage\".", + "bbox": [ + 75, + 588, + 468, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pixel Coverage of Gaussian. The \"pixel coverage\" of a Gaussian reflects the size of the Gaussian when splatted onto the screen space compared to the pixel size at the current rendering resolution. The \"pixel coverage\" $S_{k}$ of a splatted 2D Gaussian $\\mathcal{G}_{(\\mu^k,V^k)}$ is defined as the length of its horizontal or vertical axis until the low opacity level set, whichever is smaller, as shown in Fig. 3. The pixel coverage is measured in pixel count and the opacity threshold $\\sigma_T$ is set as $\\frac{1}{255}$ .", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nu; \\sigma_ {k} \\mathcal {G} _ {(\\mu^ {k}, V ^ {k})} \\left(\\mu_ {x} ^ {k} + \\frac {u}{2}, \\mu_ {y} ^ {k}\\right) = \\sigma_ {T}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 90, + 738, + 117 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} v \\colon \\sigma_ {k} \\mathcal {G} _ {(\\mu^ {k}, V ^ {k})} \\left(\\mu_ {x} ^ {k}, \\mu_ {y} ^ {k} + \\frac {v}{2}\\right) = \\sigma_ {T} \\\\ S _ {k} = \\min (u, v) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 117, + 736, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c49ff51666af448f95746b33dee309d3eceb9f9926eb835cb783b7daa8fb303d.jpg", + "image_caption": [ + "Figure 3. Pixel coverage of a 3D Gaussian is its horizontal or vertical size, whichever is smaller measured by the level set." + ], + "image_footnote": [], + "bbox": [ + 692, + 90, + 888, + 191 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/52a83eb8876fa856b15822877f55ed3ae8b65ebe1ac25089a5dce2d93c083a0b.jpg", + "image_caption": [ + "Figure 4. Missing parts caused by naive small Gaussian filtering at different resolution scales." + ], + "image_footnote": [], + "bbox": [ + 519, + 246, + 638, + 366 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/163f40aca66f15030db0b51fc3bbc7b9d64aa7e6fe0e0a505af0d8b65e8e8271.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 246, + 754, + 366 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/36968b76538c6d92057e1666774887c285d732b5501adc0acb8a42d602cd7d58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 246, + 870, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pixel coverage approximates the extent of a 2D splatted Gaussian in the spatial domain. During the rendering from a given camera direction, the color of each splatted Gaussian is constant within this pixel coverage. As a result, the coverage of this pixel approximates the inverse of the highest frequency component $f_{max} = 1 / S_k$ in this region. Compared to the sampling frequency of $f_s = 1\\mathrm{px}^{-1}$ during rasterization, a signal frequency of $f_{max} > f_s / 2$ can cause the sampling to fall below the Nyquist frequency needed to avoid aliasing.", + "bbox": [ + 496, + 425, + 890, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consequently, the Gaussians with pixel coverage $S_{k} < S_{T} = 2\\mathrm{px}$ should be filtered out during rendering to avoid aliasing. Since 3D Gaussian representation does not encode the signal of different frequencies at different Gaussians, naively filtering out the small Gaussians will result in a hole or part missing in the scene as shown in Fig. 4. To address this issue, we propose to aggregate the small Gaussians to form large Gaussians that encode the low-frequency signal. These large Gaussians would appear when the small Gaussians are filtered out.", + "bbox": [ + 496, + 575, + 890, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Aggregate to Insert Large Gaussians. All 3D Gaussians initialized from the input point cloud at the start of the training belong to the finest level $l = 1$ . They are densified by splitting and cloning as in [12], and all the densified Gaussians would inherit the same level. After the warm-up stage of the first 1,000 iterations, we introduce coarse-level Gaussians by aggregating fine-level Gaussians that are too small as visualized in Fig. 5 and described in Algorithm 1. The procedure is outlined as follows:", + "bbox": [ + 496, + 734, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1. For all levels $\\{l_m\\mid 2\\leq l_m\\leq l_{max}\\}$ , we render all 3D Gaussians from $[1,l_m - 1]$ at the $4^{l_m - 1}$ times downsam", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20926", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e30244ed326674ca9c05c95ab3b935ac8634751f2eb8c781970c4398def740a8.jpg", + "image_caption": [ + "Figure 5. Large Gaussians are created by aggregating the small Gaussians in each voxel below the pixel coverage threshold, and then enlarged by the pixel coverage multiplier." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 472, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pled resolution of all training images. All 3D Gaussians with the minimal \"pixel coverage\" $S_{k}$ smaller than the filter threshold $S_{T}$ are chosen for the aggregation.", + "bbox": [ + 94, + 284, + 468, + 329 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. The chosen 3D Gaussians are binned by a $(400 / l_{m})^{3}$ resolution voxel grid based on their positions. The attributes of all Gaussians within each voxel are aggregated to create a new Gaussian using average pooling, including position, scaling, opacity and color. More details are included in the supplementary.", + "3. Based on the average \"pixel coverage\" $S_{avg}$ of Gaussians in each voxel, the scaling of each new Gaussian created is enlarged by $S_T / S_{avg}$ so that it is of a size suitable to be rendered at $l_m$ . This new Gaussian belongs to level $l_m$ ." + ], + "bbox": [ + 76, + 329, + 468, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Not all Gaussians from the fine levels are small. Many Gaussians in the background or in the textureless regions are large and do not need to be aggregated. The number of Gaussians created is often fewer than $5\\%$ of the final total number of Gaussians.", + "bbox": [ + 75, + 496, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fa48479560dfc9a4f3b4a16dfc4226f245e4e36b1987920dd8827e73e7865cc3.jpg", + "table_caption": [ + "Algorithm 1 Aggregate Small Gaussians" + ], + "table_footnote": [], + "table_body": "
1:procedure AGGREGATEGAUSSIANS(G1:lmax)
2:for lm←2 to lmax do
3:S1:K=PixelCoverage(G1:lm-1, scale 4lmax-1)
4:Gsmall = {Gk|Sk < ST, ∀k ∈ [1:K]}
5:for n←1 to (400/lm)3 do
6:Gn = {Gk|Gk in voxel n, ∀Gk ∈ Gsmall}
7:Glmnew,n = Enlarge(Average(Gn))
8:InsertIntoScene(Glmnew,n)
9:end for
10:end for
11:end procedure
", + "bbox": [ + 84, + 606, + 473, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-Scale Training and Selective Rendering. After the large Gaussians are added, the model is trained with both the original images and the downsampled images. A maximum pixel coverage $S_{k}^{max}$ and a minimum pixel coverage $S_{k}^{min}$ of each Gaussian are stored for the selective rendering. If the rendering downsample scale equals to", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Selective Rendering Based on Pixel Coverage" + ], + "code_body": "1: procedure SELECTIVERENDER $(\\mathcal{G}_{1:K}^{1;l_{max}}$ , scale $l_{r}$ \n2: $S_{1:K} =$ PixelCoverage $(\\mathcal{G}_{1:K}^{1;l_{max}})$ \n3: $G_{1} = \\{\\mathcal{G}_{k}|S_{k} / S_{k}^{max}\\leq S_{rel}^{max},\\forall k\\}$ \n4: $G_{2} = \\{\\mathcal{G}_{k}|S_{k} / S_{k}^{min}\\geq S_{rel}^{min}\\vee S_{k}\\geq S_{T},\\forall k\\}$ \n5: $G_{l_r} = \\{\\mathcal{G}_k^l |l = l_r,\\forall k\\}$ \n6: for $\\mathcal{G}_k^l\\in G_{l_r}$ do \n7: UpdateRange $(S_k^{max},S_k^{min},S_k)$ \n8: end for \n9: return Render $(G_{1}\\cap G_{2},l_{r})$ \n10: end procedure", + "bbox": [ + 508, + 111, + 862, + 268 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3ef8eb5df30babd6aad4d601b3c4ed24737a4c2966164cf7fe7dbe60a894dad4.jpg", + "image_caption": [ + "Figure 6. Based on the rendering resolution, the current pixel coverage of a Gaussian relative to its minimum and maximum pixel coverages determines whether it is selected for rendering." + ], + "image_footnote": [], + "bbox": [ + 511, + 281, + 635, + 383 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d33891ce746d130d36ee262969176b16882a03181ef012ddc9d92f638de77e71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 281, + 743, + 383 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3f0e93130cf10cc2fef179dc7662b711d1d94d1dce1ee2bcaa7e831775bcd9d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 745, + 281, + 880, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the downsample scale when the Gaussian $\\mathcal{G}_k$ is created, its $S_{k}^{max}$ and $S_{k}^{min}$ values are updated with the new pixel coverage $S_{k}$ :", + "bbox": [ + 498, + 455, + 890, + 501 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {k} ^ {\\prime \\max } = \\max \\left(\\lambda_ {1} S _ {k} ^ {\\max }, S _ {k}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 511, + 890, + 537 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {k} ^ {\\prime \\min } = \\operatorname {m i n} \\left(\\lambda_ {2} S _ {k} ^ {\\min }, S _ {k}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 536, + 789, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are decay coefficients taking the empirical value of 0.95 and 1.05, respectively.", + "bbox": [ + 500, + 561, + 890, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During rendering at any resolution or camera distance, a Gaussian is selected for rendering if its pixel coverage $S_{k}$ on the screen satisfies the following condition:", + "bbox": [ + 498, + 592, + 890, + 637 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\frac {S _ {k}}{S _ {k} ^ {\\text {m a x}}} \\leq S _ {\\text {r e l}} ^ {\\text {m a x}}\\right) \\wedge \\left(\\frac {S _ {k}}{S _ {k} ^ {\\text {m i n}}} \\geq S _ {\\text {r e l}} ^ {\\text {m i n}} \\vee S _ {k} \\geq S _ {T}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 645, + 890, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S_{rel}^{max}$ and $S_{rel}^{min}$ are the maximum and minimum relative pixel coverage taking the empirical values of 1.5 and 0.5 respectively. If the pixel coverage of a Gaussian is too much larger than the $S_{k}^{max}$ , it is filtered out from rendering. Similarly, if it is too much smaller than the $S_{k}^{min}$ and is smaller than $S_{T}$ , it is filtered out from rendering (cf. Fig. 6). The absolute $S_{T}$ threshold is used to preserve the large Gaussians from the lower scales, as they do not cause the aliasing problem if their screen size is not sufficiently small. This selective rendering procedure is described in Algorithm 2. Additionally, even if the Gaussians from the finest level are too large or the Gaussians from the coarsest level are too small, they are not filtered to render beyond the maximum and below the minimum training resolutions.", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20927", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6fbd51d2dde1dfe3420778ad9e2444556b3e84540aa249d9da85415cce6f848a.jpg", + "table_caption": [ + "Table 1. Quantitative comparison and ablation study on the 360 dataset [3] at various downsampled scales, with time in \"ms\"." + ], + "table_footnote": [], + "table_body": "
Scale1x4x16x64x
MetricPSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓
3D Gaussian[12]27.520.14210.522.500.1379.317.790.14927.915.23N.A.103.3
3DGS + MS Train27.350.15511.323.500.1267.720.210.11522.819.38N.A.84.8
3DGS + Filter Small27.400.15310.023.810.1495.420.020.1864.817.38N.A.4.6
3DGS + Insert Large18.020.6049.718.750.5312.520.230.2562.721.53N.A.7.1
Our Full Method27.390.1559.124.820.1325.424.750.0664.925.35N.A.4.9
", + "bbox": [ + 80, + 104, + 890, + 233 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bd91888460e12433ce89f9a9e4c6432c309b40e2416dcef448e3e45f61aaa875.jpg", + "table_caption": [ + "Table 2. Quantitative comparison and ablation study on Tank and Temples dataset [13] at various downsampled scales, with time in \"ms\"." + ], + "table_footnote": [], + "table_body": "
Scale1x4x16x64x
MetricPSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓
3D Gaussian[12]23.740.0966.519.700.10511.115.610.06843.413.88N.A.82.6
3DGS + MS Train22.970.1186.021.460.0869.618.560.04937.416.54N.A.71.7
3DGS + Filter Small23.780.1005.620.120.1074.517.410.0724.414.95N.A.4.7
3DGS + Insert Large10.840.6975.111.150.7031.711.730.4471.712.62N.A.2.5
Our Full Method23.460.1117.621.920.0874.720.910.0344.819.67N.A.5.9
", + "bbox": [ + 80, + 261, + 890, + 388 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f112a1b1a8540c8a2061f7d6b4fa02b8dfac459ec0a02af4d7c35597d9aa2fa2.jpg", + "table_caption": [ + "Table 3. Quantitative comparison and ablation study on the Deep Blending dataset [8] at various downsampled scales, with time in \"ms\"." + ], + "table_footnote": [], + "table_body": "
Scale1x4x16x64x
MetricPSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓
3D Gaussian[12]29.650.0948.627.480.0667.522.060.06720.717.75N.A.59.7
3DGS + MS Train29.460.1026.628.180.0625.324.130.05514.320.03N.A.41.3
3DGS + Filter Small29.680.0956.728.260.0644.224.520.0783.618.29N.A.3.2
3DGS + Insert Large20.590.3794.620.830.3361.621.290.1432.120.10N.A.4.2
Our Full Method29.700.0967.428.430.0643.927.660.0363.425.70N.A.3.4
", + "bbox": [ + 80, + 417, + 890, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The pixel coverage range of each Gaussian allows the model to maintain multi-scale Gaussians for different levels of detail. The appropriate subset of Gaussians is chosen for rendering at different resolutions and distances. More smaller Gaussians encoding the high-frequency information are rendered at high resolution, and fewer and larger Gaussians encoding the low-frequency information are rendered at low resolution for less aliasing effect and faster speed.", + "bbox": [ + 75, + 570, + 468, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 708, + 209, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present a comprehensive evaluation of our proposed model, which is grounded on the implementation framework of the official release of the 3D Gaussian Splatting code. To achieve a similar training time as the baseline model, our models are trained for 40000 iterations with all other hyper-parameters unchanged. All rendering speed are measured on a single RTX3090 GPU. We evaluate the performance of the vanilla 3D Gaussian Splatting[12] algorithm and our model on the multi-scale 360[3], Tank And Temples[13], and Deep Blending[8] dataset, aligned with the data used by the original paper. These datasets", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cover a wide range of object centric, indoor, and ourdoor scenes.", + "bbox": [ + 496, + 570, + 890, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our evaluation focuses on the rendering quality and speed at multiple downsampling scales of 1x, 4x, 16x, and 64x derived from the test views. The rendering quality is measured in PSNR and LPIPS, while the speed is measured in per-image rendering time. This multi-scale evaluation is aimed at simulating the rendering performance in scenarios of low-resolution imaging or when captured from distant cameras. More detailed evaluations, including the results for more resolution scales and per-scene decomposition, are included in the supplementary materials due to the space constraint. Additionally, the supplementary materials include a video that offers an intuitive qualitative comparison of the two algorithms, vividly demonstrating the improvement of our algorithm in quality and speed from multiple viewpoints.", + "bbox": [ + 496, + 602, + 892, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison. As shown in Tab. 1, Tab. 2, and Tab. 3, our method can achieve substantial quality and speed improvements compared to the original 3D Gaussian", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20928", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f723dfbd9a223345a651ef437a8e7819e18e7aa3a6e3cc53296f93fbdaa5c678.jpg", + "image_caption": [ + "Figure 7. Qualitative Comparison on 360 dataset[3] for different resolution scales." + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 893, + 372 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26d4d8d03d4daf995105f5f09a4bc0161a1f186a7e0884cc8cbc7057b76aa4ea.jpg", + "image_caption": [ + "Figure 8. Qualitative Comparison on Tank and Temples dataset[13] for different resolution scales." + ], + "image_footnote": [], + "bbox": [ + 86, + 404, + 893, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Splatting [12] at lower resolutions. The quality and speed improvements become more pronounced as the resolution reduces, with the most noticeable 6-10dB PSNR and $20 - 30 \\times$ speed gain at the $64 \\times$ resolution scale. As the resolution reduces, the original splatting algorithm slows down while our method accelerates. The rendering quality and speed at the original resolution $(1 \\times)$ remain comparable, indicating the effectiveness of our multi-scale Gaussians in representing both the high and low resolutions together.", + "bbox": [ + 75, + 732, + 470, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison. We present the qualitative comparison with the original 3D Gaussian Splatting [12] shown in Fig. 7, Fig. 8, and Fig. 9. At higher resolutions $(1\\times -8\\times)$ , both ours and the original algorithm can render the novel view rather faithfully. However, as the resolution reduces further $(16\\times -64\\times)$ , the original splatting algorithm produces severe artifacts, where the foreground becomes larger and larger, dominating the pixel colors as explained in Sec. 3.2. In contrast, the images rendered by our method closely resemble the ground truth across all resolution scales.", + "bbox": [ + 496, + 732, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20929", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc2dbc57a5b2a45cad06e6c4e0919210a18c5c5bf10db833887ba469a4427e86.jpg", + "image_caption": [ + "Figure 9. Qualitative Comparison on Deep Blending dataset[8] for different resolution scales." + ], + "image_footnote": [], + "bbox": [ + 84, + 85, + 893, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablations. To evaluate the effectiveness of the different components proposed, we present the ablation quantitative results in Tab. 1, Tab. 2, and Tab. 3 and the ablation qualitative results in the supplementary. The three ablation methods evaluated are named \"3DGS+MS Train\", \"3DGS+Filter Small\", and \"3DGS+Insert Large\". \"3DGS+MS Train\" reports the result with multi-scale training on top of the original 3D Gaussian splatting. The \"3DGS+Filter Small\" reports the result with small Gaussian filtering using pixel coverage and the multi-scale training, which is needed to update the maximum and minimum pixel coverage. Similarly, the \"3DGS+Insert Large\" reports the result with large Gaussian insertion and the multi-scale training.", + "bbox": [ + 75, + 382, + 468, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ablation results reflect that multi-scale training marginally improves low-resolution rendering quality, but the rendering speed remains very slow. When filtering out the small Gaussians with multi-scale training, the speed at low resolution is increased by $20 - 30 \\times$ with minimal rendering quality loss. The speed gain is caused by the considerably fewer Gaussians rendered. When inserting the large Gaussians and training with multi-scale supervision, without the small Gaussian filtering, the rendering quality drops significantly because the details of the scene are covered with large Gaussians completely for all resolutions. However, when adding the large Gaussians together with the small Gaussian filtering, the rendering quality and speed at low resolution are enhanced significantly without jeopardizing the high-resolution quality. This indicates the effectiveness of all three components and the full method proposed.", + "bbox": [ + 75, + 580, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 76, + 843, + 199, + 858 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since all Gaussian filtering of our proposed method relies on the pixel coverage, it can only be done after the initial", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "splatting process when the coverage is calculated. Although the splatting of individual Gaussians are performed in parallel and does not take more time at lower resolution, it is still a considerable overhead when rendering at a very low resolution. Even if a very small portion of the Gaussians are used for rendering in the end, all Gaussians still need to be splatted. This is the main reason why our rendering time is not decreased linearly as the resolution decreases. In our future work, we will look into a more lightweight criteria to filter small and large Gaussians before splating them onto the screen to achieve an even faster rendering speed.", + "bbox": [ + 496, + 382, + 890, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 556, + 617, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyzed the cause of the severe aliasing artifact and speed degradation of the existing 3D Gaussian splatting. We identified the key challenge of mitigating the aliasing for 3D Gaussian splatting lies in representing the scene with Gaussians of appropriate scale. Based on this observation, we propose to calculate the pixel coverage of 3D Gaussians during splatting and use this as a criteria for selective rendering. Gaussians that are too large or too small at the current rendering resolution are filtered for anti-aliasing and speed improvements. We also proposed to insert large Gaussians by aggregating small Gaussians during training to preserve the low frequency details and prevent part missing. Our experiments on various datasets support the effectiveness of our algorithm in rendering quality and speed at both high and low resolution, mitigating the severe aliasing artifact of the original 3D Gaussian splatting.", + "bbox": [ + 496, + 580, + 890, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021).", + "bbox": [ + 496, + 844, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20930", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kurt Akeley. Reality engine graphics. In Proceedings of the 20th Annual Conference on Computer Graphics and Interactive Techniques, page 109-116, New York, NY, USA, 1993. Association for Computing Machinery. 2", + "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 1, 2", + "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2, 6, 7", + "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 2", + "[5] Kristof Beets and David L. Barron. Super-sampling anti-aliasing analyzed. 2000. 2", + "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2", + "[7] Carl Erikson. *Polygonal simplification.* Technical Report 96-016, 1996. 2", + "[8] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Trans. Graph., 37(6), 2018. 6, 8", + "[9] Tan Kim Heok and D. Daman. A review on level of detail. In Proceedings. International Conference on Computer Graphics, Imaging and Visualization, 2004. CGIV 2004., pages 70-75, 2004. 2", + "[10] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19774–19783, 2023. 3", + "[11] Jorge Jimenez, Diego Gutierrez, Jason Yang, Alexander Reshetov, Pete Demoreuille, Tobias Berghoff, Cedric Perthuis, Henry Yu, Morgan Mcguire, Timothy Lottes, Hugh Malan, and Emil Persson. Filtering approaches for real-time anti-aliasing. ACM SIGGRAPH 2011 Courses, SIGGRAPH'11, 2011. 2", + "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 6, 7", + "[13] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 6, 7", + "[14] Timothy Lottes. *Fxaa. Technical report*, Nvidia, 2011. 2", + "[15] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[16] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1, 2", + "[17] Seungtae Nam, Daniel Rho, Jong Hwan Ko, and Eunbyung Park. Mip-grid: Anti-aliased grid representations for neural radiance fields. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3", + "[18] Lance Williams. Pyramidal parametrics. SIGGRAPH Comput. Graph., 17(3):1-11, 1983. 2", + "[19] M. Zwicker, H. Pfister, J. van Baar, and M. Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS '01., pages 29-538, 2001. 2, 3, 4" + ], + "bbox": [ + 503, + 92, + 890, + 247 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20931", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/20977810-85e5-40f1-a4c6-cf52c533cece_model.json b/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/20977810-85e5-40f1-a4c6-cf52c533cece_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c4a6ec428854953e8092df09bc4279cd40222aa5 --- /dev/null +++ b/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/20977810-85e5-40f1-a4c6-cf52c533cece_model.json @@ -0,0 +1,1692 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.17, + 0.131, + 0.803, + 0.154 + ], + "angle": 0, + "content": "Multi-Scale 3D Gaussian Splating for Anti-Aliased Rendering" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.182, + 0.324, + 0.198 + ], + "angle": 0, + "content": "Zhiwen Yan" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.182, + 0.484, + 0.198 + ], + "angle": 0, + "content": "Weng Fei Low" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.182, + 0.596, + 0.197 + ], + "angle": 0, + "content": "Yu Chen" + }, + { + "type": "text", + "bbox": [ + 0.637, + 0.182, + 0.745, + 0.197 + ], + "angle": 0, + "content": "Gim Hee Lee" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.2, + 0.753, + 0.218 + ], + "angle": 0, + "content": "Department of Computer Science, National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.22, + 0.851, + 0.234 + ], + "angle": 0, + "content": "yan.zhiwen@u.nus.edu {wengfei.low, chenyu}@comp.nus.edu.sg gimhee.lee@nus.edu.sg" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.251, + 0.895, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.544, + 0.895, + 0.586 + ], + "angle": 0, + "content": "Figure 1. The rendering quality and speed of the original 3D Gaussian splatting[12] deteriorate severely at low resolutions or from distant cameras due to aliasing. Conversely, our multi-scale 3D Gaussians representation utilizes selective rendering to achieve faster \\((160\\% - 2400\\%)\\) at \\(4\\times -128\\times\\) resolution) and more accurate rendering at lower resolutions." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.592, + 0.314, + 0.608 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.902 + ], + "angle": 0, + "content": "3D Gaussians have recently emerged as a highly efficient representation for 3D reconstruction and rendering. Despite its high rendering quality and speed at high resolutions, they both deteriorate drastically when rendered at lower resolutions or from far away camera position. During low resolution or far away rendering, the pixel size of the image can fall below the Nyquist frequency compared to the screen size of each splatted 3D Gaussian and leads to aliasing effect. The rendering is also drastically slowed down by the sequential alpha blending of more splatted Gaussians per pixel. To address these issues, we propose a multi-scale 3D Gaussian splatting algorithm, which maintains Gaussians at different scales to represent the same scene. Higher-resolution images are rendered with more small Gaussians, and lower-resolution images are rendered with fewer larger Gaussians. With similar training time, our algorithm can achieve \\(13\\% -66\\%\\) PSNR and \\(160\\% -2400\\%\\) rendering speed improvement at \\(4\\times -128\\times\\) scale rendering" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.594, + 0.892, + 0.641 + ], + "angle": 0, + "content": "on Mip-NeRF360 dataset compared to the single scale 3D Gaussian splatting. More results and code are released on our project page." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.661, + 0.633, + 0.677 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.854 + ], + "angle": 0, + "content": "3D Gaussian Splatting[12] has recently emerged as a highly efficient representation for novel view synthesis. Compared to the time-consuming ray marching used in most neural radiance fields (NeRF) [2, 15, 16], a high-resolution image can be rendered in real-time by rasterizing the splatted 3D Gaussians. However, this rasterization algorithm is subjected to severe aliasing effect and speed deterioration when rendering the same scene at low resolution or from distant positions as shown in Fig. 1. This limitation significantly constrains the application of the 3D Gaussian splatting algorithm in reconstructing and rendering large-scale scenes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Aliasing effect is a consequence of inadequate sampling frequency failing to capture the continuous signal accurately. In the context of rendering, image pixels are sampled" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20923" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "with an interval of one-pixel size. The signal can be considered as the 3D scene represented implicitly as in NeRF or explicitly as in 3D Gaussians. When part of the 3D scene is represented with high details but rendered with low resolution or from distant positions, the disparity between the low sampling and high signal frequencies culminates in aliasing artifacts. A naive solution is to render at high resolution and subsequently down-scale the rendered image to a lower resolution. However, this solution is not viable for scenes containing both near and far regions which are very common. Due to the inability of 3D Gaussian splatting algorithm to accommodate varying resolutions within a single image, rendering the entire image with a even higher resolution for the sake of far away regions is neither time nor memory efficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.53 + ], + "angle": 0, + "content": "We postulate that the pronounced aliasing artifacts observed when rendering with 3D Gaussians, as opposed to other techniques such as NeRF, are primarily attributable to the splatting of small Gaussians. 3D regions with intricate details are represented with large amount of small Gaussians. When rendering these regions with low resolution or from a distant view, many splatted small Gaussians are cramped in one pixel and therefore the pixel color of this region is dominated by the front-most Gaussian, even if this Gaussian is much smaller than others and not at the center. This problem is further aggravated by the low pass filter in [12, 19] applied to each individual Gaussian with the intention to mitigate aliasing on edges at high resolutions. This problem is explained in more detail in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.473, + 0.698 + ], + "angle": 0, + "content": "In addition to the aliasing artifacts, the rendering speed of 3D Gaussians is also affected at low resolution. The number of 3D Gaussians that need to be rendered remains constant at lower resolutions, but they are more concentrated to fewer pixels. The Gaussians that are splatted to the same pixel cannot be rendered in parallel. This means that the image rendering is even slower at lower resolution in comparison with NeRF rendering time that reduces linearly with decreasing resolution. Hence, although aliasing is not a problem exclusive to 3D Gaussian splattering, it is more prominent and more difficult to tackle." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Contributions To mitigate the aliasing problem for 3D Gaussian splatting, we propose a novel multi-scale 3D Gaussians to represent the scene at different levels of detail (LOD) as shown in Fig. 2. This is inspired by the mipmap and LOD algorithms widely used in computer graphics, which pre-computes textures and polygons at different scales to be rendered under different resolutions and distances. Similarly, we add larger, coarser Gaussians for lower resolutions by aggregating the smaller and finer Gaussians from higher resolutions. Depending on the pixel coverage of the splatted Gaussians during rendering, only a subset of the Gaussians is used. A simplified explanation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.214 + ], + "angle": 0, + "content": "for this is that the coarse Gaussians are used to render low-resolution images and the fine Gaussians are used to render high-resolution images. With fewer than \\(5\\%\\) number of Gaussians added and a similar training time, our method can achieve \\(13\\% - 66\\%\\) PSNR and \\(160\\% - 2400\\%\\) rendering speed improvements at \\(4 \\times -128 \\times\\) scale rendering on MipNeRF360 dataset[2], while maintaining a comparable rendering quality and speed at \\(1 \\times\\) scale." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.228, + 0.651, + 0.243 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.253, + 0.819, + 0.27 + ], + "angle": 0, + "content": "2.1. Anti-Aliasing in Computer Graphics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.895, + 0.563 + ], + "angle": 0, + "content": "Ailiasing is a long-standing problem for computer graphics when rendering a scene to a discrete image. Traditional anti-aliasing techniques primarily target mesh representations. Supersampling Anti-Aliasing (SSAA) [5] renders the scene at a higher resolution before downscaling, leading to significantly more time and memory demain, and therefore is less used in real-time applications. The Multisample Anti-Aliasing (MSAA) [1, 5] algorithm selectively supersamples pixels on the edges, reducing resource and time consumption. This technique is not very suitable for 3D Gaussian splatting because of its requirement for regular grids and lack of support for variable sampling resolution at different pixels. The more recent Fast Approximate Anti-Aliasing (FXAA) [11, 14] is a post processing algorithm that smooths the jagged edges after the image is rendered. Unfortunately, this technique is also not suitable for Gaussian representation as the front-most Gaussian dominates the pixel color and produces chunky instead of jagged artifacts in mesh rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.896, + 0.731 + ], + "angle": 0, + "content": "In contrast to the supersampling methods mentioned above, our method takes the inspiration from hierarchical mipmap [18] and level of details (LOD) [7, 9] algorithms to address the aliasing for 3D Gaussians. Mipmap uses multi-scale textures for the rendering at different resolution or from different distances. LOD algorithm represents the models in a scene with different complexity to be rendered at different distances. Both techniques not only mitigate the aliasing effect by reducing the complexity of the scene representation, but also enhances rendering speed, particularly for large-scale scenes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.741, + 0.838, + 0.758 + ], + "angle": 0, + "content": "2.2. Anti-Aliasing in Neural Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The recent success of neural representations especially Neural Radiance Fields (NeRF) [6, 15, 16] has also inspired some works to develop algorithms against aliasing effect on neural representations beyond the traditional mesh representation. Mip-NeRF [2, 3] employ low pass filters on the positional encoding of the input spatial coordinates to reduce the scene signal frequency. Building on the hash grid representation used by InstantNGP [16] with no position encoding, Zip-NeRF [4] proposes a multi-sampling" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20924" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.089, + 0.43, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.089, + 0.85, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.297, + 0.895, + 0.355 + ], + "angle": 0, + "content": "Figure 2. Overall pipeline of our algorithm. At the early stage of training (left), small Gaussians below certain size threshold in each voxel are aggregated, enlarged and inserted into the scene at different resolution scale. During rendering (right), the multi-scale Gaussians of the appropriate \"pixel coverage\" at the current render resolution are selected for rendering. If the rendering resolution scale equals to the scale of the Gaussians, the expected \"pixel coverage\" range of the Gaussians are updated accordingly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.47, + 0.444 + ], + "angle": 0, + "content": "strategy in the conical frustum instead of the camera ray, at the cost of \\(6 \\times\\) rendering time. Similar to the mipmap algorithm in mesh texture rendering, Tri-MipNeRF [10] and MipGrid [17] proposes to use multi-scale feature grids for rendering at different resolution or distance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.471, + 0.596 + ], + "angle": 0, + "content": "Conversely, 3D Gaussian splatting [12] presents unique anti-aliasing challenges due to its distinct scene representation. It does not have any positional encoding or feature grid, and its requirement for regular grids conflicts the more flexible multi-sampling strategies. The concentration of small Gaussians in detail-rich regions exacerbates aliasing and speed issues, more so than in NeRF representations. To the best of our knowledge, we are the first to propose an anti-aliasing algorithm for scene reconstruction using 3D Gaussian splatting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.607, + 0.218, + 0.623 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.632, + 0.287, + 0.649 + ], + "angle": 0, + "content": "3.1. 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.839 + ], + "angle": 0, + "content": "3D Gaussian splatting is first proposed in EWA Splatting [19], and later used by [12] for scene reconstruction and novel view synthesis. The scene is represented by a set of \\(\\mathbf{K}\\) 3D Gaussians \\(\\{\\mathcal{G}_{\\hat{V}^k,\\hat{\\mu}^k},\\sigma_k,\\mathbf{c}_k\\mid k\\in [1,\\mathbf{K}]\\}\\) with variance \\(\\hat{V}^k\\), center \\(\\hat{\\mu}^k\\), density \\(\\sigma_{k}\\) and color \\(\\mathbf{c}^k\\). During rendering, the 3D Gaussians are splatted to the 2D screen to by the perspective transformation to form 2D Gaussians \\(\\mathcal{G}_{V^k,\\mu^k}\\). The image is then divided into \\(16\\times 16\\) regular tiles and all 2D Gaussians touching each tile are sorted based on their original depth. The color of each pixel in the tile is then rasterized from the sequential alpha blending the 2D Gaussians from front to back." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.445, + 0.865 + ], + "angle": 0, + "content": "3.2. Cause of Aliasing in 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Aliasing can occur when sampling a continuous signal \\( g(x) \\) with a discrete sampling function \\( \\delta_s(x, \\Delta x) = \\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.368, + 0.892, + 0.4 + ], + "angle": 0, + "content": "\\(\\sum_{n = -\\infty}^{\\infty}\\delta (x - n\\cdot \\Delta x)\\), where \\(\\delta\\) is a impulse function. The result of the sample in the spatial domain is:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.408, + 0.892, + 0.426 + ], + "angle": 0, + "content": "\\[\ng _ {s} (x) = \\delta_ {s} (x, \\Delta x) \\cdot g (x). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.434, + 0.892, + 0.464 + ], + "angle": 0, + "content": "This sampled function converted into the frequency domain using Fourier transform operator \\(\\mathcal{F}\\) becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.473, + 0.892, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {F} \\left[ g _ {s} (u) \\right] = \\frac {1}{\\Delta x} \\sum_ {\\substack {k = - \\infty \\\\ \\infty}} ^ {\\infty} \\delta \\left(u - \\frac {k}{\\Delta x}\\right) * \\mathcal {F} [ g (x) ] \\tag{2} \\\\ = \\frac {1}{\\Delta x} \\sum_ {k = - \\infty} ^ {\\infty} \\mathbf {G} (u - \\frac {k}{\\Delta x}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.564, + 0.892, + 0.672 + ], + "angle": 0, + "content": "When the highest frequency component \\( f_{max} \\) of the signal is greater than half of the sampling frequency \\( f_{s} = \\frac{1}{\\Delta x} \\), \\( \\mathbf{G}(u - \\frac{k}{\\Delta x}) \\) in the summation sequence would overlap with each other and causes the sampled signal to diverge from the actual signal. This phenomenon is the aliasing effect and the minimum sampling frequency needed to avoid aliasing is \\( f_{Ny} = 2 \\cdot f_{max} \\), known as the Nyquist frequency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.763 + ], + "angle": 0, + "content": "The EWA splatting [19] used by 3D Gaussian splatting [12] also tries to mitigate the aliasing problem by applying a low pass filter to the rendered color. To approximate this efficiently, it applies a Gaussian kernel \\( h(\\mathbf{x}) \\) as the low pass filter on each splatted 2D signal \\( g_{c}(\\mathbf{x}) \\) independently to produce a band limited signal:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.77, + 0.892, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} g _ {c} ^ {\\prime} (\\mathbf {x}) = g _ {c} (\\mathbf {x}) * h (\\mathbf {x}) \\\\ \\approx \\sum_ {k} \\sigma_ {k} \\mathbf {c} _ {k} T _ {k} \\int_ {\\mathcal {R} ^ {2}} q _ {k} (\\eta) h (\\mathbf {x} - \\eta) d \\eta \\tag {3} \\\\ = \\sum_ {k} \\sigma_ {k} \\mathbf {c} _ {k} T _ {k} \\cdot (q _ {k} * h) (\\mathbf {x}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.87, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}^2\\) is the range of one pixel, \\(q_{k}(\\mathbf{x})\\) is the 2D integrated Gaussian kernel, and \\(\\sigma_k,\\mathbf{c}_k,T_k\\) are the opacity," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "20925" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "color, and transmittance at each Gaussian, respectively. By combining the reconstruction Gaussian kernel \\(\\mathcal{G}_{V^k}\\) and low pass Gaussian kernel \\(\\mathcal{G}_{V^h}\\) of covariance matrix \\(V^{k}\\) and \\(V^{h}\\), the band limit function becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.161, + 0.387, + 0.227 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} g _ {c} ^ {\\prime} (\\mathbf {x}) = \\sum_ {k} \\alpha_ {k} \\cdot (\\mathcal {G} _ {V ^ {k}} * \\mathcal {G} _ {V ^ {h}}) (\\mathbf {x}) \\\\ = \\sum_ {k} \\alpha_ {k} \\cdot \\mathcal {G} _ {V ^ {k} + V ^ {h}} (\\mathbf {x}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.239, + 0.469, + 0.3 + ], + "angle": 0, + "content": "where \\(\\alpha_{k}\\) represents all coefficients invariant of \\(\\mathbf{x}\\) at each Gaussian and \\(V^{h}\\) is determined by the screen pixel size. A simple understanding of this is that the covariance of each 3D Gaussian is increased based on the screen pixel size." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.469, + 0.374 + ], + "angle": 0, + "content": "This method of applying a low pass filter to each 3D Gaussian independently helps to smooth the edges of the Gaussians when the Gaussians are not too small compared to the pixel size. However, it also gives rise to two substantial issues at low resolutions:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.375, + 0.469, + 0.451 + ], + "angle": 0, + "content": "1. \\( V^{h} \\) added to the original covariance \\( V^{k} \\) effectively increases the extent of each Gaussian, especially when \\( V^{h} \\) is large compared to \\( V^{k} \\) at low resolutions. Small Gaussians in the front dominate the color of the pixel and cause severe artifacts shown in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.451, + 0.469, + 0.526 + ], + "angle": 0, + "content": "2. The number of Gaussians involved in the sequential \\(\\sum_{k}\\) for each pixel scales increases with decreasing image resolution. Due to the incremental calculation of the transmittance \\(T_{k}\\), the rendering even slower at lower resolutions." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.375, + 0.469, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.541, + 0.208, + 0.556 + ], + "angle": 0, + "content": "4. Our Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.566, + 0.469, + 0.582 + ], + "angle": 0, + "content": "4.1. Multi-Scale Gaussians Based on Pixel Coverage" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.469, + 0.756 + ], + "angle": 0, + "content": "To mitigate the aliasing artifacts of 3D Gaussians [12] while avoiding the two problems of the EWA splatting [19], we introduce multi-scale 3D Gaussians (cf. Fig. 2) that tackle the problem on the scene-level instead of on each individual Gaussian. The 3D scene is represented with Gaussians from 4 levels of detail, corresponding to the \\(1 \\times\\), \\(4 \\times\\), \\(16 \\times\\), and \\(64 \\times\\) downsampled resolution. Small finer-level Gaussians are aggregated to create larger Gaussians for coarser levels during training. Each 3D Gaussian \\(\\mathcal{G}_k^l\\) belongs to one of the levels \\(l\\) and is included or excluded independently during the rendering based on its \"pixel coverage\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.904 + ], + "angle": 0, + "content": "Pixel Coverage of Gaussian. The \"pixel coverage\" of a Gaussian reflects the size of the Gaussian when splatted onto the screen space compared to the pixel size at the current rendering resolution. The \"pixel coverage\" \\( S_{k} \\) of a splatted 2D Gaussian \\( \\mathcal{G}_{(\\mu^k,V^k)} \\) is defined as the length of its horizontal or vertical axis until the low opacity level set, whichever is smaller, as shown in Fig. 3. The pixel coverage is measured in pixel count and the opacity threshold \\( \\sigma_T \\) is set as \\( \\frac{1}{255} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.092, + 0.74, + 0.118 + ], + "angle": 0, + "content": "\\[\nu; \\sigma_ {k} \\mathcal {G} _ {(\\mu^ {k}, V ^ {k})} \\left(\\mu_ {x} ^ {k} + \\frac {u}{2}, \\mu_ {y} ^ {k}\\right) = \\sigma_ {T}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.118, + 0.737, + 0.159 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} v \\colon \\sigma_ {k} \\mathcal {G} _ {(\\mu^ {k}, V ^ {k})} \\left(\\mu_ {x} ^ {k}, \\mu_ {y} ^ {k} + \\frac {v}{2}\\right) = \\sigma_ {T} \\\\ S _ {k} = \\min (u, v) \\\\ \\end{array}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.092, + 0.89, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.205, + 0.892, + 0.233 + ], + "angle": 0, + "content": "Figure 3. Pixel coverage of a 3D Gaussian is its horizontal or vertical size, whichever is smaller measured by the level set." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.247, + 0.639, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.247, + 0.755, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.247, + 0.872, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.378, + 0.892, + 0.405 + ], + "angle": 0, + "content": "Figure 4. Missing parts caused by naive small Gaussian filtering at different resolution scales." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.892, + 0.576 + ], + "angle": 0, + "content": "The pixel coverage approximates the extent of a 2D splatted Gaussian in the spatial domain. During the rendering from a given camera direction, the color of each splatted Gaussian is constant within this pixel coverage. As a result, the coverage of this pixel approximates the inverse of the highest frequency component \\( f_{max} = 1 / S_k \\) in this region. Compared to the sampling frequency of \\( f_s = 1\\mathrm{px}^{-1} \\) during rasterization, a signal frequency of \\( f_{max} > f_s / 2 \\) can cause the sampling to fall below the Nyquist frequency needed to avoid aliasing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.727 + ], + "angle": 0, + "content": "Consequently, the Gaussians with pixel coverage \\( S_{k} < S_{T} = 2\\mathrm{px} \\) should be filtered out during rendering to avoid aliasing. Since 3D Gaussian representation does not encode the signal of different frequencies at different Gaussians, naively filtering out the small Gaussians will result in a hole or part missing in the scene as shown in Fig. 4. To address this issue, we propose to aggregate the small Gaussians to form large Gaussians that encode the low-frequency signal. These large Gaussians would appear when the small Gaussians are filtered out." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Aggregate to Insert Large Gaussians. All 3D Gaussians initialized from the input point cloud at the start of the training belong to the finest level \\( l = 1 \\). They are densified by splitting and cloning as in [12], and all the densified Gaussians would inherit the same level. After the warm-up stage of the first 1,000 iterations, we introduce coarse-level Gaussians by aggregating fine-level Gaussians that are too small as visualized in Fig. 5 and described in Algorithm 1. The procedure is outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1. For all levels \\(\\{l_m\\mid 2\\leq l_m\\leq l_{max}\\}\\) , we render all 3D Gaussians from \\([1,l_m - 1]\\) at the \\(4^{l_m - 1}\\) times downsam" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20926" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.473, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.215, + 0.47, + 0.258 + ], + "angle": 0, + "content": "Figure 5. Large Gaussians are created by aggregating the small Gaussians in each voxel below the pixel coverage threshold, and then enlarged by the pixel coverage multiplier." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.285, + 0.469, + 0.33 + ], + "angle": 0, + "content": "pled resolution of all training images. All 3D Gaussians with the minimal \"pixel coverage\" \\( S_{k} \\) smaller than the filter threshold \\( S_{T} \\) are chosen for the aggregation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.33, + 0.469, + 0.42 + ], + "angle": 0, + "content": "2. The chosen 3D Gaussians are binned by a \\((400 / l_{m})^{3}\\) resolution voxel grid based on their positions. The attributes of all Gaussians within each voxel are aggregated to create a new Gaussian using average pooling, including position, scaling, opacity and color. More details are included in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.421, + 0.469, + 0.496 + ], + "angle": 0, + "content": "3. Based on the average \"pixel coverage\" \\( S_{avg} \\) of Gaussians in each voxel, the scaling of each new Gaussian created is enlarged by \\( S_T / S_{avg} \\) so that it is of a size suitable to be rendered at \\( l_m \\). This new Gaussian belongs to level \\( l_m \\)." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.33, + 0.469, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.47, + 0.572 + ], + "angle": 0, + "content": "Not all Gaussians from the fine levels are small. Many Gaussians in the background or in the textureless regions are large and do not need to be aggregated. The number of Gaussians created is often fewer than \\(5\\%\\) of the final total number of Gaussians." + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.586, + 0.351, + 0.601 + ], + "angle": 0, + "content": "Algorithm 1 Aggregate Small Gaussians" + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.607, + 0.475, + 0.778 + ], + "angle": 0, + "content": "
1:procedure AGGREGATEGAUSSIANS(G1:lmax)
2:for lm←2 to lmax do
3:S1:K=PixelCoverage(G1:lm-1, scale 4lmax-1)
4:Gsmall = {Gk|Sk < ST, ∀k ∈ [1:K]}
5:for n←1 to (400/lm)3 do
6:Gn = {Gk|Gk in voxel n, ∀Gk ∈ Gsmall}
7:Glmnew,n = Enlarge(Average(Gn))
8:InsertIntoScene(Glmnew,n)
9:end for
10:end for
11:end procedure
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Multi-Scale Training and Selective Rendering. After the large Gaussians are added, the model is trained with both the original images and the downsampled images. A maximum pixel coverage \\( S_{k}^{max} \\) and a minimum pixel coverage \\( S_{k}^{min} \\) of each Gaussian are stored for the selective rendering. If the rendering downsample scale equals to" + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.892, + 0.107 + ], + "angle": 0, + "content": "Algorithm 2 Selective Rendering Based on Pixel Coverage" + }, + { + "type": "algorithm", + "bbox": [ + 0.509, + 0.112, + 0.864, + 0.27 + ], + "angle": 0, + "content": "1: procedure SELECTIVERENDER \\((\\mathcal{G}_{1:K}^{1;l_{max}}\\) , scale \\(l_{r}\\) \n2: \\(S_{1:K} =\\) PixelCoverage \\((\\mathcal{G}_{1:K}^{1;l_{max}})\\) \n3: \\(G_{1} = \\{\\mathcal{G}_{k}|S_{k} / S_{k}^{max}\\leq S_{rel}^{max},\\forall k\\}\\) \n4: \\(G_{2} = \\{\\mathcal{G}_{k}|S_{k} / S_{k}^{min}\\geq S_{rel}^{min}\\vee S_{k}\\geq S_{T},\\forall k\\}\\) \n5: \\(G_{l_r} = \\{\\mathcal{G}_k^l |l = l_r,\\forall k\\}\\) \n6: for \\(\\mathcal{G}_k^l\\in G_{l_r}\\) do \n7: UpdateRange \\((S_k^{max},S_k^{min},S_k)\\) \n8: end for \n9: return Render \\((G_{1}\\cap G_{2},l_{r})\\) \n10: end procedure" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.282, + 0.637, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.282, + 0.745, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.282, + 0.882, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.4, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Figure 6. Based on the rendering resolution, the current pixel coverage of a Gaussian relative to its minimum and maximum pixel coverages determines whether it is selected for rendering." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.457, + 0.892, + 0.502 + ], + "angle": 0, + "content": "the downsample scale when the Gaussian \\(\\mathcal{G}_k\\) is created, its \\(S_{k}^{max}\\) and \\(S_{k}^{min}\\) values are updated with the new pixel coverage \\(S_{k}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.512, + 0.891, + 0.538 + ], + "angle": 0, + "content": "\\[\nS _ {k} ^ {\\prime \\max } = \\max \\left(\\lambda_ {1} S _ {k} ^ {\\max }, S _ {k}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.537, + 0.79, + 0.552 + ], + "angle": 0, + "content": "\\[\nS _ {k} ^ {\\prime \\min } = \\operatorname {m i n} \\left(\\lambda_ {2} S _ {k} ^ {\\min }, S _ {k}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.562, + 0.891, + 0.592 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are decay coefficients taking the empirical value of 0.95 and 1.05, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.593, + 0.891, + 0.638 + ], + "angle": 0, + "content": "During rendering at any resolution or camera distance, a Gaussian is selected for rendering if its pixel coverage \\( S_{k} \\) on the screen satisfies the following condition:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.646, + 0.892, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\left(\\frac {S _ {k}}{S _ {k} ^ {\\text {m a x}}} \\leq S _ {\\text {r e l}} ^ {\\text {m a x}}\\right) \\wedge \\left(\\frac {S _ {k}}{S _ {k} ^ {\\text {m i n}}} \\geq S _ {\\text {r e l}} ^ {\\text {m i n}} \\vee S _ {k} \\geq S _ {T}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( S_{rel}^{max} \\) and \\( S_{rel}^{min} \\) are the maximum and minimum relative pixel coverage taking the empirical values of 1.5 and 0.5 respectively. If the pixel coverage of a Gaussian is too much larger than the \\( S_{k}^{max} \\), it is filtered out from rendering. Similarly, if it is too much smaller than the \\( S_{k}^{min} \\) and is smaller than \\( S_{T} \\), it is filtered out from rendering (cf. Fig. 6). The absolute \\( S_{T} \\) threshold is used to preserve the large Gaussians from the lower scales, as they do not cause the aliasing problem if their screen size is not sufficiently small. This selective rendering procedure is described in Algorithm 2. Additionally, even if the Gaussians from the finest level are too large or the Gaussians from the coarsest level are too small, they are not filtered to render beyond the maximum and below the minimum training resolutions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20927" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.089, + 0.856, + 0.103 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison and ablation study on the 360 dataset [3] at various downsampled scales, with time in \"ms\"." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.105, + 0.892, + 0.234 + ], + "angle": 0, + "content": "
Scale1x4x16x64x
MetricPSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓
3D Gaussian[12]27.520.14210.522.500.1379.317.790.14927.915.23N.A.103.3
3DGS + MS Train27.350.15511.323.500.1267.720.210.11522.819.38N.A.84.8
3DGS + Filter Small27.400.15310.023.810.1495.420.020.1864.817.38N.A.4.6
3DGS + Insert Large18.020.6049.718.750.5312.520.230.2562.721.53N.A.7.1
Our Full Method27.390.1559.124.820.1325.424.750.0664.925.35N.A.4.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.246, + 0.89, + 0.26 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison and ablation study on Tank and Temples dataset [13] at various downsampled scales, with time in \"ms\"." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.262, + 0.892, + 0.39 + ], + "angle": 0, + "content": "
Scale1x4x16x64x
MetricPSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓
3D Gaussian[12]23.740.0966.519.700.10511.115.610.06843.413.88N.A.82.6
3DGS + MS Train22.970.1186.021.460.0869.618.560.04937.416.54N.A.71.7
3DGS + Filter Small23.780.1005.620.120.1074.517.410.0724.414.95N.A.4.7
3DGS + Insert Large10.840.6975.111.150.7031.711.730.4471.712.62N.A.2.5
Our Full Method23.460.1117.621.920.0874.720.910.0344.819.67N.A.5.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.403, + 0.89, + 0.417 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison and ablation study on the Deep Blending dataset [8] at various downsampled scales, with time in \"ms\"." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.419, + 0.892, + 0.546 + ], + "angle": 0, + "content": "
Scale1x4x16x64x
MetricPSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓
3D Gaussian[12]29.650.0948.627.480.0667.522.060.06720.717.75N.A.59.7
3DGS + MS Train29.460.1026.628.180.0625.324.130.05514.320.03N.A.41.3
3DGS + Filter Small29.680.0956.728.260.0644.224.520.0783.618.29N.A.3.2
3DGS + Insert Large20.590.3794.620.830.3361.621.290.1432.120.10N.A.4.2
Our Full Method29.700.0967.428.430.0643.927.660.0363.425.70N.A.3.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.571, + 0.47, + 0.692 + ], + "angle": 0, + "content": "The pixel coverage range of each Gaussian allows the model to maintain multi-scale Gaussians for different levels of detail. The appropriate subset of Gaussians is chosen for rendering at different resolutions and distances. More smaller Gaussians encoding the high-frequency information are rendered at high resolution, and fewer and larger Gaussians encoding the low-frequency information are rendered at low resolution for less aliasing effect and faster speed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.709, + 0.21, + 0.725 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this section, we present a comprehensive evaluation of our proposed model, which is grounded on the implementation framework of the official release of the 3D Gaussian Splatting code. To achieve a similar training time as the baseline model, our models are trained for 40000 iterations with all other hyper-parameters unchanged. All rendering speed are measured on a single RTX3090 GPU. We evaluate the performance of the vanilla 3D Gaussian Splatting[12] algorithm and our model on the multi-scale 360[3], Tank And Temples[13], and Deep Blending[8] dataset, aligned with the data used by the original paper. These datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.892, + 0.6 + ], + "angle": 0, + "content": "cover a wide range of object centric, indoor, and ourdoor scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.893, + 0.83 + ], + "angle": 0, + "content": "Our evaluation focuses on the rendering quality and speed at multiple downsampling scales of 1x, 4x, 16x, and 64x derived from the test views. The rendering quality is measured in PSNR and LPIPS, while the speed is measured in per-image rendering time. This multi-scale evaluation is aimed at simulating the rendering performance in scenarios of low-resolution imaging or when captured from distant cameras. More detailed evaluations, including the results for more resolution scales and per-scene decomposition, are included in the supplementary materials due to the space constraint. Additionally, the supplementary materials include a video that offers an intuitive qualitative comparison of the two algorithms, vividly demonstrating the improvement of our algorithm in quality and speed from multiple viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Quantitative Comparison. As shown in Tab. 1, Tab. 2, and Tab. 3, our method can achieve substantial quality and speed improvements compared to the original 3D Gaussian" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20928" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.088, + 0.895, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.379, + 0.728, + 0.394 + ], + "angle": 0, + "content": "Figure 7. Qualitative Comparison on 360 dataset[3] for different resolution scales." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.405, + 0.895, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.693, + 0.773, + 0.707 + ], + "angle": 0, + "content": "Figure 8. Qualitative Comparison on Tank and Temples dataset[13] for different resolution scales." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.471, + 0.87 + ], + "angle": 0, + "content": "Splatting [12] at lower resolutions. The quality and speed improvements become more pronounced as the resolution reduces, with the most noticeable 6-10dB PSNR and \\(20 - 30 \\times\\) speed gain at the \\(64 \\times\\) resolution scale. As the resolution reduces, the original splatting algorithm slows down while our method accelerates. The rendering quality and speed at the original resolution \\((1 \\times)\\) remain comparable, indicating the effectiveness of our multi-scale Gaussians in representing both the high and low resolutions together." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.899 + ], + "angle": 0, + "content": "Qualitative Comparison. We present the qualitative comparison with the original 3D Gaussian Splatting [12] shown in Fig. 7, Fig. 8, and Fig. 9. At higher resolutions \\((1\\times -8\\times)\\), both ours and the original algorithm can render the novel view rather faithfully. However, as the resolution reduces further \\((16\\times -64\\times)\\), the original splatting algorithm produces severe artifacts, where the foreground becomes larger and larger, dominating the pixel colors as explained in Sec. 3.2. In contrast, the images rendered by our method closely resemble the ground truth across all resolution scales." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20929" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.087, + 0.895, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.36, + 0.761, + 0.374 + ], + "angle": 0, + "content": "Figure 9. Qualitative Comparison on Deep Blending dataset[8] for different resolution scales." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.47, + 0.579 + ], + "angle": 0, + "content": "Ablations. To evaluate the effectiveness of the different components proposed, we present the ablation quantitative results in Tab. 1, Tab. 2, and Tab. 3 and the ablation qualitative results in the supplementary. The three ablation methods evaluated are named \"3DGS+MS Train\", \"3DGS+Filter Small\", and \"3DGS+Insert Large\". \"3DGS+MS Train\" reports the result with multi-scale training on top of the original 3D Gaussian splatting. The \"3DGS+Filter Small\" reports the result with small Gaussian filtering using pixel coverage and the multi-scale training, which is needed to update the maximum and minimum pixel coverage. Similarly, the \"3DGS+Insert Large\" reports the result with large Gaussian insertion and the multi-scale training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.824 + ], + "angle": 0, + "content": "The ablation results reflect that multi-scale training marginally improves low-resolution rendering quality, but the rendering speed remains very slow. When filtering out the small Gaussians with multi-scale training, the speed at low resolution is increased by \\(20 - 30 \\times\\) with minimal rendering quality loss. The speed gain is caused by the considerably fewer Gaussians rendered. When inserting the large Gaussians and training with multi-scale supervision, without the small Gaussian filtering, the rendering quality drops significantly because the details of the scene are covered with large Gaussians completely for all resolutions. However, when adding the large Gaussians together with the small Gaussian filtering, the rendering quality and speed at low resolution are enhanced significantly without jeopardizing the high-resolution quality. This indicates the effectiveness of all three components and the full method proposed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.844, + 0.2, + 0.859 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Since all Gaussian filtering of our proposed method relies on the pixel coverage, it can only be done after the initial" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.549 + ], + "angle": 0, + "content": "splatting process when the coverage is calculated. Although the splatting of individual Gaussians are performed in parallel and does not take more time at lower resolution, it is still a considerable overhead when rendering at a very low resolution. Even if a very small portion of the Gaussians are used for rendering in the end, all Gaussians still need to be splatted. This is the main reason why our rendering time is not decreased linearly as the resolution decreases. In our future work, we will look into a more lightweight criteria to filter small and large Gaussians before splating them onto the screen to achieve an even faster rendering speed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.558, + 0.618, + 0.573 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.823 + ], + "angle": 0, + "content": "We analyzed the cause of the severe aliasing artifact and speed degradation of the existing 3D Gaussian splatting. We identified the key challenge of mitigating the aliasing for 3D Gaussian splatting lies in representing the scene with Gaussians of appropriate scale. Based on this observation, we propose to calculate the pixel coverage of 3D Gaussians during splatting and use this as a criteria for selective rendering. Gaussians that are too large or too small at the current rendering resolution are filtered for anti-aliasing and speed improvements. We also proposed to insert large Gaussians by aggregating small Gaussians during training to preserve the low frequency details and prevent part missing. Our experiments on various datasets support the effectiveness of our algorithm in rendering quality and speed at both high and low resolution, mitigating the severe aliasing artifact of the original 3D Gaussian splatting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20930" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Kurt Akeley. Reality engine graphics. In Proceedings of the 20th Annual Conference on Computer Graphics and Interactive Techniques, page 109-116, New York, NY, USA, 1993. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.229, + 0.471, + 0.27 + ], + "angle": 0, + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.47, + 0.314 + ], + "angle": 0, + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[5] Kristof Beets and David L. Barron. Super-sampling anti-aliasing analyzed. 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.47, + 0.386 + ], + "angle": 0, + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.47, + 0.415 + ], + "angle": 0, + "content": "[7] Carl Erikson. *Polygonal simplification.* Technical Report 96-016, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[8] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Trans. Graph., 37(6), 2018. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.47, + 0.528 + ], + "angle": 0, + "content": "[9] Tan Kim Heok and D. Daman. A review on level of detail. In Proceedings. International Conference on Computer Graphics, Imaging and Visualization, 2004. CGIV 2004., pages 70-75, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.47, + 0.601 + ], + "angle": 0, + "content": "[10] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19774–19783, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.602, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[11] Jorge Jimenez, Diego Gutierrez, Jason Yang, Alexander Reshetov, Pete Demoreuille, Tobias Berghoff, Cedric Perthuis, Henry Yu, Morgan Mcguire, Timothy Lottes, Hugh Malan, and Emil Persson. Filtering approaches for real-time anti-aliasing. ACM SIGGRAPH 2011 Courses, SIGGRAPH'11, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.47, + 0.741 + ], + "angle": 0, + "content": "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.47, + 0.798 + ], + "angle": 0, + "content": "[13] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.801, + 0.449, + 0.814 + ], + "angle": 0, + "content": "[14] Timothy Lottes. *Fxaa. Technical report*, Nvidia, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[15] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[16] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[17] Seungtae Nam, Daniel Rho, Jong Hwan Ko, and Eunbyung Park. Mip-grid: Anti-aliased grid representations for neural radiance fields. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[18] Lance Williams. Pyramidal parametrics. SIGGRAPH Comput. Graph., 17(3):1-11, 1983. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[19] M. Zwicker, H. Pfister, J. van Baar, and M. Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS '01., pages 29-538, 2001. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20931" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/20977810-85e5-40f1-a4c6-cf52c533cece_origin.pdf b/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/20977810-85e5-40f1-a4c6-cf52c533cece_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e31af2e3ea2c4ddcc4976586757390972e954652 --- /dev/null +++ b/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/20977810-85e5-40f1-a4c6-cf52c533cece_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e12244f605ef4e9ee9adb908bcbb451ae36e7b03ec72059efcbcbf6e4c7d0ae +size 2738049 diff --git a/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/full.md b/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..df1cffd0364d89f4f78d2df71549ae6e80f75c20 --- /dev/null +++ b/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/full.md @@ -0,0 +1,270 @@ +# Multi-Scale 3D Gaussian Splating for Anti-Aliased Rendering + +Zhiwen Yan + +Weng Fei Low + +Yu Chen + +Gim Hee Lee + +Department of Computer Science, National University of Singapore + +yan.zhiwen@u.nus.edu {wengfei.low, chenyu}@comp.nus.edu.sg gimhee.lee@nus.edu.sg + +![](images/22407358080dc2e589a148aa68a17808d0b136374d0c8104810b69f9dbc561a9.jpg) +Figure 1. The rendering quality and speed of the original 3D Gaussian splatting[12] deteriorate severely at low resolutions or from distant cameras due to aliasing. Conversely, our multi-scale 3D Gaussians representation utilizes selective rendering to achieve faster $(160\% - 2400\%)$ at $4\times -128\times$ resolution) and more accurate rendering at lower resolutions. + +# Abstract + +3D Gaussians have recently emerged as a highly efficient representation for 3D reconstruction and rendering. Despite its high rendering quality and speed at high resolutions, they both deteriorate drastically when rendered at lower resolutions or from far away camera position. During low resolution or far away rendering, the pixel size of the image can fall below the Nyquist frequency compared to the screen size of each splatted 3D Gaussian and leads to aliasing effect. The rendering is also drastically slowed down by the sequential alpha blending of more splatted Gaussians per pixel. To address these issues, we propose a multi-scale 3D Gaussian splatting algorithm, which maintains Gaussians at different scales to represent the same scene. Higher-resolution images are rendered with more small Gaussians, and lower-resolution images are rendered with fewer larger Gaussians. With similar training time, our algorithm can achieve $13\% -66\%$ PSNR and $160\% -2400\%$ rendering speed improvement at $4\times -128\times$ scale rendering + +on Mip-NeRF360 dataset compared to the single scale 3D Gaussian splatting. More results and code are released on our project page. + +# 1. Introduction + +3D Gaussian Splatting[12] has recently emerged as a highly efficient representation for novel view synthesis. Compared to the time-consuming ray marching used in most neural radiance fields (NeRF) [2, 15, 16], a high-resolution image can be rendered in real-time by rasterizing the splatted 3D Gaussians. However, this rasterization algorithm is subjected to severe aliasing effect and speed deterioration when rendering the same scene at low resolution or from distant positions as shown in Fig. 1. This limitation significantly constrains the application of the 3D Gaussian splatting algorithm in reconstructing and rendering large-scale scenes. + +Aliasing effect is a consequence of inadequate sampling frequency failing to capture the continuous signal accurately. In the context of rendering, image pixels are sampled + +with an interval of one-pixel size. The signal can be considered as the 3D scene represented implicitly as in NeRF or explicitly as in 3D Gaussians. When part of the 3D scene is represented with high details but rendered with low resolution or from distant positions, the disparity between the low sampling and high signal frequencies culminates in aliasing artifacts. A naive solution is to render at high resolution and subsequently down-scale the rendered image to a lower resolution. However, this solution is not viable for scenes containing both near and far regions which are very common. Due to the inability of 3D Gaussian splatting algorithm to accommodate varying resolutions within a single image, rendering the entire image with a even higher resolution for the sake of far away regions is neither time nor memory efficient. + +We postulate that the pronounced aliasing artifacts observed when rendering with 3D Gaussians, as opposed to other techniques such as NeRF, are primarily attributable to the splatting of small Gaussians. 3D regions with intricate details are represented with large amount of small Gaussians. When rendering these regions with low resolution or from a distant view, many splatted small Gaussians are cramped in one pixel and therefore the pixel color of this region is dominated by the front-most Gaussian, even if this Gaussian is much smaller than others and not at the center. This problem is further aggravated by the low pass filter in [12, 19] applied to each individual Gaussian with the intention to mitigate aliasing on edges at high resolutions. This problem is explained in more detail in Sec. 3.2. + +In addition to the aliasing artifacts, the rendering speed of 3D Gaussians is also affected at low resolution. The number of 3D Gaussians that need to be rendered remains constant at lower resolutions, but they are more concentrated to fewer pixels. The Gaussians that are splatted to the same pixel cannot be rendered in parallel. This means that the image rendering is even slower at lower resolution in comparison with NeRF rendering time that reduces linearly with decreasing resolution. Hence, although aliasing is not a problem exclusive to 3D Gaussian splattering, it is more prominent and more difficult to tackle. + +Contributions To mitigate the aliasing problem for 3D Gaussian splatting, we propose a novel multi-scale 3D Gaussians to represent the scene at different levels of detail (LOD) as shown in Fig. 2. This is inspired by the mipmap and LOD algorithms widely used in computer graphics, which pre-computes textures and polygons at different scales to be rendered under different resolutions and distances. Similarly, we add larger, coarser Gaussians for lower resolutions by aggregating the smaller and finer Gaussians from higher resolutions. Depending on the pixel coverage of the splatted Gaussians during rendering, only a subset of the Gaussians is used. A simplified explanation + +for this is that the coarse Gaussians are used to render low-resolution images and the fine Gaussians are used to render high-resolution images. With fewer than $5\%$ number of Gaussians added and a similar training time, our method can achieve $13\% - 66\%$ PSNR and $160\% - 2400\%$ rendering speed improvements at $4 \times -128 \times$ scale rendering on MipNeRF360 dataset[2], while maintaining a comparable rendering quality and speed at $1 \times$ scale. + +# 2. Related Works + +# 2.1. Anti-Aliasing in Computer Graphics + +Ailiasing is a long-standing problem for computer graphics when rendering a scene to a discrete image. Traditional anti-aliasing techniques primarily target mesh representations. Supersampling Anti-Aliasing (SSAA) [5] renders the scene at a higher resolution before downscaling, leading to significantly more time and memory demain, and therefore is less used in real-time applications. The Multisample Anti-Aliasing (MSAA) [1, 5] algorithm selectively supersamples pixels on the edges, reducing resource and time consumption. This technique is not very suitable for 3D Gaussian splatting because of its requirement for regular grids and lack of support for variable sampling resolution at different pixels. The more recent Fast Approximate Anti-Aliasing (FXAA) [11, 14] is a post processing algorithm that smooths the jagged edges after the image is rendered. Unfortunately, this technique is also not suitable for Gaussian representation as the front-most Gaussian dominates the pixel color and produces chunky instead of jagged artifacts in mesh rendering. + +In contrast to the supersampling methods mentioned above, our method takes the inspiration from hierarchical mipmap [18] and level of details (LOD) [7, 9] algorithms to address the aliasing for 3D Gaussians. Mipmap uses multi-scale textures for the rendering at different resolution or from different distances. LOD algorithm represents the models in a scene with different complexity to be rendered at different distances. Both techniques not only mitigate the aliasing effect by reducing the complexity of the scene representation, but also enhances rendering speed, particularly for large-scale scenes. + +# 2.2. Anti-Aliasing in Neural Representation + +The recent success of neural representations especially Neural Radiance Fields (NeRF) [6, 15, 16] has also inspired some works to develop algorithms against aliasing effect on neural representations beyond the traditional mesh representation. Mip-NeRF [2, 3] employ low pass filters on the positional encoding of the input spatial coordinates to reduce the scene signal frequency. Building on the hash grid representation used by InstantNGP [16] with no position encoding, Zip-NeRF [4] proposes a multi-sampling + +![](images/49c1372bcd42994df6ad48a0c39401d577612edebdd55ad6e09fcfe3214ccdd2.jpg) +Figure 2. Overall pipeline of our algorithm. At the early stage of training (left), small Gaussians below certain size threshold in each voxel are aggregated, enlarged and inserted into the scene at different resolution scale. During rendering (right), the multi-scale Gaussians of the appropriate "pixel coverage" at the current render resolution are selected for rendering. If the rendering resolution scale equals to the scale of the Gaussians, the expected "pixel coverage" range of the Gaussians are updated accordingly. + +![](images/528e9aadd169d7d48229070c53928db14ac276d72d9fede1bc9d2acdf1ffda30.jpg) + +strategy in the conical frustum instead of the camera ray, at the cost of $6 \times$ rendering time. Similar to the mipmap algorithm in mesh texture rendering, Tri-MipNeRF [10] and MipGrid [17] proposes to use multi-scale feature grids for rendering at different resolution or distance. + +Conversely, 3D Gaussian splatting [12] presents unique anti-aliasing challenges due to its distinct scene representation. It does not have any positional encoding or feature grid, and its requirement for regular grids conflicts the more flexible multi-sampling strategies. The concentration of small Gaussians in detail-rich regions exacerbates aliasing and speed issues, more so than in NeRF representations. To the best of our knowledge, we are the first to propose an anti-aliasing algorithm for scene reconstruction using 3D Gaussian splatting. + +# 3. Preliminaries + +# 3.1. 3D Gaussian Splatting + +3D Gaussian splatting is first proposed in EWA Splatting [19], and later used by [12] for scene reconstruction and novel view synthesis. The scene is represented by a set of $\mathbf{K}$ 3D Gaussians $\{\mathcal{G}_{\hat{V}^k,\hat{\mu}^k},\sigma_k,\mathbf{c}_k\mid k\in [1,\mathbf{K}]\}$ with variance $\hat{V}^k$ , center $\hat{\mu}^k$ , density $\sigma_{k}$ and color $\mathbf{c}^k$ . During rendering, the 3D Gaussians are splatted to the 2D screen to by the perspective transformation to form 2D Gaussians $\mathcal{G}_{V^k,\mu^k}$ . The image is then divided into $16\times 16$ regular tiles and all 2D Gaussians touching each tile are sorted based on their original depth. The color of each pixel in the tile is then rasterized from the sequential alpha blending the 2D Gaussians from front to back. + +# 3.2. Cause of Aliasing in 3D Gaussian Splatting + +Aliasing can occur when sampling a continuous signal $g(x)$ with a discrete sampling function $\delta_s(x, \Delta x) =$ + +$\sum_{n = -\infty}^{\infty}\delta (x - n\cdot \Delta x)$ , where $\delta$ is a impulse function. The result of the sample in the spatial domain is: + +$$ +g _ {s} (x) = \delta_ {s} (x, \Delta x) \cdot g (x). \tag {1} +$$ + +This sampled function converted into the frequency domain using Fourier transform operator $\mathcal{F}$ becomes: + +$$ +\begin{array}{l} \mathcal {F} \left[ g _ {s} (u) \right] = \frac {1}{\Delta x} \sum_ {\substack {k = - \infty \\ \infty}} ^ {\infty} \delta \left(u - \frac {k}{\Delta x}\right) * \mathcal {F} [ g (x) ] \tag{2} \\ = \frac {1}{\Delta x} \sum_ {k = - \infty} ^ {\infty} \mathbf {G} (u - \frac {k}{\Delta x}). \\ \end{array} +$$ + +When the highest frequency component $f_{max}$ of the signal is greater than half of the sampling frequency $f_{s} = \frac{1}{\Delta x}$ , $\mathbf{G}(u - \frac{k}{\Delta x})$ in the summation sequence would overlap with each other and causes the sampled signal to diverge from the actual signal. This phenomenon is the aliasing effect and the minimum sampling frequency needed to avoid aliasing is $f_{Ny} = 2 \cdot f_{max}$ , known as the Nyquist frequency. + +The EWA splatting [19] used by 3D Gaussian splatting [12] also tries to mitigate the aliasing problem by applying a low pass filter to the rendered color. To approximate this efficiently, it applies a Gaussian kernel $h(\mathbf{x})$ as the low pass filter on each splatted 2D signal $g_{c}(\mathbf{x})$ independently to produce a band limited signal: + +$$ +\begin{array}{l} g _ {c} ^ {\prime} (\mathbf {x}) = g _ {c} (\mathbf {x}) * h (\mathbf {x}) \\ \approx \sum_ {k} \sigma_ {k} \mathbf {c} _ {k} T _ {k} \int_ {\mathcal {R} ^ {2}} q _ {k} (\eta) h (\mathbf {x} - \eta) d \eta \tag {3} \\ = \sum_ {k} \sigma_ {k} \mathbf {c} _ {k} T _ {k} \cdot (q _ {k} * h) (\mathbf {x}), \\ \end{array} +$$ + +where $\mathcal{R}^2$ is the range of one pixel, $q_{k}(\mathbf{x})$ is the 2D integrated Gaussian kernel, and $\sigma_k,\mathbf{c}_k,T_k$ are the opacity, + +color, and transmittance at each Gaussian, respectively. By combining the reconstruction Gaussian kernel $\mathcal{G}_{V^k}$ and low pass Gaussian kernel $\mathcal{G}_{V^h}$ of covariance matrix $V^{k}$ and $V^{h}$ , the band limit function becomes: + +$$ +\begin{array}{l} g _ {c} ^ {\prime} (\mathbf {x}) = \sum_ {k} \alpha_ {k} \cdot (\mathcal {G} _ {V ^ {k}} * \mathcal {G} _ {V ^ {h}}) (\mathbf {x}) \\ = \sum_ {k} \alpha_ {k} \cdot \mathcal {G} _ {V ^ {k} + V ^ {h}} (\mathbf {x}), \\ \end{array} +$$ + +where $\alpha_{k}$ represents all coefficients invariant of $\mathbf{x}$ at each Gaussian and $V^{h}$ is determined by the screen pixel size. A simple understanding of this is that the covariance of each 3D Gaussian is increased based on the screen pixel size. + +This method of applying a low pass filter to each 3D Gaussian independently helps to smooth the edges of the Gaussians when the Gaussians are not too small compared to the pixel size. However, it also gives rise to two substantial issues at low resolutions: + +1. $V^{h}$ added to the original covariance $V^{k}$ effectively increases the extent of each Gaussian, especially when $V^{h}$ is large compared to $V^{k}$ at low resolutions. Small Gaussians in the front dominate the color of the pixel and cause severe artifacts shown in Fig. 7. +2. The number of Gaussians involved in the sequential $\sum_{k}$ for each pixel scales increases with decreasing image resolution. Due to the incremental calculation of the transmittance $T_{k}$ , the rendering even slower at lower resolutions. + +# 4. Our Method + +# 4.1. Multi-Scale Gaussians Based on Pixel Coverage + +To mitigate the aliasing artifacts of 3D Gaussians [12] while avoiding the two problems of the EWA splatting [19], we introduce multi-scale 3D Gaussians (cf. Fig. 2) that tackle the problem on the scene-level instead of on each individual Gaussian. The 3D scene is represented with Gaussians from 4 levels of detail, corresponding to the $1 \times$ , $4 \times$ , $16 \times$ , and $64 \times$ downsampled resolution. Small finer-level Gaussians are aggregated to create larger Gaussians for coarser levels during training. Each 3D Gaussian $\mathcal{G}_k^l$ belongs to one of the levels $l$ and is included or excluded independently during the rendering based on its "pixel coverage". + +Pixel Coverage of Gaussian. The "pixel coverage" of a Gaussian reflects the size of the Gaussian when splatted onto the screen space compared to the pixel size at the current rendering resolution. The "pixel coverage" $S_{k}$ of a splatted 2D Gaussian $\mathcal{G}_{(\mu^k,V^k)}$ is defined as the length of its horizontal or vertical axis until the low opacity level set, whichever is smaller, as shown in Fig. 3. The pixel coverage is measured in pixel count and the opacity threshold $\sigma_T$ is set as $\frac{1}{255}$ . + +$$ +u; \sigma_ {k} \mathcal {G} _ {(\mu^ {k}, V ^ {k})} \left(\mu_ {x} ^ {k} + \frac {u}{2}, \mu_ {y} ^ {k}\right) = \sigma_ {T} +$$ + +$$ +\begin{array}{l} v \colon \sigma_ {k} \mathcal {G} _ {(\mu^ {k}, V ^ {k})} \left(\mu_ {x} ^ {k}, \mu_ {y} ^ {k} + \frac {v}{2}\right) = \sigma_ {T} \\ S _ {k} = \min (u, v) \\ \end{array} +$$ + +![](images/c49ff51666af448f95746b33dee309d3eceb9f9926eb835cb783b7daa8fb303d.jpg) +Figure 3. Pixel coverage of a 3D Gaussian is its horizontal or vertical size, whichever is smaller measured by the level set. + +![](images/52a83eb8876fa856b15822877f55ed3ae8b65ebe1ac25089a5dce2d93c083a0b.jpg) +Figure 4. Missing parts caused by naive small Gaussian filtering at different resolution scales. + +![](images/163f40aca66f15030db0b51fc3bbc7b9d64aa7e6fe0e0a505af0d8b65e8e8271.jpg) + +![](images/36968b76538c6d92057e1666774887c285d732b5501adc0acb8a42d602cd7d58.jpg) + +The pixel coverage approximates the extent of a 2D splatted Gaussian in the spatial domain. During the rendering from a given camera direction, the color of each splatted Gaussian is constant within this pixel coverage. As a result, the coverage of this pixel approximates the inverse of the highest frequency component $f_{max} = 1 / S_k$ in this region. Compared to the sampling frequency of $f_s = 1\mathrm{px}^{-1}$ during rasterization, a signal frequency of $f_{max} > f_s / 2$ can cause the sampling to fall below the Nyquist frequency needed to avoid aliasing. + +Consequently, the Gaussians with pixel coverage $S_{k} < S_{T} = 2\mathrm{px}$ should be filtered out during rendering to avoid aliasing. Since 3D Gaussian representation does not encode the signal of different frequencies at different Gaussians, naively filtering out the small Gaussians will result in a hole or part missing in the scene as shown in Fig. 4. To address this issue, we propose to aggregate the small Gaussians to form large Gaussians that encode the low-frequency signal. These large Gaussians would appear when the small Gaussians are filtered out. + +Aggregate to Insert Large Gaussians. All 3D Gaussians initialized from the input point cloud at the start of the training belong to the finest level $l = 1$ . They are densified by splitting and cloning as in [12], and all the densified Gaussians would inherit the same level. After the warm-up stage of the first 1,000 iterations, we introduce coarse-level Gaussians by aggregating fine-level Gaussians that are too small as visualized in Fig. 5 and described in Algorithm 1. The procedure is outlined as follows: + +1. For all levels $\{l_m\mid 2\leq l_m\leq l_{max}\}$ , we render all 3D Gaussians from $[1,l_m - 1]$ at the $4^{l_m - 1}$ times downsam + +![](images/e30244ed326674ca9c05c95ab3b935ac8634751f2eb8c781970c4398def740a8.jpg) +Figure 5. Large Gaussians are created by aggregating the small Gaussians in each voxel below the pixel coverage threshold, and then enlarged by the pixel coverage multiplier. + +pled resolution of all training images. All 3D Gaussians with the minimal "pixel coverage" $S_{k}$ smaller than the filter threshold $S_{T}$ are chosen for the aggregation. + +2. The chosen 3D Gaussians are binned by a $(400 / l_{m})^{3}$ resolution voxel grid based on their positions. The attributes of all Gaussians within each voxel are aggregated to create a new Gaussian using average pooling, including position, scaling, opacity and color. More details are included in the supplementary. +3. Based on the average "pixel coverage" $S_{avg}$ of Gaussians in each voxel, the scaling of each new Gaussian created is enlarged by $S_T / S_{avg}$ so that it is of a size suitable to be rendered at $l_m$ . This new Gaussian belongs to level $l_m$ . + +Not all Gaussians from the fine levels are small. Many Gaussians in the background or in the textureless regions are large and do not need to be aggregated. The number of Gaussians created is often fewer than $5\%$ of the final total number of Gaussians. + +Algorithm 1 Aggregate Small Gaussians + +
1:procedure AGGREGATEGAUSSIANS(G1:lmax)
2:for lm←2 to lmax do
3:S1:K=PixelCoverage(G1:lm-1, scale 4lmax-1)
4:Gsmall = {Gk|Sk < ST, ∀k ∈ [1:K]}
5:for n←1 to (400/lm)3 do
6:Gn = {Gk|Gk in voxel n, ∀Gk ∈ Gsmall}
7:Glmnew,n = Enlarge(Average(Gn))
8:InsertIntoScene(Glmnew,n)
9:end for
10:end for
11:end procedure
+ +Multi-Scale Training and Selective Rendering. After the large Gaussians are added, the model is trained with both the original images and the downsampled images. A maximum pixel coverage $S_{k}^{max}$ and a minimum pixel coverage $S_{k}^{min}$ of each Gaussian are stored for the selective rendering. If the rendering downsample scale equals to + +Algorithm 2 Selective Rendering Based on Pixel Coverage +1: procedure SELECTIVERENDER $(\mathcal{G}_{1:K}^{1;l_{max}}$ , scale $l_{r}$ +2: $S_{1:K} =$ PixelCoverage $(\mathcal{G}_{1:K}^{1;l_{max}})$ +3: $G_{1} = \{\mathcal{G}_{k}|S_{k} / S_{k}^{max}\leq S_{rel}^{max},\forall k\}$ +4: $G_{2} = \{\mathcal{G}_{k}|S_{k} / S_{k}^{min}\geq S_{rel}^{min}\vee S_{k}\geq S_{T},\forall k\}$ +5: $G_{l_r} = \{\mathcal{G}_k^l |l = l_r,\forall k\}$ +6: for $\mathcal{G}_k^l\in G_{l_r}$ do +7: UpdateRange $(S_k^{max},S_k^{min},S_k)$ +8: end for +9: return Render $(G_{1}\cap G_{2},l_{r})$ +10: end procedure + +![](images/3ef8eb5df30babd6aad4d601b3c4ed24737a4c2966164cf7fe7dbe60a894dad4.jpg) +Figure 6. Based on the rendering resolution, the current pixel coverage of a Gaussian relative to its minimum and maximum pixel coverages determines whether it is selected for rendering. + +![](images/d33891ce746d130d36ee262969176b16882a03181ef012ddc9d92f638de77e71.jpg) + +![](images/3f0e93130cf10cc2fef179dc7662b711d1d94d1dce1ee2bcaa7e831775bcd9d2.jpg) + +the downsample scale when the Gaussian $\mathcal{G}_k$ is created, its $S_{k}^{max}$ and $S_{k}^{min}$ values are updated with the new pixel coverage $S_{k}$ : + +$$ +S _ {k} ^ {\prime \max } = \max \left(\lambda_ {1} S _ {k} ^ {\max }, S _ {k}\right), \tag {5} +$$ + +$$ +S _ {k} ^ {\prime \min } = \operatorname {m i n} \left(\lambda_ {2} S _ {k} ^ {\min }, S _ {k}\right), +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are decay coefficients taking the empirical value of 0.95 and 1.05, respectively. + +During rendering at any resolution or camera distance, a Gaussian is selected for rendering if its pixel coverage $S_{k}$ on the screen satisfies the following condition: + +$$ +\left(\frac {S _ {k}}{S _ {k} ^ {\text {m a x}}} \leq S _ {\text {r e l}} ^ {\text {m a x}}\right) \wedge \left(\frac {S _ {k}}{S _ {k} ^ {\text {m i n}}} \geq S _ {\text {r e l}} ^ {\text {m i n}} \vee S _ {k} \geq S _ {T}\right), \tag {6} +$$ + +where $S_{rel}^{max}$ and $S_{rel}^{min}$ are the maximum and minimum relative pixel coverage taking the empirical values of 1.5 and 0.5 respectively. If the pixel coverage of a Gaussian is too much larger than the $S_{k}^{max}$ , it is filtered out from rendering. Similarly, if it is too much smaller than the $S_{k}^{min}$ and is smaller than $S_{T}$ , it is filtered out from rendering (cf. Fig. 6). The absolute $S_{T}$ threshold is used to preserve the large Gaussians from the lower scales, as they do not cause the aliasing problem if their screen size is not sufficiently small. This selective rendering procedure is described in Algorithm 2. Additionally, even if the Gaussians from the finest level are too large or the Gaussians from the coarsest level are too small, they are not filtered to render beyond the maximum and below the minimum training resolutions. + +Table 1. Quantitative comparison and ablation study on the 360 dataset [3] at various downsampled scales, with time in "ms". + +
Scale1x4x16x64x
MetricPSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓
3D Gaussian[12]27.520.14210.522.500.1379.317.790.14927.915.23N.A.103.3
3DGS + MS Train27.350.15511.323.500.1267.720.210.11522.819.38N.A.84.8
3DGS + Filter Small27.400.15310.023.810.1495.420.020.1864.817.38N.A.4.6
3DGS + Insert Large18.020.6049.718.750.5312.520.230.2562.721.53N.A.7.1
Our Full Method27.390.1559.124.820.1325.424.750.0664.925.35N.A.4.9
+ +Table 2. Quantitative comparison and ablation study on Tank and Temples dataset [13] at various downsampled scales, with time in "ms". + +
Scale1x4x16x64x
MetricPSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓
3D Gaussian[12]23.740.0966.519.700.10511.115.610.06843.413.88N.A.82.6
3DGS + MS Train22.970.1186.021.460.0869.618.560.04937.416.54N.A.71.7
3DGS + Filter Small23.780.1005.620.120.1074.517.410.0724.414.95N.A.4.7
3DGS + Insert Large10.840.6975.111.150.7031.711.730.4471.712.62N.A.2.5
Our Full Method23.460.1117.621.920.0874.720.910.0344.819.67N.A.5.9
+ +Table 3. Quantitative comparison and ablation study on the Deep Blending dataset [8] at various downsampled scales, with time in "ms". + +
Scale1x4x16x64x
MetricPSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓
3D Gaussian[12]29.650.0948.627.480.0667.522.060.06720.717.75N.A.59.7
3DGS + MS Train29.460.1026.628.180.0625.324.130.05514.320.03N.A.41.3
3DGS + Filter Small29.680.0956.728.260.0644.224.520.0783.618.29N.A.3.2
3DGS + Insert Large20.590.3794.620.830.3361.621.290.1432.120.10N.A.4.2
Our Full Method29.700.0967.428.430.0643.927.660.0363.425.70N.A.3.4
+ +The pixel coverage range of each Gaussian allows the model to maintain multi-scale Gaussians for different levels of detail. The appropriate subset of Gaussians is chosen for rendering at different resolutions and distances. More smaller Gaussians encoding the high-frequency information are rendered at high resolution, and fewer and larger Gaussians encoding the low-frequency information are rendered at low resolution for less aliasing effect and faster speed. + +# 5. Experiments + +In this section, we present a comprehensive evaluation of our proposed model, which is grounded on the implementation framework of the official release of the 3D Gaussian Splatting code. To achieve a similar training time as the baseline model, our models are trained for 40000 iterations with all other hyper-parameters unchanged. All rendering speed are measured on a single RTX3090 GPU. We evaluate the performance of the vanilla 3D Gaussian Splatting[12] algorithm and our model on the multi-scale 360[3], Tank And Temples[13], and Deep Blending[8] dataset, aligned with the data used by the original paper. These datasets + +cover a wide range of object centric, indoor, and ourdoor scenes. + +Our evaluation focuses on the rendering quality and speed at multiple downsampling scales of 1x, 4x, 16x, and 64x derived from the test views. The rendering quality is measured in PSNR and LPIPS, while the speed is measured in per-image rendering time. This multi-scale evaluation is aimed at simulating the rendering performance in scenarios of low-resolution imaging or when captured from distant cameras. More detailed evaluations, including the results for more resolution scales and per-scene decomposition, are included in the supplementary materials due to the space constraint. Additionally, the supplementary materials include a video that offers an intuitive qualitative comparison of the two algorithms, vividly demonstrating the improvement of our algorithm in quality and speed from multiple viewpoints. + +Quantitative Comparison. As shown in Tab. 1, Tab. 2, and Tab. 3, our method can achieve substantial quality and speed improvements compared to the original 3D Gaussian + +![](images/f723dfbd9a223345a651ef437a8e7819e18e7aa3a6e3cc53296f93fbdaa5c678.jpg) +Figure 7. Qualitative Comparison on 360 dataset[3] for different resolution scales. + +![](images/26d4d8d03d4daf995105f5f09a4bc0161a1f186a7e0884cc8cbc7057b76aa4ea.jpg) +Figure 8. Qualitative Comparison on Tank and Temples dataset[13] for different resolution scales. + +Splatting [12] at lower resolutions. The quality and speed improvements become more pronounced as the resolution reduces, with the most noticeable 6-10dB PSNR and $20 - 30 \times$ speed gain at the $64 \times$ resolution scale. As the resolution reduces, the original splatting algorithm slows down while our method accelerates. The rendering quality and speed at the original resolution $(1 \times)$ remain comparable, indicating the effectiveness of our multi-scale Gaussians in representing both the high and low resolutions together. + +Qualitative Comparison. We present the qualitative comparison with the original 3D Gaussian Splatting [12] shown in Fig. 7, Fig. 8, and Fig. 9. At higher resolutions $(1\times -8\times)$ , both ours and the original algorithm can render the novel view rather faithfully. However, as the resolution reduces further $(16\times -64\times)$ , the original splatting algorithm produces severe artifacts, where the foreground becomes larger and larger, dominating the pixel colors as explained in Sec. 3.2. In contrast, the images rendered by our method closely resemble the ground truth across all resolution scales. + +![](images/bc2dbc57a5b2a45cad06e6c4e0919210a18c5c5bf10db833887ba469a4427e86.jpg) +Figure 9. Qualitative Comparison on Deep Blending dataset[8] for different resolution scales. + +Ablations. To evaluate the effectiveness of the different components proposed, we present the ablation quantitative results in Tab. 1, Tab. 2, and Tab. 3 and the ablation qualitative results in the supplementary. The three ablation methods evaluated are named "3DGS+MS Train", "3DGS+Filter Small", and "3DGS+Insert Large". "3DGS+MS Train" reports the result with multi-scale training on top of the original 3D Gaussian splatting. The "3DGS+Filter Small" reports the result with small Gaussian filtering using pixel coverage and the multi-scale training, which is needed to update the maximum and minimum pixel coverage. Similarly, the "3DGS+Insert Large" reports the result with large Gaussian insertion and the multi-scale training. + +The ablation results reflect that multi-scale training marginally improves low-resolution rendering quality, but the rendering speed remains very slow. When filtering out the small Gaussians with multi-scale training, the speed at low resolution is increased by $20 - 30 \times$ with minimal rendering quality loss. The speed gain is caused by the considerably fewer Gaussians rendered. When inserting the large Gaussians and training with multi-scale supervision, without the small Gaussian filtering, the rendering quality drops significantly because the details of the scene are covered with large Gaussians completely for all resolutions. However, when adding the large Gaussians together with the small Gaussian filtering, the rendering quality and speed at low resolution are enhanced significantly without jeopardizing the high-resolution quality. This indicates the effectiveness of all three components and the full method proposed. + +# 6. Limitations + +Since all Gaussian filtering of our proposed method relies on the pixel coverage, it can only be done after the initial + +splatting process when the coverage is calculated. Although the splatting of individual Gaussians are performed in parallel and does not take more time at lower resolution, it is still a considerable overhead when rendering at a very low resolution. Even if a very small portion of the Gaussians are used for rendering in the end, all Gaussians still need to be splatted. This is the main reason why our rendering time is not decreased linearly as the resolution decreases. In our future work, we will look into a more lightweight criteria to filter small and large Gaussians before splating them onto the screen to achieve an even faster rendering speed. + +# 7. Conclusion + +We analyzed the cause of the severe aliasing artifact and speed degradation of the existing 3D Gaussian splatting. We identified the key challenge of mitigating the aliasing for 3D Gaussian splatting lies in representing the scene with Gaussians of appropriate scale. Based on this observation, we propose to calculate the pixel coverage of 3D Gaussians during splatting and use this as a criteria for selective rendering. Gaussians that are too large or too small at the current rendering resolution are filtered for anti-aliasing and speed improvements. We also proposed to insert large Gaussians by aggregating small Gaussians during training to preserve the low frequency details and prevent part missing. Our experiments on various datasets support the effectiveness of our algorithm in rendering quality and speed at both high and low resolution, mitigating the severe aliasing artifact of the original 3D Gaussian splatting. + +Acknowledgement. This work is supported by the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021). + +# References + +[1] Kurt Akeley. Reality engine graphics. In Proceedings of the 20th Annual Conference on Computer Graphics and Interactive Techniques, page 109-116, New York, NY, USA, 1993. Association for Computing Machinery. 2 +[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 1, 2 +[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2, 6, 7 +[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 2 +[5] Kristof Beets and David L. Barron. Super-sampling anti-aliasing analyzed. 2000. 2 +[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2 +[7] Carl Erikson. *Polygonal simplification.* Technical Report 96-016, 1996. 2 +[8] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Trans. Graph., 37(6), 2018. 6, 8 +[9] Tan Kim Heok and D. Daman. A review on level of detail. In Proceedings. International Conference on Computer Graphics, Imaging and Visualization, 2004. CGIV 2004., pages 70-75, 2004. 2 +[10] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19774–19783, 2023. 3 +[11] Jorge Jimenez, Diego Gutierrez, Jason Yang, Alexander Reshetov, Pete Demoreuille, Tobias Berghoff, Cedric Perthuis, Henry Yu, Morgan Mcguire, Timothy Lottes, Hugh Malan, and Emil Persson. Filtering approaches for real-time anti-aliasing. ACM SIGGRAPH 2011 Courses, SIGGRAPH'11, 2011. 2 +[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 6, 7 +[13] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 6, 7 +[14] Timothy Lottes. *Fxaa. Technical report*, Nvidia, 2011. 2 +[15] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[16] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires + +olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1, 2 +[17] Seungtae Nam, Daniel Rho, Jong Hwan Ko, and Eunbyung Park. Mip-grid: Anti-aliased grid representations for neural radiance fields. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3 +[18] Lance Williams. Pyramidal parametrics. SIGGRAPH Comput. Graph., 17(3):1-11, 1983. 2 +[19] M. Zwicker, H. Pfister, J. van Baar, and M. Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS '01., pages 29-538, 2001. 2, 3, 4 \ No newline at end of file diff --git a/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/images.zip b/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3bcf8022398a26da62995044a1175ab0da9a8459 --- /dev/null +++ b/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:914ea3843c802605d25c5aeab2fa38e349f91fc4fa7743227aeb80439f677a70 +size 941384 diff --git a/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/layout.json b/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cec3462a4e509e2613d8a72f696463760776cb19 --- /dev/null +++ b/2024/Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering/layout.json @@ -0,0 +1,7327 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 103, + 491, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 103, + 491, + 121 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 491, + 121 + ], + "type": "text", + "content": "Multi-Scale 3D Gaussian Splating for Anti-Aliased Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 144, + 198, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 144, + 198, + 156 + ], + "spans": [ + { + "bbox": [ + 137, + 144, + 198, + 156 + ], + "type": "text", + "content": "Zhiwen Yan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 144, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 144, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 223, + 144, + 296, + 156 + ], + "type": "text", + "content": "Weng Fei Low" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 321, + 144, + 364, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 144, + 364, + 156 + ], + "spans": [ + { + "bbox": [ + 321, + 144, + 364, + 156 + ], + "type": "text", + "content": "Yu Chen" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 389, + 144, + 455, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 144, + 455, + 156 + ], + "spans": [ + { + "bbox": [ + 389, + 144, + 455, + 156 + ], + "type": "text", + "content": "Gim Hee Lee" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 158, + 460, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 158, + 460, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 158, + 460, + 172 + ], + "type": "text", + "content": "Department of Computer Science, National University of Singapore" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 174, + 520, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 174, + 520, + 185 + ], + "spans": [ + { + "bbox": [ + 69, + 174, + 520, + 185 + ], + "type": "text", + "content": "yan.zhiwen@u.nus.edu {wengfei.low, chenyu}@comp.nus.edu.sg gimhee.lee@nus.edu.sg" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 53, + 198, + 547, + 423 + ], + "blocks": [ + { + "bbox": [ + 53, + 198, + 547, + 423 + ], + "lines": [ + { + "bbox": [ + 53, + 198, + 547, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 547, + 423 + ], + "type": "image", + "image_path": "22407358080dc2e589a148aa68a17808d0b136374d0c8104810b69f9dbc561a9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 430, + 547, + 464 + ], + "lines": [ + { + "bbox": [ + 46, + 430, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 547, + 464 + ], + "type": "text", + "content": "Figure 1. The rendering quality and speed of the original 3D Gaussian splatting[12] deteriorate severely at low resolutions or from distant cameras due to aliasing. Conversely, our multi-scale 3D Gaussians representation utilizes selective rendering to achieve faster " + }, + { + "bbox": [ + 46, + 430, + 547, + 464 + ], + "type": "inline_equation", + "content": "(160\\% - 2400\\%)" + }, + { + "bbox": [ + 46, + 430, + 547, + 464 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 430, + 547, + 464 + ], + "type": "inline_equation", + "content": "4\\times -128\\times" + }, + { + "bbox": [ + 46, + 430, + 547, + 464 + ], + "type": "text", + "content": " resolution) and more accurate rendering at lower resolutions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 468, + 192, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 468, + 192, + 481 + ], + "spans": [ + { + "bbox": [ + 143, + 468, + 192, + 481 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": "3D Gaussians have recently emerged as a highly efficient representation for 3D reconstruction and rendering. Despite its high rendering quality and speed at high resolutions, they both deteriorate drastically when rendered at lower resolutions or from far away camera position. During low resolution or far away rendering, the pixel size of the image can fall below the Nyquist frequency compared to the screen size of each splatted 3D Gaussian and leads to aliasing effect. The rendering is also drastically slowed down by the sequential alpha blending of more splatted Gaussians per pixel. To address these issues, we propose a multi-scale 3D Gaussian splatting algorithm, which maintains Gaussians at different scales to represent the same scene. Higher-resolution images are rendered with more small Gaussians, and lower-resolution images are rendered with fewer larger Gaussians. With similar training time, our algorithm can achieve " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "13\\% -66\\%" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": " PSNR and " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "160\\% -2400\\%" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": " rendering speed improvement at " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "4\\times -128\\times" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": " scale rendering" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 470, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 470, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 305, + 470, + 545, + 507 + ], + "type": "text", + "content": "on Mip-NeRF360 dataset compared to the single scale 3D Gaussian splatting. More results and code are released on our project page." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 523, + 387, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 523, + 387, + 536 + ], + "spans": [ + { + "bbox": [ + 306, + 523, + 387, + 536 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 544, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 676 + ], + "type": "text", + "content": "3D Gaussian Splatting[12] has recently emerged as a highly efficient representation for novel view synthesis. Compared to the time-consuming ray marching used in most neural radiance fields (NeRF) [2, 15, 16], a high-resolution image can be rendered in real-time by rasterizing the splatted 3D Gaussians. However, this rasterization algorithm is subjected to severe aliasing effect and speed deterioration when rendering the same scene at low resolution or from distant positions as shown in Fig. 1. This limitation significantly constrains the application of the 3D Gaussian splatting algorithm in reconstructing and rendering large-scale scenes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Aliasing effect is a consequence of inadequate sampling frequency failing to capture the continuous signal accurately. In the context of rendering, image pixels are sampled" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20923" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "with an interval of one-pixel size. The signal can be considered as the 3D scene represented implicitly as in NeRF or explicitly as in 3D Gaussians. When part of the 3D scene is represented with high details but rendered with low resolution or from distant positions, the disparity between the low sampling and high signal frequencies culminates in aliasing artifacts. A naive solution is to render at high resolution and subsequently down-scale the rendered image to a lower resolution. However, this solution is not viable for scenes containing both near and far regions which are very common. Due to the inability of 3D Gaussian splatting algorithm to accommodate varying resolutions within a single image, rendering the entire image with a even higher resolution for the sake of far away regions is neither time nor memory efficient." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 252, + 289, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 419 + ], + "type": "text", + "content": "We postulate that the pronounced aliasing artifacts observed when rendering with 3D Gaussians, as opposed to other techniques such as NeRF, are primarily attributable to the splatting of small Gaussians. 3D regions with intricate details are represented with large amount of small Gaussians. When rendering these regions with low resolution or from a distant view, many splatted small Gaussians are cramped in one pixel and therefore the pixel color of this region is dominated by the front-most Gaussian, even if this Gaussian is much smaller than others and not at the center. This problem is further aggravated by the low pass filter in [12, 19] applied to each individual Gaussian with the intention to mitigate aliasing on edges at high resolutions. This problem is explained in more detail in Sec. 3.2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 420, + 289, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 289, + 552 + ], + "type": "text", + "content": "In addition to the aliasing artifacts, the rendering speed of 3D Gaussians is also affected at low resolution. The number of 3D Gaussians that need to be rendered remains constant at lower resolutions, but they are more concentrated to fewer pixels. The Gaussians that are splatted to the same pixel cannot be rendered in parallel. This means that the image rendering is even slower at lower resolution in comparison with NeRF rendering time that reduces linearly with decreasing resolution. Hence, although aliasing is not a problem exclusive to 3D Gaussian splattering, it is more prominent and more difficult to tackle." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "Contributions To mitigate the aliasing problem for 3D Gaussian splatting, we propose a novel multi-scale 3D Gaussians to represent the scene at different levels of detail (LOD) as shown in Fig. 2. This is inspired by the mipmap and LOD algorithms widely used in computer graphics, which pre-computes textures and polygons at different scales to be rendered under different resolutions and distances. Similarly, we add larger, coarser Gaussians for lower resolutions by aggregating the smaller and finer Gaussians from higher resolutions. Depending on the pixel coverage of the splatted Gaussians during rendering, only a subset of the Gaussians is used. A simplified explanation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": "for this is that the coarse Gaussians are used to render low-resolution images and the fine Gaussians are used to render high-resolution images. With fewer than " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " number of Gaussians added and a similar training time, our method can achieve " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "13\\% - 66\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " PSNR and " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "160\\% - 2400\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " rendering speed improvements at " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "4 \\times -128 \\times" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " scale rendering on MipNeRF360 dataset[2], while maintaining a comparable rendering quality and speed at " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " scale." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 180, + 398, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 180, + 398, + 192 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 398, + 192 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 200, + 501, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 200, + 501, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 200, + 501, + 213 + ], + "type": "text", + "content": "2.1. Anti-Aliasing in Computer Graphics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 219, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 547, + 445 + ], + "type": "text", + "content": "Ailiasing is a long-standing problem for computer graphics when rendering a scene to a discrete image. Traditional anti-aliasing techniques primarily target mesh representations. Supersampling Anti-Aliasing (SSAA) [5] renders the scene at a higher resolution before downscaling, leading to significantly more time and memory demain, and therefore is less used in real-time applications. The Multisample Anti-Aliasing (MSAA) [1, 5] algorithm selectively supersamples pixels on the edges, reducing resource and time consumption. This technique is not very suitable for 3D Gaussian splatting because of its requirement for regular grids and lack of support for variable sampling resolution at different pixels. The more recent Fast Approximate Anti-Aliasing (FXAA) [11, 14] is a post processing algorithm that smooths the jagged edges after the image is rendered. Unfortunately, this technique is also not suitable for Gaussian representation as the front-most Gaussian dominates the pixel color and produces chunky instead of jagged artifacts in mesh rendering." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 447, + 548, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 548, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 548, + 578 + ], + "type": "text", + "content": "In contrast to the supersampling methods mentioned above, our method takes the inspiration from hierarchical mipmap [18] and level of details (LOD) [7, 9] algorithms to address the aliasing for 3D Gaussians. Mipmap uses multi-scale textures for the rendering at different resolution or from different distances. LOD algorithm represents the models in a scene with different complexity to be rendered at different distances. Both techniques not only mitigate the aliasing effect by reducing the complexity of the scene representation, but also enhances rendering speed, particularly for large-scale scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 586, + 512, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 512, + 600 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 512, + 600 + ], + "type": "text", + "content": "2.2. Anti-Aliasing in Neural Representation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "The recent success of neural representations especially Neural Radiance Fields (NeRF) [6, 15, 16] has also inspired some works to develop algorithms against aliasing effect on neural representations beyond the traditional mesh representation. Mip-NeRF [2, 3] employ low pass filters on the positional encoding of the input spatial coordinates to reduce the scene signal frequency. Building on the hash grid representation used by InstantNGP [16] with no position encoding, Zip-NeRF [4] proposes a multi-sampling" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20924" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 70, + 263, + 224 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 263, + 224 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 263, + 224 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 263, + 224 + ], + "type": "image", + "image_path": "49c1372bcd42994df6ad48a0c39401d577612edebdd55ad6e09fcfe3214ccdd2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 235, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 235, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 547, + 281 + ], + "type": "text", + "content": "Figure 2. Overall pipeline of our algorithm. At the early stage of training (left), small Gaussians below certain size threshold in each voxel are aggregated, enlarged and inserted into the scene at different resolution scale. During rendering (right), the multi-scale Gaussians of the appropriate \"pixel coverage\" at the current render resolution are selected for rendering. If the rendering resolution scale equals to the scale of the Gaussians, the expected \"pixel coverage\" range of the Gaussians are updated accordingly." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 265, + 70, + 520, + 224 + ], + "blocks": [ + { + "bbox": [ + 265, + 70, + 520, + 224 + ], + "lines": [ + { + "bbox": [ + 265, + 70, + 520, + 224 + ], + "spans": [ + { + "bbox": [ + 265, + 70, + 520, + 224 + ], + "type": "image", + "image_path": "528e9aadd169d7d48229070c53928db14ac276d72d9fede1bc9d2acdf1ffda30.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 291, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 287, + 351 + ], + "type": "text", + "content": "strategy in the conical frustum instead of the camera ray, at the cost of " + }, + { + "bbox": [ + 46, + 291, + 287, + 351 + ], + "type": "inline_equation", + "content": "6 \\times" + }, + { + "bbox": [ + 46, + 291, + 287, + 351 + ], + "type": "text", + "content": " rendering time. Similar to the mipmap algorithm in mesh texture rendering, Tri-MipNeRF [10] and MipGrid [17] proposes to use multi-scale feature grids for rendering at different resolution or distance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 351, + 288, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 288, + 472 + ], + "type": "text", + "content": "Conversely, 3D Gaussian splatting [12] presents unique anti-aliasing challenges due to its distinct scene representation. It does not have any positional encoding or feature grid, and its requirement for regular grids conflicts the more flexible multi-sampling strategies. The concentration of small Gaussians in detail-rich regions exacerbates aliasing and speed issues, more so than in NeRF representations. To the best of our knowledge, we are the first to propose an anti-aliasing algorithm for scene reconstruction using 3D Gaussian splatting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 480, + 133, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 133, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 133, + 493 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 500, + 175, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 175, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 175, + 514 + ], + "type": "text", + "content": "3.1. 3D Gaussian Splatting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "text", + "content": "3D Gaussian splatting is first proposed in EWA Splatting [19], and later used by [12] for scene reconstruction and novel view synthesis. The scene is represented by a set of " + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "text", + "content": " 3D Gaussians " + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_{\\hat{V}^k,\\hat{\\mu}^k},\\sigma_k,\\mathbf{c}_k\\mid k\\in [1,\\mathbf{K}]\\}" + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "text", + "content": " with variance " + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\hat{V}^k" + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "text", + "content": ", center " + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}^k" + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "text", + "content": ", density " + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^k" + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "text", + "content": ". During rendering, the 3D Gaussians are splatted to the 2D screen to by the perspective transformation to form 2D Gaussians " + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{V^k,\\mu^k}" + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "text", + "content": ". The image is then divided into " + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "inline_equation", + "content": "16\\times 16" + }, + { + "bbox": [ + 46, + 518, + 287, + 664 + ], + "type": "text", + "content": " regular tiles and all 2D Gaussians touching each tile are sorted based on their original depth. The color of each pixel in the tile is then rasterized from the sequential alpha blending the 2D Gaussians from front to back." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 671, + 272, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 272, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 272, + 685 + ], + "type": "text", + "content": "3.2. Cause of Aliasing in 3D Gaussian Splatting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Aliasing can occur when sampling a continuous signal " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "g(x)" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " with a discrete sampling function " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\delta_s(x, \\Delta x) =" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 291, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 291, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 291, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\sum_{n = -\\infty}^{\\infty}\\delta (x - n\\cdot \\Delta x)" + }, + { + "bbox": [ + 305, + 291, + 545, + 316 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 291, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 305, + 291, + 545, + 316 + ], + "type": "text", + "content": " is a impulse function. The result of the sample in the spatial domain is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 371, + 323, + 545, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 323, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 371, + 323, + 545, + 337 + ], + "type": "interline_equation", + "content": "g _ {s} (x) = \\delta_ {s} (x, \\Delta x) \\cdot g (x). \\tag {1}", + "image_path": "75fbf939f62734589e2ad3b56d55bdaeb507a48aadc07ad4dc22ce8bf25bed7a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 343, + 545, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 343, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 343, + 545, + 367 + ], + "type": "text", + "content": "This sampled function converted into the frequency domain using Fourier transform operator " + }, + { + "bbox": [ + 305, + 343, + 545, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 343, + 545, + 367 + ], + "type": "text", + "content": " becomes:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 332, + 374, + 545, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 374, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 332, + 374, + 545, + 439 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {F} \\left[ g _ {s} (u) \\right] = \\frac {1}{\\Delta x} \\sum_ {\\substack {k = - \\infty \\\\ \\infty}} ^ {\\infty} \\delta \\left(u - \\frac {k}{\\Delta x}\\right) * \\mathcal {F} [ g (x) ] \\tag{2} \\\\ = \\frac {1}{\\Delta x} \\sum_ {k = - \\infty} ^ {\\infty} \\mathbf {G} (u - \\frac {k}{\\Delta x}). \\\\ \\end{array}", + "image_path": "c27a725dc9d60d3359c531bb2402f1fbe1b5487a20c35d3c26af1bff57137765.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 446, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 446, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 305, + 446, + 545, + 532 + ], + "type": "text", + "content": "When the highest frequency component " + }, + { + "bbox": [ + 305, + 446, + 545, + 532 + ], + "type": "inline_equation", + "content": "f_{max}" + }, + { + "bbox": [ + 305, + 446, + 545, + 532 + ], + "type": "text", + "content": " of the signal is greater than half of the sampling frequency " + }, + { + "bbox": [ + 305, + 446, + 545, + 532 + ], + "type": "inline_equation", + "content": "f_{s} = \\frac{1}{\\Delta x}" + }, + { + "bbox": [ + 305, + 446, + 545, + 532 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 446, + 545, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{G}(u - \\frac{k}{\\Delta x})" + }, + { + "bbox": [ + 305, + 446, + 545, + 532 + ], + "type": "text", + "content": " in the summation sequence would overlap with each other and causes the sampled signal to diverge from the actual signal. This phenomenon is the aliasing effect and the minimum sampling frequency needed to avoid aliasing is " + }, + { + "bbox": [ + 305, + 446, + 545, + 532 + ], + "type": "inline_equation", + "content": "f_{Ny} = 2 \\cdot f_{max}" + }, + { + "bbox": [ + 305, + 446, + 545, + 532 + ], + "type": "text", + "content": ", known as the Nyquist frequency." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 532, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 604 + ], + "type": "text", + "content": "The EWA splatting [19] used by 3D Gaussian splatting [12] also tries to mitigate the aliasing problem by applying a low pass filter to the rendered color. To approximate this efficiently, it applies a Gaussian kernel " + }, + { + "bbox": [ + 304, + 532, + 545, + 604 + ], + "type": "inline_equation", + "content": "h(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 532, + 545, + 604 + ], + "type": "text", + "content": " as the low pass filter on each splatted 2D signal " + }, + { + "bbox": [ + 304, + 532, + 545, + 604 + ], + "type": "inline_equation", + "content": "g_{c}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 532, + 545, + 604 + ], + "type": "text", + "content": " independently to produce a band limited signal:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 336, + 609, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 609, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 336, + 609, + 545, + 681 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} g _ {c} ^ {\\prime} (\\mathbf {x}) = g _ {c} (\\mathbf {x}) * h (\\mathbf {x}) \\\\ \\approx \\sum_ {k} \\sigma_ {k} \\mathbf {c} _ {k} T _ {k} \\int_ {\\mathcal {R} ^ {2}} q _ {k} (\\eta) h (\\mathbf {x} - \\eta) d \\eta \\tag {3} \\\\ = \\sum_ {k} \\sigma_ {k} \\mathbf {c} _ {k} T _ {k} \\cdot (q _ {k} * h) (\\mathbf {x}), \\\\ \\end{array}", + "image_path": "b2e40fcb0e88146ed5c73ded4605185a5b702c2e150059428d6d6c011ba5102f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^2" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " is the range of one pixel, " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "q_{k}(\\mathbf{x})" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " is the 2D integrated Gaussian kernel, and " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_k,\\mathbf{c}_k,T_k" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " are the opacity," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20925" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "color, and transmittance at each Gaussian, respectively. By combining the reconstruction Gaussian kernel " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{V^k}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " and low pass Gaussian kernel " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{V^h}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " of covariance matrix " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "V^{k}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "V^{h}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ", the band limit function becomes:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 96, + 127, + 236, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 127, + 236, + 179 + ], + "spans": [ + { + "bbox": [ + 96, + 127, + 236, + 179 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} g _ {c} ^ {\\prime} (\\mathbf {x}) = \\sum_ {k} \\alpha_ {k} \\cdot (\\mathcal {G} _ {V ^ {k}} * \\mathcal {G} _ {V ^ {h}}) (\\mathbf {x}) \\\\ = \\sum_ {k} \\alpha_ {k} \\cdot \\mathcal {G} _ {V ^ {k} + V ^ {h}} (\\mathbf {x}), \\\\ \\end{array}", + "image_path": "f9c6e9b04a422fcf3471bf28ea0071b7e316df75f8bffe67582705e57f95e827.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 189, + 287, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 189, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 287, + 237 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 189, + 287, + 237 + ], + "type": "inline_equation", + "content": "\\alpha_{k}" + }, + { + "bbox": [ + 46, + 189, + 287, + 237 + ], + "type": "text", + "content": " represents all coefficients invariant of " + }, + { + "bbox": [ + 46, + 189, + 287, + 237 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 189, + 287, + 237 + ], + "type": "text", + "content": " at each Gaussian and " + }, + { + "bbox": [ + 46, + 189, + 287, + 237 + ], + "type": "inline_equation", + "content": "V^{h}" + }, + { + "bbox": [ + 46, + 189, + 287, + 237 + ], + "type": "text", + "content": " is determined by the screen pixel size. A simple understanding of this is that the covariance of each 3D Gaussian is increased based on the screen pixel size." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 237, + 287, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 287, + 296 + ], + "type": "text", + "content": "This method of applying a low pass filter to each 3D Gaussian independently helps to smooth the edges of the Gaussians when the Gaussians are not too small compared to the pixel size. However, it also gives rise to two substantial issues at low resolutions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 297, + 287, + 416 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 297, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 287, + 357 + ], + "type": "text", + "content": "1. " + }, + { + "bbox": [ + 47, + 297, + 287, + 357 + ], + "type": "inline_equation", + "content": "V^{h}" + }, + { + "bbox": [ + 47, + 297, + 287, + 357 + ], + "type": "text", + "content": " added to the original covariance " + }, + { + "bbox": [ + 47, + 297, + 287, + 357 + ], + "type": "inline_equation", + "content": "V^{k}" + }, + { + "bbox": [ + 47, + 297, + 287, + 357 + ], + "type": "text", + "content": " effectively increases the extent of each Gaussian, especially when " + }, + { + "bbox": [ + 47, + 297, + 287, + 357 + ], + "type": "inline_equation", + "content": "V^{h}" + }, + { + "bbox": [ + 47, + 297, + 287, + 357 + ], + "type": "text", + "content": " is large compared to " + }, + { + "bbox": [ + 47, + 297, + 287, + 357 + ], + "type": "inline_equation", + "content": "V^{k}" + }, + { + "bbox": [ + 47, + 297, + 287, + 357 + ], + "type": "text", + "content": " at low resolutions. Small Gaussians in the front dominate the color of the pixel and cause severe artifacts shown in Fig. 7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 357, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 287, + 416 + ], + "type": "text", + "content": "2. The number of Gaussians involved in the sequential " + }, + { + "bbox": [ + 47, + 357, + 287, + 416 + ], + "type": "inline_equation", + "content": "\\sum_{k}" + }, + { + "bbox": [ + 47, + 357, + 287, + 416 + ], + "type": "text", + "content": " for each pixel scales increases with decreasing image resolution. Due to the incremental calculation of the transmittance " + }, + { + "bbox": [ + 47, + 357, + 287, + 416 + ], + "type": "inline_equation", + "content": "T_{k}" + }, + { + "bbox": [ + 47, + 357, + 287, + 416 + ], + "type": "text", + "content": ", the rendering even slower at lower resolutions." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 428, + 127, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 428, + 127, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 428, + 127, + 440 + ], + "type": "text", + "content": "4. Our Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 448, + 287, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 448, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 448, + 287, + 460 + ], + "type": "text", + "content": "4.1. Multi-Scale Gaussians Based on Pixel Coverage" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "text", + "content": "To mitigate the aliasing artifacts of 3D Gaussians [12] while avoiding the two problems of the EWA splatting [19], we introduce multi-scale 3D Gaussians (cf. Fig. 2) that tackle the problem on the scene-level instead of on each individual Gaussian. The 3D scene is represented with Gaussians from 4 levels of detail, corresponding to the " + }, + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "inline_equation", + "content": "64 \\times" + }, + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "text", + "content": " downsampled resolution. Small finer-level Gaussians are aggregated to create larger Gaussians for coarser levels during training. Each 3D Gaussian " + }, + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_k^l" + }, + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "text", + "content": " belongs to one of the levels " + }, + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "text", + "content": " and is included or excluded independently during the rendering based on its \"pixel coverage\"." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Pixel Coverage of Gaussian. The \"pixel coverage\" of a Gaussian reflects the size of the Gaussian when splatted onto the screen space compared to the pixel size at the current rendering resolution. The \"pixel coverage\" " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " of a splatted 2D Gaussian " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{(\\mu^k,V^k)}" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " is defined as the length of its horizontal or vertical axis until the low opacity level set, whichever is smaller, as shown in Fig. 3. The pixel coverage is measured in pixel count and the opacity threshold " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\sigma_T" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " is set as " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\frac{1}{255}" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 72, + 452, + 93 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 72, + 452, + 93 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 452, + 93 + ], + "type": "interline_equation", + "content": "u; \\sigma_ {k} \\mathcal {G} _ {(\\mu^ {k}, V ^ {k})} \\left(\\mu_ {x} ^ {k} + \\frac {u}{2}, \\mu_ {y} ^ {k}\\right) = \\sigma_ {T}", + "image_path": "a8f0346cadc855a65347993ff9ef52a701369a0bd550fea1b357fa0a00051d4e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 93, + 451, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 93, + 451, + 125 + ], + "spans": [ + { + "bbox": [ + 321, + 93, + 451, + 125 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} v \\colon \\sigma_ {k} \\mathcal {G} _ {(\\mu^ {k}, V ^ {k})} \\left(\\mu_ {x} ^ {k}, \\mu_ {y} ^ {k} + \\frac {v}{2}\\right) = \\sigma_ {T} \\\\ S _ {k} = \\min (u, v) \\\\ \\end{array}", + "image_path": "2104d2bac4ae543adb4742538c5120a80d5872c37e664acf62ec4593dc6f99d2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 424, + 72, + 544, + 152 + ], + "blocks": [ + { + "bbox": [ + 424, + 72, + 544, + 152 + ], + "lines": [ + { + "bbox": [ + 424, + 72, + 544, + 152 + ], + "spans": [ + { + "bbox": [ + 424, + 72, + 544, + 152 + ], + "type": "image", + "image_path": "c49ff51666af448f95746b33dee309d3eceb9f9926eb835cb783b7daa8fb303d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 162, + 545, + 184 + ], + "lines": [ + { + "bbox": [ + 305, + 162, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 305, + 162, + 545, + 184 + ], + "type": "text", + "content": "Figure 3. Pixel coverage of a 3D Gaussian is its horizontal or vertical size, whichever is smaller measured by the level set." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 318, + 195, + 391, + 290 + ], + "blocks": [ + { + "bbox": [ + 318, + 195, + 391, + 290 + ], + "lines": [ + { + "bbox": [ + 318, + 195, + 391, + 290 + ], + "spans": [ + { + "bbox": [ + 318, + 195, + 391, + 290 + ], + "type": "image", + "image_path": "52a83eb8876fa856b15822877f55ed3ae8b65ebe1ac25089a5dce2d93c083a0b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 299, + 545, + 320 + ], + "lines": [ + { + "bbox": [ + 305, + 299, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 545, + 320 + ], + "type": "text", + "content": "Figure 4. Missing parts caused by naive small Gaussian filtering at different resolution scales." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 391, + 195, + 462, + 290 + ], + "blocks": [ + { + "bbox": [ + 391, + 195, + 462, + 290 + ], + "lines": [ + { + "bbox": [ + 391, + 195, + 462, + 290 + ], + "spans": [ + { + "bbox": [ + 391, + 195, + 462, + 290 + ], + "type": "image", + "image_path": "163f40aca66f15030db0b51fc3bbc7b9d64aa7e6fe0e0a505af0d8b65e8e8271.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 462, + 195, + 533, + 290 + ], + "blocks": [ + { + "bbox": [ + 462, + 195, + 533, + 290 + ], + "lines": [ + { + "bbox": [ + 462, + 195, + 533, + 290 + ], + "spans": [ + { + "bbox": [ + 462, + 195, + 533, + 290 + ], + "type": "image", + "image_path": "36968b76538c6d92057e1666774887c285d732b5501adc0acb8a42d602cd7d58.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 337, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 545, + 456 + ], + "type": "text", + "content": "The pixel coverage approximates the extent of a 2D splatted Gaussian in the spatial domain. During the rendering from a given camera direction, the color of each splatted Gaussian is constant within this pixel coverage. As a result, the coverage of this pixel approximates the inverse of the highest frequency component " + }, + { + "bbox": [ + 304, + 337, + 545, + 456 + ], + "type": "inline_equation", + "content": "f_{max} = 1 / S_k" + }, + { + "bbox": [ + 304, + 337, + 545, + 456 + ], + "type": "text", + "content": " in this region. Compared to the sampling frequency of " + }, + { + "bbox": [ + 304, + 337, + 545, + 456 + ], + "type": "inline_equation", + "content": "f_s = 1\\mathrm{px}^{-1}" + }, + { + "bbox": [ + 304, + 337, + 545, + 456 + ], + "type": "text", + "content": " during rasterization, a signal frequency of " + }, + { + "bbox": [ + 304, + 337, + 545, + 456 + ], + "type": "inline_equation", + "content": "f_{max} > f_s / 2" + }, + { + "bbox": [ + 304, + 337, + 545, + 456 + ], + "type": "text", + "content": " can cause the sampling to fall below the Nyquist frequency needed to avoid aliasing." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 456, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 575 + ], + "type": "text", + "content": "Consequently, the Gaussians with pixel coverage " + }, + { + "bbox": [ + 304, + 456, + 545, + 575 + ], + "type": "inline_equation", + "content": "S_{k} < S_{T} = 2\\mathrm{px}" + }, + { + "bbox": [ + 304, + 456, + 545, + 575 + ], + "type": "text", + "content": " should be filtered out during rendering to avoid aliasing. Since 3D Gaussian representation does not encode the signal of different frequencies at different Gaussians, naively filtering out the small Gaussians will result in a hole or part missing in the scene as shown in Fig. 4. To address this issue, we propose to aggregate the small Gaussians to form large Gaussians that encode the low-frequency signal. These large Gaussians would appear when the small Gaussians are filtered out." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": "Aggregate to Insert Large Gaussians. All 3D Gaussians initialized from the input point cloud at the start of the training belong to the finest level " + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": ". They are densified by splitting and cloning as in [12], and all the densified Gaussians would inherit the same level. After the warm-up stage of the first 1,000 iterations, we introduce coarse-level Gaussians by aggregating fine-level Gaussians that are too small as visualized in Fig. 5 and described in Algorithm 1. The procedure is outlined as follows:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "1. For all levels " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{l_m\\mid 2\\leq l_m\\leq l_{max}\\}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " , we render all 3D Gaussians from " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "[1,l_m - 1]" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "4^{l_m - 1}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " times downsam" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20926" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 289, + 159 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 289, + 159 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 289, + 159 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 289, + 159 + ], + "type": "image", + "image_path": "e30244ed326674ca9c05c95ab3b935ac8634751f2eb8c781970c4398def740a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 170, + 287, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 170, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 287, + 204 + ], + "type": "text", + "content": "Figure 5. Large Gaussians are created by aggregating the small Gaussians in each voxel below the pixel coverage threshold, and then enlarged by the pixel coverage multiplier." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 225, + 287, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 225, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 58, + 225, + 287, + 261 + ], + "type": "text", + "content": "pled resolution of all training images. All 3D Gaussians with the minimal \"pixel coverage\" " + }, + { + "bbox": [ + 58, + 225, + 287, + 261 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 58, + 225, + 287, + 261 + ], + "type": "text", + "content": " smaller than the filter threshold " + }, + { + "bbox": [ + 58, + 225, + 287, + 261 + ], + "type": "inline_equation", + "content": "S_{T}" + }, + { + "bbox": [ + 58, + 225, + 287, + 261 + ], + "type": "text", + "content": " are chosen for the aggregation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 261, + 287, + 392 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 261, + 287, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 261, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 287, + 332 + ], + "type": "text", + "content": "2. The chosen 3D Gaussians are binned by a " + }, + { + "bbox": [ + 47, + 261, + 287, + 332 + ], + "type": "inline_equation", + "content": "(400 / l_{m})^{3}" + }, + { + "bbox": [ + 47, + 261, + 287, + 332 + ], + "type": "text", + "content": " resolution voxel grid based on their positions. The attributes of all Gaussians within each voxel are aggregated to create a new Gaussian using average pooling, including position, scaling, opacity and color. More details are included in the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 333, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 333, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 333, + 287, + 392 + ], + "type": "text", + "content": "3. Based on the average \"pixel coverage\" " + }, + { + "bbox": [ + 47, + 333, + 287, + 392 + ], + "type": "inline_equation", + "content": "S_{avg}" + }, + { + "bbox": [ + 47, + 333, + 287, + 392 + ], + "type": "text", + "content": " of Gaussians in each voxel, the scaling of each new Gaussian created is enlarged by " + }, + { + "bbox": [ + 47, + 333, + 287, + 392 + ], + "type": "inline_equation", + "content": "S_T / S_{avg}" + }, + { + "bbox": [ + 47, + 333, + 287, + 392 + ], + "type": "text", + "content": " so that it is of a size suitable to be rendered at " + }, + { + "bbox": [ + 47, + 333, + 287, + 392 + ], + "type": "inline_equation", + "content": "l_m" + }, + { + "bbox": [ + 47, + 333, + 287, + 392 + ], + "type": "text", + "content": ". This new Gaussian belongs to level " + }, + { + "bbox": [ + 47, + 333, + 287, + 392 + ], + "type": "inline_equation", + "content": "l_m" + }, + { + "bbox": [ + 47, + 333, + 287, + 392 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "content": "Not all Gaussians from the fine levels are small. Many Gaussians in the background or in the textureless regions are large and do not need to be aggregated. The number of Gaussians created is often fewer than " + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "content": " of the final total number of Gaussians." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 52, + 480, + 290, + 616 + ], + "blocks": [ + { + "bbox": [ + 48, + 464, + 214, + 475 + ], + "lines": [ + { + "bbox": [ + 48, + 464, + 214, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 214, + 475 + ], + "type": "text", + "content": "Algorithm 1 Aggregate Small Gaussians" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 480, + 290, + 616 + ], + "lines": [ + { + "bbox": [ + 52, + 480, + 290, + 616 + ], + "spans": [ + { + "bbox": [ + 52, + 480, + 290, + 616 + ], + "type": "table", + "html": "
1:procedure AGGREGATEGAUSSIANS(G1:lmax)
2:for lm←2 to lmax do
3:S1:K=PixelCoverage(G1:lm-1, scale 4lmax-1)
4:Gsmall = {Gk|Sk < ST, ∀k ∈ [1:K]}
5:for n←1 to (400/lm)3 do
6:Gn = {Gk|Gk in voxel n, ∀Gk ∈ Gsmall}
7:Glmnew,n = Enlarge(Average(Gn))
8:InsertIntoScene(Glmnew,n)
9:end for
10:end for
11:end procedure
", + "image_path": "fa48479560dfc9a4f3b4a16dfc4226f245e4e36b1987920dd8827e73e7865cc3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "Multi-Scale Training and Selective Rendering. After the large Gaussians are added, the model is trained with both the original images and the downsampled images. A maximum pixel coverage " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "S_{k}^{max}" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " and a minimum pixel coverage " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "S_{k}^{min}" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " of each Gaussian are stored for the selective rendering. If the rendering downsample scale equals to" + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 311, + 88, + 528, + 213 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 84 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 84 + ], + "type": "text", + "content": "Algorithm 2 Selective Rendering Based on Pixel Coverage" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "lines": [ + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "spans": [ + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "text", + "content": "1: procedure SELECTIVERENDER " + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "inline_equation", + "content": "(\\mathcal{G}_{1:K}^{1;l_{max}}" + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "text", + "content": " , scale " + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "inline_equation", + "content": "l_{r}" + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "text", + "content": " \n2: " + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "inline_equation", + "content": "S_{1:K} =" + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "text", + "content": " PixelCoverage " + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "inline_equation", + "content": "(\\mathcal{G}_{1:K}^{1;l_{max}})" + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "text", + "content": " \n3: " + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "inline_equation", + "content": "G_{1} = \\{\\mathcal{G}_{k}|S_{k} / S_{k}^{max}\\leq S_{rel}^{max},\\forall k\\}" + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "inline_equation", + "content": "G_{2} = \\{\\mathcal{G}_{k}|S_{k} / S_{k}^{min}\\geq S_{rel}^{min}\\vee S_{k}\\geq S_{T},\\forall k\\}" + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "inline_equation", + "content": "G_{l_r} = \\{\\mathcal{G}_k^l |l = l_r,\\forall k\\}" + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "text", + "content": " \n6: for " + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_k^l\\in G_{l_r}" + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "text", + "content": " do \n7: UpdateRange " + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "inline_equation", + "content": "(S_k^{max},S_k^{min},S_k)" + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "text", + "content": " \n8: end for \n9: return Render " + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "inline_equation", + "content": "(G_{1}\\cap G_{2},l_{r})" + }, + { + "bbox": [ + 311, + 88, + 528, + 213 + ], + "type": "text", + "content": " \n10: end procedure" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "type": "image", + "bbox": [ + 313, + 223, + 389, + 304 + ], + "blocks": [ + { + "bbox": [ + 313, + 223, + 389, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 223, + 389, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 389, + 304 + ], + "type": "image", + "image_path": "3ef8eb5df30babd6aad4d601b3c4ed24737a4c2966164cf7fe7dbe60a894dad4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 316, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 350 + ], + "type": "text", + "content": "Figure 6. Based on the rendering resolution, the current pixel coverage of a Gaussian relative to its minimum and maximum pixel coverages determines whether it is selected for rendering." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 390, + 223, + 455, + 304 + ], + "blocks": [ + { + "bbox": [ + 390, + 223, + 455, + 304 + ], + "lines": [ + { + "bbox": [ + 390, + 223, + 455, + 304 + ], + "spans": [ + { + "bbox": [ + 390, + 223, + 455, + 304 + ], + "type": "image", + "image_path": "d33891ce746d130d36ee262969176b16882a03181ef012ddc9d92f638de77e71.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 456, + 223, + 539, + 304 + ], + "blocks": [ + { + "bbox": [ + 456, + 223, + 539, + 304 + ], + "lines": [ + { + "bbox": [ + 456, + 223, + 539, + 304 + ], + "spans": [ + { + "bbox": [ + 456, + 223, + 539, + 304 + ], + "type": "image", + "image_path": "3f0e93130cf10cc2fef179dc7662b711d1d94d1dce1ee2bcaa7e831775bcd9d2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 361, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 545, + 397 + ], + "type": "text", + "content": "the downsample scale when the Gaussian " + }, + { + "bbox": [ + 305, + 361, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_k" + }, + { + "bbox": [ + 305, + 361, + 545, + 397 + ], + "type": "text", + "content": " is created, its " + }, + { + "bbox": [ + 305, + 361, + 545, + 397 + ], + "type": "inline_equation", + "content": "S_{k}^{max}" + }, + { + "bbox": [ + 305, + 361, + 545, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 361, + 545, + 397 + ], + "type": "inline_equation", + "content": "S_{k}^{min}" + }, + { + "bbox": [ + 305, + 361, + 545, + 397 + ], + "type": "text", + "content": " values are updated with the new pixel coverage " + }, + { + "bbox": [ + 305, + 361, + 545, + 397 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 305, + 361, + 545, + 397 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 365, + 405, + 545, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 405, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 365, + 405, + 545, + 426 + ], + "type": "interline_equation", + "content": "S _ {k} ^ {\\prime \\max } = \\max \\left(\\lambda_ {1} S _ {k} ^ {\\max }, S _ {k}\\right), \\tag {5}", + "image_path": "9ceb5b0f84ad5f8cfe9ddaf48bd33a64302a0f43c282738f393cc3b7983589d5.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 367, + 425, + 483, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 425, + 483, + 437 + ], + "spans": [ + { + "bbox": [ + 367, + 425, + 483, + 437 + ], + "type": "interline_equation", + "content": "S _ {k} ^ {\\prime \\min } = \\operatorname {m i n} \\left(\\lambda_ {2} S _ {k} ^ {\\min }, S _ {k}\\right),", + "image_path": "3f250b983f41938577bdeed89e5c59f8122e08c749017d204e8a4c3db552a115.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 445, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 545, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 445, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 306, + 445, + 545, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 445, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 306, + 445, + 545, + 468 + ], + "type": "text", + "content": " are decay coefficients taking the empirical value of 0.95 and 1.05, respectively." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 469, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 469, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 305, + 469, + 545, + 505 + ], + "type": "text", + "content": "During rendering at any resolution or camera distance, a Gaussian is selected for rendering if its pixel coverage " + }, + { + "bbox": [ + 305, + 469, + 545, + 505 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 305, + 469, + 545, + 505 + ], + "type": "text", + "content": " on the screen satisfies the following condition:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 511, + 545, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 511, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 318, + 511, + 545, + 538 + ], + "type": "interline_equation", + "content": "\\left(\\frac {S _ {k}}{S _ {k} ^ {\\text {m a x}}} \\leq S _ {\\text {r e l}} ^ {\\text {m a x}}\\right) \\wedge \\left(\\frac {S _ {k}}{S _ {k} ^ {\\text {m i n}}} \\geq S _ {\\text {r e l}} ^ {\\text {m i n}} \\vee S _ {k} \\geq S _ {T}\\right), \\tag {6}", + "image_path": "0cd31358456c77037397f005caedf8d86c25c822c37e217797b5c4feec54a79e.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{rel}^{max}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{rel}^{min}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " are the maximum and minimum relative pixel coverage taking the empirical values of 1.5 and 0.5 respectively. If the pixel coverage of a Gaussian is too much larger than the " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{k}^{max}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", it is filtered out from rendering. Similarly, if it is too much smaller than the " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{k}^{min}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " and is smaller than " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{T}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", it is filtered out from rendering (cf. Fig. 6). The absolute " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{T}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " threshold is used to preserve the large Gaussians from the lower scales, as they do not cause the aliasing problem if their screen size is not sufficiently small. This selective rendering procedure is described in Algorithm 2. Additionally, even if the Gaussians from the finest level are too large or the Gaussians from the coarsest level are too small, they are not filtered to render beyond the maximum and below the minimum training resolutions." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20927" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 83, + 545, + 185 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 523, + 81 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 523, + 81 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 523, + 81 + ], + "type": "text", + "content": "Table 1. Quantitative comparison and ablation study on the 360 dataset [3] at various downsampled scales, with time in \"ms\"." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 83, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 49, + 83, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 83, + 545, + 185 + ], + "type": "table", + "html": "
Scale1x4x16x64x
MetricPSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓
3D Gaussian[12]27.520.14210.522.500.1379.317.790.14927.915.23N.A.103.3
3DGS + MS Train27.350.15511.323.500.1267.720.210.11522.819.38N.A.84.8
3DGS + Filter Small27.400.15310.023.810.1495.420.020.1864.817.38N.A.4.6
3DGS + Insert Large18.020.6049.718.750.5312.520.230.2562.721.53N.A.7.1
Our Full Method27.390.1559.124.820.1325.424.750.0664.925.35N.A.4.9
", + "image_path": "6fbd51d2dde1dfe3420778ad9e2444556b3e84540aa249d9da85415cce6f848a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 207, + 545, + 308 + ], + "blocks": [ + { + "bbox": [ + 49, + 194, + 544, + 205 + ], + "lines": [ + { + "bbox": [ + 49, + 194, + 544, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 194, + 544, + 205 + ], + "type": "text", + "content": "Table 2. Quantitative comparison and ablation study on Tank and Temples dataset [13] at various downsampled scales, with time in \"ms\"." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 207, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 49, + 207, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 207, + 545, + 308 + ], + "type": "table", + "html": "
Scale1x4x16x64x
MetricPSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓
3D Gaussian[12]23.740.0966.519.700.10511.115.610.06843.413.88N.A.82.6
3DGS + MS Train22.970.1186.021.460.0869.618.560.04937.416.54N.A.71.7
3DGS + Filter Small23.780.1005.620.120.1074.517.410.0724.414.95N.A.4.7
3DGS + Insert Large10.840.6975.111.150.7031.711.730.4471.712.62N.A.2.5
Our Full Method23.460.1117.621.920.0874.720.910.0344.819.67N.A.5.9
", + "image_path": "bd91888460e12433ce89f9a9e4c6432c309b40e2416dcef448e3e45f61aaa875.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 49, + 331, + 545, + 432 + ], + "blocks": [ + { + "bbox": [ + 49, + 319, + 544, + 330 + ], + "lines": [ + { + "bbox": [ + 49, + 319, + 544, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 544, + 330 + ], + "type": "text", + "content": "Table 3. Quantitative comparison and ablation study on the Deep Blending dataset [8] at various downsampled scales, with time in \"ms\"." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 331, + 545, + 432 + ], + "lines": [ + { + "bbox": [ + 49, + 331, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 49, + 331, + 545, + 432 + ], + "type": "table", + "html": "
Scale1x4x16x64x
MetricPSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓PSNR↑LPIPS↓Time↓
3D Gaussian[12]29.650.0948.627.480.0667.522.060.06720.717.75N.A.59.7
3DGS + MS Train29.460.1026.628.180.0625.324.130.05514.320.03N.A.41.3
3DGS + Filter Small29.680.0956.728.260.0644.224.520.0783.618.29N.A.3.2
3DGS + Insert Large20.590.3794.620.830.3361.621.290.1432.120.10N.A.4.2
Our Full Method29.700.0967.428.430.0643.927.660.0363.425.70N.A.3.4
", + "image_path": "f112a1b1a8540c8a2061f7d6b4fa02b8dfac459ec0a02af4d7c35597d9aa2fa2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 452, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 452, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 287, + 548 + ], + "type": "text", + "content": "The pixel coverage range of each Gaussian allows the model to maintain multi-scale Gaussians for different levels of detail. The appropriate subset of Gaussians is chosen for rendering at different resolutions and distances. More smaller Gaussians encoding the high-frequency information are rendered at high resolution, and fewer and larger Gaussians encoding the low-frequency information are rendered at low resolution for less aliasing effect and faster speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "In this section, we present a comprehensive evaluation of our proposed model, which is grounded on the implementation framework of the official release of the 3D Gaussian Splatting code. To achieve a similar training time as the baseline model, our models are trained for 40000 iterations with all other hyper-parameters unchanged. All rendering speed are measured on a single RTX3090 GPU. We evaluate the performance of the vanilla 3D Gaussian Splatting[12] algorithm and our model on the multi-scale 360[3], Tank And Temples[13], and Deep Blending[8] dataset, aligned with the data used by the original paper. These datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 452, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 545, + 475 + ], + "type": "text", + "content": "cover a wide range of object centric, indoor, and ourdoor scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 477, + 546, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 546, + 657 + ], + "type": "text", + "content": "Our evaluation focuses on the rendering quality and speed at multiple downsampling scales of 1x, 4x, 16x, and 64x derived from the test views. The rendering quality is measured in PSNR and LPIPS, while the speed is measured in per-image rendering time. This multi-scale evaluation is aimed at simulating the rendering performance in scenarios of low-resolution imaging or when captured from distant cameras. More detailed evaluations, including the results for more resolution scales and per-scene decomposition, are included in the supplementary materials due to the space constraint. Additionally, the supplementary materials include a video that offers an intuitive qualitative comparison of the two algorithms, vividly demonstrating the improvement of our algorithm in quality and speed from multiple viewpoints." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Quantitative Comparison. As shown in Tab. 1, Tab. 2, and Tab. 3, our method can achieve substantial quality and speed improvements compared to the original 3D Gaussian" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20928" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 547, + 295 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 547, + 295 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 547, + 295 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 547, + 295 + ], + "type": "image", + "image_path": "f723dfbd9a223345a651ef437a8e7819e18e7aa3a6e3cc53296f93fbdaa5c678.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 300, + 445, + 312 + ], + "lines": [ + { + "bbox": [ + 147, + 300, + 445, + 312 + ], + "spans": [ + { + "bbox": [ + 147, + 300, + 445, + 312 + ], + "type": "text", + "content": "Figure 7. Qualitative Comparison on 360 dataset[3] for different resolution scales." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 320, + 547, + 543 + ], + "blocks": [ + { + "bbox": [ + 53, + 320, + 547, + 543 + ], + "lines": [ + { + "bbox": [ + 53, + 320, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 53, + 320, + 547, + 543 + ], + "type": "image", + "image_path": "26d4d8d03d4daf995105f5f09a4bc0161a1f186a7e0884cc8cbc7057b76aa4ea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 548, + 473, + 559 + ], + "lines": [ + { + "bbox": [ + 118, + 548, + 473, + 559 + ], + "spans": [ + { + "bbox": [ + 118, + 548, + 473, + 559 + ], + "type": "text", + "content": "Figure 8. Qualitative Comparison on Tank and Temples dataset[13] for different resolution scales." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 580, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 288, + 689 + ], + "type": "text", + "content": "Splatting [12] at lower resolutions. The quality and speed improvements become more pronounced as the resolution reduces, with the most noticeable 6-10dB PSNR and " + }, + { + "bbox": [ + 46, + 580, + 288, + 689 + ], + "type": "inline_equation", + "content": "20 - 30 \\times" + }, + { + "bbox": [ + 46, + 580, + 288, + 689 + ], + "type": "text", + "content": " speed gain at the " + }, + { + "bbox": [ + 46, + 580, + 288, + 689 + ], + "type": "inline_equation", + "content": "64 \\times" + }, + { + "bbox": [ + 46, + 580, + 288, + 689 + ], + "type": "text", + "content": " resolution scale. As the resolution reduces, the original splatting algorithm slows down while our method accelerates. The rendering quality and speed at the original resolution " + }, + { + "bbox": [ + 46, + 580, + 288, + 689 + ], + "type": "inline_equation", + "content": "(1 \\times)" + }, + { + "bbox": [ + 46, + 580, + 288, + 689 + ], + "type": "text", + "content": " remain comparable, indicating the effectiveness of our multi-scale Gaussians in representing both the high and low resolutions together." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 580, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 712 + ], + "type": "text", + "content": "Qualitative Comparison. We present the qualitative comparison with the original 3D Gaussian Splatting [12] shown in Fig. 7, Fig. 8, and Fig. 9. At higher resolutions " + }, + { + "bbox": [ + 304, + 580, + 545, + 712 + ], + "type": "inline_equation", + "content": "(1\\times -8\\times)" + }, + { + "bbox": [ + 304, + 580, + 545, + 712 + ], + "type": "text", + "content": ", both ours and the original algorithm can render the novel view rather faithfully. However, as the resolution reduces further " + }, + { + "bbox": [ + 304, + 580, + 545, + 712 + ], + "type": "inline_equation", + "content": "(16\\times -64\\times)" + }, + { + "bbox": [ + 304, + 580, + 545, + 712 + ], + "type": "text", + "content": ", the original splatting algorithm produces severe artifacts, where the foreground becomes larger and larger, dominating the pixel colors as explained in Sec. 3.2. In contrast, the images rendered by our method closely resemble the ground truth across all resolution scales." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20929" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 68, + 547, + 281 + ], + "blocks": [ + { + "bbox": [ + 52, + 68, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 52, + 68, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 52, + 68, + 547, + 281 + ], + "type": "image", + "image_path": "bc2dbc57a5b2a45cad06e6c4e0919210a18c5c5bf10db833887ba469a4427e86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 285, + 465, + 296 + ], + "lines": [ + { + "bbox": [ + 126, + 285, + 465, + 296 + ], + "spans": [ + { + "bbox": [ + 126, + 285, + 465, + 296 + ], + "type": "text", + "content": "Figure 9. Qualitative Comparison on Deep Blending dataset[8] for different resolution scales." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 303, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 458 + ], + "type": "text", + "content": "Ablations. To evaluate the effectiveness of the different components proposed, we present the ablation quantitative results in Tab. 1, Tab. 2, and Tab. 3 and the ablation qualitative results in the supplementary. The three ablation methods evaluated are named \"3DGS+MS Train\", \"3DGS+Filter Small\", and \"3DGS+Insert Large\". \"3DGS+MS Train\" reports the result with multi-scale training on top of the original 3D Gaussian splatting. The \"3DGS+Filter Small\" reports the result with small Gaussian filtering using pixel coverage and the multi-scale training, which is needed to update the maximum and minimum pixel coverage. Similarly, the \"3DGS+Insert Large\" reports the result with large Gaussian insertion and the multi-scale training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 652 + ], + "type": "text", + "content": "The ablation results reflect that multi-scale training marginally improves low-resolution rendering quality, but the rendering speed remains very slow. When filtering out the small Gaussians with multi-scale training, the speed at low resolution is increased by " + }, + { + "bbox": [ + 46, + 460, + 287, + 652 + ], + "type": "inline_equation", + "content": "20 - 30 \\times" + }, + { + "bbox": [ + 46, + 460, + 287, + 652 + ], + "type": "text", + "content": " with minimal rendering quality loss. The speed gain is caused by the considerably fewer Gaussians rendered. When inserting the large Gaussians and training with multi-scale supervision, without the small Gaussian filtering, the rendering quality drops significantly because the details of the scene are covered with large Gaussians completely for all resolutions. However, when adding the large Gaussians together with the small Gaussian filtering, the rendering quality and speed at low resolution are enhanced significantly without jeopardizing the high-resolution quality. This indicates the effectiveness of all three components and the full method proposed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 668, + 122, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 668, + 122, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 668, + 122, + 680 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Since all Gaussian filtering of our proposed method relies on the pixel coverage, it can only be done after the initial" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 303, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 434 + ], + "type": "text", + "content": "splatting process when the coverage is calculated. Although the splatting of individual Gaussians are performed in parallel and does not take more time at lower resolution, it is still a considerable overhead when rendering at a very low resolution. Even if a very small portion of the Gaussians are used for rendering in the end, all Gaussians still need to be splatted. This is the main reason why our rendering time is not decreased linearly as the resolution decreases. In our future work, we will look into a more lightweight criteria to filter small and large Gaussians before splating them onto the screen to achieve an even faster rendering speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 460, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 651 + ], + "type": "text", + "content": "We analyzed the cause of the severe aliasing artifact and speed degradation of the existing 3D Gaussian splatting. We identified the key challenge of mitigating the aliasing for 3D Gaussian splatting lies in representing the scene with Gaussians of appropriate scale. Based on this observation, we propose to calculate the pixel coverage of 3D Gaussians during splatting and use this as a criteria for selective rendering. Gaussians that are too large or too small at the current rendering resolution are filtered for anti-aliasing and speed improvements. We also proposed to insert large Gaussians by aggregating small Gaussians during training to preserve the low frequency details and prevent part missing. Our experiments on various datasets support the effectiveness of our algorithm in rendering quality and speed at both high and low resolution, mitigating the severe aliasing artifact of the original 3D Gaussian splatting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 669, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 669, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20930" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Kurt Akeley. Reality engine graphics. In Proceedings of the 20th Annual Conference on Computer Graphics and Interactive Techniques, page 109-116, New York, NY, USA, 1993. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "text", + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 181, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 181, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 181, + 288, + 213 + ], + "type": "text", + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 215, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 287, + 248 + ], + "type": "text", + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 250, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 271 + ], + "type": "text", + "content": "[5] Kristof Beets and David L. Barron. Super-sampling anti-aliasing analyzed. 2000. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 273, + 287, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 287, + 305 + ], + "type": "text", + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 307, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 287, + 328 + ], + "type": "text", + "content": "[7] Carl Erikson. *Polygonal simplification.* Technical Report 96-016, 1996. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 331, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 287, + 373 + ], + "type": "text", + "content": "[8] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Trans. Graph., 37(6), 2018. 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "type": "text", + "content": "[9] Tan Kim Heok and D. Daman. A review on level of detail. In Proceedings. International Conference on Computer Graphics, Imaging and Visualization, 2004. CGIV 2004., pages 70-75, 2004. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 420, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 287, + 475 + ], + "type": "text", + "content": "[10] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19774–19783, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 476, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 542 + ], + "type": "text", + "content": "[11] Jorge Jimenez, Diego Gutierrez, Jason Yang, Alexander Reshetov, Pete Demoreuille, Tobias Berghoff, Cedric Perthuis, Henry Yu, Morgan Mcguire, Timothy Lottes, Hugh Malan, and Emil Persson. Filtering approaches for real-time anti-aliasing. ACM SIGGRAPH 2011 Courses, SIGGRAPH'11, 2011. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 586 + ], + "type": "text", + "content": "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "text", + "content": "[13] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 634, + 274, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 634, + 274, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 634, + 274, + 644 + ], + "type": "text", + "content": "[14] Timothy Lottes. *Fxaa. Technical report*, Nvidia, 2011. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[15] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[16] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 196 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[17] Seungtae Nam, Daniel Rho, Jong Hwan Ko, and Eunbyung Park. Mip-grid: Anti-aliased grid representations for neural radiance fields. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 141, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 163 + ], + "type": "text", + "content": "[18] Lance Williams. Pyramidal parametrics. SIGGRAPH Comput. Graph., 17(3):1-11, 1983. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 164, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 196 + ], + "type": "text", + "content": "[19] M. Zwicker, H. Pfister, J. van Baar, and M. Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS '01., pages 29-538, 2001. 2, 3, 4" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20931" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/82eacc56-0492-4ac2-b935-f0e92d5461ce_content_list.json b/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/82eacc56-0492-4ac2-b935-f0e92d5461ce_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c35d300d34107071813d4e0980b080f1b411b302 --- /dev/null +++ b/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/82eacc56-0492-4ac2-b935-f0e92d5461ce_content_list.json @@ -0,0 +1,1585 @@ +[ + { + "type": "text", + "text": "Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning", + "text_level": 1, + "bbox": [ + 114, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Menghao Zhang, Jingyu Wang\\*, Qi Qi, Haifeng Sun, Zirui Zhuang, Pengfei Ren, Ruilong Ma, Jianxin Liao \nState Key Laboratory of Networking and Switching Technology, \nBeijing University of Posts and Telecommunications", + "bbox": [ + 89, + 203, + 929, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhangmenghao, wangjingyu, qiqi8266, hfsun, zhuangzirui, rpf, maruilong@bupt.edu.cn; jxlbupt@gmail.com", + "bbox": [ + 86, + 258, + 928, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent progress in video anomaly detection suggests that the features of appearance and motion play crucial roles in distinguishing abnormal patterns from normal ones. However, we note that the effect of spatial scales of anomalies is ignored. The fact that many abnormal events occur in limited localized regions and severe background noise interferes with the learning of anomalous changes. Meanwhile, most existing methods are limited by coarse-grained modeling approaches, which are inadequate for learning highly discriminative features to discriminate subtle differences between small-scale anomalies and normal patterns. To this end, this paper addresses multi-scale video anomaly detection by multi-grained spatio-temporal representation learning. We utilize video continuity to design three proxy tasks to perform feature learning at both coarse-grained and fine-grained levels, i.e., continuity judgment, discontinuity localization, and missing frame estimation. In particular, we formulate missing frame estimation as a contrastive learning task in feature space instead of a reconstruction task in RGB space to learn highly discriminative features. Experiments show that our proposed method outperforms state-of-the-art methods on four datasets, especially in scenes with small-scale anomalies.", + "bbox": [ + 75, + 340, + 473, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Anomaly Detection (VAD) is dedicated to detecting anomalous events in videos with wide applications in public safety and intelligent surveillance. The primary challenge of VAD is the sparsity of abnormal samples, limiting the direct learning of abnormal patterns from the available data. Therefore, most previous VAD studies can be divided into the weakly supervised category [9, 29, 36, 40, 43, 50, 53] that learns with video-level annotations, or the category only learning from normal data [5, 23, 24, 39, 46]. Our work", + "bbox": [ + 75, + 741, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bdda958cde7f6fc7eaf99424daddf605a6341516f9f8ef3be8c651bad8370536.jpg", + "image_caption": [ + "Figure 1. Examples of small-scale anomalies and comparisons of results from different methods. From top to bottom: video sequence, ground truth (green regions are abnormal), result of PMem [31] (reconstruction-based method), result of RoadMap [41] (prediction-based method) and result of ours." + ], + "image_footnote": [], + "bbox": [ + 509, + 309, + 883, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "focuses on the latter, which typically models the spatiotemporal features of the normal pattern, while samples that do not conform to the model are labeled as anomalies.", + "bbox": [ + 496, + 597, + 892, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the same scene, the appearance and motion of foreground objects are the main differences between normal and abnormal patterns, and also the crucial features that the model needs to learn. We note that the effect of the spatial scale of the anomalies on feature learning is ignored. As illustrated in Fig. 1, many abnormal events occur in a limited region with only subtle differences from the normal pattern while background occupying a larger portion of the frame is the same. Therefore, severe background noise interferes with the feature learning of anomalous changes. Apart from small-scale anomalies, there are also abnormal events that span almost the entire frame, such as traffic accidents. Such changes in the scale of anomalies require the model to remain robust to anomalies of various scales.", + "bbox": [ + 496, + 643, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unfortunately, most previous methods exhibit a lower accuracy in detecting small-scale anomalies compared to larger-scale anomalies, as shown in Fig. 1. Reconstruction-", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 1, + 807, + 21 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 753, + 45 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "**Corresponding author", + "bbox": [ + 94, + 887, + 225, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17385", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "based [14, 16, 28, 30, 31, 45] and prediction-based [2, 5, 20, 23, 41] approaches are the mainstream modeling paradigms. The two approaches learn spatio-temporal features of normal patterns by reconstructing and predicting frames, respectively. However, both approaches are interfered by severe background noise since they perform reconstruction or prediction of frame in RGB space. Recent object-centric works [11, 12, 17] separate and process foregrounds and backgrounds to address the challenge of background noise. Nevertheless, modeling approaches based on prediction and binary classification struggle to learn highly discriminative features to distinguish small-scale anomalies. Some multi-scale VAD methods [41, 52] learn features at different resolutions through pooling or feature pyramid-ing but do not focus on appearance and motion features.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we aim to explore a robust multi-scale VAD method. On the one hand, the context dependence of the anomalies necessitates the model to grasp global motion patterns and long-range features within the video. On the other hand, detecting small-scale anomalies not only demands that the model avoid the background noise interference but also requires it to discern highly discriminative short-range spatio-temporal variances between frames, owing to their subtle differences from the normal pattern. Consequently, the model needs to comprehensively learn the spatio-temporal features (spatial appearance and temporal motion features) of normal pattern in a multi-grained manner to achieve robustness to multi-scale anomalies.", + "bbox": [ + 75, + 319, + 472, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we take video continuity [19] as supervision to construct three self-supervised proxy tasks, enabling VAD model to learn spatio-temporal features in both coarse-grained and fine-grained manners during training. (i) To determine whether a video sequence is continuous or not. Continuity judgment requires the model to learn the overall long-range temporal features and global motion patterns of the video in a coarse-grained manner. (ii) To locate discontinuities. Finding where the discontinuities occur drives the model to capture changes in local motion in a fine-grained manner rather than background noise. (iii) To estimate the missing frame in feature space instead of RGB space. We formulate the estimation task as a contrastive learning task to avoid the background noise interference and learn highly discriminative features of motion and appearances. By jointly solving three proxy tasks, the model can maintain robustness to anomalies of various scales.", + "bbox": [ + 75, + 517, + 472, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct experiments on four challenging datasets (Avenue [25], ShanghaiTech [27], UCF-Crime [36] and Campus [2]). The experiments show that our proposed method outperforms the State-Of-The-Art (SOTA) method, especially in scenes with small-scale anomalies.", + "bbox": [ + 75, + 776, + 470, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 96, + 853, + 421, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Three straightforward yet effective multi-grained spatiotemporal representation proxy tasks are designed, en", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "abling the video anomaly detection model to maintain robustness to multi-scale anomalies.", + "bbox": [ + 511, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A reconstruction scheme for frames based on contrastive learning is proposed, motivating the model to learn highly discriminative features from both appearance and motion, rather than simple RGB features and background noise.", + "- Our method achieves state-of-the-art performance on four datasets, especially outperforming previous methods by a large margin in scenes with small-scale anomalies." + ], + "bbox": [ + 500, + 121, + 893, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 242, + 640, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Video Anomaly Detection", + "text_level": 1, + "bbox": [ + 500, + 267, + 732, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work focuses on the video anomaly detection method that only learning from normal data, which is mainly grouped into the reconstruction-based methods [4, 16, 28, 33] and the prediction-based methods [20, 41, 48]. These two methods typically use autoencoders (AEs) [15, 51], memory-augmented AEs [13, 14, 23, 28, 31], or generative models [20, 32, 41] to reconstruct current frames or predict future frames so that frames with large reconstruction or prediction errors are recognized as anomalous. While some work attempts to introduce optical flow [23, 30], skeletal information [7, 10, 39], or utilize bi-directional prediction [5, 8, 18, 52] to learn motion and temporal features, these methods cannot be effective for small-scale anomalies due to disturbance from background noise. Recent object-centric methods [12, 17, 39] attempt to separate background and objects to avoid the disturbance of background noise. Although such approach can focus on the foreground objects, the way these methods model the object in terms of classification or RGB prediction lacks the exploration of fine-grained features. In addition, some efforts [1, 24] to synthesize virtual anomaly data fail to work for such anomalies owing to the lack of small-scale virtual data.", + "bbox": [ + 496, + 291, + 893, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from these methods, we expect to enable the model to learn different scales of variation and fine-grained features of the normal pattern via self-supervised spatiotemporal representation learning.", + "bbox": [ + 496, + 625, + 893, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Self-Supervised Learning in VAD", + "text_level": 1, + "bbox": [ + 500, + 695, + 792, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised learning VAD methods typically perform contrastive learning or solve pretext tasks. For instance, Wang et al. [42] performs contrastive learning with clustered attention mechanism. However, it requires customized data enhancement strategies. For pretext tasks, Yang et al. [46] learn spatio-temporal representations by keyframe-based event restoration, while the work [11] designs multiple proxy tasks in an object-centric way to detect anomalies in different aspects. However, the former [46] suffers from severe background noise interference due to reconstructing events in RGB space, while the latter [11] lacks the ability to learn fine-grained features as its tasks are primarily", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17386", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/de84bc4cb7564ce77f45c44a7cad59e9c75ec4148da1917b691eb0db2efc5e23.jpg", + "image_caption": [ + "Figure 2. Method overview. The overall framework consists of a shared backbone network and three branches. The three branches perform continuity judgment, discontinuity localization, and missing frame estimation tasks to learn the spatio-temporal features of normal patterns at coarse- and fine-grained levels, respectively. The feature representation $f$ generated after self-supervised training can effectively detect anomalies of various scales. Blue arrows and grey arrows represent training and inference processes, respectively. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 880, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "binary classifications.", + "bbox": [ + 76, + 462, + 222, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we design proxy tasks based on video continuity [19] to learn spatio-temporal features of normal events in a multi-grained manner to perceive anomalies of various scales. In addition, we formulate the reconstruction of frames as a contrastive learning task to learning highly discriminative features instead of background noise.", + "bbox": [ + 75, + 486, + 468, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Multi-Scale Learning", + "text_level": 1, + "bbox": [ + 76, + 611, + 277, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-scale learning typically considers information from different scales or granularities. Similar to our work, Xiao et al. [44] hierarchically learns the spatio-temporal features through event mask prediction task. However, event-level features are not sufficient for detecting subtle variances between frames. There is also some multi-scale works [41, 52] in VAD. For instance, Wang et al. [41] obtain image features of different resolutions to perform prediction through different pooling operations, while Zhong et al. [52] obtain features at different scales through feature pyramid. Although pooling and feature pyramids can provide feature representations of different resolutions, they cannot specifically capture the motion and appearance features which is crucial for anomaly detection. In contrast, we design the self-supervised proxy tasks to direct the model's attention towards changes in appearance and motion between frames, enabling it to learn features in a multi-grained manner.", + "bbox": [ + 75, + 643, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 460, + 589, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 489, + 609, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed multi-scale video anomaly detection method attempts to learn spatio-temporal features of normal patterns at both coarse-grained and fine-grained levels by jointly solving multiple proxy tasks of spatio-temporal representation learning. For a normal video clip $V$ , our approach is devoted to learning a video encoder $F$ that transforms the video clip $V$ into the scale-aware high-quality spatio-temporal feature $f$ . We define $f$ as the high-quality feature in terms of anomalous information of different scales if it is able to easily address the following questions: (1) Whether there are missing frames in video clips $V$ , i.e., whether feature representation $f$ remains sensitive to large-scale anomalies. (2) If the video clip $V$ is not continuous, where are the missing frames, i.e., whether feature representation $f$ remains sensitive to small-scale anomalies. (3) If video clip $V$ is discontinuous, what are missing frames, i.e., whether the feature representation $f$ contains high-level motion feature and contextual feature.", + "bbox": [ + 496, + 518, + 890, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall architecture of our proposed method is illustrated in Fig. 2. The network consists of three branches and a video encoder $F$ , all of the branches share the same backbone encoder $F$ . Given two nonoverlapping video clips: continuous clip $V_{c}$ , discontinuous clip $V_{d}$ , and the corresponding missing frame $I_{m}$ , each of the three branches solves a proxy task and operates together to optimize the", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17387", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "backbone. The first branch is used to discriminate whether a video clip is continuous or not. The second branch is used to locate the position of the missing frame in the discontinuous clip $V_{d}$ . The last branch is designed to learn high-level features through contrastive learning mechanism, so that the feature representation of the discontinuous clip $V_{d}$ can effectively approach the feature representation of the missing frame $I_{m}$ . After joint optimization of self-supervised tasks, the backbone $F$ can generate scale-aware feature representations $f$ for anomaly inference.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Data Acquisition", + "text_level": 1, + "bbox": [ + 76, + 252, + 243, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To perform the proposed scale-aware proxy tasks, we select continuous video clips and discontinuous clips containing only one missing frame $I_{m}$ , respectively. The continuous clip contains $T$ continuous frames, while the discontinuous clip has a length of $T - 1$ . Given an initial continuous clip of length $T$ , we uniformly select a frame from the interval $[1, T - 2]$ as the missing frame $I_{m}$ to form the breakpoint (counting from 0). And the leading and alternate of the missing frames are concatenated together to form discontinuous video clip $V_{d}$ . The restriction of missing frame inside the clip sequence ensures that the construction of missing frames and discontinuous clips remains consistent.", + "bbox": [ + 75, + 275, + 468, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, we formulate the continuity discrimination task as a binary classification task while the missing frame localization task is formulated as a $T - 2$ class classification task labeled by $I_{m}$ . The estimation task is formulated as a contrastive learning task. Both initial and continuous clips are formed by randomly sampling non-overlapping clips of length $T$ from the same video.", + "bbox": [ + 75, + 455, + 472, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Network Architecture", + "text_level": 1, + "bbox": [ + 76, + 571, + 282, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ a video encoder (defaults a 3D-ConvNet) as the backbone network $F$ and use it as a shared base for three proxy tasks. Each task introduces into its top a projection network, denoted $P_{1}, P_{2}$ , and $P_{3}$ , which are used to process the video clip embeddings $f$ extracted from the backbone network. All projection heads contain spatio-temporal averaging pooling layers to ensure that the dimensionality of the deep feature embeddings remains consistent, even though the temporal lengths may be different.", + "bbox": [ + 75, + 595, + 468, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In continuity judgment and missing frame localization, we use fully connected layers as classifiers at the end of $P_{1}$ and $P_{2}$ for performing classification operations. As for the missing frame estimation task, we use $P_{3}$ to embed the video features $f$ extracted by the backbone network into the low-dimensional features to get the representation, i.e., the feature representation of the missing frame.", + "bbox": [ + 75, + 731, + 468, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Proxy Tasks and Joint Learning", + "text_level": 1, + "bbox": [ + 76, + 847, + 359, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task 1: Continuity Judgment. The continuity judgment is intended to help the model learn spatio-temporal features", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and motion patterns in the video in a coarse-grained manner to perceive large-scale anomalies. In this task, in order for the model to learn by video features rather than other cues such as context, the positive and negative samples in each batch in our training come from the same video scene. With N samples in a training batch, cross-entropy loss $\\mathcal{L}_{CE}$ is used to optimize the model, and the loss for continuity judgment is $\\mathcal{L}_1$ :", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {1} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\mathcal {L} _ {C E} \\left(P _ {1} \\left(F \\left(v _ {d} ^ {i}\\right)\\right) _ {1} + P _ {1} \\left(F \\left(v _ {c} ^ {i}\\right)\\right) _ {0}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 223, + 890, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(\\cdot)$ represents the process of feature processing by the network, $F$ and $P_{1}$ represent the shared backbone network and the continuity judgment projection head, respectively, and $v_{c}^{i}$ and $v_{d}^{i}$ represent the $i_{th}$ continuous and discontinuous video clip, respectively.", + "bbox": [ + 496, + 277, + 890, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task 2: Discontinuity Localization. While the continuity judgment task only learns spatio-temporal features in a coarse-grained manner, the discontinuous localization task expects to learn in a fine-grained manner in order to perceive small-scale anomalies. Compared to binary labels, labeling with missing frames $I_{m}$ in this task can drive the model to learn changes in motion between adjacent frames in a fine-grained manner. Similarly, the discontinuous localization loss $\\mathcal{L}_2$ can be expressed as an average of the cross-entropy loss of $N$ samples in each batch:", + "bbox": [ + 496, + 353, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {C E} \\left(P _ {2} \\left(F \\left(v _ {d} ^ {i}\\right)\\right) _ {I _ {i}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 515, + 890, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{2}$ represents the discontinuity localization projection head and $I_{i}$ represents the missing frame position corresponding to discontinuous clip $v_{d}^{i}$ .", + "bbox": [ + 496, + 568, + 890, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task 3: Missing Frame Estimation. In order to further learn the highly discriminative spatio-temporal features of the video, we estimate missing frames in terms of both appearance and motion. Meanwhile, we choose to estimate frames in feature space rather than RGB space to avoid background noise interference. We design a contrastive learning mechanism to perform this estimation task based on the following facts. First, since adjacent frames and missing frames are temporally connected, adjacent frames and missing frames contain more similar motions compared to non-adjacent frames from the same scene but farther away. Second, video frames from the same scene share more similar appearance features with missing frames than video frames from other scenes.", + "bbox": [ + 496, + 613, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Accordingly, we learn the motion information in the video by triplet loss [34] taking the discontinuous video clips $v_{d}^{i}$ as the anchor point, the missing frames $I_{m}$ as the positive samples, and the continuous video clips $v_{c}^{i}$ as the negative samples. In addition, we take discontinuous clips", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17388", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$v_{d}^{i}$ as anchor points, continuous clips $v_{c}^{i}$ from the same scene as positive samples, and video clips $\\{v^{j}\\}_{i\\neq j}$ from different scenes as negative samples, to learn the appearance information in the video by contrastive loss [19]. Defining the cosine similarity computation operation as $sim(\\cdot ,\\cdot)$ , and the hyperparameter that balances the relative contributions of the triplet loss and the contrastive loss as $\\omega \\in [0,1]$ , the missing frame estimation loss is denoted as:", + "bbox": [ + 76, + 90, + 472, + 213 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {3} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} (\\omega \\times \\max (0, \\gamma - \\left(p _ {i} ^ {+} - p _ {i} ^ {-}\\right)) \\tag {3} \\\\ - (1 - \\omega) \\log \\left(\\frac {q _ {i} ^ {+}}{q _ {i} ^ {+} + \\sum_ {j = 1 , j \\neq i} ^ {N} q _ {i , j ^ {-}}}\\right)), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 224, + 468, + 311 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} ^ {+} = \\mathrm {s i m} \\left(P _ {3} (V _ {d} ^ {i}, I _ {i})\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 328, + 468, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} ^ {-} = \\operatorname {s i m} \\left(P _ {3} \\left(V _ {d} ^ {i}, V _ {c} ^ {i}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 369, + 468, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq _ {i} ^ {+} = \\exp \\left(\\operatorname {s i m} \\left(P _ {3} \\left(V _ {d} ^ {i}, V _ {c} ^ {i}\\right)\\right) / \\tau\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 411, + 468, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} q _ {i, j} ^ {-} = \\exp \\left(\\sin \\left(P _ {3} \\left(v _ {d} ^ {i}, v _ {d} ^ {j}\\right)\\right) / \\tau\\right) \\tag {7} \\\\ + \\exp \\left(\\mathrm {s i m} \\left(P _ {3} (v _ {d} ^ {i}, v _ {c} ^ {j})\\right) / \\tau\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 449, + 468, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_i^+$ and $p_i^-$ denote the similarity between positive sample pairs and negative sample pairs in the triplet loss, respectively, $q_i^+$ denotes a single positive sample pair in the contrastive loss, $q_{i,j}^{-}$ denotes the similarity of negative sample pairs in the contrastive loss, and $\\tau$ is the temperature factor in the contrastive loss. Increasing $\\omega$ allows for more focus on motion information, while decreasing $\\omega$ allows for more focus on appearance information. Overall, the model is incentivized to learn fine-grained motion changes and contextual features, including background information and object appearance.", + "bbox": [ + 75, + 503, + 468, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Joint Optimization. The above three proxy tasks learn spatio-temporal features in videos from different levels, and we jointly optimize our shared video encoder $F$ with the multi-loss function.", + "bbox": [ + 75, + 671, + 468, + 731 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {F} = w _ {1} \\mathcal {L} _ {1} + w _ {2} \\mathcal {L} _ {2} + w _ {3} \\mathcal {L} _ {3}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 744, + 468, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{1}, w_{2}, w_{3} \\in [0,1]$ are the individual weights on the $\\mathcal{L}_1, \\mathcal{L}_2$ and $\\mathcal{L}_3$ . Joint optimization helps the model to learn video spatio-temporal features from both coarse- and fine-grained scales to perceive anomalies of various scales.", + "bbox": [ + 76, + 773, + 468, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. VAD Inference", + "text_level": 1, + "bbox": [ + 76, + 845, + 227, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the inference, we use $T$ frames as the processing unit. Following the setup of the SOTA method [20, 41, 46], we", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "choose not to compute anomaly scores using the error between each estimated frame $\\hat{I}_t$ ( $1 \\leq t \\leq T$ ) and the corresponding ground truth $I_t$ . Instead, we opt to employ the Peak Signal-to-Noise Ratio (PSNR) corresponding to the frame with the largest Mean-Square Error (MSE) among the $T$ frames and the ground truth. This approach serves as the detection metric for the video sequence, as follow:", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M S E} \\left(\\boldsymbol {I} _ {t}, \\hat {\\boldsymbol {I}} _ {t}\\right) = \\frac {1}{H \\cdot W} \\sum_ {i, j} \\left\\| \\boldsymbol {I} _ {t} (i, j) - \\hat {\\boldsymbol {I}} _ {t} (i, j) \\right\\| _ {2} ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 204, + 890, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP S N R _ {t} = 1 0 \\log_ {1 0} \\frac {\\operatorname {M A X} _ {\\boldsymbol {I} _ {t}} ^ {2}}{\\operatorname {M S E} \\left(\\boldsymbol {I} _ {t} , \\hat {\\boldsymbol {I}} _ {t}\\right)}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 250, + 890, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H$ and $W$ are the height and width of the frames, respectively, $I_{t}(i,j) / \\hat{I}_{t}(i,j)$ represents the RGB value of the $(i,j)$ -th pixel in $I_{t} / \\hat{I}_{t}$ , and $\\mathrm{MAX}_{I_t}$ represents the maximum value of the image pixel in $I_{t}$ . Finally, we calculate the anomaly score $S_{t}$ for the $t$ -th frame $I_{t}$ by normalizing the PSNR values over all $T$ frames:", + "bbox": [ + 500, + 297, + 890, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {t} = 1 - \\frac {P S N R _ {t} - \\operatorname* {m i n} _ {t ^ {\\prime}} P S N R _ {t ^ {\\prime}}}{\\operatorname* {m a x} _ {t ^ {\\prime}} P S N R _ {t ^ {\\prime}} - \\operatorname* {m i n} _ {t ^ {\\prime}} P S N R _ {t ^ {\\prime}}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 393, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $1 \\leq t' \\leq T$ . A higher score $S_{t}$ indicates that the frame $I_{t}$ is more likely to be anomalous.", + "bbox": [ + 498, + 431, + 890, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 474, + 624, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 500, + 777, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-Scene Anomaly Dataset Avenue [25] contains 16 training videos and 21 test videos with 47 abnormal events, including running and throwing. The scenes in this dataset are single and the anomalies are mainly large-scale. ShanghaiTech [27] has 13 scenes with complex lighting conditions and different perspectives. In addition, the dataset includes anomalies caused by sudden movements, such as chasing and arguing. The different perspectives and the unfixed position of the camera lead to a large variation of both the object scale and the anomaly scale in the scene. This is a significant challenge to the robustness of anomaly detection methods to scale variations. UCF-Crime [36] comprises 13 anomaly types, spanning a total of 128 hours of video footage. This dataset is complex because of the unconstrained backgrounds and anomalies of various scales. The training set contains video-level annotations, whereas the testing set includes frame-level annotations.", + "bbox": [ + 496, + 522, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scene-Dependent Anomaly Dataset. Campus [2] is currently the most challenging dataset in its field with 43 scenes, 28 classes of anomalous events and 16 hours of videos. Especially, it contains scene-dependent anomalies, which means an normal event may be abnormal in another scene. Detecting scene-dependent anomalies requires the model to understand the scene and learn highly discriminative features rather than overfit.", + "bbox": [ + 496, + 779, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17389", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/482eaa015823174570d0fcac6ab38d360ee8e07ad3d27e057e1b6dd40f98498f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLength TShanghaiTechCampus
AUCAUC
MicroMacroMicroMacro
Continuity Judgment878.182.360.163.2
1277.281.160.363.9
1675.179.258.260.9
Discontinuity Localization882.685.562.165.4
1283.487.262.665.1
1682.185.561.965.0
Missing Frame Estimation879.883.263.566.8
1281.284.965.268.9
1680.584.064.968.0
Overall Framework884.288.866.369.2
1285.189.867.370.9
1683.988.366.069.2
", + "bbox": [ + 81, + 88, + 464, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics The area under the ROC curve (AUC) serves as a commonly used metric for evaluation and comparison. A higher AUC score indicates a better anomaly detection capability. Following previous research [12], we evaluate both the Micro and Macro versions of AUC.", + "bbox": [ + 75, + 367, + 468, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 452, + 294, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train and evaluate our method with an NVIDIA RTX 3090 GPU. In the training phase, we resize the resolution of all input video clips to $256 \\times 256$ pixels, while the values of the pixels in all frames are normalized to the range [0, 1]. For the pre-training of the three proxy tasks, we utilize AdamW as the optimizer while the length of the continuous video clips is set to $T = 12$ frames. The initial learning rate is set to 0.0003 and is gradually decayed following the scheme of cosine annealing. For the hyperparameters, we set $\\omega = 0.5$ in Equation 3, $w_{1} = w_{2} = w_{3} = 1$ in Equation 8, and the temperature factor $\\tau = 0.1$ in the contrastive loss. In our reported experimental performance, the shared backbone network $F$ is implemented as I3D-RGB [3]. In addition, we incorporate the designed proxy tasks into the backbone of the SOTA methods [2, 11], the experimental results are also reported. Following the existing method setup [11, 24, 33], we perform the proxy task at the object-level with the multi-task backbone [11] and at the frame-level with the other backbones.", + "bbox": [ + 75, + 474, + 468, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 771, + 230, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering the ShanghaiTech and Campus datasets contain diverse perspectives and scenes, and large variations in scales of anomaly, we conduct exhaustive ablation experiments on the ShanghaiTech and Campus datasets.", + "bbox": [ + 75, + 794, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sensitivity to the Length of Video Clips. The length $T$ of the input clips is an important setting in our proposed framework. With a small $T$ , both continuity judgments and dis", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/eb6e57fab9fe008ef0843f1591e97f70a2137fcf64ddfbef0a9547a3548241a0.jpg", + "table_caption": [ + "Table 1. Comparison of proposed framework and individual proxy tasks between different lengths $T$ of input clips. We report Micro and Macro AUC (%) on ShanghaiTech and Campus datasets. The best performing results are marked in bold." + ], + "table_footnote": [], + "table_body": "
IDCJDLMFEFPShanghaiTechCampus
AUCAUC
MicroMacroMicroMacro
1---✓70.271.258.158.6
2✓---77.281.160.363.9
3-✓--83.487.262.665.1
4--✓-81.284.965.268.9
5✓✓--84.187.862.865.4
6✓-✓-82.385.665.369.2
7-✓✓-84.589.566.270.1
8✓✓✓-85.189.867.370.9
9✓✓-✓84.287.662.865.6
", + "bbox": [ + 504, + 88, + 887, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Ablation experiments on the contributions of each proxy task. We report the AUC (\\%) scores on ShanghaiTech and Campus datasets. 'CJ', 'DL' and 'MFE' stand for the three proxy tasks of continuity judgment, discontinuity localization and missing frame estimation, respectively. In addition we add experiments on frame prediction for comparison, 'FP' represents frame prediction.", + "bbox": [ + 496, + 255, + 890, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "continuity localization will be so ambiguous as to be difficult. While with a larger $T$ , the model may not need to learn high-quality spatio-temporal features to solve the task due to enough spatio-temporal information in the feed. Table 1 shows the sensitivity of each proxy task to $T$ . For the discontinuous localization and missing frame estimation tasks, $T = 12$ presents better performance. Although at $T = 8$ , the continuity judgment yields competitive results on the ShanghaiTech dataset, the overall framework achieves the best outcomes with jointly optimized at $T = 12$ .", + "bbox": [ + 496, + 354, + 890, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Continuity Judgment. The results of the ablation experiments for each proxy task are shown in Table 2, where we take the frame prediction [20] performed on the same backbone $F$ as the baseline and train it with gradient loss and intensity loss (ID 1). Performing the continuity judgment task alone improves the performance of the model compared to frame prediction (ID 2). This suggests that continuity judgment task motivates the model to generate higher quality representations than frame prediction.", + "bbox": [ + 496, + 507, + 890, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Discontinuity Localization. Compared to other proxy tasks and baseline, the performance of the discontinuous localization task alone can be significantly improved, especially on ShanghaiTech which contains anomalies of various scales (ID 3, ID 5 and ID 7). As this task samples discontinuous positions uniformly along the time axis, it motivates the model to capture subtle variations between frames, obtaining fine-grained representations.", + "bbox": [ + 496, + 643, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Missing Frame Estimation. Performing missing frame estimation alone is effective in improving performance on Campus compared to baseline and other proxy tasks, which contains scene-dependent anomalies (ID 4). This improvement illustrates that our proposed contrastive learning scheme can help the model understand the scene. Compared to performing the frame prediction task to learn RGB features (ID 1 and ID 9), our proposed scheme learns highly discriminative motion features and semantically ef", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17390", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fa70421f712d3c5b65d754d1c03eb722cee419da0f9105a408951e6e2e7334cc.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Comparison with SOTA methods of the Micro and Macro AUC (%) on Avenue and ShanghaiTech datasets. The best performing results are marked in bold.", + "$\\nabla$ : Methods apply virtual dataset for training.", + "*: Methods utilize multi-task model ([11] or [12]) as backbone." + ], + "table_body": "
yearMethodAvenueShanghaiTech
AUCAUC
MicroMacroMicroMacro
2018Liu et al. [20]85.1-72.8-
Liu et al. [22]84.4---
Sultani et al. [36]--76.5-
2019Gong et al. [14]83.8-71.2-
Lee et al. [18]90.0-76.2-
Ionescu et al. [17]87.490.478.784.9
2020Park et al. [31]88.5-70.5-
Sun et al. [37]89.6-74.7-
Lu et al. [26]85.8-75.9-
Wang et al. [42]87.0-79.3-
Yu et al. [47]89.6-74.8-
2021Liu et al. [23]91.1-76.2-
Lv et al. [28]89.5-73.8-
Georgescu et al. [11]*91.591.982.489.3
Georgescu et al. [12]*92.390.482.789.3
2022Wang et al. [41]88.3-76.6-
Zaheer et al. [49]74.2-79.6-
Chen et al. [5]90.3-78.1-
Zhong et al. [52]89.0-74.5-
Cho et al. [6]88.0-76.3-
Yang et al. [45]89.9-74.7-
Ristea et al. [33]*92.991.983.689.5
Acsintoae et al. [1]▼*93.093.283.790.5
2023Yang et al. [46]89.9-73.8-
Cao et al. [2]86.8-79.2-
Liu et al. [21]92.8-78.8-
Singh et al. [35]86.0-76.6-
Sun et al. [39]92.4-83.0-
Sun et al. [38]91.5-78.6-
Liu et al. [24]▼91.892.383.887.8
Liu et al. [24] ▼*93.693.985.091.4
Ours92.492.985.189.8
Ours▼93.292.586.291.0
Ours*(Estimation Task)92.693.085.591.8
Ours*(Three Tasks)93.693.886.892.4
Ours▼*(Three Tasks)94.394.587.593.0
", + "bbox": [ + 81, + 87, + 464, + 588 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5d3546b814017dd527849bde0af4a868fe26cd7c170d5eaaed047b97f2eac24d.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Results of Micro and Macro AUC(%) on UCF-Crime dataset. The best performing results are marked in bold." + ], + "table_body": "
MethodReferenceMicro AUCMacro AUC
Park et al. [31]CVPR2068.972.4
Georgescu et al. [11]CVPR2174.678.2
Wang et al. [41]TNNLS2272.976.8
Sun et al. [39]CVPR2375.578.3
Ours-80.683.9
", + "bbox": [ + 81, + 685, + 465, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fective contexts, including the appearance of objects and background (ID 4 and ID 8).", + "bbox": [ + 76, + 824, + 468, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Joint Optimization. Although a single task may be able to achieve competitive performance, Experiments have shown that when proxy tasks are executed in pairs (ID", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/58e128a9a2fed2772ffea3a9c3f635cd0bfd35f30008dd83153f851790cf81d9.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Comparison with SOTA methods of the Micro and Macro AUC(%) on Campus dataset. The best performing results are marked in bold." + ], + "table_body": "
MethodReferenceMiceo AUCMacro AUC
Liu et al. [20]CVPR1857.960.2
Gong et al. [14]CVPR1961.962.5
Ionescu et al. [17]CVPR1959.363.4
Park et al. [31]CVPR2062.563.6
Liu et al. [23]ICCV2163.7-
Lv et al. [28]CVPR2164.4-
Wang et al. [41]TNNLS2261.964.2
Cao et al. [2]△CVPR2368.2-
Ours-67.370.9
Ours△-70.172.2
", + "bbox": [ + 504, + 87, + 890, + 241 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\triangle$ :Version that utilize scene-conditioned model [2] as backbone.", + "bbox": [ + 501, + 292, + 880, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5, ID 6 and ID 7) and all tasks are executed together (ID 8), the performance of the model continues to improve to an optimum. This joint improvement is attributed to the fact that different proxy tasks learn spatio-temporal features at different levels, and they complement each other in motion and contextual feature learning.", + "bbox": [ + 498, + 334, + 890, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparisons with State-Of-The-Arts", + "text_level": 1, + "bbox": [ + 500, + 436, + 821, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the proposed framework with SOTA methods in terms of Micro and Macro AUC(\\%). It is noteworthy that the current state-of-the-art methods [1, 24, 33] utilize the multi-task framework proposed by Georgescu et al. ([11] and [12], respectively) as the backbone, where [1] and [24] employ virtual data for training. Therefore, we evaluate different configurations of the proposed method on the Avenue and ShanghaiTech datasets, as shown in Table 3. With the multi-task backbone [11] (marked with * in Table 3), we assess two versions: one involves replacing only the prediction component with our designed missing estimation task, while keeping the other proxy tasks unchanged. The other version replaces both proxy tasks in the motion branch with continuity judgment and discontinuity localization tasks in addition to incorporating the missing estimation task. In addition, we report the performance of applying virtual data for training (marked with $\\nabla$ in Table 3).", + "bbox": [ + 496, + 460, + 890, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Avenue. As shown in Table 3, our method achieved the highest AUC scores, obtaining a Micro AUC of $94.3\\%$ and a Macro AUC of $94.5\\%$ . Without the backbone of multi-task [11, 12] and virtual data [24], our method still scored the best performance with $92.9\\%$ Macro AUC.", + "bbox": [ + 496, + 718, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on ShanghaiTech. Our proposed method yields a Micro AUC of $87.5\\%$ and a Macro AUC of $93.0\\%$ on the ShanghaiTech dataset. With the backbone of multi-task [11, 12], we improve the Micro AUC by $3.1\\%$ compared to Georgescu et al. [11] in the setup with only the missing estimation task, and by $4.4\\%$ when incorporating three designed proxy tasks. Without the backbone of multi-task", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17391", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a234b3d579b70549459df2c7994935fb761c60a799895e7e9098136854df9a1c.jpg", + "image_caption": [ + "Figure 3. Examples of detection for anomalies of different scales from the ShanghaiTech dataset, including ground truth and prediction or reconstruction error maps from different methods (from top to bottom: ROADMAP[41], DEDDnet[52], USTN-DSC[46], and our method). Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 460, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[11, 12] and virtual data [24], our method still achieves a Micro AUC score of $85.1\\%$ , which is comparable to the method [24] that employed the multi-task backbone and virtual training data. In comparison with the multi-scale VAD methods ([41] and [52]), our method achieves a significant improvement of $8.5\\%$ and $10.6\\%$ , respectively.", + "bbox": [ + 75, + 484, + 467, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on UCF-Crime. Due to the absence of published results (methods only learning from normal data) on the UCF-Crime dataset, we implement the code from the existing literature [11, 31, 39, 41]. As shown in Table 4, our proposed method achieves a significant improvement compared to the second best method by $5.1\\%$ in terms of Micro AUC and $5.6\\%$ in terms of Macro AUC.", + "bbox": [ + 75, + 577, + 467, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on Campus. We evaluate an additional version on the Campus dataset with the scene-conditioned model (provided by Cao et al. [2]) as the backbone, as it is necessary to incorporate scene features for detecting scene-dependent anomalies. As shown in Table 5, our method achieves the highest Micro AUC score of $70.1\\%$ and Macro AUC score of $72.2\\%$ . Without the scene-conditioned model, our method still obtain a Micro AUC of $67.3\\%$ , surpassing other RGB reconstruction-based [14, 28, 31] or RGB prediction-based methods [20, 23, 41].", + "bbox": [ + 75, + 683, + 467, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 845, + 259, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 1 shows the anomaly curves of the test video and the comparison of our method with PMem [31] (reconstruction-", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/531e5c2fdf059771824956ac7c8841e96d3c08419d0f70952f4c22459a4459db.jpg", + "image_caption": [ + "Figure 4. Visualization of salient regions of test frames by the proposed method. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 509, + 89, + 625, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0689b7c8e5d1a25f72a479d59d49f68678c849016004bc9c1ffe8b3f9e18261b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 89, + 754, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e3baa39affc129a32a1531fb8896679998a5fbe4ead68b3e85c6aebd63fe2570.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 89, + 880, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "based method) and RoadMap [41] (prediction-based method). When there are small-scale anomalies in the sampled videos, the anomaly score curves of the other two methods have obvious missed detections, while our method can accurately perceive the small-scale anomalies. In addition, Fig. 3 illustrates more examples, including the comparison with the multi-scale VAD methods [41, 52] and the comparison with self-supervised VAD method [46]. It can be observed that for normal regions in video frames, our method is able to estimate them well, while for abnormal event regions large errors occur. While other methods generate large errors for some normal regions, especially for frames with small-scale anomalies.", + "bbox": [ + 496, + 257, + 890, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualizations of the salient regions of the frames by our method are illustrated in Fig. 4. For both large-scale anomalies (1-th row) and small-scale anomalies (2-th row), our model can accurately focus on the anomalous regions.", + "bbox": [ + 498, + 454, + 890, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 527, + 617, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we observe that anomalous spatial scales affect the feature learning of the model. Small-scale anomalies require the model to learn highly discriminative features in a fine-grained manner rather than background noise. To detect multi-scale anomalies, we design proxy tasks supervised by video continuity that motivate the model to learn spatio-temporal features in both coarse-grained and fine-grained manners. Furthermore, instead of performing RGB reconstruction and prediction, we estimation frames in feature space through contrastive learning to learn highly discriminative features of appearance and motion. Experiments conducted on four challenging benchmark datasets validate the effectiveness of our proposed method.", + "bbox": [ + 496, + 553, + 890, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grants (62101064, 62171057, 62201072, U23B2001, 62001054, 62071067), the National Postdoctoral Program for Innovative Talents under Grant BX20230052, China Postdoctoral Science Foundation (2023TQ0039), the Ministry of Education and China Mobile Joint Fund (MCM20200202, MCM20180101), Beijing University of Posts and Telecommunications-China Mobile Research Institute Joint Innovation Center.", + "bbox": [ + 496, + 750, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17392", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. *Ubnormal: New benchmark for supervised open-set video anomaly detection.* In *CVPR*, pages 20111-20121, 2022. 2, 7", + "[2] Congqi Cao, Yue Lu, Peng Wang, and Yanning Zhang. A new comprehensive benchmark for semi-supervised video anomaly detection and anticipation. In CVPR, pages 20392-20401, 2023. 2, 5, 6, 7, 8", + "[3] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, pages 4724-4733, 2017. 6", + "[4] Yunpeng Chang, Zhigang Tu, Wei Xie, and Junsong Yuan. Clustering driven deep autoencoder for video anomaly detection. In ECCV, pages 329-345, 2020. 2", + "[5] Chengwei Chen, Yuan Xie, Shaohui Lin, Angela Yao, Guannan Jiang, Wei Zhang, Yanyun Qu, Ruizhi Qiao, Bo Ren, and Lizhuang Ma. Comprehensive regularization in a bidirectional predictive network for video anomaly detection. In AAAI, pages 230-238, 2022. 1, 2, 7", + "[6] MyeongAh Cho, Taeoh Kim, Woo Jin Kim, Suhwan Cho, and Sangyoun Lee. Unsupervised video anomaly detection via normalizing flows with implicit latent features. Pattern Recognit., 129:108703, 2022. 7", + "[7] Romero F. A. B. de Morais, Vuong Le, Truyen Tran, Budhaditya Saha, Moussa Reda Mansour, and Svetha Venkatesh. Learning regularity in skeleton trajectories for anomaly detection in videos. In CVPR, pages 11996-12004, 2019. 2", + "[8] Zhiwen Fang, Jiafei Liang, Joey Tianyi Zhou, Yang Xiao, and Feng Yang. Anomaly detection with bidirectional consistency in videos. IEEE Trans. Neural Networks Learn. Syst., 33(3):1079-1092, 2022. 2", + "[9] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. MIST: multiple instance self-training framework for video anomaly detection. In CVPR, pages 14009-14018, 2021. 1", + "[10] Alessandro Flaborea, Luca Collorone, Guido Maria D'Amely di Melendugno, Stefano D'Arrigo, Bardh Prenkaj, and Fabio Galasso. Multimodal motion conditioned diffusion model for skeleton-based video anomaly detection. In (ICCV), pages 10318-10329, 2023. 2", + "[11] Mariana-Iuliana Georgescu, Antonio Barbalau, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. Anomaly detection in video via self-supervised and multi-task learning. In CVPR, pages 12742-12752, 2021. 2, 6, 7, 8", + "[12] Mariana-Iuliana Georgescu, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4505-4523, 2022. 2, 6, 7, 8", + "[13] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In ICCV, pages 1705-1714, 2019. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In ICCV, pages 1705-1714, 2019. 2, 7, 8", + "[15] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal regularity in video sequences. In CVPR, pages 733-742, 2016. 2", + "[16] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In ICCV, pages 8771-8780, 2021. 2", + "[17] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In CVPR, pages 7842–7851, 2019. 2, 7", + "[18] Sangmin Lee, Hak Gu Kim, and Yong Man Ro. BMAN: bidirectional multi-scale aggregation networks for abnormal event detection. IEEE Trans. Image Process., 29:2395-2408, 2019. 2, 7", + "[19] Hanwen Liang, Niamul Quader, Zhixiang Chi, Lizhe Chen, Peng Dai, Juwei Lu, and Yang Wang. Self-supervised spatiotemporal representation learning by exploiting video continuity. In AAAI, pages 1564-1573, 2022. 2, 3, 5", + "[20] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In CVPR, pages 6536-6545, 2018. 2, 5, 6, 7, 8", + "[21] Wenrui Liu, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Diversity-measurable anomaly detection. In CVPR, pages 12147-12156, 2023. 7", + "[22] Yusha Liu, Chun-Liang Li, and Barnabás Póczos. Classifier two sample test for video anomaly detections. In BMVC, page 71. BMVA Press, 2018. 7", + "[23] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In ICCV, pages 13568-13577, 2021. 1, 2, 7, 8", + "[24] Zuhao Liu, Xiao-Ming Wu, Dian Zheng, Kun-Yu Lin, and Wei-Shi Zheng. Generating anomalies for video anomaly detection with prompt-based feature mapping. In CVPR, pages 24500-24510. IEEE, 2023. 1, 2, 6, 7, 8", + "[25] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 FPS in MATLAB. In ICCV, pages 2720-2727, 2013. 2, 5", + "[26] Yiwei Lu, Frank Yu, Mahesh Kumar Krishna Reddy, and Yang Wang. Few-shot scene-adaptive anomaly detection. In ECCV, pages 125-141, 2020. 7", + "[27] Weixin Luo, Wen Liu, and Shenghua Gao. A revisit of sparse coding based anomaly detection in stacked RNN framework. In ICCV, pages 341-349, 2017. 2, 5", + "[28] Hui Lv, Chen Chen, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Learning normal dynamics in videos with meta prototype network. In CVPR, pages 15425-15434, 2021. 2, 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17393", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In CVPR, pages 8022-8031, 2023. 1", + "[30] Trong-Nguyen Nguyen and Jean Meunier. Anomaly detection in video sequence with appearance-motion correspondence. In ICCV, pages 1273-1283, 2019. 2", + "[31] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In CVPR, pages 14360-14369, 2020. 1, 2, 7, 8", + "[32] Mahdyar Ravanbakhsh, Enver Sangineto, Moin Nabi, and Nicu Sebe. Training adversarial discriminators for cross-channel abnormal event detection in crowds. In WACV, pages 1896-1904, 2019. 2", + "[33] Nicolae-Catalin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In CVPR, pages 13566-13576, 2022. 2, 6, 7", + "[34] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823, 2015. 4", + "[35] Ashish Singh, Michael J Jones, and Erik G Learned-Miller. Eval: Explainable video anomaly localization. In CVPR, pages 18717-18726, 2023. 7", + "[36] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, pages 6479-6488, 2018. 1, 2, 5, 7", + "[37] Che Sun, Yunde Jia, Yao Hu, and Yuwei Wu. Scene-aware context reasoning for unsupervised abnormal event detection in videos. In ACM Multimedia, pages 184-192, 2020. 7", + "[38] Che Sun, Chenrui Shi, Yunde Jia, and Yuwei Wu. Learning event-relevant factors for video anomaly detection. In AAAI, pages 2384-2392, 2023. 7", + "[39] Shengyang Sun and Xiaojin Gong. Hierarchical semantic contrast for scene-aware video anomaly detection. In CVPR, pages 22846-22856, 2023. 1, 2, 7, 8", + "[40] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W. Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In ICCV, pages 4955-4966, 2021. 1", + "[41] Xuanzhao Wang, Zhengping Che, Bo Jiang, Ning Xiao, Ke Yang, Jian Tang, Jieping Ye, Jingyu Wang, and Qi Qi. Robust unsupervised video anomaly detection by multipath frame prediction. IEEE Trans. Neural Networks Learn. Syst., 33 (6):2301-2312, 2022. 1, 2, 3, 5, 7, 8", + "[42] Ziming Wang, Yuexian Zou, and Zeming Zhang. Cluster attention contrast for video anomaly detection. In ACM Multimedia, pages 2463-2471, 2020. 2, 7", + "[43] Jie Wu, Wei Zhang, Guanbin Li, Wenhao Wu, Xiao Tan, Yingying Li, Errui Ding, and Liang Lin. Weakly-supervised spatio-temporal anomaly detection in surveillance video. In IJCAI, pages 1172-1178, 2021. 1", + "[44] Fanyi Xiao, Kaustav Kundu, Joseph Tighe, and Davide Modolo. Hierarchical self-supervised representation learning for movie understanding. In CVPR, pages 9717-9726, 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Zhiwei Yang, Peng Wu, Jing Liu, and Xiaotao Liu. Dynamic local aggregation network with adaptive clusterer for anomaly detection. In ECCV, pages 404-421, 2022. 2, 7", + "[46] Zhiwei Yang, Jing Liu, Zhaoyang Wu, Peng Wu, and Xiaotao Liu. Video event restoration based on keyframes for video anomaly detection. In CVPR, pages 14592-14601, 2023. 1, 2, 5, 7, 8", + "[47] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze test helps: Effective video anomaly detection via learning to complete video events. In ACM Multimedia, pages 583-591, 2020. 7", + "[48] Jongmin Yu, Younkwan Lee, Kin Choong Yow, Moongu Jeon, and Witold Pedrycz. Abnormal event detection and localization via adversarial event prediction. IEEE Trans. Neural Networks Learn. Syst., 33(8):3572-3586, 2022. 2", + "[49] Muhammad Zaigham Zaheer, Arif Mahmood, Muhammad Haris Khan, Mattia Segù, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In CVPR, pages 14724–14734, 2022. 7", + "[50] Chen Zhang, Guorong Li, Yuankai Qi, Shuhui Wang, Laiyun Qing, Qingming Huang, and Ming-Hsuan Yang. Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection. In CVPR, pages 16271-16280, 2023. 1", + "[51] Yiru Zhao, Bing Deng, Chen Shen, Yao Liu, Hongtao Lu, and Xian-Sheng Hua. Spatio-temporal autoencoder for video anomaly detection. In ACM Multimedia, pages 1933-1941, 2017. 2", + "[52] Yuanhong Zhong, Xia Chen, Yongting Hu, Panliang Tang, and Fan Ren. Bidirectional spatio-temporal feature learning with multiscale evaluation for video anomaly detection. IEEE Trans. Circuits Syst. Video Technol., 32(12):8285-8296, 2022. 2, 3, 7, 8", + "[53] Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection. In AAAI, pages 3769-3777. AAAI, 2023. 1" + ], + "bbox": [ + 501, + 92, + 890, + 612 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17394", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/82eacc56-0492-4ac2-b935-f0e92d5461ce_model.json b/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/82eacc56-0492-4ac2-b935-f0e92d5461ce_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1c33be33b24b3c06d9a197eb1d719b6629fbf307 --- /dev/null +++ b/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/82eacc56-0492-4ac2-b935-f0e92d5461ce_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.002, + 0.808, + 0.022 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.754, + 0.046 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.13, + 0.856, + 0.177 + ], + "angle": 0, + "content": "Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.204, + 0.93, + 0.257 + ], + "angle": 0, + "content": "Menghao Zhang, Jingyu Wang\\*, Qi Qi, Haifeng Sun, Zirui Zhuang, Pengfei Ren, Ruilong Ma, Jianxin Liao \nState Key Laboratory of Networking and Switching Technology, \nBeijing University of Posts and Telecommunications" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.26, + 0.929, + 0.275 + ], + "angle": 0, + "content": "zhangmenghao, wangjingyu, qiqi8266, hfsun, zhuangzirui, rpf, maruilong@bupt.edu.cn; jxlbupt@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.474, + 0.691 + ], + "angle": 0, + "content": "Recent progress in video anomaly detection suggests that the features of appearance and motion play crucial roles in distinguishing abnormal patterns from normal ones. However, we note that the effect of spatial scales of anomalies is ignored. The fact that many abnormal events occur in limited localized regions and severe background noise interferes with the learning of anomalous changes. Meanwhile, most existing methods are limited by coarse-grained modeling approaches, which are inadequate for learning highly discriminative features to discriminate subtle differences between small-scale anomalies and normal patterns. To this end, this paper addresses multi-scale video anomaly detection by multi-grained spatio-temporal representation learning. We utilize video continuity to design three proxy tasks to perform feature learning at both coarse-grained and fine-grained levels, i.e., continuity judgment, discontinuity localization, and missing frame estimation. In particular, we formulate missing frame estimation as a contrastive learning task in feature space instead of a reconstruction task in RGB space to learn highly discriminative features. Experiments show that our proposed method outperforms state-of-the-art methods on four datasets, especially in scenes with small-scale anomalies." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.717, + 0.21, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Video Anomaly Detection (VAD) is dedicated to detecting anomalous events in videos with wide applications in public safety and intelligent surveillance. The primary challenge of VAD is the sparsity of abnormal samples, limiting the direct learning of abnormal patterns from the available data. Therefore, most previous VAD studies can be divided into the weakly supervised category [9, 29, 36, 40, 43, 50, 53] that learns with video-level annotations, or the category only learning from normal data [5, 23, 24, 39, 46]. Our work" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.31, + 0.885, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.5, + 0.895, + 0.571 + ], + "angle": 0, + "content": "Figure 1. Examples of small-scale anomalies and comparisons of results from different methods. From top to bottom: video sequence, ground truth (green regions are abnormal), result of PMem [31] (reconstruction-based method), result of RoadMap [41] (prediction-based method) and result of ours." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.893, + 0.643 + ], + "angle": 0, + "content": "focuses on the latter, which typically models the spatiotemporal features of the normal pattern, while samples that do not conform to the model are labeled as anomalies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.854 + ], + "angle": 0, + "content": "In the same scene, the appearance and motion of foreground objects are the main differences between normal and abnormal patterns, and also the crucial features that the model needs to learn. We note that the effect of the spatial scale of the anomalies on feature learning is ignored. As illustrated in Fig. 1, many abnormal events occur in a limited region with only subtle differences from the normal pattern while background occupying a larger portion of the frame is the same. Therefore, severe background noise interferes with the feature learning of anomalous changes. Apart from small-scale anomalies, there are also abnormal events that span almost the entire frame, such as traffic accidents. Such changes in the scale of anomalies require the model to remain robust to anomalies of various scales." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Unfortunately, most previous methods exhibit a lower accuracy in detecting small-scale anomalies compared to larger-scale anomalies, as shown in Fig. 1. Reconstruction-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.227, + 0.901 + ], + "angle": 0, + "content": "**Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17385" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "based [14, 16, 28, 30, 31, 45] and prediction-based [2, 5, 20, 23, 41] approaches are the mainstream modeling paradigms. The two approaches learn spatio-temporal features of normal patterns by reconstructing and predicting frames, respectively. However, both approaches are interfered by severe background noise since they perform reconstruction or prediction of frame in RGB space. Recent object-centric works [11, 12, 17] separate and process foregrounds and backgrounds to address the challenge of background noise. Nevertheless, modeling approaches based on prediction and binary classification struggle to learn highly discriminative features to distinguish small-scale anomalies. Some multi-scale VAD methods [41, 52] learn features at different resolutions through pooling or feature pyramid-ing but do not focus on appearance and motion features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.473, + 0.515 + ], + "angle": 0, + "content": "In this paper, we aim to explore a robust multi-scale VAD method. On the one hand, the context dependence of the anomalies necessitates the model to grasp global motion patterns and long-range features within the video. On the other hand, detecting small-scale anomalies not only demands that the model avoid the background noise interference but also requires it to discern highly discriminative short-range spatio-temporal variances between frames, owing to their subtle differences from the normal pattern. Consequently, the model needs to comprehensively learn the spatio-temporal features (spatial appearance and temporal motion features) of normal pattern in a multi-grained manner to achieve robustness to multi-scale anomalies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.473, + 0.775 + ], + "angle": 0, + "content": "To this end, we take video continuity [19] as supervision to construct three self-supervised proxy tasks, enabling VAD model to learn spatio-temporal features in both coarse-grained and fine-grained manners during training. (i) To determine whether a video sequence is continuous or not. Continuity judgment requires the model to learn the overall long-range temporal features and global motion patterns of the video in a coarse-grained manner. (ii) To locate discontinuities. Finding where the discontinuities occur drives the model to capture changes in local motion in a fine-grained manner rather than background noise. (iii) To estimate the missing frame in feature space instead of RGB space. We formulate the estimation task as a contrastive learning task to avoid the background noise interference and learn highly discriminative features of motion and appearances. By jointly solving three proxy tasks, the model can maintain robustness to anomalies of various scales." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.472, + 0.852 + ], + "angle": 0, + "content": "We conduct experiments on four challenging datasets (Avenue [25], ShanghaiTech [27], UCF-Crime [36] and Campus [2]). The experiments show that our proposed method outperforms the State-Of-The-Art (SOTA) method, especially in scenes with small-scale anomalies." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.854, + 0.423, + 0.868 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "- Three straightforward yet effective multi-grained spatiotemporal representation proxy tasks are designed, en" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "abling the video anomaly detection model to maintain robustness to multi-scale anomalies." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.894, + 0.182 + ], + "angle": 0, + "content": "- A reconstruction scheme for frames based on contrastive learning is proposed, motivating the model to learn highly discriminative features from both appearance and motion, rather than simple RGB features and background noise." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.183, + 0.895, + 0.228 + ], + "angle": 0, + "content": "- Our method achieves state-of-the-art performance on four datasets, especially outperforming previous methods by a large margin in scenes with small-scale anomalies." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.122, + 0.895, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.243, + 0.642, + 0.258 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.268, + 0.733, + 0.284 + ], + "angle": 0, + "content": "2.1. Video Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.895, + 0.625 + ], + "angle": 0, + "content": "Our work focuses on the video anomaly detection method that only learning from normal data, which is mainly grouped into the reconstruction-based methods [4, 16, 28, 33] and the prediction-based methods [20, 41, 48]. These two methods typically use autoencoders (AEs) [15, 51], memory-augmented AEs [13, 14, 23, 28, 31], or generative models [20, 32, 41] to reconstruct current frames or predict future frames so that frames with large reconstruction or prediction errors are recognized as anomalous. While some work attempts to introduce optical flow [23, 30], skeletal information [7, 10, 39], or utilize bi-directional prediction [5, 8, 18, 52] to learn motion and temporal features, these methods cannot be effective for small-scale anomalies due to disturbance from background noise. Recent object-centric methods [12, 17, 39] attempt to separate background and objects to avoid the disturbance of background noise. Although such approach can focus on the foreground objects, the way these methods model the object in terms of classification or RGB prediction lacks the exploration of fine-grained features. In addition, some efforts [1, 24] to synthesize virtual anomaly data fail to work for such anomalies owing to the lack of small-scale virtual data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.894, + 0.686 + ], + "angle": 0, + "content": "Different from these methods, we expect to enable the model to learn different scales of variation and fine-grained features of the normal pattern via self-supervised spatiotemporal representation learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.794, + 0.712 + ], + "angle": 0, + "content": "2.2. Self-Supervised Learning in VAD" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Self-supervised learning VAD methods typically perform contrastive learning or solve pretext tasks. For instance, Wang et al. [42] performs contrastive learning with clustered attention mechanism. However, it requires customized data enhancement strategies. For pretext tasks, Yang et al. [46] learn spatio-temporal representations by keyframe-based event restoration, while the work [11] designs multiple proxy tasks in an object-centric way to detect anomalies in different aspects. However, the former [46] suffers from severe background noise interference due to reconstructing events in RGB space, while the latter [11] lacks the ability to learn fine-grained features as its tasks are primarily" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "17386" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.09, + 0.881, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.38, + 0.893, + 0.437 + ], + "angle": 0, + "content": "Figure 2. Method overview. The overall framework consists of a shared backbone network and three branches. The three branches perform continuity judgment, discontinuity localization, and missing frame estimation tasks to learn the spatio-temporal features of normal patterns at coarse- and fine-grained levels, respectively. The feature representation \\( f \\) generated after self-supervised training can effectively detect anomalies of various scales. Blue arrows and grey arrows represent training and inference processes, respectively. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.463, + 0.223, + 0.478 + ], + "angle": 0, + "content": "binary classifications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.578 + ], + "angle": 0, + "content": "In this work, we design proxy tasks based on video continuity [19] to learn spatio-temporal features of normal events in a multi-grained manner to perceive anomalies of various scales. In addition, we formulate the reconstruction of frames as a contrastive learning task to learning highly discriminative features instead of background noise." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.612, + 0.279, + 0.629 + ], + "angle": 0, + "content": "2.3. Multi-Scale Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Multi-scale learning typically considers information from different scales or granularities. Similar to our work, Xiao et al. [44] hierarchically learns the spatio-temporal features through event mask prediction task. However, event-level features are not sufficient for detecting subtle variances between frames. There is also some multi-scale works [41, 52] in VAD. For instance, Wang et al. [41] obtain image features of different resolutions to perform prediction through different pooling operations, while Zhong et al. [52] obtain features at different scales through feature pyramid. Although pooling and feature pyramids can provide feature representations of different resolutions, they cannot specifically capture the motion and appearance features which is crucial for anomaly detection. In contrast, we design the self-supervised proxy tasks to direct the model's attention towards changes in appearance and motion between frames, enabling it to learn features in a multi-grained manner." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.461, + 0.591, + 0.476 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.491, + 0.61, + 0.505 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Our proposed multi-scale video anomaly detection method attempts to learn spatio-temporal features of normal patterns at both coarse-grained and fine-grained levels by jointly solving multiple proxy tasks of spatio-temporal representation learning. For a normal video clip \\( V \\), our approach is devoted to learning a video encoder \\( F \\) that transforms the video clip \\( V \\) into the scale-aware high-quality spatio-temporal feature \\( f \\). We define \\( f \\) as the high-quality feature in terms of anomalous information of different scales if it is able to easily address the following questions: (1) Whether there are missing frames in video clips \\( V \\), i.e., whether feature representation \\( f \\) remains sensitive to large-scale anomalies. (2) If the video clip \\( V \\) is not continuous, where are the missing frames, i.e., whether feature representation \\( f \\) remains sensitive to small-scale anomalies. (3) If video clip \\( V \\) is discontinuous, what are missing frames, i.e., whether the feature representation \\( f \\) contains high-level motion feature and contextual feature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The overall architecture of our proposed method is illustrated in Fig. 2. The network consists of three branches and a video encoder \\( F \\), all of the branches share the same backbone encoder \\( F \\). Given two nonoverlapping video clips: continuous clip \\( V_{c} \\), discontinuous clip \\( V_{d} \\), and the corresponding missing frame \\( I_{m} \\), each of the three branches solves a proxy task and operates together to optimize the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17387" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "backbone. The first branch is used to discriminate whether a video clip is continuous or not. The second branch is used to locate the position of the missing frame in the discontinuous clip \\( V_{d} \\). The last branch is designed to learn high-level features through contrastive learning mechanism, so that the feature representation of the discontinuous clip \\( V_{d} \\) can effectively approach the feature representation of the missing frame \\( I_{m} \\). After joint optimization of self-supervised tasks, the backbone \\( F \\) can generate scale-aware feature representations \\( f \\) for anomaly inference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.253, + 0.245, + 0.269 + ], + "angle": 0, + "content": "3.2. Data Acquisition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.47, + 0.456 + ], + "angle": 0, + "content": "To perform the proposed scale-aware proxy tasks, we select continuous video clips and discontinuous clips containing only one missing frame \\( I_{m} \\), respectively. The continuous clip contains \\( T \\) continuous frames, while the discontinuous clip has a length of \\( T - 1 \\). Given an initial continuous clip of length \\( T \\), we uniformly select a frame from the interval \\( [1, T - 2] \\) as the missing frame \\( I_{m} \\) to form the breakpoint (counting from 0). And the leading and alternate of the missing frames are concatenated together to form discontinuous video clip \\( V_{d} \\). The restriction of missing frame inside the clip sequence ensures that the construction of missing frames and discontinuous clips remains consistent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.473, + 0.563 + ], + "angle": 0, + "content": "As shown in Fig. 2, we formulate the continuity discrimination task as a binary classification task while the missing frame localization task is formulated as a \\( T - 2 \\) class classification task labeled by \\( I_{m} \\). The estimation task is formulated as a contrastive learning task. Both initial and continuous clips are formed by randomly sampling non-overlapping clips of length \\( T \\) from the same video." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.283, + 0.587 + ], + "angle": 0, + "content": "3.3. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.732 + ], + "angle": 0, + "content": "We employ a video encoder (defaults a 3D-ConvNet) as the backbone network \\( F \\) and use it as a shared base for three proxy tasks. Each task introduces into its top a projection network, denoted \\( P_{1}, P_{2} \\), and \\( P_{3} \\), which are used to process the video clip embeddings \\( f \\) extracted from the backbone network. All projection heads contain spatio-temporal averaging pooling layers to ensure that the dimensionality of the deep feature embeddings remains consistent, even though the temporal lengths may be different." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.839 + ], + "angle": 0, + "content": "In continuity judgment and missing frame localization, we use fully connected layers as classifiers at the end of \\( P_{1} \\) and \\( P_{2} \\) for performing classification operations. As for the missing frame estimation task, we use \\( P_{3} \\) to embed the video features \\( f \\) extracted by the backbone network into the low-dimensional features to get the representation, i.e., the feature representation of the missing frame." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.36, + 0.865 + ], + "angle": 0, + "content": "3.4. Proxy Tasks and Joint Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Task 1: Continuity Judgment. The continuity judgment is intended to help the model learn spatio-temporal features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "and motion patterns in the video in a coarse-grained manner to perceive large-scale anomalies. In this task, in order for the model to learn by video features rather than other cues such as context, the positive and negative samples in each batch in our training come from the same video scene. With N samples in a training batch, cross-entropy loss \\(\\mathcal{L}_{CE}\\) is used to optimize the model, and the loss for continuity judgment is \\(\\mathcal{L}_1\\):" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.224, + 0.892, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {1} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\mathcal {L} _ {C E} \\left(P _ {1} \\left(F \\left(v _ {d} ^ {i}\\right)\\right) _ {1} + P _ {1} \\left(F \\left(v _ {c} ^ {i}\\right)\\right) _ {0}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.354 + ], + "angle": 0, + "content": "where \\((\\cdot)\\) represents the process of feature processing by the network, \\(F\\) and \\(P_{1}\\) represent the shared backbone network and the continuity judgment projection head, respectively, and \\(v_{c}^{i}\\) and \\(v_{d}^{i}\\) represent the \\(i_{th}\\) continuous and discontinuous video clip, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.505 + ], + "angle": 0, + "content": "Task 2: Discontinuity Localization. While the continuity judgment task only learns spatio-temporal features in a coarse-grained manner, the discontinuous localization task expects to learn in a fine-grained manner in order to perceive small-scale anomalies. Compared to binary labels, labeling with missing frames \\( I_{m} \\) in this task can drive the model to learn changes in motion between adjacent frames in a fine-grained manner. Similarly, the discontinuous localization loss \\( \\mathcal{L}_2 \\) can be expressed as an average of the cross-entropy loss of \\( N \\) samples in each batch:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.516, + 0.892, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {C E} \\left(P _ {2} \\left(F \\left(v _ {d} ^ {i}\\right)\\right) _ {I _ {i}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.891, + 0.614 + ], + "angle": 0, + "content": "where \\(P_{2}\\) represents the discontinuity localization projection head and \\(I_{i}\\) represents the missing frame position corresponding to discontinuous clip \\(v_{d}^{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Task 3: Missing Frame Estimation. In order to further learn the highly discriminative spatio-temporal features of the video, we estimate missing frames in terms of both appearance and motion. Meanwhile, we choose to estimate frames in feature space rather than RGB space to avoid background noise interference. We design a contrastive learning mechanism to perform this estimation task based on the following facts. First, since adjacent frames and missing frames are temporally connected, adjacent frames and missing frames contain more similar motions compared to non-adjacent frames from the same scene but farther away. Second, video frames from the same scene share more similar appearance features with missing frames than video frames from other scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Accordingly, we learn the motion information in the video by triplet loss [34] taking the discontinuous video clips \\( v_{d}^{i} \\) as the anchor point, the missing frames \\( I_{m} \\) as the positive samples, and the continuous video clips \\( v_{c}^{i} \\) as the negative samples. In addition, we take discontinuous clips" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17388" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.214 + ], + "angle": 0, + "content": "\\(v_{d}^{i}\\) as anchor points, continuous clips \\(v_{c}^{i}\\) from the same scene as positive samples, and video clips \\(\\{v^{j}\\}_{i\\neq j}\\) from different scenes as negative samples, to learn the appearance information in the video by contrastive loss [19]. Defining the cosine similarity computation operation as \\(sim(\\cdot ,\\cdot)\\), and the hyperparameter that balances the relative contributions of the triplet loss and the contrastive loss as \\(\\omega \\in [0,1]\\), the missing frame estimation loss is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.226, + 0.47, + 0.313 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {3} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} (\\omega \\times \\max (0, \\gamma - \\left(p _ {i} ^ {+} - p _ {i} ^ {-}\\right)) \\tag {3} \\\\ - (1 - \\omega) \\log \\left(\\frac {q _ {i} ^ {+}}{q _ {i} ^ {+} + \\sum_ {j = 1 , j \\neq i} ^ {N} q _ {i , j ^ {-}}}\\right)), \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.329, + 0.47, + 0.35 + ], + "angle": 0, + "content": "\\[\np _ {i} ^ {+} = \\mathrm {s i m} \\left(P _ {3} (V _ {d} ^ {i}, I _ {i})\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.371, + 0.47, + 0.39 + ], + "angle": 0, + "content": "\\[\np _ {i} ^ {-} = \\operatorname {s i m} \\left(P _ {3} \\left(V _ {d} ^ {i}, V _ {c} ^ {i}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.412, + 0.47, + 0.432 + ], + "angle": 0, + "content": "\\[\nq _ {i} ^ {+} = \\exp \\left(\\operatorname {s i m} \\left(P _ {3} \\left(V _ {d} ^ {i}, V _ {c} ^ {i}\\right)\\right) / \\tau\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.45, + 0.469, + 0.498 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} q _ {i, j} ^ {-} = \\exp \\left(\\sin \\left(P _ {3} \\left(v _ {d} ^ {i}, v _ {d} ^ {j}\\right)\\right) / \\tau\\right) \\tag {7} \\\\ + \\exp \\left(\\mathrm {s i m} \\left(P _ {3} (v _ {d} ^ {i}, v _ {c} ^ {j})\\right) / \\tau\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.672 + ], + "angle": 0, + "content": "where \\( p_i^+ \\) and \\( p_i^- \\) denote the similarity between positive sample pairs and negative sample pairs in the triplet loss, respectively, \\( q_i^+ \\) denotes a single positive sample pair in the contrastive loss, \\( q_{i,j}^{-} \\) denotes the similarity of negative sample pairs in the contrastive loss, and \\( \\tau \\) is the temperature factor in the contrastive loss. Increasing \\( \\omega \\) allows for more focus on motion information, while decreasing \\( \\omega \\) allows for more focus on appearance information. Overall, the model is incentivized to learn fine-grained motion changes and contextual features, including background information and object appearance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Joint Optimization. The above three proxy tasks learn spatio-temporal features in videos from different levels, and we jointly optimize our shared video encoder \\( F \\) with the multi-loss function." + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.746, + 0.47, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {F} = w _ {1} \\mathcal {L} _ {1} + w _ {2} \\mathcal {L} _ {2} + w _ {3} \\mathcal {L} _ {3}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.775, + 0.469, + 0.837 + ], + "angle": 0, + "content": "where \\(w_{1}, w_{2}, w_{3} \\in [0,1]\\) are the individual weights on the \\(\\mathcal{L}_1, \\mathcal{L}_2\\) and \\(\\mathcal{L}_3\\). Joint optimization helps the model to learn video spatio-temporal features from both coarse- and fine-grained scales to perceive anomalies of various scales." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.228, + 0.862 + ], + "angle": 0, + "content": "3.5. VAD Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In the inference, we use \\(T\\) frames as the processing unit. Following the setup of the SOTA method [20, 41, 46], we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "choose not to compute anomaly scores using the error between each estimated frame \\(\\hat{I}_t\\) (\\(1 \\leq t \\leq T\\)) and the corresponding ground truth \\(I_t\\). Instead, we opt to employ the Peak Signal-to-Noise Ratio (PSNR) corresponding to the frame with the largest Mean-Square Error (MSE) among the \\(T\\) frames and the ground truth. This approach serves as the detection metric for the video sequence, as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.205, + 0.892, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M S E} \\left(\\boldsymbol {I} _ {t}, \\hat {\\boldsymbol {I}} _ {t}\\right) = \\frac {1}{H \\cdot W} \\sum_ {i, j} \\left\\| \\boldsymbol {I} _ {t} (i, j) - \\hat {\\boldsymbol {I}} _ {t} (i, j) \\right\\| _ {2} ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.251, + 0.892, + 0.297 + ], + "angle": 0, + "content": "\\[\nP S N R _ {t} = 1 0 \\log_ {1 0} \\frac {\\operatorname {M A X} _ {\\boldsymbol {I} _ {t}} ^ {2}}{\\operatorname {M S E} \\left(\\boldsymbol {I} _ {t} , \\hat {\\boldsymbol {I}} _ {t}\\right)}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.299, + 0.892, + 0.389 + ], + "angle": 0, + "content": "where \\(H\\) and \\(W\\) are the height and width of the frames, respectively, \\(I_{t}(i,j) / \\hat{I}_{t}(i,j)\\) represents the RGB value of the \\((i,j)\\)-th pixel in \\(I_{t} / \\hat{I}_{t}\\), and \\(\\mathrm{MAX}_{I_t}\\) represents the maximum value of the image pixel in \\(I_{t}\\). Finally, we calculate the anomaly score \\(S_{t}\\) for the \\(t\\)-th frame \\(I_{t}\\) by normalizing the PSNR values over all \\(T\\) frames:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.394, + 0.892, + 0.427 + ], + "angle": 0, + "content": "\\[\nS _ {t} = 1 - \\frac {P S N R _ {t} - \\operatorname* {m i n} _ {t ^ {\\prime}} P S N R _ {t ^ {\\prime}}}{\\operatorname* {m a x} _ {t ^ {\\prime}} P S N R _ {t ^ {\\prime}} - \\operatorname* {m i n} _ {t ^ {\\prime}} P S N R _ {t ^ {\\prime}}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.433, + 0.892, + 0.463 + ], + "angle": 0, + "content": "where \\(1 \\leq t' \\leq T\\). A higher score \\(S_{t}\\) indicates that the frame \\(I_{t}\\) is more likely to be anomalous." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.476, + 0.625, + 0.493 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.5, + 0.779, + 0.515 + ], + "angle": 0, + "content": "4.1. Dataset and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Cross-Scene Anomaly Dataset Avenue [25] contains 16 training videos and 21 test videos with 47 abnormal events, including running and throwing. The scenes in this dataset are single and the anomalies are mainly large-scale. ShanghaiTech [27] has 13 scenes with complex lighting conditions and different perspectives. In addition, the dataset includes anomalies caused by sudden movements, such as chasing and arguing. The different perspectives and the unfixed position of the camera lead to a large variation of both the object scale and the anomaly scale in the scene. This is a significant challenge to the robustness of anomaly detection methods to scale variations. UCF-Crime [36] comprises 13 anomaly types, spanning a total of 128 hours of video footage. This dataset is complex because of the unconstrained backgrounds and anomalies of various scales. The training set contains video-level annotations, whereas the testing set includes frame-level annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Scene-Dependent Anomaly Dataset. Campus [2] is currently the most challenging dataset in its field with 43 scenes, 28 classes of anomalous events and 16 hours of videos. Especially, it contains scene-dependent anomalies, which means an normal event may be abnormal in another scene. Detecting scene-dependent anomalies requires the model to understand the scene and learn highly discriminative features rather than overfit." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17389" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.083, + 0.089, + 0.465, + 0.285 + ], + "angle": 0, + "content": "
MethodLength TShanghaiTechCampus
AUCAUC
MicroMacroMicroMacro
Continuity Judgment878.182.360.163.2
1277.281.160.363.9
1675.179.258.260.9
Discontinuity Localization882.685.562.165.4
1283.487.262.665.1
1682.185.561.965.0
Missing Frame Estimation879.883.263.566.8
1281.284.965.268.9
1680.584.064.968.0
Overall Framework884.288.866.369.2
1285.189.867.370.9
1683.988.366.069.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.297, + 0.47, + 0.352 + ], + "angle": 0, + "content": "Table 1. Comparison of proposed framework and individual proxy tasks between different lengths \\( T \\) of input clips. We report Micro and Macro AUC (%) on ShanghaiTech and Campus datasets. The best performing results are marked in bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.469, + 0.443 + ], + "angle": 0, + "content": "Evaluation Metrics The area under the ROC curve (AUC) serves as a commonly used metric for evaluation and comparison. A higher AUC score indicates a better anomaly detection capability. Following previous research [12], we evaluate both the Micro and Macro versions of AUC." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.453, + 0.295, + 0.469 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.763 + ], + "angle": 0, + "content": "We train and evaluate our method with an NVIDIA RTX 3090 GPU. In the training phase, we resize the resolution of all input video clips to \\(256 \\times 256\\) pixels, while the values of the pixels in all frames are normalized to the range [0, 1]. For the pre-training of the three proxy tasks, we utilize AdamW as the optimizer while the length of the continuous video clips is set to \\(T = 12\\) frames. The initial learning rate is set to 0.0003 and is gradually decayed following the scheme of cosine annealing. For the hyperparameters, we set \\(\\omega = 0.5\\) in Equation 3, \\(w_{1} = w_{2} = w_{3} = 1\\) in Equation 8, and the temperature factor \\(\\tau = 0.1\\) in the contrastive loss. In our reported experimental performance, the shared backbone network \\(F\\) is implemented as I3D-RGB [3]. In addition, we incorporate the designed proxy tasks into the backbone of the SOTA methods [2, 11], the experimental results are also reported. Following the existing method setup [11, 24, 33], we perform the proxy task at the object-level with the multi-task backbone [11] and at the frame-level with the other backbones." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.231, + 0.789 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.856 + ], + "angle": 0, + "content": "Considering the ShanghaiTech and Campus datasets contain diverse perspectives and scenes, and large variations in scales of anomaly, we conduct exhaustive ablation experiments on the ShanghaiTech and Campus datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Sensitivity to the Length of Video Clips. The length \\( T \\) of the input clips is an important setting in our proposed framework. With a small \\( T \\), both continuity judgments and dis" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.888, + 0.245 + ], + "angle": 0, + "content": "
IDCJDLMFEFPShanghaiTechCampus
AUCAUC
MicroMacroMicroMacro
1---✓70.271.258.158.6
2✓---77.281.160.363.9
3-✓--83.487.262.665.1
4--✓-81.284.965.268.9
5✓✓--84.187.862.865.4
6✓-✓-82.385.665.369.2
7-✓✓-84.589.566.270.1
8✓✓✓-85.189.867.370.9
9✓✓-✓84.287.662.865.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Table 2. Ablation experiments on the contributions of each proxy task. We report the AUC (\\%) scores on ShanghaiTech and Campus datasets. 'CJ', 'DL' and 'MFE' stand for the three proxy tasks of continuity judgment, discontinuity localization and missing frame estimation, respectively. In addition we add experiments on frame prediction for comparison, 'FP' represents frame prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.892, + 0.507 + ], + "angle": 0, + "content": "continuity localization will be so ambiguous as to be difficult. While with a larger \\( T \\), the model may not need to learn high-quality spatio-temporal features to solve the task due to enough spatio-temporal information in the feed. Table 1 shows the sensitivity of each proxy task to \\( T \\). For the discontinuous localization and missing frame estimation tasks, \\( T = 12 \\) presents better performance. Although at \\( T = 8 \\), the continuity judgment yields competitive results on the ShanghaiTech dataset, the overall framework achieves the best outcomes with jointly optimized at \\( T = 12 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Effect of Continuity Judgment. The results of the ablation experiments for each proxy task are shown in Table 2, where we take the frame prediction [20] performed on the same backbone \\( F \\) as the baseline and train it with gradient loss and intensity loss (ID 1). Performing the continuity judgment task alone improves the performance of the model compared to frame prediction (ID 2). This suggests that continuity judgment task motivates the model to generate higher quality representations than frame prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Effect of Discontinuity Localization. Compared to other proxy tasks and baseline, the performance of the discontinuous localization task alone can be significantly improved, especially on ShanghaiTech which contains anomalies of various scales (ID 3, ID 5 and ID 7). As this task samples discontinuous positions uniformly along the time axis, it motivates the model to capture subtle variations between frames, obtaining fine-grained representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect of Missing Frame Estimation. Performing missing frame estimation alone is effective in improving performance on Campus compared to baseline and other proxy tasks, which contains scene-dependent anomalies (ID 4). This improvement illustrates that our proposed contrastive learning scheme can help the model understand the scene. Compared to performing the frame prediction task to learn RGB features (ID 1 and ID 9), our proposed scheme learns highly discriminative motion features and semantically ef" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17390" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.088, + 0.465, + 0.589 + ], + "angle": 0, + "content": "
yearMethodAvenueShanghaiTech
AUCAUC
MicroMacroMicroMacro
2018Liu et al. [20]85.1-72.8-
Liu et al. [22]84.4---
Sultani et al. [36]--76.5-
2019Gong et al. [14]83.8-71.2-
Lee et al. [18]90.0-76.2-
Ionescu et al. [17]87.490.478.784.9
2020Park et al. [31]88.5-70.5-
Sun et al. [37]89.6-74.7-
Lu et al. [26]85.8-75.9-
Wang et al. [42]87.0-79.3-
Yu et al. [47]89.6-74.8-
2021Liu et al. [23]91.1-76.2-
Lv et al. [28]89.5-73.8-
Georgescu et al. [11]*91.591.982.489.3
Georgescu et al. [12]*92.390.482.789.3
2022Wang et al. [41]88.3-76.6-
Zaheer et al. [49]74.2-79.6-
Chen et al. [5]90.3-78.1-
Zhong et al. [52]89.0-74.5-
Cho et al. [6]88.0-76.3-
Yang et al. [45]89.9-74.7-
Ristea et al. [33]*92.991.983.689.5
Acsintoae et al. [1]▼*93.093.283.790.5
2023Yang et al. [46]89.9-73.8-
Cao et al. [2]86.8-79.2-
Liu et al. [21]92.8-78.8-
Singh et al. [35]86.0-76.6-
Sun et al. [39]92.4-83.0-
Sun et al. [38]91.5-78.6-
Liu et al. [24]▼91.892.383.887.8
Liu et al. [24] ▼*93.693.985.091.4
Ours92.492.985.189.8
Ours▼93.292.586.291.0
Ours*(Estimation Task)92.693.085.591.8
Ours*(Three Tasks)93.693.886.892.4
Ours▼*(Three Tasks)94.394.587.593.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.598, + 0.47, + 0.64 + ], + "angle": 0, + "content": "Table 3. Comparison with SOTA methods of the Micro and Macro AUC (%) on Avenue and ShanghaiTech datasets. The best performing results are marked in bold." + }, + { + "type": "table_footnote", + "bbox": [ + 0.081, + 0.642, + 0.348, + 0.654 + ], + "angle": 0, + "content": "\\(\\nabla\\) : Methods apply virtual dataset for training." + }, + { + "type": "table_footnote", + "bbox": [ + 0.081, + 0.655, + 0.457, + 0.668 + ], + "angle": 0, + "content": "*: Methods utilize multi-task model ([11] or [12]) as backbone." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.686, + 0.466, + 0.768 + ], + "angle": 0, + "content": "
MethodReferenceMicro AUCMacro AUC
Park et al. [31]CVPR2068.972.4
Georgescu et al. [11]CVPR2174.678.2
Wang et al. [41]TNNLS2272.976.8
Sun et al. [39]CVPR2375.578.3
Ours-80.683.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.779, + 0.469, + 0.809 + ], + "angle": 0, + "content": "Table 4. Results of Micro and Macro AUC(%) on UCF-Crime dataset. The best performing results are marked in bold." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.855 + ], + "angle": 0, + "content": "fective contexts, including the appearance of objects and background (ID 4 and ID 8)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Effect of Joint Optimization. Although a single task may be able to achieve competitive performance, Experiments have shown that when proxy tasks are executed in pairs (ID" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.088, + 0.891, + 0.242 + ], + "angle": 0, + "content": "
MethodReferenceMiceo AUCMacro AUC
Liu et al. [20]CVPR1857.960.2
Gong et al. [14]CVPR1961.962.5
Ionescu et al. [17]CVPR1959.363.4
Park et al. [31]CVPR2062.563.6
Liu et al. [23]ICCV2163.7-
Lv et al. [28]CVPR2164.4-
Wang et al. [41]TNNLS2261.964.2
Cao et al. [2]△CVPR2368.2-
Ours-67.370.9
Ours△-70.172.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.251, + 0.892, + 0.292 + ], + "angle": 0, + "content": "Table 5. Comparison with SOTA methods of the Micro and Macro AUC(%) on Campus dataset. The best performing results are marked in bold." + }, + { + "type": "table_footnote", + "bbox": [ + 0.503, + 0.293, + 0.882, + 0.307 + ], + "angle": 0, + "content": "\\(\\triangle\\) :Version that utilize scene-conditioned model [2] as backbone." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.335, + 0.892, + 0.426 + ], + "angle": 0, + "content": "5, ID 6 and ID 7) and all tasks are executed together (ID 8), the performance of the model continues to improve to an optimum. This joint improvement is attributed to the fact that different proxy tasks learn spatio-temporal features at different levels, and they complement each other in motion and contextual feature learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.438, + 0.822, + 0.453 + ], + "angle": 0, + "content": "4.4. Comparisons with State-Of-The-Arts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.718 + ], + "angle": 0, + "content": "We compare the proposed framework with SOTA methods in terms of Micro and Macro AUC(\\%). It is noteworthy that the current state-of-the-art methods [1, 24, 33] utilize the multi-task framework proposed by Georgescu et al. ([11] and [12], respectively) as the backbone, where [1] and [24] employ virtual data for training. Therefore, we evaluate different configurations of the proposed method on the Avenue and ShanghaiTech datasets, as shown in Table 3. With the multi-task backbone [11] (marked with * in Table 3), we assess two versions: one involves replacing only the prediction component with our designed missing estimation task, while keeping the other proxy tasks unchanged. The other version replaces both proxy tasks in the motion branch with continuity judgment and discontinuity localization tasks in addition to incorporating the missing estimation task. In addition, we report the performance of applying virtual data for training (marked with \\(\\nabla\\) in Table 3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Results on Avenue. As shown in Table 3, our method achieved the highest AUC scores, obtaining a Micro AUC of \\(94.3\\%\\) and a Macro AUC of \\(94.5\\%\\). Without the backbone of multi-task [11, 12] and virtual data [24], our method still scored the best performance with \\(92.9\\%\\) Macro AUC." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results on ShanghaiTech. Our proposed method yields a Micro AUC of \\(87.5\\%\\) and a Macro AUC of \\(93.0\\%\\) on the ShanghaiTech dataset. With the backbone of multi-task [11, 12], we improve the Micro AUC by \\(3.1\\%\\) compared to Georgescu et al. [11] in the setup with only the missing estimation task, and by \\(4.4\\%\\) when incorporating three designed proxy tasks. Without the backbone of multi-task" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17391" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.089, + 0.462, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.388, + 0.47, + 0.457 + ], + "angle": 0, + "content": "Figure 3. Examples of detection for anomalies of different scales from the ShanghaiTech dataset, including ground truth and prediction or reconstruction error maps from different methods (from top to bottom: ROADMAP[41], DEDDnet[52], USTN-DSC[46], and our method). Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.468, + 0.576 + ], + "angle": 0, + "content": "[11, 12] and virtual data [24], our method still achieves a Micro AUC score of \\(85.1\\%\\), which is comparable to the method [24] that employed the multi-task backbone and virtual training data. In comparison with the multi-scale VAD methods ([41] and [52]), our method achieves a significant improvement of \\(8.5\\%\\) and \\(10.6\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.468, + 0.682 + ], + "angle": 0, + "content": "Results on UCF-Crime. Due to the absence of published results (methods only learning from normal data) on the UCF-Crime dataset, we implement the code from the existing literature [11, 31, 39, 41]. As shown in Table 4, our proposed method achieves a significant improvement compared to the second best method by \\(5.1\\%\\) in terms of Micro AUC and \\(5.6\\%\\) in terms of Macro AUC." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.468, + 0.835 + ], + "angle": 0, + "content": "Results on Campus. We evaluate an additional version on the Campus dataset with the scene-conditioned model (provided by Cao et al. [2]) as the backbone, as it is necessary to incorporate scene features for detecting scene-dependent anomalies. As shown in Table 5, our method achieves the highest Micro AUC score of \\(70.1\\%\\) and Macro AUC score of \\(72.2\\%\\). Without the scene-conditioned model, our method still obtain a Micro AUC of \\(67.3\\%\\), surpassing other RGB reconstruction-based [14, 28, 31] or RGB prediction-based methods [20, 23, 41]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.26, + 0.862 + ], + "angle": 0, + "content": "4.5. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Fig. 1 shows the anomaly curves of the test video and the comparison of our method with PMem [31] (reconstruction-" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.09, + 0.627, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.09, + 0.755, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.09, + 0.882, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.214, + 0.892, + 0.242 + ], + "angle": 0, + "content": "Figure 4. Visualization of salient regions of test frames by the proposed method. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.452 + ], + "angle": 0, + "content": "based method) and RoadMap [41] (prediction-based method). When there are small-scale anomalies in the sampled videos, the anomaly score curves of the other two methods have obvious missed detections, while our method can accurately perceive the small-scale anomalies. In addition, Fig. 3 illustrates more examples, including the comparison with the multi-scale VAD methods [41, 52] and the comparison with self-supervised VAD method [46]. It can be observed that for normal regions in video frames, our method is able to estimate them well, while for abnormal event regions large errors occur. While other methods generate large errors for some normal regions, especially for frames with small-scale anomalies." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.455, + 0.892, + 0.515 + ], + "angle": 0, + "content": "Visualizations of the salient regions of the frames by our method are illustrated in Fig. 4. For both large-scale anomalies (1-th row) and small-scale anomalies (2-th row), our model can accurately focus on the anomalous regions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.529, + 0.619, + 0.544 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.749 + ], + "angle": 0, + "content": "In this paper, we observe that anomalous spatial scales affect the feature learning of the model. Small-scale anomalies require the model to learn highly discriminative features in a fine-grained manner rather than background noise. To detect multi-scale anomalies, we design proxy tasks supervised by video continuity that motivate the model to learn spatio-temporal features in both coarse-grained and fine-grained manners. Furthermore, instead of performing RGB reconstruction and prediction, we estimation frames in feature space through contrastive learning to learn highly discriminative features of appearance and motion. Experiments conducted on four challenging benchmark datasets validate the effectiveness of our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grants (62101064, 62171057, 62201072, U23B2001, 62001054, 62071067), the National Postdoctoral Program for Innovative Talents under Grant BX20230052, China Postdoctoral Science Foundation (2023TQ0039), the Ministry of Education and China Mobile Joint Fund (MCM20200202, MCM20180101), Beijing University of Posts and Telecommunications-China Mobile Research Institute Joint Innovation Center." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17392" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. *Ubnormal: New benchmark for supervised open-set video anomaly detection.* In *CVPR*, pages 20111-20121, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.254 + ], + "angle": 0, + "content": "[2] Congqi Cao, Yue Lu, Peng Wang, and Yanning Zhang. A new comprehensive benchmark for semi-supervised video anomaly detection and anticipation. In CVPR, pages 20392-20401, 2023. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.47, + 0.297 + ], + "angle": 0, + "content": "[3] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, pages 4724-4733, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.34 + ], + "angle": 0, + "content": "[4] Yunpeng Chang, Zhigang Tu, Wei Xie, and Junsong Yuan. Clustering driven deep autoencoder for video anomaly detection. In ECCV, pages 329-345, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.47, + 0.409 + ], + "angle": 0, + "content": "[5] Chengwei Chen, Yuan Xie, Shaohui Lin, Angela Yao, Guannan Jiang, Wei Zhang, Yanyun Qu, Ruizhi Qiao, Bo Ren, and Lizhuang Ma. Comprehensive regularization in a bidirectional predictive network for video anomaly detection. In AAAI, pages 230-238, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.466 + ], + "angle": 0, + "content": "[6] MyeongAh Cho, Taeoh Kim, Woo Jin Kim, Suhwan Cho, and Sangyoun Lee. Unsupervised video anomaly detection via normalizing flows with implicit latent features. Pattern Recognit., 129:108703, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[7] Romero F. A. B. de Morais, Vuong Le, Truyen Tran, Budhaditya Saha, Moussa Reda Mansour, and Svetha Venkatesh. Learning regularity in skeleton trajectories for anomaly detection in videos. In CVPR, pages 11996-12004, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.47, + 0.578 + ], + "angle": 0, + "content": "[8] Zhiwen Fang, Jiafei Liang, Joey Tianyi Zhou, Yang Xiao, and Feng Yang. Anomaly detection with bidirectional consistency in videos. IEEE Trans. Neural Networks Learn. Syst., 33(3):1079-1092, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[9] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. MIST: multiple instance self-training framework for video anomaly detection. In CVPR, pages 14009-14018, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.47, + 0.691 + ], + "angle": 0, + "content": "[10] Alessandro Flaborea, Luca Collorone, Guido Maria D'Amely di Melendugno, Stefano D'Arrigo, Bardh Prenkaj, and Fabio Galasso. Multimodal motion conditioned diffusion model for skeleton-based video anomaly detection. In (ICCV), pages 10318-10329, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.76 + ], + "angle": 0, + "content": "[11] Mariana-Iuliana Georgescu, Antonio Barbalau, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. Anomaly detection in video via self-supervised and multi-task learning. In CVPR, pages 12742-12752, 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[12] Mariana-Iuliana Georgescu, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4505-4523, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In ICCV, pages 1705-1714, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[14] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In ICCV, pages 1705-1714, 2019. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[15] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal regularity in video sequences. In CVPR, pages 733-742, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[16] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In ICCV, pages 8771-8780, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[17] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In CVPR, pages 7842–7851, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[18] Sangmin Lee, Hak Gu Kim, and Yong Man Ro. BMAN: bidirectional multi-scale aggregation networks for abnormal event detection. IEEE Trans. Image Process., 29:2395-2408, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[19] Hanwen Liang, Niamul Quader, Zhixiang Chi, Lizhe Chen, Peng Dai, Juwei Lu, and Yang Wang. Self-supervised spatiotemporal representation learning by exploiting video continuity. In AAAI, pages 1564-1573, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[20] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In CVPR, pages 6536-6545, 2018. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[21] Wenrui Liu, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Diversity-measurable anomaly detection. In CVPR, pages 12147-12156, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[22] Yusha Liu, Chun-Liang Li, and Barnabás Póczos. Classifier two sample test for video anomaly detections. In BMVC, page 71. BMVA Press, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[23] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In ICCV, pages 13568-13577, 2021. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[24] Zuhao Liu, Xiao-Ming Wu, Dian Zheng, Kun-Yu Lin, and Wei-Shi Zheng. Generating anomalies for video anomaly detection with prompt-based feature mapping. In CVPR, pages 24500-24510. IEEE, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[25] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 FPS in MATLAB. In ICCV, pages 2720-2727, 2013. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[26] Yiwei Lu, Frank Yu, Mahesh Kumar Krishna Reddy, and Yang Wang. Few-shot scene-adaptive anomaly detection. In ECCV, pages 125-141, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[27] Weixin Luo, Wen Liu, and Shenghua Gao. A revisit of sparse coding based anomaly detection in stacked RNN framework. In ICCV, pages 341-349, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Hui Lv, Chen Chen, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Learning normal dynamics in videos with meta prototype network. In CVPR, pages 15425-15434, 2021. 2, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17393" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In CVPR, pages 8022-8031, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.151, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[30] Trong-Nguyen Nguyen and Jean Meunier. Anomaly detection in video sequence with appearance-motion correspondence. In ICCV, pages 1273-1283, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.194, + 0.469, + 0.235 + ], + "angle": 0, + "content": "[31] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In CVPR, pages 14360-14369, 2020. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.238, + 0.469, + 0.292 + ], + "angle": 0, + "content": "[32] Mahdyar Ravanbakhsh, Enver Sangineto, Moin Nabi, and Nicu Sebe. Training adversarial discriminators for cross-channel abnormal event detection in crowds. In WACV, pages 1896-1904, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.295, + 0.469, + 0.364 + ], + "angle": 0, + "content": "[33] Nicolae-Catalin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In CVPR, pages 13566-13576, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.367, + 0.469, + 0.408 + ], + "angle": 0, + "content": "[34] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.411, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[35] Ashish Singh, Michael J Jones, and Erik G Learned-Miller. Eval: Explainable video anomaly localization. In CVPR, pages 18717-18726, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.454, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[36] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, pages 6479-6488, 2018. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.498, + 0.469, + 0.539 + ], + "angle": 0, + "content": "[37] Che Sun, Yunde Jia, Yao Hu, and Yuwei Wu. Scene-aware context reasoning for unsupervised abnormal event detection in videos. In ACM Multimedia, pages 184-192, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.542, + 0.469, + 0.583 + ], + "angle": 0, + "content": "[38] Che Sun, Chenrui Shi, Yunde Jia, and Yuwei Wu. Learning event-relevant factors for video anomaly detection. In AAAI, pages 2384-2392, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.586, + 0.469, + 0.627 + ], + "angle": 0, + "content": "[39] Shengyang Sun and Xiaojin Gong. Hierarchical semantic contrast for scene-aware video anomaly detection. In CVPR, pages 22846-22856, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.63, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[40] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W. Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In ICCV, pages 4955-4966, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.687, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[41] Xuanzhao Wang, Zhengping Che, Bo Jiang, Ning Xiao, Ke Yang, Jian Tang, Jieping Ye, Jingyu Wang, and Qi Qi. Robust unsupervised video anomaly detection by multipath frame prediction. IEEE Trans. Neural Networks Learn. Syst., 33 (6):2301-2312, 2022. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.758, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[42] Ziming Wang, Yuexian Zou, and Zeming Zhang. Cluster attention contrast for video anomaly detection. In ACM Multimedia, pages 2463-2471, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[43] Jie Wu, Wei Zhang, Guanbin Li, Wenhao Wu, Xiao Tan, Yingying Li, Errui Ding, and Liang Lin. Weakly-supervised spatio-temporal anomaly detection in surveillance video. In IJCAI, pages 1172-1178, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[44] Fanyi Xiao, Kaustav Kundu, Joseph Tighe, and Davide Modolo. Hierarchical self-supervised representation learning for movie understanding. In CVPR, pages 9717-9726, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[45] Zhiwei Yang, Peng Wu, Jing Liu, and Xiaotao Liu. Dynamic local aggregation network with adaptive clusterer for anomaly detection. In ECCV, pages 404-421, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[46] Zhiwei Yang, Jing Liu, Zhaoyang Wu, Peng Wu, and Xiaotao Liu. Video event restoration based on keyframes for video anomaly detection. In CVPR, pages 14592-14601, 2023. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[47] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze test helps: Effective video anomaly detection via learning to complete video events. In ACM Multimedia, pages 583-591, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[48] Jongmin Yu, Younkwan Lee, Kin Choong Yow, Moongu Jeon, and Witold Pedrycz. Abnormal event detection and localization via adversarial event prediction. IEEE Trans. Neural Networks Learn. Syst., 33(8):3572-3586, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.306, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[49] Muhammad Zaigham Zaheer, Arif Mahmood, Muhammad Haris Khan, Mattia Segù, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In CVPR, pages 14724–14734, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[50] Chen Zhang, Guorong Li, Yuankai Qi, Shuhui Wang, Laiyun Qing, Qingming Huang, and Ming-Hsuan Yang. Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection. In CVPR, pages 16271-16280, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[51] Yiru Zhao, Bing Deng, Chen Shen, Yao Liu, Hongtao Lu, and Xian-Sheng Hua. Spatio-temporal autoencoder for video anomaly detection. In ACM Multimedia, pages 1933-1941, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[52] Yuanhong Zhong, Xia Chen, Yongting Hu, Panliang Tang, and Fan Ren. Bidirectional spatio-temporal feature learning with multiscale evaluation for video anomaly detection. IEEE Trans. Circuits Syst. Video Technol., 32(12):8285-8296, 2022. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.56, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[53] Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection. In AAAI, pages 3769-3777. AAAI, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17394" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/82eacc56-0492-4ac2-b935-f0e92d5461ce_origin.pdf b/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/82eacc56-0492-4ac2-b935-f0e92d5461ce_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7fb42230461d6cbb53977afa31b1cde395ae0e9c --- /dev/null +++ b/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/82eacc56-0492-4ac2-b935-f0e92d5461ce_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0aa615986bc15f6480a0697fa45b0f39cb2f5d0663a0f5a734ece466fa40bd0 +size 976271 diff --git a/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/full.md b/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a17ceb36d1dd1ae6071ea05b4129a3320b79d431 --- /dev/null +++ b/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/full.md @@ -0,0 +1,319 @@ +# Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning + +Menghao Zhang, Jingyu Wang\*, Qi Qi, Haifeng Sun, Zirui Zhuang, Pengfei Ren, Ruilong Ma, Jianxin Liao +State Key Laboratory of Networking and Switching Technology, +Beijing University of Posts and Telecommunications + +zhangmenghao, wangjingyu, qiqi8266, hfsun, zhuangzirui, rpf, maruilong@bupt.edu.cn; jxlbupt@gmail.com + +# Abstract + +Recent progress in video anomaly detection suggests that the features of appearance and motion play crucial roles in distinguishing abnormal patterns from normal ones. However, we note that the effect of spatial scales of anomalies is ignored. The fact that many abnormal events occur in limited localized regions and severe background noise interferes with the learning of anomalous changes. Meanwhile, most existing methods are limited by coarse-grained modeling approaches, which are inadequate for learning highly discriminative features to discriminate subtle differences between small-scale anomalies and normal patterns. To this end, this paper addresses multi-scale video anomaly detection by multi-grained spatio-temporal representation learning. We utilize video continuity to design three proxy tasks to perform feature learning at both coarse-grained and fine-grained levels, i.e., continuity judgment, discontinuity localization, and missing frame estimation. In particular, we formulate missing frame estimation as a contrastive learning task in feature space instead of a reconstruction task in RGB space to learn highly discriminative features. Experiments show that our proposed method outperforms state-of-the-art methods on four datasets, especially in scenes with small-scale anomalies. + +# 1. Introduction + +Video Anomaly Detection (VAD) is dedicated to detecting anomalous events in videos with wide applications in public safety and intelligent surveillance. The primary challenge of VAD is the sparsity of abnormal samples, limiting the direct learning of abnormal patterns from the available data. Therefore, most previous VAD studies can be divided into the weakly supervised category [9, 29, 36, 40, 43, 50, 53] that learns with video-level annotations, or the category only learning from normal data [5, 23, 24, 39, 46]. Our work + +![](images/bdda958cde7f6fc7eaf99424daddf605a6341516f9f8ef3be8c651bad8370536.jpg) +Figure 1. Examples of small-scale anomalies and comparisons of results from different methods. From top to bottom: video sequence, ground truth (green regions are abnormal), result of PMem [31] (reconstruction-based method), result of RoadMap [41] (prediction-based method) and result of ours. + +focuses on the latter, which typically models the spatiotemporal features of the normal pattern, while samples that do not conform to the model are labeled as anomalies. + +In the same scene, the appearance and motion of foreground objects are the main differences between normal and abnormal patterns, and also the crucial features that the model needs to learn. We note that the effect of the spatial scale of the anomalies on feature learning is ignored. As illustrated in Fig. 1, many abnormal events occur in a limited region with only subtle differences from the normal pattern while background occupying a larger portion of the frame is the same. Therefore, severe background noise interferes with the feature learning of anomalous changes. Apart from small-scale anomalies, there are also abnormal events that span almost the entire frame, such as traffic accidents. Such changes in the scale of anomalies require the model to remain robust to anomalies of various scales. + +Unfortunately, most previous methods exhibit a lower accuracy in detecting small-scale anomalies compared to larger-scale anomalies, as shown in Fig. 1. Reconstruction- + +based [14, 16, 28, 30, 31, 45] and prediction-based [2, 5, 20, 23, 41] approaches are the mainstream modeling paradigms. The two approaches learn spatio-temporal features of normal patterns by reconstructing and predicting frames, respectively. However, both approaches are interfered by severe background noise since they perform reconstruction or prediction of frame in RGB space. Recent object-centric works [11, 12, 17] separate and process foregrounds and backgrounds to address the challenge of background noise. Nevertheless, modeling approaches based on prediction and binary classification struggle to learn highly discriminative features to distinguish small-scale anomalies. Some multi-scale VAD methods [41, 52] learn features at different resolutions through pooling or feature pyramid-ing but do not focus on appearance and motion features. + +In this paper, we aim to explore a robust multi-scale VAD method. On the one hand, the context dependence of the anomalies necessitates the model to grasp global motion patterns and long-range features within the video. On the other hand, detecting small-scale anomalies not only demands that the model avoid the background noise interference but also requires it to discern highly discriminative short-range spatio-temporal variances between frames, owing to their subtle differences from the normal pattern. Consequently, the model needs to comprehensively learn the spatio-temporal features (spatial appearance and temporal motion features) of normal pattern in a multi-grained manner to achieve robustness to multi-scale anomalies. + +To this end, we take video continuity [19] as supervision to construct three self-supervised proxy tasks, enabling VAD model to learn spatio-temporal features in both coarse-grained and fine-grained manners during training. (i) To determine whether a video sequence is continuous or not. Continuity judgment requires the model to learn the overall long-range temporal features and global motion patterns of the video in a coarse-grained manner. (ii) To locate discontinuities. Finding where the discontinuities occur drives the model to capture changes in local motion in a fine-grained manner rather than background noise. (iii) To estimate the missing frame in feature space instead of RGB space. We formulate the estimation task as a contrastive learning task to avoid the background noise interference and learn highly discriminative features of motion and appearances. By jointly solving three proxy tasks, the model can maintain robustness to anomalies of various scales. + +We conduct experiments on four challenging datasets (Avenue [25], ShanghaiTech [27], UCF-Crime [36] and Campus [2]). The experiments show that our proposed method outperforms the State-Of-The-Art (SOTA) method, especially in scenes with small-scale anomalies. + +Our contributions can be summarized as follows: + +- Three straightforward yet effective multi-grained spatiotemporal representation proxy tasks are designed, en + +abling the video anomaly detection model to maintain robustness to multi-scale anomalies. + +- A reconstruction scheme for frames based on contrastive learning is proposed, motivating the model to learn highly discriminative features from both appearance and motion, rather than simple RGB features and background noise. +- Our method achieves state-of-the-art performance on four datasets, especially outperforming previous methods by a large margin in scenes with small-scale anomalies. + +# 2. Related Work + +# 2.1. Video Anomaly Detection + +Our work focuses on the video anomaly detection method that only learning from normal data, which is mainly grouped into the reconstruction-based methods [4, 16, 28, 33] and the prediction-based methods [20, 41, 48]. These two methods typically use autoencoders (AEs) [15, 51], memory-augmented AEs [13, 14, 23, 28, 31], or generative models [20, 32, 41] to reconstruct current frames or predict future frames so that frames with large reconstruction or prediction errors are recognized as anomalous. While some work attempts to introduce optical flow [23, 30], skeletal information [7, 10, 39], or utilize bi-directional prediction [5, 8, 18, 52] to learn motion and temporal features, these methods cannot be effective for small-scale anomalies due to disturbance from background noise. Recent object-centric methods [12, 17, 39] attempt to separate background and objects to avoid the disturbance of background noise. Although such approach can focus on the foreground objects, the way these methods model the object in terms of classification or RGB prediction lacks the exploration of fine-grained features. In addition, some efforts [1, 24] to synthesize virtual anomaly data fail to work for such anomalies owing to the lack of small-scale virtual data. + +Different from these methods, we expect to enable the model to learn different scales of variation and fine-grained features of the normal pattern via self-supervised spatiotemporal representation learning. + +# 2.2. Self-Supervised Learning in VAD + +Self-supervised learning VAD methods typically perform contrastive learning or solve pretext tasks. For instance, Wang et al. [42] performs contrastive learning with clustered attention mechanism. However, it requires customized data enhancement strategies. For pretext tasks, Yang et al. [46] learn spatio-temporal representations by keyframe-based event restoration, while the work [11] designs multiple proxy tasks in an object-centric way to detect anomalies in different aspects. However, the former [46] suffers from severe background noise interference due to reconstructing events in RGB space, while the latter [11] lacks the ability to learn fine-grained features as its tasks are primarily + +![](images/de84bc4cb7564ce77f45c44a7cad59e9c75ec4148da1917b691eb0db2efc5e23.jpg) +Figure 2. Method overview. The overall framework consists of a shared backbone network and three branches. The three branches perform continuity judgment, discontinuity localization, and missing frame estimation tasks to learn the spatio-temporal features of normal patterns at coarse- and fine-grained levels, respectively. The feature representation $f$ generated after self-supervised training can effectively detect anomalies of various scales. Blue arrows and grey arrows represent training and inference processes, respectively. Best viewed in color. + +binary classifications. + +In this work, we design proxy tasks based on video continuity [19] to learn spatio-temporal features of normal events in a multi-grained manner to perceive anomalies of various scales. In addition, we formulate the reconstruction of frames as a contrastive learning task to learning highly discriminative features instead of background noise. + +# 2.3. Multi-Scale Learning + +Multi-scale learning typically considers information from different scales or granularities. Similar to our work, Xiao et al. [44] hierarchically learns the spatio-temporal features through event mask prediction task. However, event-level features are not sufficient for detecting subtle variances between frames. There is also some multi-scale works [41, 52] in VAD. For instance, Wang et al. [41] obtain image features of different resolutions to perform prediction through different pooling operations, while Zhong et al. [52] obtain features at different scales through feature pyramid. Although pooling and feature pyramids can provide feature representations of different resolutions, they cannot specifically capture the motion and appearance features which is crucial for anomaly detection. In contrast, we design the self-supervised proxy tasks to direct the model's attention towards changes in appearance and motion between frames, enabling it to learn features in a multi-grained manner. + +# 3. Method + +# 3.1. Overview + +Our proposed multi-scale video anomaly detection method attempts to learn spatio-temporal features of normal patterns at both coarse-grained and fine-grained levels by jointly solving multiple proxy tasks of spatio-temporal representation learning. For a normal video clip $V$ , our approach is devoted to learning a video encoder $F$ that transforms the video clip $V$ into the scale-aware high-quality spatio-temporal feature $f$ . We define $f$ as the high-quality feature in terms of anomalous information of different scales if it is able to easily address the following questions: (1) Whether there are missing frames in video clips $V$ , i.e., whether feature representation $f$ remains sensitive to large-scale anomalies. (2) If the video clip $V$ is not continuous, where are the missing frames, i.e., whether feature representation $f$ remains sensitive to small-scale anomalies. (3) If video clip $V$ is discontinuous, what are missing frames, i.e., whether the feature representation $f$ contains high-level motion feature and contextual feature. + +The overall architecture of our proposed method is illustrated in Fig. 2. The network consists of three branches and a video encoder $F$ , all of the branches share the same backbone encoder $F$ . Given two nonoverlapping video clips: continuous clip $V_{c}$ , discontinuous clip $V_{d}$ , and the corresponding missing frame $I_{m}$ , each of the three branches solves a proxy task and operates together to optimize the + +backbone. The first branch is used to discriminate whether a video clip is continuous or not. The second branch is used to locate the position of the missing frame in the discontinuous clip $V_{d}$ . The last branch is designed to learn high-level features through contrastive learning mechanism, so that the feature representation of the discontinuous clip $V_{d}$ can effectively approach the feature representation of the missing frame $I_{m}$ . After joint optimization of self-supervised tasks, the backbone $F$ can generate scale-aware feature representations $f$ for anomaly inference. + +# 3.2. Data Acquisition + +To perform the proposed scale-aware proxy tasks, we select continuous video clips and discontinuous clips containing only one missing frame $I_{m}$ , respectively. The continuous clip contains $T$ continuous frames, while the discontinuous clip has a length of $T - 1$ . Given an initial continuous clip of length $T$ , we uniformly select a frame from the interval $[1, T - 2]$ as the missing frame $I_{m}$ to form the breakpoint (counting from 0). And the leading and alternate of the missing frames are concatenated together to form discontinuous video clip $V_{d}$ . The restriction of missing frame inside the clip sequence ensures that the construction of missing frames and discontinuous clips remains consistent. + +As shown in Fig. 2, we formulate the continuity discrimination task as a binary classification task while the missing frame localization task is formulated as a $T - 2$ class classification task labeled by $I_{m}$ . The estimation task is formulated as a contrastive learning task. Both initial and continuous clips are formed by randomly sampling non-overlapping clips of length $T$ from the same video. + +# 3.3. Network Architecture + +We employ a video encoder (defaults a 3D-ConvNet) as the backbone network $F$ and use it as a shared base for three proxy tasks. Each task introduces into its top a projection network, denoted $P_{1}, P_{2}$ , and $P_{3}$ , which are used to process the video clip embeddings $f$ extracted from the backbone network. All projection heads contain spatio-temporal averaging pooling layers to ensure that the dimensionality of the deep feature embeddings remains consistent, even though the temporal lengths may be different. + +In continuity judgment and missing frame localization, we use fully connected layers as classifiers at the end of $P_{1}$ and $P_{2}$ for performing classification operations. As for the missing frame estimation task, we use $P_{3}$ to embed the video features $f$ extracted by the backbone network into the low-dimensional features to get the representation, i.e., the feature representation of the missing frame. + +# 3.4. Proxy Tasks and Joint Learning + +Task 1: Continuity Judgment. The continuity judgment is intended to help the model learn spatio-temporal features + +and motion patterns in the video in a coarse-grained manner to perceive large-scale anomalies. In this task, in order for the model to learn by video features rather than other cues such as context, the positive and negative samples in each batch in our training come from the same video scene. With N samples in a training batch, cross-entropy loss $\mathcal{L}_{CE}$ is used to optimize the model, and the loss for continuity judgment is $\mathcal{L}_1$ : + +$$ +\mathcal {L} _ {1} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\mathcal {L} _ {C E} \left(P _ {1} \left(F \left(v _ {d} ^ {i}\right)\right) _ {1} + P _ {1} \left(F \left(v _ {c} ^ {i}\right)\right) _ {0}\right)\right), \tag {1} +$$ + +where $(\cdot)$ represents the process of feature processing by the network, $F$ and $P_{1}$ represent the shared backbone network and the continuity judgment projection head, respectively, and $v_{c}^{i}$ and $v_{d}^{i}$ represent the $i_{th}$ continuous and discontinuous video clip, respectively. + +Task 2: Discontinuity Localization. While the continuity judgment task only learns spatio-temporal features in a coarse-grained manner, the discontinuous localization task expects to learn in a fine-grained manner in order to perceive small-scale anomalies. Compared to binary labels, labeling with missing frames $I_{m}$ in this task can drive the model to learn changes in motion between adjacent frames in a fine-grained manner. Similarly, the discontinuous localization loss $\mathcal{L}_2$ can be expressed as an average of the cross-entropy loss of $N$ samples in each batch: + +$$ +\mathcal {L} _ {2} = \frac {1}{N} \sum_ {i = 1} ^ {N} \mathcal {L} _ {C E} \left(P _ {2} \left(F \left(v _ {d} ^ {i}\right)\right) _ {I _ {i}}\right), \tag {2} +$$ + +where $P_{2}$ represents the discontinuity localization projection head and $I_{i}$ represents the missing frame position corresponding to discontinuous clip $v_{d}^{i}$ . + +Task 3: Missing Frame Estimation. In order to further learn the highly discriminative spatio-temporal features of the video, we estimate missing frames in terms of both appearance and motion. Meanwhile, we choose to estimate frames in feature space rather than RGB space to avoid background noise interference. We design a contrastive learning mechanism to perform this estimation task based on the following facts. First, since adjacent frames and missing frames are temporally connected, adjacent frames and missing frames contain more similar motions compared to non-adjacent frames from the same scene but farther away. Second, video frames from the same scene share more similar appearance features with missing frames than video frames from other scenes. + +Accordingly, we learn the motion information in the video by triplet loss [34] taking the discontinuous video clips $v_{d}^{i}$ as the anchor point, the missing frames $I_{m}$ as the positive samples, and the continuous video clips $v_{c}^{i}$ as the negative samples. In addition, we take discontinuous clips + +$v_{d}^{i}$ as anchor points, continuous clips $v_{c}^{i}$ from the same scene as positive samples, and video clips $\{v^{j}\}_{i\neq j}$ from different scenes as negative samples, to learn the appearance information in the video by contrastive loss [19]. Defining the cosine similarity computation operation as $sim(\cdot ,\cdot)$ , and the hyperparameter that balances the relative contributions of the triplet loss and the contrastive loss as $\omega \in [0,1]$ , the missing frame estimation loss is denoted as: + +$$ +\begin{array}{l} \mathcal {L} _ {3} = \frac {1}{N} \sum_ {i = 1} ^ {N} (\omega \times \max (0, \gamma - \left(p _ {i} ^ {+} - p _ {i} ^ {-}\right)) \tag {3} \\ - (1 - \omega) \log \left(\frac {q _ {i} ^ {+}}{q _ {i} ^ {+} + \sum_ {j = 1 , j \neq i} ^ {N} q _ {i , j ^ {-}}}\right)), \\ \end{array} +$$ + +$$ +p _ {i} ^ {+} = \mathrm {s i m} \left(P _ {3} (V _ {d} ^ {i}, I _ {i})\right), \tag {4} +$$ + +$$ +p _ {i} ^ {-} = \operatorname {s i m} \left(P _ {3} \left(V _ {d} ^ {i}, V _ {c} ^ {i}\right)\right), \tag {5} +$$ + +$$ +q _ {i} ^ {+} = \exp \left(\operatorname {s i m} \left(P _ {3} \left(V _ {d} ^ {i}, V _ {c} ^ {i}\right)\right) / \tau\right), \tag {6} +$$ + +$$ +\begin{array}{l} q _ {i, j} ^ {-} = \exp \left(\sin \left(P _ {3} \left(v _ {d} ^ {i}, v _ {d} ^ {j}\right)\right) / \tau\right) \tag {7} \\ + \exp \left(\mathrm {s i m} \left(P _ {3} (v _ {d} ^ {i}, v _ {c} ^ {j})\right) / \tau\right), \\ \end{array} +$$ + +where $p_i^+$ and $p_i^-$ denote the similarity between positive sample pairs and negative sample pairs in the triplet loss, respectively, $q_i^+$ denotes a single positive sample pair in the contrastive loss, $q_{i,j}^{-}$ denotes the similarity of negative sample pairs in the contrastive loss, and $\tau$ is the temperature factor in the contrastive loss. Increasing $\omega$ allows for more focus on motion information, while decreasing $\omega$ allows for more focus on appearance information. Overall, the model is incentivized to learn fine-grained motion changes and contextual features, including background information and object appearance. + +Joint Optimization. The above three proxy tasks learn spatio-temporal features in videos from different levels, and we jointly optimize our shared video encoder $F$ with the multi-loss function. + +$$ +\mathcal {L} _ {F} = w _ {1} \mathcal {L} _ {1} + w _ {2} \mathcal {L} _ {2} + w _ {3} \mathcal {L} _ {3}, \tag {8} +$$ + +where $w_{1}, w_{2}, w_{3} \in [0,1]$ are the individual weights on the $\mathcal{L}_1, \mathcal{L}_2$ and $\mathcal{L}_3$ . Joint optimization helps the model to learn video spatio-temporal features from both coarse- and fine-grained scales to perceive anomalies of various scales. + +# 3.5. VAD Inference + +In the inference, we use $T$ frames as the processing unit. Following the setup of the SOTA method [20, 41, 46], we + +choose not to compute anomaly scores using the error between each estimated frame $\hat{I}_t$ ( $1 \leq t \leq T$ ) and the corresponding ground truth $I_t$ . Instead, we opt to employ the Peak Signal-to-Noise Ratio (PSNR) corresponding to the frame with the largest Mean-Square Error (MSE) among the $T$ frames and the ground truth. This approach serves as the detection metric for the video sequence, as follow: + +$$ +\operatorname {M S E} \left(\boldsymbol {I} _ {t}, \hat {\boldsymbol {I}} _ {t}\right) = \frac {1}{H \cdot W} \sum_ {i, j} \left\| \boldsymbol {I} _ {t} (i, j) - \hat {\boldsymbol {I}} _ {t} (i, j) \right\| _ {2} ^ {2}, \tag {9} +$$ + +$$ +P S N R _ {t} = 1 0 \log_ {1 0} \frac {\operatorname {M A X} _ {\boldsymbol {I} _ {t}} ^ {2}}{\operatorname {M S E} \left(\boldsymbol {I} _ {t} , \hat {\boldsymbol {I}} _ {t}\right)}, \tag {10} +$$ + +where $H$ and $W$ are the height and width of the frames, respectively, $I_{t}(i,j) / \hat{I}_{t}(i,j)$ represents the RGB value of the $(i,j)$ -th pixel in $I_{t} / \hat{I}_{t}$ , and $\mathrm{MAX}_{I_t}$ represents the maximum value of the image pixel in $I_{t}$ . Finally, we calculate the anomaly score $S_{t}$ for the $t$ -th frame $I_{t}$ by normalizing the PSNR values over all $T$ frames: + +$$ +S _ {t} = 1 - \frac {P S N R _ {t} - \operatorname* {m i n} _ {t ^ {\prime}} P S N R _ {t ^ {\prime}}}{\operatorname* {m a x} _ {t ^ {\prime}} P S N R _ {t ^ {\prime}} - \operatorname* {m i n} _ {t ^ {\prime}} P S N R _ {t ^ {\prime}}}, \tag {11} +$$ + +where $1 \leq t' \leq T$ . A higher score $S_{t}$ indicates that the frame $I_{t}$ is more likely to be anomalous. + +# 4. Experiment + +# 4.1. Dataset and Evaluation Metrics + +Cross-Scene Anomaly Dataset Avenue [25] contains 16 training videos and 21 test videos with 47 abnormal events, including running and throwing. The scenes in this dataset are single and the anomalies are mainly large-scale. ShanghaiTech [27] has 13 scenes with complex lighting conditions and different perspectives. In addition, the dataset includes anomalies caused by sudden movements, such as chasing and arguing. The different perspectives and the unfixed position of the camera lead to a large variation of both the object scale and the anomaly scale in the scene. This is a significant challenge to the robustness of anomaly detection methods to scale variations. UCF-Crime [36] comprises 13 anomaly types, spanning a total of 128 hours of video footage. This dataset is complex because of the unconstrained backgrounds and anomalies of various scales. The training set contains video-level annotations, whereas the testing set includes frame-level annotations. + +Scene-Dependent Anomaly Dataset. Campus [2] is currently the most challenging dataset in its field with 43 scenes, 28 classes of anomalous events and 16 hours of videos. Especially, it contains scene-dependent anomalies, which means an normal event may be abnormal in another scene. Detecting scene-dependent anomalies requires the model to understand the scene and learn highly discriminative features rather than overfit. + +
MethodLength TShanghaiTechCampus
AUCAUC
MicroMacroMicroMacro
Continuity Judgment878.182.360.163.2
1277.281.160.363.9
1675.179.258.260.9
Discontinuity Localization882.685.562.165.4
1283.487.262.665.1
1682.185.561.965.0
Missing Frame Estimation879.883.263.566.8
1281.284.965.268.9
1680.584.064.968.0
Overall Framework884.288.866.369.2
1285.189.867.370.9
1683.988.366.069.2
+ +Evaluation Metrics The area under the ROC curve (AUC) serves as a commonly used metric for evaluation and comparison. A higher AUC score indicates a better anomaly detection capability. Following previous research [12], we evaluate both the Micro and Macro versions of AUC. + +# 4.2. Implementation Details + +We train and evaluate our method with an NVIDIA RTX 3090 GPU. In the training phase, we resize the resolution of all input video clips to $256 \times 256$ pixels, while the values of the pixels in all frames are normalized to the range [0, 1]. For the pre-training of the three proxy tasks, we utilize AdamW as the optimizer while the length of the continuous video clips is set to $T = 12$ frames. The initial learning rate is set to 0.0003 and is gradually decayed following the scheme of cosine annealing. For the hyperparameters, we set $\omega = 0.5$ in Equation 3, $w_{1} = w_{2} = w_{3} = 1$ in Equation 8, and the temperature factor $\tau = 0.1$ in the contrastive loss. In our reported experimental performance, the shared backbone network $F$ is implemented as I3D-RGB [3]. In addition, we incorporate the designed proxy tasks into the backbone of the SOTA methods [2, 11], the experimental results are also reported. Following the existing method setup [11, 24, 33], we perform the proxy task at the object-level with the multi-task backbone [11] and at the frame-level with the other backbones. + +# 4.3. Ablation Study + +Considering the ShanghaiTech and Campus datasets contain diverse perspectives and scenes, and large variations in scales of anomaly, we conduct exhaustive ablation experiments on the ShanghaiTech and Campus datasets. + +Sensitivity to the Length of Video Clips. The length $T$ of the input clips is an important setting in our proposed framework. With a small $T$ , both continuity judgments and dis + +Table 1. Comparison of proposed framework and individual proxy tasks between different lengths $T$ of input clips. We report Micro and Macro AUC (%) on ShanghaiTech and Campus datasets. The best performing results are marked in bold. + +
IDCJDLMFEFPShanghaiTechCampus
AUCAUC
MicroMacroMicroMacro
1---✓70.271.258.158.6
2✓---77.281.160.363.9
3-✓--83.487.262.665.1
4--✓-81.284.965.268.9
5✓✓--84.187.862.865.4
6✓-✓-82.385.665.369.2
7-✓✓-84.589.566.270.1
8✓✓✓-85.189.867.370.9
9✓✓-✓84.287.662.865.6
+ +Table 2. Ablation experiments on the contributions of each proxy task. We report the AUC (\%) scores on ShanghaiTech and Campus datasets. 'CJ', 'DL' and 'MFE' stand for the three proxy tasks of continuity judgment, discontinuity localization and missing frame estimation, respectively. In addition we add experiments on frame prediction for comparison, 'FP' represents frame prediction. + +continuity localization will be so ambiguous as to be difficult. While with a larger $T$ , the model may not need to learn high-quality spatio-temporal features to solve the task due to enough spatio-temporal information in the feed. Table 1 shows the sensitivity of each proxy task to $T$ . For the discontinuous localization and missing frame estimation tasks, $T = 12$ presents better performance. Although at $T = 8$ , the continuity judgment yields competitive results on the ShanghaiTech dataset, the overall framework achieves the best outcomes with jointly optimized at $T = 12$ . + +Effect of Continuity Judgment. The results of the ablation experiments for each proxy task are shown in Table 2, where we take the frame prediction [20] performed on the same backbone $F$ as the baseline and train it with gradient loss and intensity loss (ID 1). Performing the continuity judgment task alone improves the performance of the model compared to frame prediction (ID 2). This suggests that continuity judgment task motivates the model to generate higher quality representations than frame prediction. + +Effect of Discontinuity Localization. Compared to other proxy tasks and baseline, the performance of the discontinuous localization task alone can be significantly improved, especially on ShanghaiTech which contains anomalies of various scales (ID 3, ID 5 and ID 7). As this task samples discontinuous positions uniformly along the time axis, it motivates the model to capture subtle variations between frames, obtaining fine-grained representations. + +Effect of Missing Frame Estimation. Performing missing frame estimation alone is effective in improving performance on Campus compared to baseline and other proxy tasks, which contains scene-dependent anomalies (ID 4). This improvement illustrates that our proposed contrastive learning scheme can help the model understand the scene. Compared to performing the frame prediction task to learn RGB features (ID 1 and ID 9), our proposed scheme learns highly discriminative motion features and semantically ef + +
yearMethodAvenueShanghaiTech
AUCAUC
MicroMacroMicroMacro
2018Liu et al. [20]85.1-72.8-
Liu et al. [22]84.4---
Sultani et al. [36]--76.5-
2019Gong et al. [14]83.8-71.2-
Lee et al. [18]90.0-76.2-
Ionescu et al. [17]87.490.478.784.9
2020Park et al. [31]88.5-70.5-
Sun et al. [37]89.6-74.7-
Lu et al. [26]85.8-75.9-
Wang et al. [42]87.0-79.3-
Yu et al. [47]89.6-74.8-
2021Liu et al. [23]91.1-76.2-
Lv et al. [28]89.5-73.8-
Georgescu et al. [11]*91.591.982.489.3
Georgescu et al. [12]*92.390.482.789.3
2022Wang et al. [41]88.3-76.6-
Zaheer et al. [49]74.2-79.6-
Chen et al. [5]90.3-78.1-
Zhong et al. [52]89.0-74.5-
Cho et al. [6]88.0-76.3-
Yang et al. [45]89.9-74.7-
Ristea et al. [33]*92.991.983.689.5
Acsintoae et al. [1]▼*93.093.283.790.5
2023Yang et al. [46]89.9-73.8-
Cao et al. [2]86.8-79.2-
Liu et al. [21]92.8-78.8-
Singh et al. [35]86.0-76.6-
Sun et al. [39]92.4-83.0-
Sun et al. [38]91.5-78.6-
Liu et al. [24]▼91.892.383.887.8
Liu et al. [24] ▼*93.693.985.091.4
Ours92.492.985.189.8
Ours▼93.292.586.291.0
Ours*(Estimation Task)92.693.085.591.8
Ours*(Three Tasks)93.693.886.892.4
Ours▼*(Three Tasks)94.394.587.593.0
+ +Table 3. Comparison with SOTA methods of the Micro and Macro AUC (%) on Avenue and ShanghaiTech datasets. The best performing results are marked in bold. +$\nabla$ : Methods apply virtual dataset for training. +*: Methods utilize multi-task model ([11] or [12]) as backbone. + +
MethodReferenceMicro AUCMacro AUC
Park et al. [31]CVPR2068.972.4
Georgescu et al. [11]CVPR2174.678.2
Wang et al. [41]TNNLS2272.976.8
Sun et al. [39]CVPR2375.578.3
Ours-80.683.9
+ +Table 4. Results of Micro and Macro AUC(%) on UCF-Crime dataset. The best performing results are marked in bold. + +fective contexts, including the appearance of objects and background (ID 4 and ID 8). + +Effect of Joint Optimization. Although a single task may be able to achieve competitive performance, Experiments have shown that when proxy tasks are executed in pairs (ID + +
MethodReferenceMiceo AUCMacro AUC
Liu et al. [20]CVPR1857.960.2
Gong et al. [14]CVPR1961.962.5
Ionescu et al. [17]CVPR1959.363.4
Park et al. [31]CVPR2062.563.6
Liu et al. [23]ICCV2163.7-
Lv et al. [28]CVPR2164.4-
Wang et al. [41]TNNLS2261.964.2
Cao et al. [2]△CVPR2368.2-
Ours-67.370.9
Ours△-70.172.2
+ +Table 5. Comparison with SOTA methods of the Micro and Macro AUC(%) on Campus dataset. The best performing results are marked in bold. + +$\triangle$ :Version that utilize scene-conditioned model [2] as backbone. + +5, ID 6 and ID 7) and all tasks are executed together (ID 8), the performance of the model continues to improve to an optimum. This joint improvement is attributed to the fact that different proxy tasks learn spatio-temporal features at different levels, and they complement each other in motion and contextual feature learning. + +# 4.4. Comparisons with State-Of-The-Arts + +We compare the proposed framework with SOTA methods in terms of Micro and Macro AUC(\%). It is noteworthy that the current state-of-the-art methods [1, 24, 33] utilize the multi-task framework proposed by Georgescu et al. ([11] and [12], respectively) as the backbone, where [1] and [24] employ virtual data for training. Therefore, we evaluate different configurations of the proposed method on the Avenue and ShanghaiTech datasets, as shown in Table 3. With the multi-task backbone [11] (marked with * in Table 3), we assess two versions: one involves replacing only the prediction component with our designed missing estimation task, while keeping the other proxy tasks unchanged. The other version replaces both proxy tasks in the motion branch with continuity judgment and discontinuity localization tasks in addition to incorporating the missing estimation task. In addition, we report the performance of applying virtual data for training (marked with $\nabla$ in Table 3). + +Results on Avenue. As shown in Table 3, our method achieved the highest AUC scores, obtaining a Micro AUC of $94.3\%$ and a Macro AUC of $94.5\%$ . Without the backbone of multi-task [11, 12] and virtual data [24], our method still scored the best performance with $92.9\%$ Macro AUC. + +Results on ShanghaiTech. Our proposed method yields a Micro AUC of $87.5\%$ and a Macro AUC of $93.0\%$ on the ShanghaiTech dataset. With the backbone of multi-task [11, 12], we improve the Micro AUC by $3.1\%$ compared to Georgescu et al. [11] in the setup with only the missing estimation task, and by $4.4\%$ when incorporating three designed proxy tasks. Without the backbone of multi-task + +![](images/a234b3d579b70549459df2c7994935fb761c60a799895e7e9098136854df9a1c.jpg) +Figure 3. Examples of detection for anomalies of different scales from the ShanghaiTech dataset, including ground truth and prediction or reconstruction error maps from different methods (from top to bottom: ROADMAP[41], DEDDnet[52], USTN-DSC[46], and our method). Best viewed in color. + +[11, 12] and virtual data [24], our method still achieves a Micro AUC score of $85.1\%$ , which is comparable to the method [24] that employed the multi-task backbone and virtual training data. In comparison with the multi-scale VAD methods ([41] and [52]), our method achieves a significant improvement of $8.5\%$ and $10.6\%$ , respectively. + +Results on UCF-Crime. Due to the absence of published results (methods only learning from normal data) on the UCF-Crime dataset, we implement the code from the existing literature [11, 31, 39, 41]. As shown in Table 4, our proposed method achieves a significant improvement compared to the second best method by $5.1\%$ in terms of Micro AUC and $5.6\%$ in terms of Macro AUC. + +Results on Campus. We evaluate an additional version on the Campus dataset with the scene-conditioned model (provided by Cao et al. [2]) as the backbone, as it is necessary to incorporate scene features for detecting scene-dependent anomalies. As shown in Table 5, our method achieves the highest Micro AUC score of $70.1\%$ and Macro AUC score of $72.2\%$ . Without the scene-conditioned model, our method still obtain a Micro AUC of $67.3\%$ , surpassing other RGB reconstruction-based [14, 28, 31] or RGB prediction-based methods [20, 23, 41]. + +# 4.5. Qualitative Results + +Fig. 1 shows the anomaly curves of the test video and the comparison of our method with PMem [31] (reconstruction- + +![](images/531e5c2fdf059771824956ac7c8841e96d3c08419d0f70952f4c22459a4459db.jpg) +Figure 4. Visualization of salient regions of test frames by the proposed method. Best viewed in color. + +![](images/0689b7c8e5d1a25f72a479d59d49f68678c849016004bc9c1ffe8b3f9e18261b.jpg) + +![](images/e3baa39affc129a32a1531fb8896679998a5fbe4ead68b3e85c6aebd63fe2570.jpg) + +based method) and RoadMap [41] (prediction-based method). When there are small-scale anomalies in the sampled videos, the anomaly score curves of the other two methods have obvious missed detections, while our method can accurately perceive the small-scale anomalies. In addition, Fig. 3 illustrates more examples, including the comparison with the multi-scale VAD methods [41, 52] and the comparison with self-supervised VAD method [46]. It can be observed that for normal regions in video frames, our method is able to estimate them well, while for abnormal event regions large errors occur. While other methods generate large errors for some normal regions, especially for frames with small-scale anomalies. + +Visualizations of the salient regions of the frames by our method are illustrated in Fig. 4. For both large-scale anomalies (1-th row) and small-scale anomalies (2-th row), our model can accurately focus on the anomalous regions. + +# 5. Conclusion + +In this paper, we observe that anomalous spatial scales affect the feature learning of the model. Small-scale anomalies require the model to learn highly discriminative features in a fine-grained manner rather than background noise. To detect multi-scale anomalies, we design proxy tasks supervised by video continuity that motivate the model to learn spatio-temporal features in both coarse-grained and fine-grained manners. Furthermore, instead of performing RGB reconstruction and prediction, we estimation frames in feature space through contrastive learning to learn highly discriminative features of appearance and motion. Experiments conducted on four challenging benchmark datasets validate the effectiveness of our proposed method. + +Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grants (62101064, 62171057, 62201072, U23B2001, 62001054, 62071067), the National Postdoctoral Program for Innovative Talents under Grant BX20230052, China Postdoctoral Science Foundation (2023TQ0039), the Ministry of Education and China Mobile Joint Fund (MCM20200202, MCM20180101), Beijing University of Posts and Telecommunications-China Mobile Research Institute Joint Innovation Center. + +# References + +[1] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. *Ubnormal: New benchmark for supervised open-set video anomaly detection.* In *CVPR*, pages 20111-20121, 2022. 2, 7 +[2] Congqi Cao, Yue Lu, Peng Wang, and Yanning Zhang. A new comprehensive benchmark for semi-supervised video anomaly detection and anticipation. In CVPR, pages 20392-20401, 2023. 2, 5, 6, 7, 8 +[3] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, pages 4724-4733, 2017. 6 +[4] Yunpeng Chang, Zhigang Tu, Wei Xie, and Junsong Yuan. Clustering driven deep autoencoder for video anomaly detection. In ECCV, pages 329-345, 2020. 2 +[5] Chengwei Chen, Yuan Xie, Shaohui Lin, Angela Yao, Guannan Jiang, Wei Zhang, Yanyun Qu, Ruizhi Qiao, Bo Ren, and Lizhuang Ma. Comprehensive regularization in a bidirectional predictive network for video anomaly detection. In AAAI, pages 230-238, 2022. 1, 2, 7 +[6] MyeongAh Cho, Taeoh Kim, Woo Jin Kim, Suhwan Cho, and Sangyoun Lee. Unsupervised video anomaly detection via normalizing flows with implicit latent features. Pattern Recognit., 129:108703, 2022. 7 +[7] Romero F. A. B. de Morais, Vuong Le, Truyen Tran, Budhaditya Saha, Moussa Reda Mansour, and Svetha Venkatesh. Learning regularity in skeleton trajectories for anomaly detection in videos. In CVPR, pages 11996-12004, 2019. 2 +[8] Zhiwen Fang, Jiafei Liang, Joey Tianyi Zhou, Yang Xiao, and Feng Yang. Anomaly detection with bidirectional consistency in videos. IEEE Trans. Neural Networks Learn. Syst., 33(3):1079-1092, 2022. 2 +[9] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. MIST: multiple instance self-training framework for video anomaly detection. In CVPR, pages 14009-14018, 2021. 1 +[10] Alessandro Flaborea, Luca Collorone, Guido Maria D'Amely di Melendugno, Stefano D'Arrigo, Bardh Prenkaj, and Fabio Galasso. Multimodal motion conditioned diffusion model for skeleton-based video anomaly detection. In (ICCV), pages 10318-10329, 2023. 2 +[11] Mariana-Iuliana Georgescu, Antonio Barbalau, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. Anomaly detection in video via self-supervised and multi-task learning. In CVPR, pages 12742-12752, 2021. 2, 6, 7, 8 +[12] Mariana-Iuliana Georgescu, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4505-4523, 2022. 2, 6, 7, 8 +[13] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In ICCV, pages 1705-1714, 2019. 2 + +[14] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In ICCV, pages 1705-1714, 2019. 2, 7, 8 +[15] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal regularity in video sequences. In CVPR, pages 733-742, 2016. 2 +[16] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In ICCV, pages 8771-8780, 2021. 2 +[17] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In CVPR, pages 7842–7851, 2019. 2, 7 +[18] Sangmin Lee, Hak Gu Kim, and Yong Man Ro. BMAN: bidirectional multi-scale aggregation networks for abnormal event detection. IEEE Trans. Image Process., 29:2395-2408, 2019. 2, 7 +[19] Hanwen Liang, Niamul Quader, Zhixiang Chi, Lizhe Chen, Peng Dai, Juwei Lu, and Yang Wang. Self-supervised spatiotemporal representation learning by exploiting video continuity. In AAAI, pages 1564-1573, 2022. 2, 3, 5 +[20] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In CVPR, pages 6536-6545, 2018. 2, 5, 6, 7, 8 +[21] Wenrui Liu, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Diversity-measurable anomaly detection. In CVPR, pages 12147-12156, 2023. 7 +[22] Yusha Liu, Chun-Liang Li, and Barnabás Póczos. Classifier two sample test for video anomaly detections. In BMVC, page 71. BMVA Press, 2018. 7 +[23] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In ICCV, pages 13568-13577, 2021. 1, 2, 7, 8 +[24] Zuhao Liu, Xiao-Ming Wu, Dian Zheng, Kun-Yu Lin, and Wei-Shi Zheng. Generating anomalies for video anomaly detection with prompt-based feature mapping. In CVPR, pages 24500-24510. IEEE, 2023. 1, 2, 6, 7, 8 +[25] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 FPS in MATLAB. In ICCV, pages 2720-2727, 2013. 2, 5 +[26] Yiwei Lu, Frank Yu, Mahesh Kumar Krishna Reddy, and Yang Wang. Few-shot scene-adaptive anomaly detection. In ECCV, pages 125-141, 2020. 7 +[27] Weixin Luo, Wen Liu, and Shenghua Gao. A revisit of sparse coding based anomaly detection in stacked RNN framework. In ICCV, pages 341-349, 2017. 2, 5 +[28] Hui Lv, Chen Chen, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Learning normal dynamics in videos with meta prototype network. In CVPR, pages 15425-15434, 2021. 2, 7, 8 + +[29] Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In CVPR, pages 8022-8031, 2023. 1 +[30] Trong-Nguyen Nguyen and Jean Meunier. Anomaly detection in video sequence with appearance-motion correspondence. In ICCV, pages 1273-1283, 2019. 2 +[31] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In CVPR, pages 14360-14369, 2020. 1, 2, 7, 8 +[32] Mahdyar Ravanbakhsh, Enver Sangineto, Moin Nabi, and Nicu Sebe. Training adversarial discriminators for cross-channel abnormal event detection in crowds. In WACV, pages 1896-1904, 2019. 2 +[33] Nicolae-Catalin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In CVPR, pages 13566-13576, 2022. 2, 6, 7 +[34] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823, 2015. 4 +[35] Ashish Singh, Michael J Jones, and Erik G Learned-Miller. Eval: Explainable video anomaly localization. In CVPR, pages 18717-18726, 2023. 7 +[36] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, pages 6479-6488, 2018. 1, 2, 5, 7 +[37] Che Sun, Yunde Jia, Yao Hu, and Yuwei Wu. Scene-aware context reasoning for unsupervised abnormal event detection in videos. In ACM Multimedia, pages 184-192, 2020. 7 +[38] Che Sun, Chenrui Shi, Yunde Jia, and Yuwei Wu. Learning event-relevant factors for video anomaly detection. In AAAI, pages 2384-2392, 2023. 7 +[39] Shengyang Sun and Xiaojin Gong. Hierarchical semantic contrast for scene-aware video anomaly detection. In CVPR, pages 22846-22856, 2023. 1, 2, 7, 8 +[40] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W. Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In ICCV, pages 4955-4966, 2021. 1 +[41] Xuanzhao Wang, Zhengping Che, Bo Jiang, Ning Xiao, Ke Yang, Jian Tang, Jieping Ye, Jingyu Wang, and Qi Qi. Robust unsupervised video anomaly detection by multipath frame prediction. IEEE Trans. Neural Networks Learn. Syst., 33 (6):2301-2312, 2022. 1, 2, 3, 5, 7, 8 +[42] Ziming Wang, Yuexian Zou, and Zeming Zhang. Cluster attention contrast for video anomaly detection. In ACM Multimedia, pages 2463-2471, 2020. 2, 7 +[43] Jie Wu, Wei Zhang, Guanbin Li, Wenhao Wu, Xiao Tan, Yingying Li, Errui Ding, and Liang Lin. Weakly-supervised spatio-temporal anomaly detection in surveillance video. In IJCAI, pages 1172-1178, 2021. 1 +[44] Fanyi Xiao, Kaustav Kundu, Joseph Tighe, and Davide Modolo. Hierarchical self-supervised representation learning for movie understanding. In CVPR, pages 9717-9726, 2022. 3 + +[45] Zhiwei Yang, Peng Wu, Jing Liu, and Xiaotao Liu. Dynamic local aggregation network with adaptive clusterer for anomaly detection. In ECCV, pages 404-421, 2022. 2, 7 +[46] Zhiwei Yang, Jing Liu, Zhaoyang Wu, Peng Wu, and Xiaotao Liu. Video event restoration based on keyframes for video anomaly detection. In CVPR, pages 14592-14601, 2023. 1, 2, 5, 7, 8 +[47] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze test helps: Effective video anomaly detection via learning to complete video events. In ACM Multimedia, pages 583-591, 2020. 7 +[48] Jongmin Yu, Younkwan Lee, Kin Choong Yow, Moongu Jeon, and Witold Pedrycz. Abnormal event detection and localization via adversarial event prediction. IEEE Trans. Neural Networks Learn. Syst., 33(8):3572-3586, 2022. 2 +[49] Muhammad Zaigham Zaheer, Arif Mahmood, Muhammad Haris Khan, Mattia Segù, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In CVPR, pages 14724–14734, 2022. 7 +[50] Chen Zhang, Guorong Li, Yuankai Qi, Shuhui Wang, Laiyun Qing, Qingming Huang, and Ming-Hsuan Yang. Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection. In CVPR, pages 16271-16280, 2023. 1 +[51] Yiru Zhao, Bing Deng, Chen Shen, Yao Liu, Hongtao Lu, and Xian-Sheng Hua. Spatio-temporal autoencoder for video anomaly detection. In ACM Multimedia, pages 1933-1941, 2017. 2 +[52] Yuanhong Zhong, Xia Chen, Yongting Hu, Panliang Tang, and Fan Ren. Bidirectional spatio-temporal feature learning with multiscale evaluation for video anomaly detection. IEEE Trans. Circuits Syst. Video Technol., 32(12):8285-8296, 2022. 2, 3, 7, 8 +[53] Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection. In AAAI, pages 3769-3777. AAAI, 2023. 1 \ No newline at end of file diff --git a/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/images.zip b/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d5b41c9113f5c2e1bf38433d9182b94732394427 --- /dev/null +++ b/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3db16e17368652c3a80b27e5c3eafac0b7bcf64314facbf772fd2c91a0e48619 +size 554410 diff --git a/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/layout.json b/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b9531967cb02c6539bdb6e6318d80675d4797937 --- /dev/null +++ b/2024/Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning/layout.json @@ -0,0 +1,9107 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "type": "text", + "content": "Multi-Scale Video Anomaly Detection by Multi-Grained Spatio-Temporal Representation Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 161, + 569, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 161, + 569, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 569, + 203 + ], + "type": "text", + "content": "Menghao Zhang, Jingyu Wang\\*, Qi Qi, Haifeng Sun, Zirui Zhuang, Pengfei Ren, Ruilong Ma, Jianxin Liao \nState Key Laboratory of Networking and Switching Technology, \nBeijing University of Posts and Telecommunications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 205, + 568, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 568, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 568, + 217 + ], + "type": "text", + "content": "zhangmenghao, wangjingyu, qiqi8266, hfsun, zhuangzirui, rpf, maruilong@bupt.edu.cn; jxlbupt@gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 270, + 290, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 547 + ], + "type": "text", + "content": "Recent progress in video anomaly detection suggests that the features of appearance and motion play crucial roles in distinguishing abnormal patterns from normal ones. However, we note that the effect of spatial scales of anomalies is ignored. The fact that many abnormal events occur in limited localized regions and severe background noise interferes with the learning of anomalous changes. Meanwhile, most existing methods are limited by coarse-grained modeling approaches, which are inadequate for learning highly discriminative features to discriminate subtle differences between small-scale anomalies and normal patterns. To this end, this paper addresses multi-scale video anomaly detection by multi-grained spatio-temporal representation learning. We utilize video continuity to design three proxy tasks to perform feature learning at both coarse-grained and fine-grained levels, i.e., continuity judgment, discontinuity localization, and missing frame estimation. In particular, we formulate missing frame estimation as a contrastive learning task in feature space instead of a reconstruction task in RGB space to learn highly discriminative features. Experiments show that our proposed method outperforms state-of-the-art methods on four datasets, especially in scenes with small-scale anomalies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "type": "text", + "content": "Video Anomaly Detection (VAD) is dedicated to detecting anomalous events in videos with wide applications in public safety and intelligent surveillance. The primary challenge of VAD is the sparsity of abnormal samples, limiting the direct learning of abnormal patterns from the available data. Therefore, most previous VAD studies can be divided into the weakly supervised category [9, 29, 36, 40, 43, 50, 53] that learns with video-level annotations, or the category only learning from normal data [5, 23, 24, 39, 46]. Our work" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 245, + 541, + 388 + ], + "blocks": [ + { + "bbox": [ + 312, + 245, + 541, + 388 + ], + "lines": [ + { + "bbox": [ + 312, + 245, + 541, + 388 + ], + "spans": [ + { + "bbox": [ + 312, + 245, + 541, + 388 + ], + "type": "image", + "image_path": "bdda958cde7f6fc7eaf99424daddf605a6341516f9f8ef3be8c651bad8370536.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 396, + 547, + 452 + ], + "lines": [ + { + "bbox": [ + 305, + 396, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 547, + 452 + ], + "type": "text", + "content": "Figure 1. Examples of small-scale anomalies and comparisons of results from different methods. From top to bottom: video sequence, ground truth (green regions are abnormal), result of PMem [31] (reconstruction-based method), result of RoadMap [41] (prediction-based method) and result of ours." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 473, + 546, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 546, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 546, + 509 + ], + "type": "text", + "content": "focuses on the latter, which typically models the spatiotemporal features of the normal pattern, while samples that do not conform to the model are labeled as anomalies." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 676 + ], + "type": "text", + "content": "In the same scene, the appearance and motion of foreground objects are the main differences between normal and abnormal patterns, and also the crucial features that the model needs to learn. We note that the effect of the spatial scale of the anomalies on feature learning is ignored. As illustrated in Fig. 1, many abnormal events occur in a limited region with only subtle differences from the normal pattern while background occupying a larger portion of the frame is the same. Therefore, severe background noise interferes with the feature learning of anomalous changes. Apart from small-scale anomalies, there are also abnormal events that span almost the entire frame, such as traffic accidents. Such changes in the scale of anomalies require the model to remain robust to anomalies of various scales." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Unfortunately, most previous methods exhibit a lower accuracy in detecting small-scale anomalies compared to larger-scale anomalies, as shown in Fig. 1. Reconstruction-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 1, + 494, + 17 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 1, + 494, + 17 + ], + "spans": [ + { + "bbox": [ + 145, + 1, + 494, + 17 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 703, + 138, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 138, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 138, + 713 + ], + "type": "text", + "content": "**Corresponding author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17385" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "based [14, 16, 28, 30, 31, 45] and prediction-based [2, 5, 20, 23, 41] approaches are the mainstream modeling paradigms. The two approaches learn spatio-temporal features of normal patterns by reconstructing and predicting frames, respectively. However, both approaches are interfered by severe background noise since they perform reconstruction or prediction of frame in RGB space. Recent object-centric works [11, 12, 17] separate and process foregrounds and backgrounds to address the challenge of background noise. Nevertheless, modeling approaches based on prediction and binary classification struggle to learn highly discriminative features to distinguish small-scale anomalies. Some multi-scale VAD methods [41, 52] learn features at different resolutions through pooling or feature pyramid-ing but do not focus on appearance and motion features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 253, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 289, + 407 + ], + "type": "text", + "content": "In this paper, we aim to explore a robust multi-scale VAD method. On the one hand, the context dependence of the anomalies necessitates the model to grasp global motion patterns and long-range features within the video. On the other hand, detecting small-scale anomalies not only demands that the model avoid the background noise interference but also requires it to discern highly discriminative short-range spatio-temporal variances between frames, owing to their subtle differences from the normal pattern. Consequently, the model needs to comprehensively learn the spatio-temporal features (spatial appearance and temporal motion features) of normal pattern in a multi-grained manner to achieve robustness to multi-scale anomalies." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 410, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 289, + 613 + ], + "type": "text", + "content": "To this end, we take video continuity [19] as supervision to construct three self-supervised proxy tasks, enabling VAD model to learn spatio-temporal features in both coarse-grained and fine-grained manners during training. (i) To determine whether a video sequence is continuous or not. Continuity judgment requires the model to learn the overall long-range temporal features and global motion patterns of the video in a coarse-grained manner. (ii) To locate discontinuities. Finding where the discontinuities occur drives the model to capture changes in local motion in a fine-grained manner rather than background noise. (iii) To estimate the missing frame in feature space instead of RGB space. We formulate the estimation task as a contrastive learning task to avoid the background noise interference and learn highly discriminative features of motion and appearances. By jointly solving three proxy tasks, the model can maintain robustness to anomalies of various scales." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 615, + 288, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 288, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 288, + 674 + ], + "type": "text", + "content": "We conduct experiments on four challenging datasets (Avenue [25], ShanghaiTech [27], UCF-Crime [36] and Campus [2]). The experiments show that our proposed method outperforms the State-Of-The-Art (SOTA) method, especially in scenes with small-scale anomalies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 676, + 258, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 676, + 258, + 687 + ], + "spans": [ + { + "bbox": [ + 59, + 676, + 258, + 687 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "- Three straightforward yet effective multi-grained spatiotemporal representation proxy tasks are designed, en" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 95 + ], + "type": "text", + "content": "abling the video anomaly detection model to maintain robustness to multi-scale anomalies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 96, + 547, + 180 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 96, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 547, + 144 + ], + "type": "text", + "content": "- A reconstruction scheme for frames based on contrastive learning is proposed, motivating the model to learn highly discriminative features from both appearance and motion, rather than simple RGB features and background noise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 144, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 547, + 180 + ], + "type": "text", + "content": "- Our method achieves state-of-the-art performance on four datasets, especially outperforming previous methods by a large margin in scenes with small-scale anomalies." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 192, + 392, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 192, + 392, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 392, + 204 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 212, + 448, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 212, + 448, + 224 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 448, + 224 + ], + "type": "text", + "content": "2.1. Video Anomaly Detection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 231, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 547, + 495 + ], + "type": "text", + "content": "Our work focuses on the video anomaly detection method that only learning from normal data, which is mainly grouped into the reconstruction-based methods [4, 16, 28, 33] and the prediction-based methods [20, 41, 48]. These two methods typically use autoencoders (AEs) [15, 51], memory-augmented AEs [13, 14, 23, 28, 31], or generative models [20, 32, 41] to reconstruct current frames or predict future frames so that frames with large reconstruction or prediction errors are recognized as anomalous. While some work attempts to introduce optical flow [23, 30], skeletal information [7, 10, 39], or utilize bi-directional prediction [5, 8, 18, 52] to learn motion and temporal features, these methods cannot be effective for small-scale anomalies due to disturbance from background noise. Recent object-centric methods [12, 17, 39] attempt to separate background and objects to avoid the disturbance of background noise. Although such approach can focus on the foreground objects, the way these methods model the object in terms of classification or RGB prediction lacks the exploration of fine-grained features. In addition, some efforts [1, 24] to synthesize virtual anomaly data fail to work for such anomalies owing to the lack of small-scale virtual data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 495, + 547, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 543 + ], + "type": "text", + "content": "Different from these methods, we expect to enable the model to learn different scales of variation and fine-grained features of the normal pattern via self-supervised spatiotemporal representation learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 551, + 485, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 485, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 485, + 563 + ], + "type": "text", + "content": "2.2. Self-Supervised Learning in VAD" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "Self-supervised learning VAD methods typically perform contrastive learning or solve pretext tasks. For instance, Wang et al. [42] performs contrastive learning with clustered attention mechanism. However, it requires customized data enhancement strategies. For pretext tasks, Yang et al. [46] learn spatio-temporal representations by keyframe-based event restoration, while the work [11] designs multiple proxy tasks in an object-centric way to detect anomalies in different aspects. However, the former [46] suffers from severe background noise interference due to reconstructing events in RGB space, while the latter [11] lacks the ability to learn fine-grained features as its tasks are primarily" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17386" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 539, + 289 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 539, + 289 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 539, + 289 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 539, + 289 + ], + "type": "image", + "image_path": "de84bc4cb7564ce77f45c44a7cad59e9c75ec4148da1917b691eb0db2efc5e23.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "type": "text", + "content": "Figure 2. Method overview. The overall framework consists of a shared backbone network and three branches. The three branches perform continuity judgment, discontinuity localization, and missing frame estimation tasks to learn the spatio-temporal features of normal patterns at coarse- and fine-grained levels, respectively. The feature representation " + }, + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "type": "text", + "content": " generated after self-supervised training can effectively detect anomalies of various scales. Blue arrows and grey arrows represent training and inference processes, respectively. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 366, + 136, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 366, + 136, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 136, + 378 + ], + "type": "text", + "content": "binary classifications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 385, + 287, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 457 + ], + "type": "text", + "content": "In this work, we design proxy tasks based on video continuity [19] to learn spatio-temporal features of normal events in a multi-grained manner to perceive anomalies of various scales. In addition, we formulate the reconstruction of frames as a contrastive learning task to learning highly discriminative features instead of background noise." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 484, + 170, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 170, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 170, + 498 + ], + "type": "text", + "content": "2.3. Multi-Scale Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "content": "Multi-scale learning typically considers information from different scales or granularities. Similar to our work, Xiao et al. [44] hierarchically learns the spatio-temporal features through event mask prediction task. However, event-level features are not sufficient for detecting subtle variances between frames. There is also some multi-scale works [41, 52] in VAD. For instance, Wang et al. [41] obtain image features of different resolutions to perform prediction through different pooling operations, while Zhong et al. [52] obtain features at different scales through feature pyramid. Although pooling and feature pyramids can provide feature representations of different resolutions, they cannot specifically capture the motion and appearance features which is crucial for anomaly detection. In contrast, we design the self-supervised proxy tasks to direct the model's attention towards changes in appearance and motion between frames, enabling it to learn features in a multi-grained manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 365, + 361, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 361, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 361, + 376 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 388, + 373, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 373, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 373, + 399 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "text", + "content": "Our proposed multi-scale video anomaly detection method attempts to learn spatio-temporal features of normal patterns at both coarse-grained and fine-grained levels by jointly solving multiple proxy tasks of spatio-temporal representation learning. For a normal video clip " + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "text", + "content": ", our approach is devoted to learning a video encoder " + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "text", + "content": " that transforms the video clip " + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "text", + "content": " into the scale-aware high-quality spatio-temporal feature " + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "text", + "content": " as the high-quality feature in terms of anomalous information of different scales if it is able to easily address the following questions: (1) Whether there are missing frames in video clips " + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "text", + "content": ", i.e., whether feature representation " + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "text", + "content": " remains sensitive to large-scale anomalies. (2) If the video clip " + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "text", + "content": " is not continuous, where are the missing frames, i.e., whether feature representation " + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "text", + "content": " remains sensitive to small-scale anomalies. (3) If video clip " + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "text", + "content": " is discontinuous, what are missing frames, i.e., whether the feature representation " + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 411, + 545, + 625 + ], + "type": "text", + "content": " contains high-level motion feature and contextual feature." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "The overall architecture of our proposed method is illustrated in Fig. 2. The network consists of three branches and a video encoder " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ", all of the branches share the same backbone encoder " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ". Given two nonoverlapping video clips: continuous clip " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "V_{c}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ", discontinuous clip " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "V_{d}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ", and the corresponding missing frame " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ", each of the three branches solves a proxy task and operates together to optimize the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17387" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "backbone. The first branch is used to discriminate whether a video clip is continuous or not. The second branch is used to locate the position of the missing frame in the discontinuous clip " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "V_{d}" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": ". The last branch is designed to learn high-level features through contrastive learning mechanism, so that the feature representation of the discontinuous clip " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "V_{d}" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " can effectively approach the feature representation of the missing frame " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": ". After joint optimization of self-supervised tasks, the backbone " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " can generate scale-aware feature representations " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " for anomaly inference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 200, + 149, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 200, + 149, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 149, + 213 + ], + "type": "text", + "content": "3.2. Data Acquisition" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "text", + "content": "To perform the proposed scale-aware proxy tasks, we select continuous video clips and discontinuous clips containing only one missing frame " + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "text", + "content": ", respectively. The continuous clip contains " + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "text", + "content": " continuous frames, while the discontinuous clip has a length of " + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "inline_equation", + "content": "T - 1" + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "text", + "content": ". Given an initial continuous clip of length " + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "text", + "content": ", we uniformly select a frame from the interval " + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "inline_equation", + "content": "[1, T - 2]" + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "text", + "content": " as the missing frame " + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "text", + "content": " to form the breakpoint (counting from 0). And the leading and alternate of the missing frames are concatenated together to form discontinuous video clip " + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "inline_equation", + "content": "V_{d}" + }, + { + "bbox": [ + 46, + 218, + 287, + 361 + ], + "type": "text", + "content": ". The restriction of missing frame inside the clip sequence ensures that the construction of missing frames and discontinuous clips remains consistent." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "text", + "content": "As shown in Fig. 2, we formulate the continuity discrimination task as a binary classification task while the missing frame localization task is formulated as a " + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "inline_equation", + "content": "T - 2" + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "text", + "content": " class classification task labeled by " + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "text", + "content": ". The estimation task is formulated as a contrastive learning task. Both initial and continuous clips are formed by randomly sampling non-overlapping clips of length " + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "text", + "content": " from the same video." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 453, + 173, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 173, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 173, + 464 + ], + "type": "text", + "content": "3.3. Network Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": "We employ a video encoder (defaults a 3D-ConvNet) as the backbone network " + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": " and use it as a shared base for three proxy tasks. Each task introduces into its top a projection network, denoted " + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "inline_equation", + "content": "P_{1}, P_{2}" + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "inline_equation", + "content": "P_{3}" + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": ", which are used to process the video clip embeddings " + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": " extracted from the backbone network. All projection heads contain spatio-temporal averaging pooling layers to ensure that the dimensionality of the deep feature embeddings remains consistent, even though the temporal lengths may be different." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "text", + "content": "In continuity judgment and missing frame localization, we use fully connected layers as classifiers at the end of " + }, + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "text", + "content": " for performing classification operations. As for the missing frame estimation task, we use " + }, + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "inline_equation", + "content": "P_{3}" + }, + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "text", + "content": " to embed the video features " + }, + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "text", + "content": " extracted by the backbone network into the low-dimensional features to get the representation, i.e., the feature representation of the missing frame." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 671, + 220, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 220, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 220, + 685 + ], + "type": "text", + "content": "3.4. Proxy Tasks and Joint Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Task 1: Continuity Judgment. The continuity judgment is intended to help the model learn spatio-temporal features" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "and motion patterns in the video in a coarse-grained manner to perceive large-scale anomalies. In this task, in order for the model to learn by video features rather than other cues such as context, the positive and negative samples in each batch in our training come from the same video scene. With N samples in a training batch, cross-entropy loss " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " is used to optimize the model, and the loss for continuity judgment is " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 177, + 545, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 177, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 315, + 177, + 545, + 220 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {1} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\mathcal {L} _ {C E} \\left(P _ {1} \\left(F \\left(v _ {d} ^ {i}\\right)\\right) _ {1} + P _ {1} \\left(F \\left(v _ {c} ^ {i}\\right)\\right) _ {0}\\right)\\right), \\tag {1}", + "image_path": "e90aa483486514b4976bc99f8849909a78c2bb6e7e917e13d503a9dc8a5b8465.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "content": " represents the process of feature processing by the network, " + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "content": " represent the shared backbone network and the continuity judgment projection head, respectively, and " + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "inline_equation", + "content": "v_{c}^{i}" + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "inline_equation", + "content": "v_{d}^{i}" + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "content": " represent the " + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "content": " continuous and discontinuous video clip, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "text", + "content": "Task 2: Discontinuity Localization. While the continuity judgment task only learns spatio-temporal features in a coarse-grained manner, the discontinuous localization task expects to learn in a fine-grained manner in order to perceive small-scale anomalies. Compared to binary labels, labeling with missing frames " + }, + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "text", + "content": " in this task can drive the model to learn changes in motion between adjacent frames in a fine-grained manner. Similarly, the discontinuous localization loss " + }, + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "text", + "content": " can be expressed as an average of the cross-entropy loss of " + }, + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "text", + "content": " samples in each batch:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 349, + 408, + 545, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 408, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 349, + 408, + 545, + 441 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {C E} \\left(P _ {2} \\left(F \\left(v _ {d} ^ {i}\\right)\\right) _ {I _ {i}}\\right), \\tag {2}", + "image_path": "5aaf20158a1d8e37196eb00f7edc8e40c82142e2e8af9c0e888834c4e3ed3fe9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 450, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 450, + 545, + 486 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 304, + 450, + 545, + 486 + ], + "type": "text", + "content": " represents the discontinuity localization projection head and " + }, + { + "bbox": [ + 304, + 450, + 545, + 486 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 304, + 450, + 545, + 486 + ], + "type": "text", + "content": " represents the missing frame position corresponding to discontinuous clip " + }, + { + "bbox": [ + 304, + 450, + 545, + 486 + ], + "type": "inline_equation", + "content": "v_{d}^{i}" + }, + { + "bbox": [ + 304, + 450, + 545, + 486 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 486, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 652 + ], + "type": "text", + "content": "Task 3: Missing Frame Estimation. In order to further learn the highly discriminative spatio-temporal features of the video, we estimate missing frames in terms of both appearance and motion. Meanwhile, we choose to estimate frames in feature space rather than RGB space to avoid background noise interference. We design a contrastive learning mechanism to perform this estimation task based on the following facts. First, since adjacent frames and missing frames are temporally connected, adjacent frames and missing frames contain more similar motions compared to non-adjacent frames from the same scene but farther away. Second, video frames from the same scene share more similar appearance features with missing frames than video frames from other scenes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Accordingly, we learn the motion information in the video by triplet loss [34] taking the discontinuous video clips " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "v_{d}^{i}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " as the anchor point, the missing frames " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " as the positive samples, and the continuous video clips " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "v_{c}^{i}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " as the negative samples. In addition, we take discontinuous clips" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17388" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "v_{d}^{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "text", + "content": " as anchor points, continuous clips " + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "v_{c}^{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "text", + "content": " from the same scene as positive samples, and video clips " + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "\\{v^{j}\\}_{i\\neq j}" + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "text", + "content": " from different scenes as negative samples, to learn the appearance information in the video by contrastive loss [19]. Defining the cosine similarity computation operation as " + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "sim(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "text", + "content": ", and the hyperparameter that balances the relative contributions of the triplet loss and the contrastive loss as " + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "\\omega \\in [0,1]" + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "text", + "content": ", the missing frame estimation loss is denoted as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 178, + 287, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 178, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 63, + 178, + 287, + 247 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {3} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} (\\omega \\times \\max (0, \\gamma - \\left(p _ {i} ^ {+} - p _ {i} ^ {-}\\right)) \\tag {3} \\\\ - (1 - \\omega) \\log \\left(\\frac {q _ {i} ^ {+}}{q _ {i} ^ {+} + \\sum_ {j = 1 , j \\neq i} ^ {N} q _ {i , j ^ {-}}}\\right)), \\\\ \\end{array}", + "image_path": "60d8bcc57ac41f74c71b3bcb406edbe77179f5f9503e1ba730ab32638d2e2a6d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 260, + 287, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 260, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 116, + 260, + 287, + 277 + ], + "type": "interline_equation", + "content": "p _ {i} ^ {+} = \\mathrm {s i m} \\left(P _ {3} (V _ {d} ^ {i}, I _ {i})\\right), \\tag {4}", + "image_path": "8c1e84f0b38e03485e98d56172e2d2ccb0fe4962f53c17ddf42ee7ac582bcf57.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 293, + 287, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 293, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 114, + 293, + 287, + 308 + ], + "type": "interline_equation", + "content": "p _ {i} ^ {-} = \\operatorname {s i m} \\left(P _ {3} \\left(V _ {d} ^ {i}, V _ {c} ^ {i}\\right)\\right), \\tag {5}", + "image_path": "757c7d8317206a5f60b5adce05672be3ef9fa252a1abbe8ece5a3172036ed09d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 95, + 326, + 287, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 326, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 95, + 326, + 287, + 342 + ], + "type": "interline_equation", + "content": "q _ {i} ^ {+} = \\exp \\left(\\operatorname {s i m} \\left(P _ {3} \\left(V _ {d} ^ {i}, V _ {c} ^ {i}\\right)\\right) / \\tau\\right), \\tag {6}", + "image_path": "a725fde79bf0eb9e12d3956c8e09d3aed108f116c6383d71836f9a05b31ed311.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 356, + 287, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 356, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 91, + 356, + 287, + 394 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} q _ {i, j} ^ {-} = \\exp \\left(\\sin \\left(P _ {3} \\left(v _ {d} ^ {i}, v _ {d} ^ {j}\\right)\\right) / \\tau\\right) \\tag {7} \\\\ + \\exp \\left(\\mathrm {s i m} \\left(P _ {3} (v _ {d} ^ {i}, v _ {c} ^ {j})\\right) / \\tau\\right), \\\\ \\end{array}", + "image_path": "b98b09331182e11ed7f471afa8ec340424a860fb1a4d6c6a22b88fde3b783cb4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "inline_equation", + "content": "p_i^+" + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "inline_equation", + "content": "p_i^-" + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "text", + "content": " denote the similarity between positive sample pairs and negative sample pairs in the triplet loss, respectively, " + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "inline_equation", + "content": "q_i^+" + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "text", + "content": " denotes a single positive sample pair in the contrastive loss, " + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "inline_equation", + "content": "q_{i,j}^{-}" + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "text", + "content": " denotes the similarity of negative sample pairs in the contrastive loss, and " + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "text", + "content": " is the temperature factor in the contrastive loss. Increasing " + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "text", + "content": " allows for more focus on motion information, while decreasing " + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "text", + "content": " allows for more focus on appearance information. Overall, the model is incentivized to learn fine-grained motion changes and contextual features, including background information and object appearance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 532, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 579 + ], + "type": "text", + "content": "Joint Optimization. The above three proxy tasks learn spatio-temporal features in videos from different levels, and we jointly optimize our shared video encoder " + }, + { + "bbox": [ + 46, + 532, + 287, + 579 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 532, + 287, + 579 + ], + "type": "text", + "content": " with the multi-loss function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 590, + 287, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 590, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 104, + 590, + 287, + 604 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {F} = w _ {1} \\mathcal {L} _ {1} + w _ {2} \\mathcal {L} _ {2} + w _ {3} \\mathcal {L} _ {3}, \\tag {8}", + "image_path": "cddce7f7b005f0a81d3234ef8cfd1518bf43bb9b92996b8b17aa7ae8c7971722.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 613, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 287, + 662 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 613, + 287, + 662 + ], + "type": "inline_equation", + "content": "w_{1}, w_{2}, w_{3} \\in [0,1]" + }, + { + "bbox": [ + 47, + 613, + 287, + 662 + ], + "type": "text", + "content": " are the individual weights on the " + }, + { + "bbox": [ + 47, + 613, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1, \\mathcal{L}_2" + }, + { + "bbox": [ + 47, + 613, + 287, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 613, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_3" + }, + { + "bbox": [ + 47, + 613, + 287, + 662 + ], + "type": "text", + "content": ". Joint optimization helps the model to learn video spatio-temporal features from both coarse- and fine-grained scales to perceive anomalies of various scales." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 670, + 139, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 139, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 139, + 682 + ], + "type": "text", + "content": "3.5. VAD Inference" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "In the inference, we use " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " frames as the processing unit. Following the setup of the SOTA method [20, 41, 46], we" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "choose not to compute anomaly scores using the error between each estimated frame " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\hat{I}_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "1 \\leq t \\leq T" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ") and the corresponding ground truth " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". Instead, we opt to employ the Peak Signal-to-Noise Ratio (PSNR) corresponding to the frame with the largest Mean-Square Error (MSE) among the " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " frames and the ground truth. This approach serves as the detection metric for the video sequence, as follow:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 162, + 545, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 545, + 192 + ], + "type": "interline_equation", + "content": "\\operatorname {M S E} \\left(\\boldsymbol {I} _ {t}, \\hat {\\boldsymbol {I}} _ {t}\\right) = \\frac {1}{H \\cdot W} \\sum_ {i, j} \\left\\| \\boldsymbol {I} _ {t} (i, j) - \\hat {\\boldsymbol {I}} _ {t} (i, j) \\right\\| _ {2} ^ {2}, \\tag {9}", + "image_path": "70ec25751a7b43974d8866aba2d70be61b086c1e74ba6c0edc320b48a19c06ae.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 351, + 198, + 545, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 198, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 351, + 198, + 545, + 235 + ], + "type": "interline_equation", + "content": "P S N R _ {t} = 1 0 \\log_ {1 0} \\frac {\\operatorname {M A X} _ {\\boldsymbol {I} _ {t}} ^ {2}}{\\operatorname {M S E} \\left(\\boldsymbol {I} _ {t} , \\hat {\\boldsymbol {I}} _ {t}\\right)}, \\tag {10}", + "image_path": "19609a3d93df622f7f098023552fd6b9e61446272187c3622c55e7e9773059a4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "text", + "content": " are the height and width of the frames, respectively, " + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "inline_equation", + "content": "I_{t}(i,j) / \\hat{I}_{t}(i,j)" + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "text", + "content": " represents the RGB value of the " + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "text", + "content": "-th pixel in " + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "inline_equation", + "content": "I_{t} / \\hat{I}_{t}" + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{MAX}_{I_t}" + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "text", + "content": " represents the maximum value of the image pixel in " + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "text", + "content": ". Finally, we calculate the anomaly score " + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "text", + "content": "-th frame " + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "text", + "content": " by normalizing the PSNR values over all " + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 306, + 236, + 545, + 308 + ], + "type": "text", + "content": " frames:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 323, + 312, + 545, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 312, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 323, + 312, + 545, + 338 + ], + "type": "interline_equation", + "content": "S _ {t} = 1 - \\frac {P S N R _ {t} - \\operatorname* {m i n} _ {t ^ {\\prime}} P S N R _ {t ^ {\\prime}}}{\\operatorname* {m a x} _ {t ^ {\\prime}} P S N R _ {t ^ {\\prime}} - \\operatorname* {m i n} _ {t ^ {\\prime}} P S N R _ {t ^ {\\prime}}}, \\tag {11}", + "image_path": "0ef4fd8cddb0d312a4d679cffdf42cdb3c637de20815e3bba5de963b0f979ca7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 342, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 545, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 342, + 545, + 366 + ], + "type": "inline_equation", + "content": "1 \\leq t' \\leq T" + }, + { + "bbox": [ + 305, + 342, + 545, + 366 + ], + "type": "text", + "content": ". A higher score " + }, + { + "bbox": [ + 305, + 342, + 545, + 366 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 305, + 342, + 545, + 366 + ], + "type": "text", + "content": " indicates that the frame " + }, + { + "bbox": [ + 305, + 342, + 545, + 366 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 305, + 342, + 545, + 366 + ], + "type": "text", + "content": " is more likely to be anomalous." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 376, + 382, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 382, + 390 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 382, + 390 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 396, + 476, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 476, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 476, + 407 + ], + "type": "text", + "content": "4.1. Dataset and Evaluation Metrics" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 414, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 617 + ], + "type": "text", + "content": "Cross-Scene Anomaly Dataset Avenue [25] contains 16 training videos and 21 test videos with 47 abnormal events, including running and throwing. The scenes in this dataset are single and the anomalies are mainly large-scale. ShanghaiTech [27] has 13 scenes with complex lighting conditions and different perspectives. In addition, the dataset includes anomalies caused by sudden movements, such as chasing and arguing. The different perspectives and the unfixed position of the camera lead to a large variation of both the object scale and the anomaly scale in the scene. This is a significant challenge to the robustness of anomaly detection methods to scale variations. UCF-Crime [36] comprises 13 anomaly types, spanning a total of 128 hours of video footage. This dataset is complex because of the unconstrained backgrounds and anomalies of various scales. The training set contains video-level annotations, whereas the testing set includes frame-level annotations." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": "Scene-Dependent Anomaly Dataset. Campus [2] is currently the most challenging dataset in its field with 43 scenes, 28 classes of anomalous events and 16 hours of videos. Especially, it contains scene-dependent anomalies, which means an normal event may be abnormal in another scene. Detecting scene-dependent anomalies requires the model to understand the scene and learn highly discriminative features rather than overfit." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17389" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 284, + 225 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 284, + 225 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 284, + 225 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 284, + 225 + ], + "type": "table", + "html": "
MethodLength TShanghaiTechCampus
AUCAUC
MicroMacroMicroMacro
Continuity Judgment878.182.360.163.2
1277.281.160.363.9
1675.179.258.260.9
Discontinuity Localization882.685.562.165.4
1283.487.262.665.1
1682.185.561.965.0
Missing Frame Estimation879.883.263.566.8
1281.284.965.268.9
1680.584.064.968.0
Overall Framework884.288.866.369.2
1285.189.867.370.9
1683.988.366.069.2
", + "image_path": "482eaa015823174570d0fcac6ab38d360ee8e07ad3d27e057e1b6dd40f98498f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 291, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 287, + 350 + ], + "type": "text", + "content": "Evaluation Metrics The area under the ROC curve (AUC) serves as a commonly used metric for evaluation and comparison. A higher AUC score indicates a better anomaly detection capability. Following previous research [12], we evaluate both the Micro and Macro versions of AUC." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 358, + 180, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 180, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 180, + 371 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "text", + "content": "We train and evaluate our method with an NVIDIA RTX 3090 GPU. In the training phase, we resize the resolution of all input video clips to " + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "text", + "content": " pixels, while the values of the pixels in all frames are normalized to the range [0, 1]. For the pre-training of the three proxy tasks, we utilize AdamW as the optimizer while the length of the continuous video clips is set to " + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "inline_equation", + "content": "T = 12" + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "text", + "content": " frames. The initial learning rate is set to 0.0003 and is gradually decayed following the scheme of cosine annealing. For the hyperparameters, we set " + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\omega = 0.5" + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "text", + "content": " in Equation 3, " + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "inline_equation", + "content": "w_{1} = w_{2} = w_{3} = 1" + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "text", + "content": " in Equation 8, and the temperature factor " + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "text", + "content": " in the contrastive loss. In our reported experimental performance, the shared backbone network " + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "text", + "content": " is implemented as I3D-RGB [3]. In addition, we incorporate the designed proxy tasks into the backbone of the SOTA methods [2, 11], the experimental results are also reported. Following the existing method setup [11, 24, 33], we perform the proxy task at the object-level with the multi-task backbone [11] and at the frame-level with the other backbones." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 611, + 141, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 141, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 141, + 624 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": "Considering the ShanghaiTech and Campus datasets contain diverse perspectives and scenes, and large variations in scales of anomaly, we conduct exhaustive ablation experiments on the ShanghaiTech and Campus datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Sensitivity to the Length of Video Clips. The length " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " of the input clips is an important setting in our proposed framework. With a small " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": ", both continuity judgments and dis" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 543, + 194 + ], + "blocks": [ + { + "bbox": [ + 46, + 235, + 287, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 235, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 287, + 278 + ], + "type": "text", + "content": "Table 1. Comparison of proposed framework and individual proxy tasks between different lengths " + }, + { + "bbox": [ + 46, + 235, + 287, + 278 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 235, + 287, + 278 + ], + "type": "text", + "content": " of input clips. We report Micro and Macro AUC (%) on ShanghaiTech and Campus datasets. The best performing results are marked in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 543, + 194 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 543, + 194 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 543, + 194 + ], + "type": "table", + "html": "
IDCJDLMFEFPShanghaiTechCampus
AUCAUC
MicroMacroMicroMacro
1---✓70.271.258.158.6
2✓---77.281.160.363.9
3-✓--83.487.262.665.1
4--✓-81.284.965.268.9
5✓✓--84.187.862.865.4
6✓-✓-82.385.665.369.2
7-✓✓-84.589.566.270.1
8✓✓✓-85.189.867.370.9
9✓✓-✓84.287.662.865.6
", + "image_path": "eb6e57fab9fe008ef0843f1591e97f70a2137fcf64ddfbef0a9547a3548241a0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 202, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 269 + ], + "type": "text", + "content": "Table 2. Ablation experiments on the contributions of each proxy task. We report the AUC (\\%) scores on ShanghaiTech and Campus datasets. 'CJ', 'DL' and 'MFE' stand for the three proxy tasks of continuity judgment, discontinuity localization and missing frame estimation, respectively. In addition we add experiments on frame prediction for comparison, 'FP' represents frame prediction." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 281, + 545, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 401 + ], + "type": "text", + "content": "continuity localization will be so ambiguous as to be difficult. While with a larger " + }, + { + "bbox": [ + 304, + 281, + 545, + 401 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 281, + 545, + 401 + ], + "type": "text", + "content": ", the model may not need to learn high-quality spatio-temporal features to solve the task due to enough spatio-temporal information in the feed. Table 1 shows the sensitivity of each proxy task to " + }, + { + "bbox": [ + 304, + 281, + 545, + 401 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 281, + 545, + 401 + ], + "type": "text", + "content": ". For the discontinuous localization and missing frame estimation tasks, " + }, + { + "bbox": [ + 304, + 281, + 545, + 401 + ], + "type": "inline_equation", + "content": "T = 12" + }, + { + "bbox": [ + 304, + 281, + 545, + 401 + ], + "type": "text", + "content": " presents better performance. Although at " + }, + { + "bbox": [ + 304, + 281, + 545, + 401 + ], + "type": "inline_equation", + "content": "T = 8" + }, + { + "bbox": [ + 304, + 281, + 545, + 401 + ], + "type": "text", + "content": ", the continuity judgment yields competitive results on the ShanghaiTech dataset, the overall framework achieves the best outcomes with jointly optimized at " + }, + { + "bbox": [ + 304, + 281, + 545, + 401 + ], + "type": "inline_equation", + "content": "T = 12" + }, + { + "bbox": [ + 304, + 281, + 545, + 401 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 402, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 510 + ], + "type": "text", + "content": "Effect of Continuity Judgment. The results of the ablation experiments for each proxy task are shown in Table 2, where we take the frame prediction [20] performed on the same backbone " + }, + { + "bbox": [ + 304, + 402, + 545, + 510 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 402, + 545, + 510 + ], + "type": "text", + "content": " as the baseline and train it with gradient loss and intensity loss (ID 1). Performing the continuity judgment task alone improves the performance of the model compared to frame prediction (ID 2). This suggests that continuity judgment task motivates the model to generate higher quality representations than frame prediction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": "Effect of Discontinuity Localization. Compared to other proxy tasks and baseline, the performance of the discontinuous localization task alone can be significantly improved, especially on ShanghaiTech which contains anomalies of various scales (ID 3, ID 5 and ID 7). As this task samples discontinuous positions uniformly along the time axis, it motivates the model to capture subtle variations between frames, obtaining fine-grained representations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Effect of Missing Frame Estimation. Performing missing frame estimation alone is effective in improving performance on Campus compared to baseline and other proxy tasks, which contains scene-dependent anomalies (ID 4). This improvement illustrates that our proposed contrastive learning scheme can help the model understand the scene. Compared to performing the frame prediction task to learn RGB features (ID 1 and ID 9), our proposed scheme learns highly discriminative motion features and semantically ef" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17390" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 69, + 284, + 466 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 284, + 466 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 284, + 466 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 284, + 466 + ], + "type": "table", + "html": "
yearMethodAvenueShanghaiTech
AUCAUC
MicroMacroMicroMacro
2018Liu et al. [20]85.1-72.8-
Liu et al. [22]84.4---
Sultani et al. [36]--76.5-
2019Gong et al. [14]83.8-71.2-
Lee et al. [18]90.0-76.2-
Ionescu et al. [17]87.490.478.784.9
2020Park et al. [31]88.5-70.5-
Sun et al. [37]89.6-74.7-
Lu et al. [26]85.8-75.9-
Wang et al. [42]87.0-79.3-
Yu et al. [47]89.6-74.8-
2021Liu et al. [23]91.1-76.2-
Lv et al. [28]89.5-73.8-
Georgescu et al. [11]*91.591.982.489.3
Georgescu et al. [12]*92.390.482.789.3
2022Wang et al. [41]88.3-76.6-
Zaheer et al. [49]74.2-79.6-
Chen et al. [5]90.3-78.1-
Zhong et al. [52]89.0-74.5-
Cho et al. [6]88.0-76.3-
Yang et al. [45]89.9-74.7-
Ristea et al. [33]*92.991.983.689.5
Acsintoae et al. [1]▼*93.093.283.790.5
2023Yang et al. [46]89.9-73.8-
Cao et al. [2]86.8-79.2-
Liu et al. [21]92.8-78.8-
Singh et al. [35]86.0-76.6-
Sun et al. [39]92.4-83.0-
Sun et al. [38]91.5-78.6-
Liu et al. [24]▼91.892.383.887.8
Liu et al. [24] ▼*93.693.985.091.4
Ours92.492.985.189.8
Ours▼93.292.586.291.0
Ours*(Estimation Task)92.693.085.591.8
Ours*(Three Tasks)93.693.886.892.4
Ours▼*(Three Tasks)94.394.587.593.0
", + "image_path": "fa70421f712d3c5b65d754d1c03eb722cee419da0f9105a408951e6e2e7334cc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 473, + 287, + 506 + ], + "lines": [ + { + "bbox": [ + 47, + 473, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 473, + 287, + 506 + ], + "type": "text", + "content": "Table 3. Comparison with SOTA methods of the Micro and Macro AUC (%) on Avenue and ShanghaiTech datasets. The best performing results are marked in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 49, + 508, + 212, + 517 + ], + "lines": [ + { + "bbox": [ + 49, + 508, + 212, + 517 + ], + "spans": [ + { + "bbox": [ + 49, + 508, + 212, + 517 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 49, + 508, + 212, + 517 + ], + "type": "text", + "content": " : Methods apply virtual dataset for training." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 49, + 518, + 279, + 529 + ], + "lines": [ + { + "bbox": [ + 49, + 518, + 279, + 529 + ], + "spans": [ + { + "bbox": [ + 49, + 518, + 279, + 529 + ], + "type": "text", + "content": "*: Methods utilize multi-task model ([11] or [12]) as backbone." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 543, + 285, + 608 + ], + "blocks": [ + { + "bbox": [ + 50, + 543, + 285, + 608 + ], + "lines": [ + { + "bbox": [ + 50, + 543, + 285, + 608 + ], + "spans": [ + { + "bbox": [ + 50, + 543, + 285, + 608 + ], + "type": "table", + "html": "
MethodReferenceMicro AUCMacro AUC
Park et al. [31]CVPR2068.972.4
Georgescu et al. [11]CVPR2174.678.2
Wang et al. [41]TNNLS2272.976.8
Sun et al. [39]CVPR2375.578.3
Ours-80.683.9
", + "image_path": "5d3546b814017dd527849bde0af4a868fe26cd7c170d5eaaed047b97f2eac24d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 616, + 287, + 640 + ], + "lines": [ + { + "bbox": [ + 47, + 616, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 287, + 640 + ], + "type": "text", + "content": "Table 4. Results of Micro and Macro AUC(%) on UCF-Crime dataset. The best performing results are marked in bold." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 653, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 677 + ], + "type": "text", + "content": "fective contexts, including the appearance of objects and background (ID 4 and ID 8)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Effect of Joint Optimization. Although a single task may be able to achieve competitive performance, Experiments have shown that when proxy tasks are executed in pairs (ID" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 69, + 545, + 191 + ], + "blocks": [ + { + "bbox": [ + 309, + 69, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 309, + 69, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 545, + 191 + ], + "type": "table", + "html": "
MethodReferenceMiceo AUCMacro AUC
Liu et al. [20]CVPR1857.960.2
Gong et al. [14]CVPR1961.962.5
Ionescu et al. [17]CVPR1959.363.4
Park et al. [31]CVPR2062.563.6
Liu et al. [23]ICCV2163.7-
Lv et al. [28]CVPR2164.4-
Wang et al. [41]TNNLS2261.964.2
Cao et al. [2]△CVPR2368.2-
Ours-67.370.9
Ours△-70.172.2
", + "image_path": "58e128a9a2fed2772ffea3a9c3f635cd0bfd35f30008dd83153f851790cf81d9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 198, + 545, + 231 + ], + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 231 + ], + "type": "text", + "content": "Table 5. Comparison with SOTA methods of the Micro and Macro AUC(%) on Campus dataset. The best performing results are marked in bold." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 232, + 539, + 243 + ], + "lines": [ + { + "bbox": [ + 307, + 232, + 539, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 539, + 243 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 307, + 232, + 539, + 243 + ], + "type": "text", + "content": " :Version that utilize scene-conditioned model [2] as backbone." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 265, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 265, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 305, + 265, + 545, + 337 + ], + "type": "text", + "content": "5, ID 6 and ID 7) and all tasks are executed together (ID 8), the performance of the model continues to improve to an optimum. This joint improvement is attributed to the fact that different proxy tasks learn spatio-temporal features at different levels, and they complement each other in motion and contextual feature learning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 346, + 503, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 503, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 503, + 358 + ], + "type": "text", + "content": "4.4. Comparisons with State-Of-The-Arts" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 365, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 568 + ], + "type": "text", + "content": "We compare the proposed framework with SOTA methods in terms of Micro and Macro AUC(\\%). It is noteworthy that the current state-of-the-art methods [1, 24, 33] utilize the multi-task framework proposed by Georgescu et al. ([11] and [12], respectively) as the backbone, where [1] and [24] employ virtual data for training. Therefore, we evaluate different configurations of the proposed method on the Avenue and ShanghaiTech datasets, as shown in Table 3. With the multi-task backbone [11] (marked with * in Table 3), we assess two versions: one involves replacing only the prediction component with our designed missing estimation task, while keeping the other proxy tasks unchanged. The other version replaces both proxy tasks in the motion branch with continuity judgment and discontinuity localization tasks in addition to incorporating the missing estimation task. In addition, we report the performance of applying virtual data for training (marked with " + }, + { + "bbox": [ + 304, + 365, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 304, + 365, + 545, + 568 + ], + "type": "text", + "content": " in Table 3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "type": "text", + "content": "Results on Avenue. As shown in Table 3, our method achieved the highest AUC scores, obtaining a Micro AUC of " + }, + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "type": "inline_equation", + "content": "94.3\\%" + }, + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "type": "text", + "content": " and a Macro AUC of " + }, + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "type": "inline_equation", + "content": "94.5\\%" + }, + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "type": "text", + "content": ". Without the backbone of multi-task [11, 12] and virtual data [24], our method still scored the best performance with " + }, + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "type": "inline_equation", + "content": "92.9\\%" + }, + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "type": "text", + "content": " Macro AUC." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Results on ShanghaiTech. Our proposed method yields a Micro AUC of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "87.5\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and a Macro AUC of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "93.0\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " on the ShanghaiTech dataset. With the backbone of multi-task [11, 12], we improve the Micro AUC by " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " compared to Georgescu et al. [11] in the setup with only the missing estimation task, and by " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " when incorporating three designed proxy tasks. Without the backbone of multi-task" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17391" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 282, + 297 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 282, + 297 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 282, + 297 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 282, + 297 + ], + "type": "image", + "image_path": "a234b3d579b70549459df2c7994935fb761c60a799895e7e9098136854df9a1c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 307, + 287, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 307, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 287, + 361 + ], + "type": "text", + "content": "Figure 3. Examples of detection for anomalies of different scales from the ShanghaiTech dataset, including ground truth and prediction or reconstruction error maps from different methods (from top to bottom: ROADMAP[41], DEDDnet[52], USTN-DSC[46], and our method). Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 384, + 286, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 286, + 456 + ], + "type": "text", + "content": "[11, 12] and virtual data [24], our method still achieves a Micro AUC score of " + }, + { + "bbox": [ + 46, + 384, + 286, + 456 + ], + "type": "inline_equation", + "content": "85.1\\%" + }, + { + "bbox": [ + 46, + 384, + 286, + 456 + ], + "type": "text", + "content": ", which is comparable to the method [24] that employed the multi-task backbone and virtual training data. In comparison with the multi-scale VAD methods ([41] and [52]), our method achieves a significant improvement of " + }, + { + "bbox": [ + 46, + 384, + 286, + 456 + ], + "type": "inline_equation", + "content": "8.5\\%" + }, + { + "bbox": [ + 46, + 384, + 286, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 384, + 286, + 456 + ], + "type": "inline_equation", + "content": "10.6\\%" + }, + { + "bbox": [ + 46, + 384, + 286, + 456 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 457, + 286, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 286, + 540 + ], + "type": "text", + "content": "Results on UCF-Crime. Due to the absence of published results (methods only learning from normal data) on the UCF-Crime dataset, we implement the code from the existing literature [11, 31, 39, 41]. As shown in Table 4, our proposed method achieves a significant improvement compared to the second best method by " + }, + { + "bbox": [ + 46, + 457, + 286, + 540 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 46, + 457, + 286, + 540 + ], + "type": "text", + "content": " in terms of Micro AUC and " + }, + { + "bbox": [ + 46, + 457, + 286, + 540 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 46, + 457, + 286, + 540 + ], + "type": "text", + "content": " in terms of Macro AUC." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 541, + 286, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 286, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 286, + 661 + ], + "type": "text", + "content": "Results on Campus. We evaluate an additional version on the Campus dataset with the scene-conditioned model (provided by Cao et al. [2]) as the backbone, as it is necessary to incorporate scene features for detecting scene-dependent anomalies. As shown in Table 5, our method achieves the highest Micro AUC score of " + }, + { + "bbox": [ + 46, + 541, + 286, + 661 + ], + "type": "inline_equation", + "content": "70.1\\%" + }, + { + "bbox": [ + 46, + 541, + 286, + 661 + ], + "type": "text", + "content": " and Macro AUC score of " + }, + { + "bbox": [ + 46, + 541, + 286, + 661 + ], + "type": "inline_equation", + "content": "72.2\\%" + }, + { + "bbox": [ + 46, + 541, + 286, + 661 + ], + "type": "text", + "content": ". Without the scene-conditioned model, our method still obtain a Micro AUC of " + }, + { + "bbox": [ + 46, + 541, + 286, + 661 + ], + "type": "inline_equation", + "content": "67.3\\%" + }, + { + "bbox": [ + 46, + 541, + 286, + 661 + ], + "type": "text", + "content": ", surpassing other RGB reconstruction-based [14, 28, 31] or RGB prediction-based methods [20, 23, 41]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 159, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 159, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 159, + 682 + ], + "type": "text", + "content": "4.5. Qualitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "Fig. 1 shows the anomaly curves of the test video and the comparison of our method with PMem [31] (reconstruction-" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 312, + 71, + 383, + 156 + ], + "blocks": [ + { + "bbox": [ + 312, + 71, + 383, + 156 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 383, + 156 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 383, + 156 + ], + "type": "image", + "image_path": "531e5c2fdf059771824956ac7c8841e96d3c08419d0f70952f4c22459a4459db.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 169, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 305, + 169, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 305, + 169, + 545, + 191 + ], + "type": "text", + "content": "Figure 4. Visualization of salient regions of test frames by the proposed method. Best viewed in color." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 390, + 71, + 462, + 156 + ], + "blocks": [ + { + "bbox": [ + 390, + 71, + 462, + 156 + ], + "lines": [ + { + "bbox": [ + 390, + 71, + 462, + 156 + ], + "spans": [ + { + "bbox": [ + 390, + 71, + 462, + 156 + ], + "type": "image", + "image_path": "0689b7c8e5d1a25f72a479d59d49f68678c849016004bc9c1ffe8b3f9e18261b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 468, + 71, + 539, + 157 + ], + "blocks": [ + { + "bbox": [ + 468, + 71, + 539, + 157 + ], + "lines": [ + { + "bbox": [ + 468, + 71, + 539, + 157 + ], + "spans": [ + { + "bbox": [ + 468, + 71, + 539, + 157 + ], + "type": "image", + "image_path": "e3baa39affc129a32a1531fb8896679998a5fbe4ead68b3e85c6aebd63fe2570.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 204, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 357 + ], + "type": "text", + "content": "based method) and RoadMap [41] (prediction-based method). When there are small-scale anomalies in the sampled videos, the anomaly score curves of the other two methods have obvious missed detections, while our method can accurately perceive the small-scale anomalies. In addition, Fig. 3 illustrates more examples, including the comparison with the multi-scale VAD methods [41, 52] and the comparison with self-supervised VAD method [46]. It can be observed that for normal regions in video frames, our method is able to estimate them well, while for abnormal event regions large errors occur. While other methods generate large errors for some normal regions, especially for frames with small-scale anomalies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 360, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 360, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 360, + 545, + 407 + ], + "type": "text", + "content": "Visualizations of the salient regions of the frames by our method are illustrated in Fig. 4. For both large-scale anomalies (1-th row) and small-scale anomalies (2-th row), our model can accurately focus on the anomalous regions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 418, + 378, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 378, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 378, + 430 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 438, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 593 + ], + "type": "text", + "content": "In this paper, we observe that anomalous spatial scales affect the feature learning of the model. Small-scale anomalies require the model to learn highly discriminative features in a fine-grained manner rather than background noise. To detect multi-scale anomalies, we design proxy tasks supervised by video continuity that motivate the model to learn spatio-temporal features in both coarse-grained and fine-grained manners. Furthermore, instead of performing RGB reconstruction and prediction, we estimation frames in feature space through contrastive learning to learn highly discriminative features of appearance and motion. Experiments conducted on four challenging benchmark datasets validate the effectiveness of our proposed method." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grants (62101064, 62171057, 62201072, U23B2001, 62001054, 62071067), the National Postdoctoral Program for Innovative Talents under Grant BX20230052, China Postdoctoral Science Foundation (2023TQ0039), the Ministry of Education and China Mobile Joint Fund (MCM20200202, MCM20180101), Beijing University of Posts and Telecommunications-China Mobile Research Institute Joint Innovation Center." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17392" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. *Ubnormal: New benchmark for supervised open-set video anomaly detection.* In *CVPR*, pages 20111-20121, 2022. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "text", + "content": "[2] Congqi Cao, Yue Lu, Peng Wang, and Yanning Zhang. A new comprehensive benchmark for semi-supervised video anomaly detection and anticipation. In CVPR, pages 20392-20401, 2023. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 287, + 235 + ], + "type": "text", + "content": "[3] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, pages 4724-4733, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 269 + ], + "type": "text", + "content": "[4] Yunpeng Chang, Zhigang Tu, Wei Xie, and Junsong Yuan. Clustering driven deep autoencoder for video anomaly detection. In ECCV, pages 329-345, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "text", + "content": "[5] Chengwei Chen, Yuan Xie, Shaohui Lin, Angela Yao, Guannan Jiang, Wei Zhang, Yanyun Qu, Ruizhi Qiao, Bo Ren, and Lizhuang Ma. Comprehensive regularization in a bidirectional predictive network for video anomaly detection. In AAAI, pages 230-238, 2022. 1, 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 369 + ], + "type": "text", + "content": "[6] MyeongAh Cho, Taeoh Kim, Woo Jin Kim, Suhwan Cho, and Sangyoun Lee. Unsupervised video anomaly detection via normalizing flows with implicit latent features. Pattern Recognit., 129:108703, 2022. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "text", + "content": "[7] Romero F. A. B. de Morais, Vuong Le, Truyen Tran, Budhaditya Saha, Moussa Reda Mansour, and Svetha Venkatesh. Learning regularity in skeleton trajectories for anomaly detection in videos. In CVPR, pages 11996-12004, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 287, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 457 + ], + "type": "text", + "content": "[8] Zhiwen Fang, Jiafei Liang, Joey Tianyi Zhou, Yang Xiao, and Feng Yang. Anomaly detection with bidirectional consistency in videos. IEEE Trans. Neural Networks Learn. Syst., 33(3):1079-1092, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 459, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 287, + 491 + ], + "type": "text", + "content": "[9] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. MIST: multiple instance self-training framework for video anomaly detection. In CVPR, pages 14009-14018, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 492, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 547 + ], + "type": "text", + "content": "[10] Alessandro Flaborea, Luca Collorone, Guido Maria D'Amely di Melendugno, Stefano D'Arrigo, Bardh Prenkaj, and Fabio Galasso. Multimodal motion conditioned diffusion model for skeleton-based video anomaly detection. In (ICCV), pages 10318-10329, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "type": "text", + "content": "[11] Mariana-Iuliana Georgescu, Antonio Barbalau, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. Anomaly detection in video via self-supervised and multi-task learning. In CVPR, pages 12742-12752, 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "text", + "content": "[12] Mariana-Iuliana Georgescu, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4505-4523, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[13] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In ICCV, pages 1705-1714, 2019. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[14] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In ICCV, pages 1705-1714, 2019. 2, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "type": "text", + "content": "[15] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal regularity in video sequences. In CVPR, pages 733-742, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "text", + "content": "[16] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In ICCV, pages 8771-8780, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "text", + "content": "[17] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In CVPR, pages 7842–7851, 2019. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 275, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 318 + ], + "type": "text", + "content": "[18] Sangmin Lee, Hak Gu Kim, and Yong Man Ro. BMAN: bidirectional multi-scale aggregation networks for abnormal event detection. IEEE Trans. Image Process., 29:2395-2408, 2019. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 319, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 364 + ], + "type": "text", + "content": "[19] Hanwen Liang, Niamul Quader, Zhixiang Chi, Lizhe Chen, Peng Dai, Juwei Lu, and Yang Wang. Self-supervised spatiotemporal representation learning by exploiting video continuity. In AAAI, pages 1564-1573, 2022. 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "text", + "content": "[20] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In CVPR, pages 6536-6545, 2018. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "type": "text", + "content": "[21] Wenrui Liu, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Diversity-measurable anomaly detection. In CVPR, pages 12147-12156, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 433, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 465 + ], + "type": "text", + "content": "[22] Yusha Liu, Chun-Liang Li, and Barnabás Póczos. Classifier two sample test for video anomaly detections. In BMVC, page 71. BMVA Press, 2018. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 467, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 520 + ], + "type": "text", + "content": "[23] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In ICCV, pages 13568-13577, 2021. 1, 2, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "text", + "content": "[24] Zuhao Liu, Xiao-Ming Wu, Dian Zheng, Kun-Yu Lin, and Wei-Shi Zheng. Generating anomalies for video anomaly detection with prompt-based feature mapping. In CVPR, pages 24500-24510. IEEE, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "type": "text", + "content": "[25] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 FPS in MATLAB. In ICCV, pages 2720-2727, 2013. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "text", + "content": "[26] Yiwei Lu, Frank Yu, Mahesh Kumar Krishna Reddy, and Yang Wang. Few-shot scene-adaptive anomaly detection. In ECCV, pages 125-141, 2020. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 668 + ], + "type": "text", + "content": "[27] Weixin Luo, Wen Liu, and Shenghua Gao. A revisit of sparse coding based anomaly detection in stacked RNN framework. In ICCV, pages 341-349, 2017. 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[28] Hui Lv, Chen Chen, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Learning normal dynamics in videos with meta prototype network. In CVPR, pages 15425-15434, 2021. 2, 7, 8" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17393" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In CVPR, pages 8022-8031, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 119, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 119, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 119, + 287, + 152 + ], + "type": "text", + "content": "[30] Trong-Nguyen Nguyen and Jean Meunier. Anomaly detection in video sequence with appearance-motion correspondence. In ICCV, pages 1273-1283, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 153, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 153, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 49, + 153, + 287, + 186 + ], + "type": "text", + "content": "[31] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In CVPR, pages 14360-14369, 2020. 1, 2, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 188, + 287, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 188, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 49, + 188, + 287, + 231 + ], + "type": "text", + "content": "[32] Mahdyar Ravanbakhsh, Enver Sangineto, Moin Nabi, and Nicu Sebe. Training adversarial discriminators for cross-channel abnormal event detection in crowds. In WACV, pages 1896-1904, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 233, + 287, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 233, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 49, + 233, + 287, + 288 + ], + "type": "text", + "content": "[33] Nicolae-Catalin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In CVPR, pages 13566-13576, 2022. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 290, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 290, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 49, + 290, + 287, + 323 + ], + "type": "text", + "content": "[34] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823, 2015. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 325, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 325, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 49, + 325, + 287, + 357 + ], + "type": "text", + "content": "[35] Ashish Singh, Michael J Jones, and Erik G Learned-Miller. Eval: Explainable video anomaly localization. In CVPR, pages 18717-18726, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 359, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 359, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 359, + 287, + 392 + ], + "type": "text", + "content": "[36] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, pages 6479-6488, 2018. 1, 2, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 394, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 394, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 49, + 394, + 287, + 426 + ], + "type": "text", + "content": "[37] Che Sun, Yunde Jia, Yao Hu, and Yuwei Wu. Scene-aware context reasoning for unsupervised abnormal event detection in videos. In ACM Multimedia, pages 184-192, 2020. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 429, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 429, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 49, + 429, + 287, + 461 + ], + "type": "text", + "content": "[38] Che Sun, Chenrui Shi, Yunde Jia, and Yuwei Wu. Learning event-relevant factors for video anomaly detection. In AAAI, pages 2384-2392, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 464, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 464, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 49, + 464, + 287, + 496 + ], + "type": "text", + "content": "[39] Shengyang Sun and Xiaojin Gong. Hierarchical semantic contrast for scene-aware video anomaly detection. In CVPR, pages 22846-22856, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 498, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 498, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 49, + 498, + 287, + 542 + ], + "type": "text", + "content": "[40] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W. Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In ICCV, pages 4955-4966, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 544, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 544, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 49, + 544, + 287, + 597 + ], + "type": "text", + "content": "[41] Xuanzhao Wang, Zhengping Che, Bo Jiang, Ning Xiao, Ke Yang, Jian Tang, Jieping Ye, Jingyu Wang, and Qi Qi. Robust unsupervised video anomaly detection by multipath frame prediction. IEEE Trans. Neural Networks Learn. Syst., 33 (6):2301-2312, 2022. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 600, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 600, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 49, + 600, + 287, + 632 + ], + "type": "text", + "content": "[42] Ziming Wang, Yuexian Zou, and Zeming Zhang. Cluster attention contrast for video anomaly detection. In ACM Multimedia, pages 2463-2471, 2020. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 287, + 678 + ], + "type": "text", + "content": "[43] Jie Wu, Wei Zhang, Guanbin Li, Wenhao Wu, Xiao Tan, Yingying Li, Errui Ding, and Liang Lin. Weakly-supervised spatio-temporal anomaly detection in surveillance video. In IJCAI, pages 1172-1178, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "text", + "content": "[44] Fanyi Xiao, Kaustav Kundu, Joseph Tighe, and Davide Modolo. Hierarchical self-supervised representation learning for movie understanding. In CVPR, pages 9717-9726, 2022. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 485 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[45] Zhiwei Yang, Peng Wu, Jing Liu, and Xiaotao Liu. Dynamic local aggregation network with adaptive clusterer for anomaly detection. In ECCV, pages 404-421, 2022. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 108, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 149 + ], + "type": "text", + "content": "[46] Zhiwei Yang, Jing Liu, Zhaoyang Wu, Peng Wu, and Xiaotao Liu. Video event restoration based on keyframes for video anomaly detection. In CVPR, pages 14592-14601, 2023. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "text", + "content": "[47] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze test helps: Effective video anomaly detection via learning to complete video events. In ACM Multimedia, pages 583-591, 2020. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 198, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 239 + ], + "type": "text", + "content": "[48] Jongmin Yu, Younkwan Lee, Kin Choong Yow, Moongu Jeon, and Witold Pedrycz. Abnormal event detection and localization via adversarial event prediction. IEEE Trans. Neural Networks Learn. Syst., 33(8):3572-3586, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 242, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 545, + 285 + ], + "type": "text", + "content": "[49] Muhammad Zaigham Zaheer, Arif Mahmood, Muhammad Haris Khan, Mattia Segù, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In CVPR, pages 14724–14734, 2022. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 287, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 340 + ], + "type": "text", + "content": "[50] Chen Zhang, Guorong Li, Yuankai Qi, Shuhui Wang, Laiyun Qing, Qingming Huang, and Ming-Hsuan Yang. Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection. In CVPR, pages 16271-16280, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "type": "text", + "content": "[51] Yiru Zhao, Bing Deng, Chen Shen, Yao Liu, Hongtao Lu, and Xian-Sheng Hua. Spatio-temporal autoencoder for video anomaly detection. In ACM Multimedia, pages 1933-1941, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 388, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 441 + ], + "type": "text", + "content": "[52] Yuanhong Zhong, Xia Chen, Yongting Hu, Panliang Tang, and Fan Ren. Bidirectional spatio-temporal feature learning with multiscale evaluation for video anomaly detection. IEEE Trans. Circuits Syst. Video Technol., 32(12):8285-8296, 2022. 2, 3, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 443, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 443, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 443, + 545, + 485 + ], + "type": "text", + "content": "[53] Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection. In AAAI, pages 3769-3777. AAAI, 2023. 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17394" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/04406ae4-0193-431d-9be9-c73afd625413_content_list.json b/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/04406ae4-0193-431d-9be9-c73afd625413_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8ece97e804ac3963991690732239a6adef2775b8 --- /dev/null +++ b/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/04406ae4-0193-431d-9be9-c73afd625413_content_list.json @@ -0,0 +1,1857 @@ +[ + { + "type": "text", + "text": "Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization", + "text_level": 1, + "bbox": [ + 109, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lahav Lipson", + "bbox": [ + 297, + 181, + 411, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Princeton University", + "bbox": [ + 272, + 200, + 437, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "llipson@princeton.edu", + "bbox": [ + 263, + 219, + 449, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jia Deng", + "bbox": [ + 573, + 181, + 647, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Princeton University", + "bbox": [ + 527, + 200, + 692, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jiadeng@princeton.edu", + "bbox": [ + 516, + 219, + 702, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce a new system for Multi-Session SLAM, which tracks camera motion across multiple disjoint videos under a single global reference. Our approach couples the prediction of optical flow with solver layers to estimate camera pose. The backbone is trained end-to-end using a novel differentiable solver for wide-baseline two-view pose. The full system can connect disjoint sequences, perform visual odometry, and global optimization. Compared to existing approaches, our design is accurate and robust to catastrophic failures. Code is available at https://github.com/princeton-v1/MultiSlam_DiffPose", + "bbox": [ + 76, + 301, + 473, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 479, + 209, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simultaneous Localization and Mapping (SLAM) is the task of estimating camera motion and a 3D map from video. The standard setup assumes a single continuous video. However, video data in the wild often consists of not a single continuous stream, but rather multiple disjoint sessions, either deliberately such as in collaborative mapping when multiple robots perform joint rapid 3D reconstruction, or inadvertently due to visual discontinuities in the video stream which can result from camera failures, extreme parallax, rapid turns, auto-exposure lag, dark areas, or extreme occlusion by dynamic objects. Handling such disjoint videos is important for many applications in AR and robotics, and gives rise to the task of Multi-Session SLAM.", + "bbox": [ + 75, + 506, + 468, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In Multi-Session SLAM, the input consists of multiple disjoint video sequences and the goal is to estimate camera poses for all video frames under a single global reference. This is in contrast to \"single-video SLAM\", whose input is a single continuous video. In this work, we focus on the monocular, RGB-only Multi-Session SLAM setting.", + "bbox": [ + 75, + 703, + 468, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several approaches have been proposed to deal with Multi-Session SLAM, however existing solutions typically require additional sensor data in order to remove gauge freedoms and make tracking easier [16, 28, 37]. Only a small number of methods, notably CCM-SLAM [36] and ORBSLAM3 [4], support Multi-Session SLAM from monocular video alone, due to the difficulty of aligning disjoint", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4891576479077935d55e89ddf5bb9d1bf95a249293d965daacfe5934c762b57b.jpg", + "image_caption": [ + "Figure 1. Our method estimates camera pose from multiple disconnected video streams." + ], + "image_footnote": [], + "bbox": [ + 504, + 268, + 893, + 377 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sequences under the 7-DOF gauge freedoms in monocular video. However, these approaches are based on classical feature descriptors, making them less accurate on average compared to recent designs based on deep networks.", + "bbox": [ + 496, + 444, + 890, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the standard SLAM setting, Teed and Deng [44] proposed to use a deep optical flow network (RAFT [43]) to track 2D motion, while jointly updating camera poses with a bundle adjustment layer. The method, DROID-SLAM, is accurate and avoids tracking failures, but the design assumes a continuous video stream and is not capable of the wide-baseline matching and non-local optimization necessary for Multi-Session SLAM. Deep Patch Visual Odometry [45] (DPVO) introduced a sparse visual-odometry-only analog of DROID-SLAM which achieves similar accuracy on single-video SLAM, but at much lower cost. However, DPVO also does not support Multi-Session SLAM for the same reasons as DROID-SLAM.", + "bbox": [ + 496, + 506, + 892, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a method for Multi-Session SLAM, capable of both wide-baseline relative pose and visual odometry using a single backbone architecture inspired by [45]. We introduce a differentiable solver layer which minimizes the symmetric epipolar distance (SED) from bi-directional optical flow. From this we construct a method for two-view pose which is capable of matching from far-apart views. This same design can be repurposed for visual odometry by swapping out the solver for bundle adjustment. By employing a unified backbone architecture for both tasks, we enable a simple approach to Multi-Session SLAM.", + "bbox": [ + 496, + 703, + 890, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We evaluate our approach on challenging real-world datasets: EuRoC-MAV [2] and ETH3D [39]. Our sys", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19626", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tem is more accurate than prior approaches, and is robust to catastrophic failures. We also evaluate our two-view pose method in isolation on the Scannet and Megadeph datasets, and show that it is competitive with transformer-based matching networks. For pairs of far-apart views, our method is capable estimating accurate relative pose.", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our backbone predicts iterative updates to optical flow coupled with a differentiable solver layer for estimating camera pose. This framework, based on RAFT [43], has worked exceptionally well for Visual Odometry [45] and SLAM [44]. By leveraging this idea for wide-baseline matching, we can extend these methods to the Multi-Session setting without introducing substantial complexity.", + "bbox": [ + 75, + 181, + 467, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 299, + 218, + 314 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most prior works treat Visual Simultaneous Localization and Mapping (Visual SLAM) as an optimization problem solving for a 3D scene model and camera trajectory which best explains the visual measurements [3].", + "bbox": [ + 75, + 316, + 467, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Indirect approaches to SLAM perform keypoint matching as a pre-processing step, then estimate a 3D point cloud and camera poses by optimizing the 2D reprojection error over all matches [4, 20, 26, 27, 33, 38]. Reprojection-error is easier to optimize than photometric alignment, making indirect methods robust to lower camera $\\text{hz}$ [44]. Keypoint matching also enables a straightforward approach to estimating two-view relative pose, a necessary step for Visual Multi-Session SLAM [14, 19, 25, 33, 46]. However, keypoint-based SLAM is less robust to low-texture environments compared to those which use photometric alignment [9, 10] or optical flow [44, 45].", + "bbox": [ + 75, + 381, + 467, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semi-Indirect approaches similarly optimize 2D reprojection error like indirect methods, but without requiring matches as input [42, 44, 45]. Instead, these approaches alternate between predicting optical flow residuals and performing bundle adjustment. Semi-indirect methods do not require repeatable keypoints across images, making them robust to low-texture settings while retaining the easier reprojection-error objective.", + "bbox": [ + 75, + 565, + 467, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach is most similar to Deep Patch Visual Odometry [45] (DPVO), which is a sparse analog of DROID-SLAM. DPVO predicts sparse optical flow instead of dense, performing similarly to DROID-SLAM while running faster and using half the memory.", + "bbox": [ + 75, + 686, + 467, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Differentiable solver layers for camera pose estimation have been used in order to learn outlier rejection with data-driven training. For two-view relative pose, Ranftl and Koltun [30] used a deep network to learn an iteratively reweighted least-squares algorithm, which solved a weighted variant of the 8-point-algorithm [14] using confidences predicted by the network. [30] required matched points as input, whereas our approach can work from images alone. Roessle and Nießner [31] proposed an", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "end-to-end architecture which used the weighted 8-point-algorithm from [30] on top of matches produced using Superglue [34], and supervised directly on the predicted pose. [31] works well, but is unable to outperform existing methods that use minimal solvers with LO-RANSAC [18] instead of differentiable solvers. In contrast, our approach is not built on top of an existing SOTA 2-view matcher; our design is also iterative, applying a recurrent module and solver layers multiple times to refine the prediction.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Session SLAM is the task of performing SLAM on multiple trajectories of the same scene. Like SLAM, Multi-Session SLAM is an online task where camera motion is estimated from a stream of images. However, in the Multi-Session setting, there are known breaks in the data-stream over which the small-baseline assumption no longer holds. While local optimization is sufficient for motion tracking from video, estimating wide-baseline camera pose is often non-convex [40] and more challenging due to large viewpoint changes. In monocular visual SLAM, the scale of each sequence is also ambiguous; aligning two trajectories requires estimating 7 degrees-of-freedom (translation, rotation, scale) for all sequences, excluding the first which can be considered the reference.", + "bbox": [ + 496, + 231, + 890, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ORB-SLAM3 [4] performs Multi-Session SLAM by matching between ORB [32] descriptors. While tracking camera motion over each sequence, ORB-SLAM3 continually updates an Atlas of the scene - a lookup table between keypoints and their 3D map point in the reference frame of their original sequence. To align disjoint sequences, candidate cross-sequence image pairs are identified using image-retrieval [12], the 3D map points are obtained from the Atlas, and the relative transformation is found using the Umeyama algorithm [46]. Several methods perform Multi-Session Visual Inertial SLAM [17, 28], however we focus on the visual-only monocular setting where the scale of each session is ambiguous and must be estimated when joining sequences.", + "bbox": [ + 496, + 443, + 890, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CCM-SLAM [35, 36] and SLAMM [7] perform Multi-Session visual SLAM and are built on top of ORBSLAM [27] as well, but are optimized for limited bandwidth and distributed processing, while ORB-SLAM3 performs better and is optimized for accuracy and speed.", + "bbox": [ + 496, + 656, + 890, + 732 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 500, + 738, + 607, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overview: We propose a backbone for matching between $\\geq 2$ views, and then construct a method for multi-session SLAM upon it. Our backbone approaches matching as optical flow; it borrows several ideas from RAFT [43], such as iteratively predicting flow residuals using a recurrent network, and the correlation feature pyramid. Our method maintains a running estimate of both camera pose and bi-directional optical flow, and uses the update operator to refine them both. We provide an overview in Fig. 2. An", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19627", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b2f41baebe36e1b30f6fab21abdcb137a14a168bc840a97406c1622d93fdf413.jpg", + "image_caption": [ + "Figure 2. Overview of our backbone. Given a group of $\\geq 2$ frames, our method jointly estimates bi-directional optical flow and camera poses. This module is applicable to (1) wide-baseline two-view matching and (2) visual odometry. We use our backbone for Multi-Session SLAM, which requires the ability to perform both (1) and (2). For each image, we select a set of 2D anchor points and initialize their depth and the camera poses trivially. Our approach then iteratively refines the matches for each anchor, similar to RAFT [43], while updating the camera poses. We alternate between matching and pose updates, where each one informs the update to the other. This entire procedure is repeated several times until convergence." + ], + "image_footnote": [], + "bbox": [ + 101, + 87, + 869, + 218 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "invariant of our backbone is that the matches are always clamped to plausible values given the pose estimates and an assumption of rigid scene geometry, therefore WLOG the matches can be considered a depth estimate.", + "bbox": [ + 75, + 330, + 468, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At initialization, our backbone produces dense correlation feature pyramids for each image and context features which remain fixed throughout the forward pass. The update operator uses an RNN to predict an update to the matches, and a differentiable solver to update poses. Our backbone treats the two-view and multi-view settings differently: In the two-view setting, the solver updates the poses to minimize the symmetric epipolar distance (SED). In the multi-view setting, the solver updates both poses and depth to minimize the reprojection error. After the solver layer, the matches are adjusted to agree with the poses/depth.", + "bbox": [ + 75, + 393, + 470, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Initialization", + "text_level": 1, + "bbox": [ + 76, + 569, + 212, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Extraction & Feature Pyramid: Similar to RAFT, our method separately extracts context and correlation feature for each image. The context features are provided as input to the recurrent update operator, whereas the correlation features are used to evaluate the visual similarity between any two pixels using a dot product. The context feature maps are produced at $1/8$ resolution using a residual network. We then apply a linear-self-attention residual for long-range feature sharing. The correlation feature maps are also produced using a residual network, but with several exit ramps to produce features at $1/2$ , $1/4$ , and $1/8$ of the input image resolution. We also average-pool the last one three additional times to produce a correlation feature pyramid with 6 levels. We depict their architecture in the Supplement.", + "bbox": [ + 75, + 593, + 468, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Anchor-Point Selection: Our method predicts sparse optical flow, where matches have one end anchored and the other end free to move in $\\mathbb{R}^2$ . We use a mix of detector [8]-chosen and randomly-chosen anchor points. Each point is assigned an initial match in the other image(s) at the same", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pixel coordinate. For the rest of the paper, we will refer to the $k^{th}$ anchor point as $a_{k}$ , and its match in image $j$ as $m_{kj}$ .", + "bbox": [ + 496, + 330, + 890, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We index the context feature map at each $a_{k}$ to produce a unique context feature vector $ctx_{k} \\in \\mathbb{R}^{384}$ , which is used by the update operator. Beyond this point, the full context feature map is not used and is discarded to save memory.", + "bbox": [ + 496, + 363, + 890, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Correlation: Similar to other RAFT-based methods, we use correlation features to assess the visual alignment/similarity given by the current matching estimate. For each $(a_{k},m_{kj})$ pair, we bilinearly sample both feature pyramids for their respective frames at locations $a_{k}$ and $m_{kj}$ and take their inner product at each level, producing", + "bbox": [ + 496, + 428, + 890, + 518 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\langle f _ {k} ^ {i}, f _ {k j} ^ {i} \\rangle \\in \\mathbb {R}: i = 1 \\dots 6 \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 523, + 890, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To provide additional spatial context, we also perturb both $a_{k}$ and $m_{kj}$ in $3 \\times 3$ and $7 \\times 7$ grids, respectively, and calculate eq. 1 for all pairs. The resulting feature vector $\\mathbf{C}_{kj} \\in \\mathbb{R}^{(6 \\times 3 \\times 3 \\times 7 \\times 7)}$ is then passed to the update operator.", + "bbox": [ + 496, + 547, + 890, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Update Operator", + "text_level": 1, + "bbox": [ + 500, + 619, + 668, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The update operator produces a revision to all $m_{kj}$ and the estimated relative pose. It consists of three stages, depicted in Fig. 3: (1) An RNN predicts a 2D update to $m_{kj}$ and an associated confidence weight $w_{kj}$ . (2) We solve for a pose estimate which is consistent with the newly predicted matches and confidence. (3) We adjust the matches to be physically plausible assuming rigid scene geometry. In the two-view setting, we clamp the matches to the epipolar lines, and in the multi-view setting we reproject the anchor points using the depth and poses from the solver.", + "bbox": [ + 496, + 643, + 890, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Through the recurrent iterations, our method maintains a running state for each anchor-match $(a_{k},m_{kj})$ pair, consisting of a hidden state vector $h_{kj}$ , a confidence weight $w_{kj}$ and a match location $m_{kj}$ in frame $j$ . This state is continually updated, during which $m_{kj}$ should ideally approach the true match location of $a_{k}$ in $j$ , and $w_{kj}$ should approach 1. This is what we observe empirically.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19628", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ab4da5dbd964d6bd8253334e6fbcea6680d6f1c2050d25785bf408754cd779a8.jpg", + "image_caption": [ + "Figure 3. A single update iteration (two-view). For each anchor-match pair, we predict an update to the matches using the RNN. We then solve for an update to the camera poses which minimizes the symmetric epipolar distance (SED), producing a new set of epipolar lines. Finally, we clamp the matches back to the best-fit epipolar lines, and repeat the whole process again. In Fig. 8 and the Supplement, we" + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 869, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "visualize these iterations on real-world images.", + "bbox": [ + 76, + 313, + 356, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "RNN Module: The learnable component of the update operator is the recurrent network which predicts updates to $h_{kj}$ , $m_{kj}$ and $w_{kj}$ for all anchor points. We initialize this", + "bbox": [ + 76, + 339, + 480, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "operator in Fig. 4.", + "bbox": [ + 76, + 385, + 199, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The input to the RNN are the context features, the previous hidden state and the correlation features generated from the current matching estimate, all of which are added together and normalized with layer norm. A self-attention residual is also applied to edges with the same source and destination frame, followed by three gated-residual-units, whose architecture is depicted in the Supplement. The output is an updated hidden state, from which we predict an updated match and confidence using the flow head and confidence head, respectively, which are implemented as two-layer MLPs. The confidence head includes a softmax to", + "bbox": [ + 76, + 400, + 470, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\text{estriel} w_{kj}\\in (0,1)$", + "bbox": [ + 76, + 565, + 217, + 580 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7f5f718f01f4a5b00c7f12616c4796de2611b4357652e20d602376f9f7d7eeda.jpg", + "image_caption": [ + "Correlation Pyramids" + ], + "image_footnote": [], + "bbox": [ + 37, + 561, + 80, + 608 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/115da04c3ca4fafc617bdfec51ba3c455883c48ed58f3f880cfbeeee325f1cf8.jpg", + "image_caption": [ + "Figure 4. The RNN Module. For each anchor-match pair $(a_{k},m_{kj})$ , it predicts an update to $m_{kj}$ and an associated confidence $w_{kj}$ . It also updates a hidden state $h_{kj} \\in \\mathbb{R}^{384}$ . Internally, the RNN shares features via attention between updates with the same incoming and outgoing frame." + ], + "image_footnote": [], + "bbox": [ + 40, + 621, + 80, + 655 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e973e8a48aedfc16da1ad981d8296568bf46dc4fa8e1f90400f327c3395b790.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 598, + 464, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Two-View Solver and Pre-Conditioning", + "text_level": 1, + "bbox": [ + 76, + 756, + 415, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our two-view solver aims to align the relative pose between two frames to the predicted matches by minimizing the distance between each match and its epipolar line. This strategy is accurate, however the objective function is nonconvex, meaning it will only converge to the global minimum if it is initialized close to it. In contrast, the 8-point-algorithm, a common approach to this task, has the opposite problem; it does not suffer from local minimum since it is a", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "homogeneous least squares, but the solution is sub-optic. In Fig. 6, we visualize how our solver either converges within a fraction of a degree, or not at all, meanwhile", + "bbox": [ + 480, + 311, + 862, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "8-point-algorithm is more robust but less accurate.", + "bbox": [ + 498, + 385, + 831, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The approach our method uses is to pre-condition the pose estimate using a weighted dense variant of the 8-point algorithm [30], and then run our solver layer to refine the pose. This combined strategy obtains the best of both worlds, since the pre-conditioning will typically initialize the pose within the basin of convergence of our solver, which then refines the prediction further. We visualize this basin on the TartanAir [48] dataset using groundtruth matches in Fig. 6.", + "bbox": [ + 480, + 401, + 890, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-conditioning: We solve the homogeneous least squares problem [14]", + "bbox": [ + 498, + 541, + 890, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\mathbf {F}} \\overline {{| | d i a g (\\vec {w}) \\mathbf {M} \\vec {\\mathbf {F}} | | ^ {2}}} \\quad s. t. | | \\mathbf {F} | | ^ {2} = 1 \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 578, + 890, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{M}$ and $\\vec{w}$ are constructed from the anchor-match pairs and confidence predictions. The points are normalized to $[-1, 1]$ beforehand. Afterwards, we reshape and uncalibrate $\\mathbf{F}$ to obtain the essential matrix.", + "bbox": [ + 496, + 608, + 890, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We obtain the four relative pose candidates following the procedure detailed in [13] and in the supplement. During training, we select the pose candidate closest to the ground-truth on the $SE(3)$ manifold. During inference, we select the candidate by testing for chirality [13].", + "bbox": [ + 496, + 669, + 890, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SED Solver Layer: The solver layer seeks to minimize the distance between each match $m_{kj}$ , and the epipolar line induced by the current pose estimate and its respective anchor point $a_k$ . The solver objective is bi-directional, since every image contains a unique set of anchor points; the objective is also known as the titular symmetric epipolar distance (SED) [11]. We parameterize the pose update as small rotations to the translation direction and orientation: $\\xi_{\\mathbf{t}}, \\xi_{\\mathbf{R}} \\in \\mathfrak{so}(3)$ . These are the free variables in eq. 3. Let $K(i)$ be the set of anchor indices for frame $i$ , $\\mathbf{R}$ and $\\mathbf{t}$ be the", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19629", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/61ddc7591b5c3aa86415c52e218a9eb77ac13fef93b52cef8712fe977e19148d.jpg", + "image_caption": [ + "Figure 5. The flow of gradients through our solver. The updated matches from the RNN are supervised directly, and then detached from the gradient tape. The solver output is supervised with a pose loss. In the backward pass, gradients from the pose loss are used to supervise the confidence head in order to learn outlier rejection." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 457, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "output of the pre-conditioning stage, and $epi(x,R,t)$ compute the epipolar line for a point $x$ given the relative pose $(R,t)$ . Formally, the minimization objective in our solver is", + "bbox": [ + 76, + 321, + 468, + 367 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} l _ {k j} = e p i \\left(a _ {k}, \\left(e ^ {\\xi_ {\\mathbf {R}}} \\mathbf {R}\\right) _ {i \\rightarrow j}, \\left(e ^ {\\xi_ {\\mathbf {t}}} \\mathbf {t}\\right) _ {i \\rightarrow j}\\right): k \\in K (i) \\\\ E _ {i \\rightarrow j} = \\sum_ {k \\in K (i)} w _ {k j} \\cdot \\| \\operatorname {e r r} \\left(m _ {k j}, l _ {k j}\\right) \\| _ {2} ^ {2} \\tag {3} \\\\ \\stackrel {\\smile}{S E D} = \\operatorname *{arg min}_{\\xi_{\\mathbf{R}},\\xi_{\\mathbf{t}}}\\left(E_{i\\to j} + E_{j\\to i}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 377, + 468, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $err$ computes the 2D point-to-line error. The predicted weights $w_{kj}$ are included to allow the network to down-weight the contribution of any match which it deems unreliable. This solver layer is implemented in Pytorch, enabling gradients to propagate backward from the pose loss to the confidence head using autograd. We depict the gradient flow in Fig. 5.", + "bbox": [ + 75, + 474, + 468, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To minimize eq. 3, we employ the Levenberg-Marquardt algorithm. This requires computing the Jacobian for all terms. We provide the full derivations in the Supplement, but define the residual function $err$ in its entirety here.", + "bbox": [ + 75, + 580, + 468, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $l$ be the epipolar line produced by $(a, e^{\\xi_{\\mathbf{R}}}\\mathbf{R}, e^{\\xi_{\\mathbf{t}}}\\mathbf{t})$ . We can express each term of eq. 3 as:", + "bbox": [ + 75, + 643, + 468, + 675 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {e r r} (m, l) = \\left(\\frac {l _ {x} m _ {x} + l _ {y} m _ {y} + l _ {z}}{l _ {x} ^ {2} + l _ {y} ^ {2}}\\right) \\left[ \\begin{array}{l} l _ {x} \\\\ l _ {y} \\end{array} \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 685, + 468, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the epipolar line as", + "bbox": [ + 76, + 734, + 233, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {E} = \\left(e ^ {\\xi_ {\\mathrm {R}}} \\mathbf {R}\\right) ^ {\\top} \\left[ e ^ {\\xi_ {\\mathrm {t}}} \\mathbf {t} \\right] _ {\\times} \\\\ l = \\mathbf {F} \\left[ \\begin{array}{l} a _ {x} \\\\ a _ {y} \\\\ 1 \\end{array} \\right] = \\left[ \\left(K _ {2} ^ {\\top}\\right) ^ {- 1} \\mathbf {E} K _ {1} ^ {\\top} \\right] \\left[ \\begin{array}{l} a _ {x} \\\\ a _ {y} \\\\ 1 \\end{array} \\right] \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 758, + 468, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{F}$ and $\\mathbf{E}$ are the fundamental and essential matrices.", + "bbox": [ + 76, + 838, + 465, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Adapting our backbone to Visual Odometry", + "text_level": 1, + "bbox": [ + 76, + 862, + 449, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To adapt our backbone to VO, we make several changes:", + "bbox": [ + 76, + 885, + 450, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/27405bfbdf6641ceb6b0276b35feefc787b9b8ea6be9bd0659c4ca2c0881034f.jpg", + "image_caption": [ + "Figure 6. Convergence of our solver, given ground-truth matches (from TartanAir [48]). Our SED solver is accurate, but only converges to the global minimum when initialized close to it. We initialize the pose using the 8-pt-algorithm to ensure it is within the SED convergence basin." + ], + "image_footnote": [], + "bbox": [ + 516, + 92, + 877, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-view Solver (BA): Our local optimizer minimizes reprojection error and treats depth as a separate variable. This is identical to the bundle adjustment from DPVO. Formally, let $F$ be the set of connected frames, $G$ be the global poses, $d_{k}$ be the depth estimate for anchor $k$ , and $\\Pi(\\cdot)$ be the $3D \\rightarrow 2D$ projection function. The bundle adjustment objective is:", + "bbox": [ + 498, + 304, + 890, + 409 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {G, \\mathbf {d}} {\\arg \\min } \\sum_ {(i, j) \\in F} \\sum_ {k \\in K (i)} w _ {k j} \\cdot \\left\\| \\Pi \\left[ G _ {j} ^ {- 1} G _ {i} \\Pi^ {- 1} \\left(a _ {k}, d _ {k}\\right) \\right] - m _ {k j} \\right\\| ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 421, + 906, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The preconditioning stage is not necessary as we can linearly-extrapolate the camera poses from previous frames to achieve good initialization.", + "bbox": [ + 498, + 470, + 890, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Clamping: Since epipolar lines do not make sense in the multi-view setting, we reset the optical flow using the pose and depth from the bundle adjustment.", + "bbox": [ + 498, + 520, + 890, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RNN changes: We include mechanisms in the RNN to share latent features between updates if they stem from the same anchor point. Specifically, we use the message passing and temporal convolutions from DPVO [45].", + "bbox": [ + 498, + 570, + 890, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Trajectory Alignment", + "text_level": 1, + "bbox": [ + 500, + 641, + 702, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the Multi-Session SLAM problem, we estimate a relative Sim(3) between pairs of disjoint trajectories in order to align them. All anchor points already have a depth estimate $d_{k}$ as a result of the bundle adjustment (eq. 6) from the VO system. We use our two-view method to estimate the relative rotation and translation direction, and compute the translation magnitude and relative scaling by comparing the $d_{k}$ 's to the depth from the triangulated two-view matches.", + "bbox": [ + 498, + 665, + 890, + 786 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Estimate relative rotation, translation direction, and matches. We first retrieve a candidate image pair $(i,j)$ using NetVLAD [1], one from each trajectory, and apply our two-view model to estimate their relative rotation and translation direction.", + "2) Align the depth from the two-view and SLAM operators. We previously defined $d_{k}$ as the depth output of the VO" + ], + "bbox": [ + 500, + 790, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19630", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "system. Let $d_k'$ be the triangulated depth obtained from our two-view matches:", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd _ {k} ^ {\\prime} = \\left\\{\\begin{array}{l l}\\operatorname {t r i a n g} \\left(\\mathbf {R} _ {i \\rightarrow j}, \\mathbf {t} _ {i \\rightarrow j}, a _ {k}, m _ {k j}\\right)&: k \\in K (i)\\\\\\operatorname {t r i a n g} \\left(\\mathbf {R} _ {j \\rightarrow i}, \\mathbf {t} _ {j \\rightarrow i} a _ {k}, m _ {k i}\\right)&: k \\in K (j)\\end{array}\\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 123, + 468, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We solve", + "bbox": [ + 76, + 172, + 142, + 185 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {s _ {i} \\in \\mathbb {R} > 0} {\\arg \\max } \\sum_ {k \\in K (i)} \\mathbf {1} \\left[ \\frac {1}{\\lambda} < \\frac {d _ {k}}{s _ {i} \\cdot d _ {k} ^ {\\prime}} < \\lambda \\right] \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 181, + 468, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to recover the translation magnitude, where the hyperparameter $\\lambda = 1.05$ . In layman terms, eq. 9 seeks what translation magnitude would align the most triangulated points to the existing 3D map? We can obtain a reasonably good solution to eq 9 by brute-force checking $s_i = d_k / d_k'$ $\\forall k\\in K(i)$ , similar to RANSAC. If the number of inliers is too small, we retry on a new candidate pair. Conversely, the scale-difference is $s_i / s_j$ , where $s_j$ is estimated the same way, but for anchor points in $K(j)$ :", + "bbox": [ + 76, + 229, + 468, + 364 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {s _ {j} \\in \\mathbb {R} > 0} {\\arg \\max } \\sum_ {k \\in K (j)} \\mathbf {1} \\left[ \\frac {1}{\\lambda} < \\frac {d _ {k}}{s _ {j} \\cdot d _ {k} ^ {\\prime}} < \\lambda \\right] \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 364, + 468, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the rotation $\\mathbf{R}_{j\\rightarrow i}$ and translation $\\mathbf{t}_{j\\rightarrow i}$ from the two-view solver layers, the resulting transformation can be used to align the two trajectories:", + "bbox": [ + 76, + 419, + 468, + 463 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {j \\rightarrow i} = \\left[\\begin{array}{c c}\\frac {s _ {i}}{s _ {j}} \\mathbf {R}&s _ {j} \\mathbf {t}\\\\0&1\\end{array}\\right] \\in \\operatorname {S i m} (3) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 468, + 468, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The graphs are then merged by concatenating all buffers.", + "bbox": [ + 76, + 506, + 450, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Training", + "text_level": 1, + "bbox": [ + 76, + 530, + 181, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our backbone separately for VO and for two-view pose. It is trained using a matching loss and a pose loss. The pose loss for the two-view training", + "bbox": [ + 76, + 553, + 468, + 599 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p o s e 2 V}} = \\sum_ {t = 1} ^ {1 2} \\cos^ {- 1} \\left(\\bar {\\mathbf {t}} ^ {t} \\cdot \\bar {\\mathbf {t}} ^ {\\bar {\\mathbf {g}}} ^ {t}\\right) + \\alpha \\left\\| \\left[ \\left(\\mathbf {R} ^ {t}\\right) ^ {T} \\mathbf {R} ^ {\\mathbf {g}} ^ {t} \\right] \\right\\| _ {S O (3)} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 609, + 468, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "penalizes angle error for both the predicted translation direction and predicted orientation. The pose loss for the VO training is the SE3 manifold distance between the predicted and ground-truth poses:", + "bbox": [ + 76, + 660, + 468, + 719 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p o s e V O}} = \\sum_ {t = 1} ^ {T} \\left\\| \\left[ \\left(\\mathbf {G} ^ {t}\\right) ^ {- 1} \\mathbf {G} ^ {\\mathbf {g t}} \\right] \\right\\| _ {S E (3)} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 720, + 468, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The matching loss", + "bbox": [ + 76, + 766, + 200, + 779 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a t c h i n g}} = \\frac {1}{| N |} \\sum_ {a} ^ {N} \\sum_ {t = 1} ^ {T} \\left\\| m _ {a} ^ {t} - m _ {a} ^ {\\mathbf {g t}} \\right\\| _ {2} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 780, + 468, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is a standard endpoint-error [24, 43], applied only on pixels which have valid depth and are verified to be visible in both images. Our update operator is applied 12 times per training example; supervision is applied after every update (See Fig. 5). The final loss is $\\mathcal{L} = \\mathcal{L}_{\\text{pose}} + \\beta \\cdot \\mathcal{L}_{\\text{matching}}$ .", + "bbox": [ + 76, + 825, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/29bd9d0c7eec35b33a57b9d37784db0e9d3d963bd66c78ce6335dfb35a6bad10.jpg", + "image_caption": [ + "Stage 1: Image Retrieval" + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 629, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/667dacec7e81c5a3476d64d62d609f784f7d21e2d0471acfa6c8a9a155509477.jpg", + "image_caption": [ + "Stage 2: Sim(3) Estimation", + "Stage 3: Graph Merging" + ], + "image_footnote": [], + "bbox": [ + 635, + 90, + 767, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/89462d1f56b69e3ee1367046a5087e4dec77904aab2a5d46fae3b8ab21e837ad.jpg", + "image_caption": [ + "Figure 7. High-level overview of our approach to aligning disjoint trajectories. We use NetVLAD [1] to retrieve two co-visible images, one from each trajectory, and apply our two-view update operator to estimate a relative Sim(3) alignment, following the procedure detailed in Sec 3.5. We then transform the second trajectory into the reference frame of the first, and merge all buffers." + ], + "image_footnote": [], + "bbox": [ + 774, + 97, + 883, + 203 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following prior work [23], we pre-train our two-view method on synthetic homographies (without pose loss) for two epochs on the Oxford-Paris 1M Distractors dataset [29]. We then tune our model on a $50/50$ mixture of Scannet [6] and Megadepth [21] for 100,000 steps and a batch size of 120, using 10 A6000 GPUs for 5 days. The VO backbone is trained using the procedure from [45] on TartanAir.", + "bbox": [ + 498, + 340, + 890, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.7. Multi-Session SLAM System", + "text_level": 1, + "bbox": [ + 500, + 460, + 754, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overview: To perform Multi-Session SLAM, we use our VO-trained backbone to perform visual odometry and global optimization; our two-view backbone connects disjoint trajectories following the procedure in Sec 3.5. Our full system builds the pose graph incrementally: new video frames inserted into the graph as they are received, and unjoined trajectories are aligned and merged as soon as a connection is found.", + "bbox": [ + 496, + 486, + 890, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VO Frontend: Our VO frontend extracts features and estimates camera poses and depth for incoming frames, operating on a sliding window covering the most recently observed 20 keyframes. Keyframing occurs retroactively on the $4^{th}$ oldest keyframe if it is found to be redundant. New poses are initialized using a linear-motion model, and depth is copied from the previous frame. Like DPVO [45], we initialize after 8 frames with significant motion are observed.", + "bbox": [ + 496, + 614, + 890, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Global Optimization: We perform global optimization by introducing proximity factors using the existing pose and depth measurements, following the approach from DROID-SLAM [44], and update the matches and poses/depth using the update operator. We only run the backend periodically, and after joining trajectories.", + "bbox": [ + 496, + 742, + 890, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Trajectory Joining: We query the database of NetVLAD descriptors to find co-visible frames. If several pairs are retrieved, we compute relative poses in a mini-batch and use the estimate with the highest inlier-ratio.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19631", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/60b32ef2eaa78fe043510dcc2636baa972208dbe0e9d87d8a5040b4b47413719.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Overall ApproachPose error AUC [%] ↑
@5°@10°@20°
Superglue [34]Matching16.2(17.7)33.8(35.6)51.8(54.2)
LightGlue [23]16.5(19.4)33.4(36.9)50.1(53.5)
LoFTR [41]↓22.1(25.7)40.1(45.0)47.6(61.4)
MatchFormer [47]RANSAC [15]24.3(27.3)43.9(47.6)61.4(64.9)
ASpanFormer [5](LO-RANSAC [18])25.6(28.4)46.0(48.8)63.3(65.8)
Roessle&Nießner [31]Matching→W8PA20.741.661.7
Roessle&Nießner [31]Matching→W8PA→BA25.747.266.4
OursOptical-Flow\n←\nClamp ← Solver30.951.167.6
", + "bbox": [ + 81, + 89, + 467, + 239 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a8635de030a474c883167004848236e7df002e5eeda0b1190445814da91f04b9.jpg", + "table_caption": [ + "Table 1. Two-view results on Scannet [6]. Existing methods perform matching as pre-processing, then use a solver; our method is iterative. Results in parenthesis use the LO-RANSAC [18] estimator. \"W8PA $\\rightarrow$ BA\" = weighted 8-point-algo. & bundle-adjustment. Our design outperforms prior approaches. We use the same model weights as in Tab. 2." + ], + "table_footnote": [], + "table_body": "
Overall ApproachPose error AUC [%] ↑
@5°@10°@20°
Superglue [34]Matching49.7(65.8)67.1(78.7)80.6(87.5)
LightGlue [23]49.9(66.7)67.0(79.3)80.1(87.9)
LoFTR [41]↓52.8(66.4)69.2(78.6)81.2(86.5)
MatchFormer [47]RANSAC [15]53.3(66.5)69.7(78.9)81.8(87.5)
ASpanFormer [5](LO-RANSAC [18])55.3(69.4)71.5(81.1)83.1(88.9)
Roessle&Nießner [31]Matching→W8PA46.962.876.3
Roessle&Nießner [31]Matching→W8PA→BA61.274.985.0
OursOptical-Flow60.472.781.3
Clamp ← Solver
", + "bbox": [ + 81, + 343, + 465, + 493 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Two-view results on Megadepth [21]. Existing methods perform matching as pre-processing, then use a solver; our method is iterative. \"W8PA $\\rightarrow$ BA\" = weighted 8-point-algo. & bundle-adjustment. Results in parenthesis use the LO-RANSAC [18] estimator. Our radically different design leads to better results indoors (Tab. 1), but is less competitive in photo-tourism settings where high-volume matching is easier and therefore RANSAC-based methods work better. We use the same model weights as in Tab. 1.", + "bbox": [ + 75, + 503, + 468, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 638, + 207, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the performance of our two-view method in isolation, and our Multi-Session SLAM system as a whole.", + "bbox": [ + 75, + 664, + 468, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Two-View Evaluation", + "text_level": 1, + "bbox": [ + 76, + 700, + 277, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our two-view method on two popular relative pose benchmarks: The Scannet [6] 1500 and MegaDepth [21] 1500 test datasets. We compare to existing matching networks [5, 23, 34, 41, 47], which are typically in service of improving COLMAP [38]-based SfM pipelines [22, 33]. In contrast, our approach is in service of Multi-Session SLAM. We report pose error AUC, as is done in prior matching work (higher is better). We use the same configuration and model weights on both datasets.", + "bbox": [ + 75, + 718, + 468, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Most existing methods perform matching as a preprocessing step using a deep network, then estimate pose with a minimal solver using RANSAC [15, 18]. Roessle", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and Nießner [31] replace the RANSAC optimizer with a weighted-8-point algorithm and bundle adjustment. In contrast, our two-view method couples the prediction of pose and matches, refining the predictions over many iterations. We ablate over our two-view model in Tab. 3.", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scannet: We report two-view results on Scannet [6] in Tab. 1. Our approach outperforms existing methods on Scannet, which contains fewer salient keypoints and many texture-less surfaces and motion blur.", + "bbox": [ + 496, + 170, + 890, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MegaDepth: We report two-view results on Megadepth [21] in Tab. 2. On Megadepth, the matching problem is substantially easier since photo-tourism images contain many salient keypoints, resulting in better performance across all methods; our approach is on-par with existing approaches using the OpenCV RANSAC pose estimator [15], but not the LO-RANSAC [18] estimator.", + "bbox": [ + 496, + 233, + 890, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Multi-Session SLAM evaluation", + "text_level": 1, + "bbox": [ + 498, + 345, + 781, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our approach on Multi-Session SLAM on the EuRoC-MAV [2] and ETH3D [39] datasets, since they provide camera poses under a single global reference. Following the evaluation for single-video SLAM, we align the final predictions to the ground truth by computing a global 7-DOF alignment to account for the guage freedoms. We report RMSE ATE[m] in metric units. We sample 96 anchors in each video frame using a mix of random and Superpoint [8] keypoints. Our two-view method uses those same anchors to join pairs of trajectories. We use the same weights and configuration on both datasets.", + "bbox": [ + 496, + 364, + 890, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "EuRoC-MAV: In Tab. 9 and Fig. 10, we report results on the EuRoC MAV dataset [2]. Following the evaluation from [4], we compare on four groups of disjoint trajectories across the three distinct environments: 3 in the Machine Hall, 5 in the Machine Hall, 3 in Vicon 1, and 3 in Vicon 2. Ground-truth poses are obtained using a laser tracker and Vicon cameras. These video sequences are long, with the entirety of the Machine-Hall sequence being 11.5 minutes of video (13.7K frames) spread across 5 sequences. Our approach achieves significantly lower error than ORBSLAM3 on all groups, including 12x lower error on the Vicon 2 sequences (0.022 vs 0.284). Video is recorded at 20-FPS; all methods run in real-time. We also compare to CCM-SLAM [36], which only provides results on the on MH01-03 sequence, the mono-inertial ORB-SLAM3, and the mono-inertial VINS [28]. We outperform the other approaches on all trajectory groups.", + "bbox": [ + 496, + 534, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ETH3D: In Tab. 4, we report results on the training set from ETH3D [39]. We compare across 5 unique scenes composed of multiple trajectories: 1-4 from Sofa, 3&4 from Table, 1-3 from Plant Scene, 1&2 from Einstein, and 2&3 from Planar. Sequences are excluded if they belong to the official ETH3D test set (with no ground-truth) or are from a different scene altogether. The sequences are trimmed", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19632", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/39ee103b7e694f32df19e18b1a2b150cf3540eb70b49e77258168d64285f4f64.jpg", + "image_caption": [ + "Figure 8. Qualitative results on Scannet. Our two-view variant is able to estimate accurate relative poses across wide camera baselines. It initializes all matches with uniform depth and identity relative pose. Progressive applications of our update operator lead to more accurate matches and higher predicted confidence. Some image pairs (row 2) take more iterations to converge than others (row 1)." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 890, + 239 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a6f21b60a13fa1e893b8034709b4fe09dd45c4f114444ffc158dfe4ae53f913a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Scene nameMH01-03MH01-05V101-103V201-203Cam.
# Disjoint Trajectories3533hz
Ours0.0220.0360.0300.02220
Mono-Visual
CCM-SLAM [36]0.077---38
Mono-Visual
ORB-SLAM3 [4]0.0300.0580.0580.28420
Mono-Visual
VINS [28]-0.210---
Mono-Inertial
ORB-SLAM30.0370.0650.0400.04820
Mono-Inertial
", + "bbox": [ + 81, + 309, + 467, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2055a6290fe878c490de84d2166b9784dbc2a86d85c363ac87fb554965774010.jpg", + "image_caption": [ + "Figure 9. Multi-Session SLAM evaluation on EuRoC MAV [2] using RMSE ATE[m] $\\downarrow$ . We perform similarly to ORB-SLAM3 (within $2mm$ ) in the Machine Hall, but perform best in the Vicon 1 and Vicon 2 rooms. Results for baseline approaches were obtained from [4].", + "Figure 10. Our prediction on the Vicon1 [2] sequences. Our method is accurate and robust to chaotic motion." + ], + "image_footnote": [], + "bbox": [ + 86, + 554, + 455, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from a single, longer video; to make joining the trajectories non-trivial, we reverse every other sequence, ensuring that there is a sufficient disconnect between subsequent videos. This is a novel benchmark, so [28, 36] do not report results.", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b1e87adcb4b98685c5612a2112637968e489724b56ce20a867656af0cd1ff604.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ablation ExperimentPose error AUC @10°[%] ↑
Megadepth ValScannet Val
BaselineFull model53.333.9
ArchitectureShaller correlation pyramid44.327.1
No attention in update-op48.630.5
No ReLU-attn in feat-extractor47.429.7
SolverNo pre-conditioning35.520.6
No SED solver23.813.1
No clamping step21.910.7
", + "bbox": [ + 503, + 310, + 888, + 434 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/562569f4080a5ff80200bb038c3de4177d6c4654a1e6efc55ca75371dd1f0c97.jpg", + "table_caption": [ + "Table 3. Two-view ablation experiments on two validation sets. Our proposed RNN and feature extractor changes from DPVO improve the result. Removing components from our proposed solver design decreases performance, as does removing the reprojection step." + ], + "table_footnote": [], + "table_body": "
Scene nameSofaTablePlant SceneEinsteinPlanar
# Disjoint Trajectories42322
ORB-SLAM3 [4] \nMono-VisualFAIL \n(no init)0.018FAIL \n(init→lost)FAIL \n(init→lost)0.010
Ours \nMono-Visual0.0130.0100.0090.0050.035
", + "bbox": [ + 503, + 526, + 888, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Multi-Session SLAM evaluation on ETH3D [39] using RMSE ATE[m] $\\downarrow$ . Our method is robust to catastrophic failures.", + "bbox": [ + 498, + 622, + 890, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ORB-SLAM3 fails on Sofa, Plant Scene, and Einstein for various reasons (couldn't initialize, or lost the feature tracks). Our approach outperforms ORB-SLAM3 on 4/5 groups, and does not fail on any. ORB-SLAM3 only succeeds on 2/5 groups.", + "bbox": [ + 498, + 665, + 890, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 758, + 617, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce a new method for mono-visual Multi-Session SLAM. Our system utilizes a novel backbone which can estimate two-view pose and perform visual odometry. We leverage a novel differentiable solver which minimizes the symmetric epipolar distance. We compare against existing approaches and show strong performance across several datasets. This work was partially supported by IARPA and the National Science Foundation.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19633", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pa-jdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5297-5307, 2016. 5, 6", + "[2] Michael Burri, Janosch Nikolic, Pascal Gohl, Thomas Schneider, Joern Rehder, Sammy Omari, Markus W Achtelik, and Roland Siegwart. The euroc micro aerial vehicle datasets. The International Journal of Robotics Research, 35 (10):1157-1163, 2016. 1, 7, 8", + "[3] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 2", + "[4] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 1, 2, 7, 8", + "[5] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming-min Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 7", + "[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6, 7", + "[7] Hayyan Afeef Daoud, Aznul Qalid Md. Sabri, Chu Kiong Loo, and Ali Mohammed Mansoor. Slamm: Visual monocular slam with continuous mapping using multiple maps. PloS one, 13(4):e0195878, 2018. 2", + "[8] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 3, 7", + "[9] Jakob Engel, Thomas Schöps, and Daniel Cremers. Lsd-slam: Large-scale direct monocular slam. In European conference on computer vision, pages 834-849. Springer, 2014. 2", + "[10] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. IEEE transactions on pattern analysis and machine intelligence, 40(3):611-625, 2017. 2", + "[11] Mohammed E Fathy, Ashraf S Hussein, and Mohammed F Tolba. Fundamental matrix estimation: A study of error criteria. Pattern Recognition Letters, 32(2):383-391, 2011. 4", + "[12] Dorian Gálvez-López and Juan D Tardos. Bags of binary words for fast place recognition in image sequences. IEEE Transactions on Robotics, 28(5):1188-1197, 2012. 2", + "[13] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, New York, NY, USA, 2 edition, 2003. 4, 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Richard I Hartley. In defense of the eight-point algorithm. IEEE Transactions on pattern analysis and machine intelligence, 19(6):580-593, 1997. 2, 4", + "[15] Itseez. Open source computer vision library. https://github.com/itseez/opencv, 2015.7", + "[16] Mathieu Labbe and François Michaud. Rtab-map as an opensource lidar and visual simultaneous localization and mapping library for large-scale and long-term online operation. Journal of field robotics, 36(2):416-446, 2019. 1", + "[17] Mathieu Labbe and François Michaud. Multi-session visual slam for illumination-invariant re-localization in indoor environments. Frontiers in Robotics and AI, 9:801886, 2022. 2", + "[18] Viktor Larsson and contributors. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 2, 7", + "[19] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 2", + "[20] Stefan Leutenegger, Paul Furgale, Vincent Rabaud, Margarita Chli, Kurt Konolige, and Roland Siegwart. Keyframe-based visual-inertial slam using nonlinear optimization. Proceedings of Robotics Science and Systems (RSS) 2013, 2013. 2", + "[21] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Computer Vision and Pattern Recognition (CVPR), 2018. 6, 7", + "[22] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987-5997, 2021. 7", + "[23] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In ICCV, 2023. 6, 7", + "[24] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 6", + "[25] Xiao Xin Lu. A review of solutions for perspective-n-point problem in camera pose estimation. In Journal of Physics: Conference Series, page 052009. IOP Publishing, 2018. 2", + "[26] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 2", + "[27] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 2", + "[28] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics, 34(4):1004-1020, 2018. 1, 2, 7, 8", + "[29] F. Radenović, A. Iscen, G. Tolias, Y. Avrithis, and O. Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In CVPR, 2018. 6" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19634", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Rene Ranftl and Vladlen Koltun. Deep fundamental matrix estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2, 4", + "[31] Barbara Roessle and Matthias Nießner. End2end multi-view feature matching with differentiable pose optimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 477-487, 2023. 2, 7", + "[32] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011. 2", + "[33] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019. 2, 7", + "[34] 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. 2, 7", + "[35] Patrik Schmuck and Margarita Chli. Multi-uav collaborative monocular slam. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3863-3870. IEEE, 2017. 2", + "[36] Patrik Schmuck and Margarita Chli. Ccm-slam: Robust and efficient centralized collaborative monocular simultaneous localization and mapping for robotic teams. Journal of Field Robotics, 36(4):763-781, 2019. 1, 2, 7, 8", + "[37] Thomas Schneider, Marcin Dymczyk, Marius Fehr, Kevin Egger, Simon Lynen, Igor Gilitschenski, and Roland Siegwart. maplab: An open framework for research in visual-inertial mapping and localization. IEEE Robotics and Automation Letters, 3(3):1418-1425, 2018. 1", + "[38] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 7", + "[39] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 7, 8", + "[40] Hauke Strasdat, J Montiel, and Andrew J Davison. Scale drift-aware large scale monocular slam. Robotics: science and Systems VI, 2(3):7, 2010. 2", + "[41] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 7", + "[42] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 2", + "[43] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 1, 2, 3, 6" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1, 2, 6", + "[45] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 5, 6", + "[46] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 2", + "[47] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, pages 2746-2762, 2022. 7", + "[48] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 4, 5" + ], + "bbox": [ + 501, + 90, + 892, + 401 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19635", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/04406ae4-0193-431d-9be9-c73afd625413_model.json b/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/04406ae4-0193-431d-9be9-c73afd625413_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0013adf218ff1b9ff1c4b3ae8b59d90e5f965859 --- /dev/null +++ b/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/04406ae4-0193-431d-9be9-c73afd625413_model.json @@ -0,0 +1,2442 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.861, + 0.154 + ], + "angle": 0, + "content": "Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.182, + 0.413, + 0.199 + ], + "angle": 0, + "content": "Lahav Lipson" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.201, + 0.439, + 0.217 + ], + "angle": 0, + "content": "Princeton University" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.22, + 0.45, + 0.233 + ], + "angle": 0, + "content": "llipson@princeton.edu" + }, + { + "type": "text", + "bbox": [ + 0.575, + 0.182, + 0.648, + 0.199 + ], + "angle": 0, + "content": "Jia Deng" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.201, + 0.694, + 0.217 + ], + "angle": 0, + "content": "Princeton University" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.22, + 0.703, + 0.234 + ], + "angle": 0, + "content": "jiadeng@princeton.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.474, + 0.468 + ], + "angle": 0, + "content": "We introduce a new system for Multi-Session SLAM, which tracks camera motion across multiple disjoint videos under a single global reference. Our approach couples the prediction of optical flow with solver layers to estimate camera pose. The backbone is trained end-to-end using a novel differentiable solver for wide-baseline two-view pose. The full system can connect disjoint sequences, perform visual odometry, and global optimization. Compared to existing approaches, our design is accurate and robust to catastrophic failures. Code is available at https://github.com/princeton-v1/MultiSlam_DiffPose" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.481, + 0.21, + 0.496 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.703 + ], + "angle": 0, + "content": "Simultaneous Localization and Mapping (SLAM) is the task of estimating camera motion and a 3D map from video. The standard setup assumes a single continuous video. However, video data in the wild often consists of not a single continuous stream, but rather multiple disjoint sessions, either deliberately such as in collaborative mapping when multiple robots perform joint rapid 3D reconstruction, or inadvertently due to visual discontinuities in the video stream which can result from camera failures, extreme parallax, rapid turns, auto-exposure lag, dark areas, or extreme occlusion by dynamic objects. Handling such disjoint videos is important for many applications in AR and robotics, and gives rise to the task of Multi-Session SLAM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.795 + ], + "angle": 0, + "content": "In Multi-Session SLAM, the input consists of multiple disjoint video sequences and the goal is to estimate camera poses for all video frames under a single global reference. This is in contrast to \"single-video SLAM\", whose input is a single continuous video. In this work, we focus on the monocular, RGB-only Multi-Session SLAM setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Several approaches have been proposed to deal with Multi-Session SLAM, however existing solutions typically require additional sensor data in order to remove gauge freedoms and make tracking easier [16, 28, 37]. Only a small number of methods, notably CCM-SLAM [36] and ORBSLAM3 [4], support Multi-Session SLAM from monocular video alone, due to the difficulty of aligning disjoint" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.27, + 0.895, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.389, + 0.892, + 0.416 + ], + "angle": 0, + "content": "Figure 1. Our method estimates camera pose from multiple disconnected video streams." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.506 + ], + "angle": 0, + "content": "sequences under the 7-DOF gauge freedoms in monocular video. However, these approaches are based on classical feature descriptors, making them less accurate on average compared to recent designs based on deep networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.702 + ], + "angle": 0, + "content": "In the standard SLAM setting, Teed and Deng [44] proposed to use a deep optical flow network (RAFT [43]) to track 2D motion, while jointly updating camera poses with a bundle adjustment layer. The method, DROID-SLAM, is accurate and avoids tracking failures, but the design assumes a continuous video stream and is not capable of the wide-baseline matching and non-local optimization necessary for Multi-Session SLAM. Deep Patch Visual Odometry [45] (DPVO) introduced a sparse visual-odometry-only analog of DROID-SLAM which achieves similar accuracy on single-video SLAM, but at much lower cost. However, DPVO also does not support Multi-Session SLAM for the same reasons as DROID-SLAM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.87 + ], + "angle": 0, + "content": "We propose a method for Multi-Session SLAM, capable of both wide-baseline relative pose and visual odometry using a single backbone architecture inspired by [45]. We introduce a differentiable solver layer which minimizes the symmetric epipolar distance (SED) from bi-directional optical flow. From this we construct a method for two-view pose which is capable of matching from far-apart views. This same design can be repurposed for visual odometry by swapping out the solver for bundle adjustment. By employing a unified backbone architecture for both tasks, we enable a simple approach to Multi-Session SLAM." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We evaluate our approach on challenging real-world datasets: EuRoC-MAV [2] and ETH3D [39]. Our sys" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19626" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "tem is more accurate than prior approaches, and is robust to catastrophic failures. We also evaluate our two-view pose method in isolation on the Scannet and Megadeph datasets, and show that it is competitive with transformer-based matching networks. For pairs of far-apart views, our method is capable estimating accurate relative pose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.468, + 0.288 + ], + "angle": 0, + "content": "Our backbone predicts iterative updates to optical flow coupled with a differentiable solver layer for estimating camera pose. This framework, based on RAFT [43], has worked exceptionally well for Visual Odometry [45] and SLAM [44]. By leveraging this idea for wide-baseline matching, we can extend these methods to the Multi-Session setting without introducing substantial complexity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.3, + 0.219, + 0.315 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.468, + 0.379 + ], + "angle": 0, + "content": "Most prior works treat Visual Simultaneous Localization and Mapping (Visual SLAM) as an optimization problem solving for a 3D scene model and camera trajectory which best explains the visual measurements [3]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.468, + 0.563 + ], + "angle": 0, + "content": "Indirect approaches to SLAM perform keypoint matching as a pre-processing step, then estimate a 3D point cloud and camera poses by optimizing the 2D reprojection error over all matches [4, 20, 26, 27, 33, 38]. Reprojection-error is easier to optimize than photometric alignment, making indirect methods robust to lower camera \\( \\text{hz} \\) [44]. Keypoint matching also enables a straightforward approach to estimating two-view relative pose, a necessary step for Visual Multi-Session SLAM [14, 19, 25, 33, 46]. However, keypoint-based SLAM is less robust to low-texture environments compared to those which use photometric alignment [9, 10] or optical flow [44, 45]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.468, + 0.686 + ], + "angle": 0, + "content": "Semi-Indirect approaches similarly optimize 2D reprojection error like indirect methods, but without requiring matches as input [42, 44, 45]. Instead, these approaches alternate between predicting optical flow residuals and performing bundle adjustment. Semi-indirect methods do not require repeatable keypoints across images, making them robust to low-texture settings while retaining the easier reprojection-error objective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.468, + 0.761 + ], + "angle": 0, + "content": "Our approach is most similar to Deep Patch Visual Odometry [45] (DPVO), which is a sparse analog of DROID-SLAM. DPVO predicts sparse optical flow instead of dense, performing similarly to DROID-SLAM while running faster and using half the memory." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Differentiable solver layers for camera pose estimation have been used in order to learn outlier rejection with data-driven training. For two-view relative pose, Ranftl and Koltun [30] used a deep network to learn an iteratively reweighted least-squares algorithm, which solved a weighted variant of the 8-point-algorithm [14] using confidences predicted by the network. [30] required matched points as input, whereas our approach can work from images alone. Roessle and Nießner [31] proposed an" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "end-to-end architecture which used the weighted 8-point-algorithm from [30] on top of matches produced using Superglue [34], and supervised directly on the predicted pose. [31] works well, but is unable to outperform existing methods that use minimal solvers with LO-RANSAC [18] instead of differentiable solvers. In contrast, our approach is not built on top of an existing SOTA 2-view matcher; our design is also iterative, applying a recurrent module and solver layers multiple times to refine the prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Multi-Session SLAM is the task of performing SLAM on multiple trajectories of the same scene. Like SLAM, Multi-Session SLAM is an online task where camera motion is estimated from a stream of images. However, in the Multi-Session setting, there are known breaks in the data-stream over which the small-baseline assumption no longer holds. While local optimization is sufficient for motion tracking from video, estimating wide-baseline camera pose is often non-convex [40] and more challenging due to large viewpoint changes. In monocular visual SLAM, the scale of each sequence is also ambiguous; aligning two trajectories requires estimating 7 degrees-of-freedom (translation, rotation, scale) for all sequences, excluding the first which can be considered the reference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.655 + ], + "angle": 0, + "content": "ORB-SLAM3 [4] performs Multi-Session SLAM by matching between ORB [32] descriptors. While tracking camera motion over each sequence, ORB-SLAM3 continually updates an Atlas of the scene - a lookup table between keypoints and their 3D map point in the reference frame of their original sequence. To align disjoint sequences, candidate cross-sequence image pairs are identified using image-retrieval [12], the 3D map points are obtained from the Atlas, and the relative transformation is found using the Umeyama algorithm [46]. Several methods perform Multi-Session Visual Inertial SLAM [17, 28], however we focus on the visual-only monocular setting where the scale of each session is ambiguous and must be estimated when joining sequences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.733 + ], + "angle": 0, + "content": "CCM-SLAM [35, 36] and SLAMM [7] perform Multi-Session visual SLAM and are built on top of ORBSLAM [27] as well, but are optimized for limited bandwidth and distributed processing, while ORB-SLAM3 performs better and is optimized for accuracy and speed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.739, + 0.608, + 0.755 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Overview: We propose a backbone for matching between \\(\\geq 2\\) views, and then construct a method for multi-session SLAM upon it. Our backbone approaches matching as optical flow; it borrows several ideas from RAFT [43], such as iteratively predicting flow residuals using a recurrent network, and the correlation feature pyramid. Our method maintains a running estimate of both camera pose and bi-directional optical flow, and uses the update operator to refine them both. We provide an overview in Fig. 2. An" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19627" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.088, + 0.87, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.233, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 2. Overview of our backbone. Given a group of \\(\\geq 2\\) frames, our method jointly estimates bi-directional optical flow and camera poses. This module is applicable to (1) wide-baseline two-view matching and (2) visual odometry. We use our backbone for Multi-Session SLAM, which requires the ability to perform both (1) and (2). For each image, we select a set of 2D anchor points and initialize their depth and the camera poses trivially. Our approach then iteratively refines the matches for each anchor, similar to RAFT [43], while updating the camera poses. We alternate between matching and pose updates, where each one informs the update to the other. This entire procedure is repeated several times until convergence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.47, + 0.392 + ], + "angle": 0, + "content": "invariant of our backbone is that the matches are always clamped to plausible values given the pose estimates and an assumption of rigid scene geometry, therefore WLOG the matches can be considered a depth estimate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.471, + 0.56 + ], + "angle": 0, + "content": "At initialization, our backbone produces dense correlation feature pyramids for each image and context features which remain fixed throughout the forward pass. The update operator uses an RNN to predict an update to the matches, and a differentiable solver to update poses. Our backbone treats the two-view and multi-view settings differently: In the two-view setting, the solver updates the poses to minimize the symmetric epipolar distance (SED). In the multi-view setting, the solver updates both poses and depth to minimize the reprojection error. After the solver layer, the matches are adjusted to agree with the poses/depth." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.57, + 0.214, + 0.585 + ], + "angle": 0, + "content": "3.1. Initialization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.47, + 0.82 + ], + "angle": 0, + "content": "Feature Extraction & Feature Pyramid: Similar to RAFT, our method separately extracts context and correlation feature for each image. The context features are provided as input to the recurrent update operator, whereas the correlation features are used to evaluate the visual similarity between any two pixels using a dot product. The context feature maps are produced at \\(1/8\\) resolution using a residual network. We then apply a linear-self-attention residual for long-range feature sharing. The correlation feature maps are also produced using a residual network, but with several exit ramps to produce features at \\(1/2\\), \\(1/4\\), and \\(1/8\\) of the input image resolution. We also average-pool the last one three additional times to produce a correlation feature pyramid with 6 levels. We depict their architecture in the Supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Anchor-Point Selection: Our method predicts sparse optical flow, where matches have one end anchored and the other end free to move in \\(\\mathbb{R}^2\\). We use a mix of detector [8]-chosen and randomly-chosen anchor points. Each point is assigned an initial match in the other image(s) at the same" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.363 + ], + "angle": 0, + "content": "pixel coordinate. For the rest of the paper, we will refer to the \\(k^{th}\\) anchor point as \\(a_{k}\\), and its match in image \\(j\\) as \\(m_{kj}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.424 + ], + "angle": 0, + "content": "We index the context feature map at each \\(a_{k}\\) to produce a unique context feature vector \\(ctx_{k} \\in \\mathbb{R}^{384}\\), which is used by the update operator. Beyond this point, the full context feature map is not used and is discarded to save memory." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.52 + ], + "angle": 0, + "content": "Correlation: Similar to other RAFT-based methods, we use correlation features to assess the visual alignment/similarity given by the current matching estimate. For each \\((a_{k},m_{kj})\\) pair, we bilinearly sample both feature pyramids for their respective frames at locations \\(a_{k}\\) and \\(m_{kj}\\) and take their inner product at each level, producing" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.525, + 0.892, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\langle f _ {k} ^ {i}, f _ {k j} ^ {i} \\rangle \\in \\mathbb {R}: i = 1 \\dots 6 \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.61 + ], + "angle": 0, + "content": "To provide additional spatial context, we also perturb both \\(a_{k}\\) and \\(m_{kj}\\) in \\(3 \\times 3\\) and \\(7 \\times 7\\) grids, respectively, and calculate eq. 1 for all pairs. The resulting feature vector \\(\\mathbf{C}_{kj} \\in \\mathbb{R}^{(6 \\times 3 \\times 3 \\times 7 \\times 7)}\\) is then passed to the update operator." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.669, + 0.636 + ], + "angle": 0, + "content": "3.2. Update Operator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.794 + ], + "angle": 0, + "content": "The update operator produces a revision to all \\( m_{kj} \\) and the estimated relative pose. It consists of three stages, depicted in Fig. 3: (1) An RNN predicts a 2D update to \\( m_{kj} \\) and an associated confidence weight \\( w_{kj} \\). (2) We solve for a pose estimate which is consistent with the newly predicted matches and confidence. (3) We adjust the matches to be physically plausible assuming rigid scene geometry. In the two-view setting, we clamp the matches to the epipolar lines, and in the multi-view setting we reproject the anchor points using the depth and poses from the solver." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Through the recurrent iterations, our method maintains a running state for each anchor-match \\((a_{k},m_{kj})\\) pair, consisting of a hidden state vector \\(h_{kj}\\), a confidence weight \\(w_{kj}\\) and a match location \\(m_{kj}\\) in frame \\(j\\). This state is continually updated, during which \\(m_{kj}\\) should ideally approach the true match location of \\(a_{k}\\) in \\(j\\), and \\(w_{kj}\\) should approach 1. This is what we observe empirically." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19628" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.092, + 0.87, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.272, + 0.895, + 0.313 + ], + "angle": 0, + "content": "Figure 3. A single update iteration (two-view). For each anchor-match pair, we predict an update to the matches using the RNN. We then solve for an update to the camera poses which minimizes the symmetric epipolar distance (SED), producing a new set of epipolar lines. Finally, we clamp the matches back to the best-fit epipolar lines, and repeat the whole process again. In Fig. 8 and the Supplement, we" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.314, + 0.357, + 0.339 + ], + "angle": 0, + "content": "visualize these iterations on real-world images." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.34, + 0.482, + 0.386 + ], + "angle": 0, + "content": "RNN Module: The learnable component of the update operator is the recurrent network which predicts updates to \\( h_{kj} \\), \\( m_{kj} \\) and \\( w_{kj} \\) for all anchor points. We initialize this" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.386, + 0.2, + 0.401 + ], + "angle": 0, + "content": "operator in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.401, + 0.472, + 0.566 + ], + "angle": 0, + "content": "The input to the RNN are the context features, the previous hidden state and the correlation features generated from the current matching estimate, all of which are added together and normalized with layer norm. A self-attention residual is also applied to edges with the same source and destination frame, followed by three gated-residual-units, whose architecture is depicted in the Supplement. The output is an updated hidden state, from which we predict an updated match and confidence using the flow head and confidence head, respectively, which are implemented as two-layer MLPs. The confidence head includes a softmax to" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.566, + 0.218, + 0.582 + ], + "angle": 0, + "content": "\\(\\text{estriel} w_{kj}\\in (0,1)\\)" + }, + { + "type": "image", + "bbox": [ + 0.039, + 0.563, + 0.081, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.591, + 0.159, + 0.598 + ], + "angle": 0, + "content": "Correlation Pyramids" + }, + { + "type": "image", + "bbox": [ + 0.042, + 0.622, + 0.081, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.599, + 0.465, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.742 + ], + "angle": 0, + "content": "Figure 4. The RNN Module. For each anchor-match pair \\((a_{k},m_{kj})\\), it predicts an update to \\(m_{kj}\\) and an associated confidence \\(w_{kj}\\). It also updates a hidden state \\(h_{kj} \\in \\mathbb{R}^{384}\\). Internally, the RNN shares features via attention between updates with the same incoming and outgoing frame." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.416, + 0.773 + ], + "angle": 0, + "content": "3.3. Two-View Solver and Pre-Conditioning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our two-view solver aims to align the relative pose between two frames to the predicted matches by minimizing the distance between each match and its epipolar line. This strategy is accurate, however the objective function is nonconvex, meaning it will only converge to the global minimum if it is initialized close to it. In contrast, the 8-point-algorithm, a common approach to this task, has the opposite problem; it does not suffer from local minimum since it is a" + }, + { + "type": "text", + "bbox": [ + 0.481, + 0.313, + 0.864, + 0.384 + ], + "angle": 0, + "content": "homogeneous least squares, but the solution is sub-optic. In Fig. 6, we visualize how our solver either converges within a fraction of a degree, or not at all, meanwhile" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.386, + 0.832, + 0.4 + ], + "angle": 0, + "content": "8-point-algorithm is more robust but less accurate." + }, + { + "type": "text", + "bbox": [ + 0.481, + 0.402, + 0.892, + 0.537 + ], + "angle": 0, + "content": "The approach our method uses is to pre-condition the pose estimate using a weighted dense variant of the 8-point algorithm [30], and then run our solver layer to refine the pose. This combined strategy obtains the best of both worlds, since the pre-conditioning will typically initialize the pose within the basin of convergence of our solver, which then refines the prediction further. We visualize this basin on the TartanAir [48] dataset using groundtruth matches in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.542, + 0.892, + 0.573 + ], + "angle": 0, + "content": "Pre-conditioning: We solve the homogeneous least squares problem [14]" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.579, + 0.892, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\mathbf {F}} \\overline {{| | d i a g (\\vec {w}) \\mathbf {M} \\vec {\\mathbf {F}} | | ^ {2}}} \\quad s. t. | | \\mathbf {F} | | ^ {2} = 1 \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.669 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}\\) and \\(\\vec{w}\\) are constructed from the anchor-match pairs and confidence predictions. The points are normalized to \\([-1, 1]\\) beforehand. Afterwards, we reshape and uncalibrate \\(\\mathbf{F}\\) to obtain the essential matrix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.746 + ], + "angle": 0, + "content": "We obtain the four relative pose candidates following the procedure detailed in [13] and in the supplement. During training, we select the pose candidate closest to the ground-truth on the \\(SE(3)\\) manifold. During inference, we select the candidate by testing for chirality [13]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "SED Solver Layer: The solver layer seeks to minimize the distance between each match \\(m_{kj}\\), and the epipolar line induced by the current pose estimate and its respective anchor point \\(a_k\\). The solver objective is bi-directional, since every image contains a unique set of anchor points; the objective is also known as the titular symmetric epipolar distance (SED) [11]. We parameterize the pose update as small rotations to the translation direction and orientation: \\(\\xi_{\\mathbf{t}}, \\xi_{\\mathbf{R}} \\in \\mathfrak{so}(3)\\). These are the free variables in eq. 3. Let \\(K(i)\\) be the set of anchor indices for frame \\(i\\), \\(\\mathbf{R}\\) and \\(\\mathbf{t}\\) be the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19629" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.458, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.237, + 0.473, + 0.307 + ], + "angle": 0, + "content": "Figure 5. The flow of gradients through our solver. The updated matches from the RNN are supervised directly, and then detached from the gradient tape. The solver output is supervised with a pose loss. In the backward pass, gradients from the pose loss are used to supervise the confidence head in order to learn outlier rejection." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.322, + 0.47, + 0.368 + ], + "angle": 0, + "content": "output of the pre-conditioning stage, and \\( epi(x,R,t) \\) compute the epipolar line for a point \\( x \\) given the relative pose \\( (R,t) \\). Formally, the minimization objective in our solver is" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.378, + 0.469, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} l _ {k j} = e p i \\left(a _ {k}, \\left(e ^ {\\xi_ {\\mathbf {R}}} \\mathbf {R}\\right) _ {i \\rightarrow j}, \\left(e ^ {\\xi_ {\\mathbf {t}}} \\mathbf {t}\\right) _ {i \\rightarrow j}\\right): k \\in K (i) \\\\ E _ {i \\rightarrow j} = \\sum_ {k \\in K (i)} w _ {k j} \\cdot \\| \\operatorname {e r r} \\left(m _ {k j}, l _ {k j}\\right) \\| _ {2} ^ {2} \\tag {3} \\\\ \\stackrel {\\smile}{S E D} = \\operatorname *{arg min}_{\\xi_{\\mathbf{R}},\\xi_{\\mathbf{t}}}\\left(E_{i\\to j} + E_{j\\to i}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.469, + 0.581 + ], + "angle": 0, + "content": "where \\(err\\) computes the 2D point-to-line error. The predicted weights \\(w_{kj}\\) are included to allow the network to down-weight the contribution of any match which it deems unreliable. This solver layer is implemented in Pytorch, enabling gradients to propagate backward from the pose loss to the confidence head using autograd. We depict the gradient flow in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.469, + 0.642 + ], + "angle": 0, + "content": "To minimize eq. 3, we employ the Levenberg-Marquardt algorithm. This requires computing the Jacobian for all terms. We provide the full derivations in the Supplement, but define the residual function \\( err \\) in its entirety here." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.676 + ], + "angle": 0, + "content": "Let \\( l \\) be the epipolar line produced by \\( (a, e^{\\xi_{\\mathbf{R}}}\\mathbf{R}, e^{\\xi_{\\mathbf{t}}}\\mathbf{t}) \\). We can express each term of eq. 3 as:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.686, + 0.47, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\operatorname {e r r} (m, l) = \\left(\\frac {l _ {x} m _ {x} + l _ {y} m _ {y} + l _ {z}}{l _ {x} ^ {2} + l _ {y} ^ {2}}\\right) \\left[ \\begin{array}{l} l _ {x} \\\\ l _ {y} \\end{array} \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.736, + 0.235, + 0.751 + ], + "angle": 0, + "content": "and the epipolar line as" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.76, + 0.469, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {E} = \\left(e ^ {\\xi_ {\\mathrm {R}}} \\mathbf {R}\\right) ^ {\\top} \\left[ e ^ {\\xi_ {\\mathrm {t}}} \\mathbf {t} \\right] _ {\\times} \\\\ l = \\mathbf {F} \\left[ \\begin{array}{l} a _ {x} \\\\ a _ {y} \\\\ 1 \\end{array} \\right] = \\left[ \\left(K _ {2} ^ {\\top}\\right) ^ {- 1} \\mathbf {E} K _ {1} ^ {\\top} \\right] \\left[ \\begin{array}{l} a _ {x} \\\\ a _ {y} \\\\ 1 \\end{array} \\right] \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.466, + 0.853 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}\\) and \\(\\mathbf{E}\\) are the fundamental and essential matrices." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.863, + 0.45, + 0.88 + ], + "angle": 0, + "content": "3.4. Adapting our backbone to Visual Odometry" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.452, + 0.901 + ], + "angle": 0, + "content": "To adapt our backbone to VO, we make several changes:" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.093, + 0.878, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.218, + 0.892, + 0.288 + ], + "angle": 0, + "content": "Figure 6. Convergence of our solver, given ground-truth matches (from TartanAir [48]). Our SED solver is accurate, but only converges to the global minimum when initialized close to it. We initialize the pose using the 8-pt-algorithm to ensure it is within the SED convergence basin." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.305, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Multi-view Solver (BA): Our local optimizer minimizes reprojection error and treats depth as a separate variable. This is identical to the bundle adjustment from DPVO. Formally, let \\( F \\) be the set of connected frames, \\( G \\) be the global poses, \\( d_{k} \\) be the depth estimate for anchor \\( k \\), and \\( \\Pi(\\cdot) \\) be the \\( 3D \\rightarrow 2D \\) projection function. The bundle adjustment objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.422, + 0.908, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\underset {G, \\mathbf {d}} {\\arg \\min } \\sum_ {(i, j) \\in F} \\sum_ {k \\in K (i)} w _ {k j} \\cdot \\left\\| \\Pi \\left[ G _ {j} ^ {- 1} G _ {i} \\Pi^ {- 1} \\left(a _ {k}, d _ {k}\\right) \\right] - m _ {k j} \\right\\| ^ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.471, + 0.892, + 0.516 + ], + "angle": 0, + "content": "The preconditioning stage is not necessary as we can linearly-extrapolate the camera poses from previous frames to achieve good initialization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.521, + 0.892, + 0.566 + ], + "angle": 0, + "content": "Clamping: Since epipolar lines do not make sense in the multi-view setting, we reset the optical flow using the pose and depth from the bundle adjustment." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.571, + 0.892, + 0.632 + ], + "angle": 0, + "content": "RNN changes: We include mechanisms in the RNN to share latent features between updates if they stem from the same anchor point. Specifically, we use the message passing and temporal convolutions from DPVO [45]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.642, + 0.704, + 0.659 + ], + "angle": 0, + "content": "3.5. Trajectory Alignment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.666, + 0.892, + 0.787 + ], + "angle": 0, + "content": "In the Multi-Session SLAM problem, we estimate a relative Sim(3) between pairs of disjoint trajectories in order to align them. All anchor points already have a depth estimate \\( d_{k} \\) as a result of the bundle adjustment (eq. 6) from the VO system. We use our two-view method to estimate the relative rotation and translation direction, and compute the translation magnitude and relative scaling by comparing the \\( d_{k} \\)'s to the depth from the triangulated two-view matches." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.791, + 0.892, + 0.865 + ], + "angle": 0, + "content": "1) Estimate relative rotation, translation direction, and matches. We first retrieve a candidate image pair \\((i,j)\\) using NetVLAD [1], one from each trajectory, and apply our two-view model to estimate their relative rotation and translation direction." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "2) Align the depth from the two-view and SLAM operators. We previously defined \\( d_{k} \\) as the depth output of the VO" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.791, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19630" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "system. Let \\( d_k' \\) be the triangulated depth obtained from our two-view matches:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.124, + 0.469, + 0.164 + ], + "angle": 0, + "content": "\\[\nd _ {k} ^ {\\prime} = \\left\\{\\begin{array}{l l}\\operatorname {t r i a n g} \\left(\\mathbf {R} _ {i \\rightarrow j}, \\mathbf {t} _ {i \\rightarrow j}, a _ {k}, m _ {k j}\\right)&: k \\in K (i)\\\\\\operatorname {t r i a n g} \\left(\\mathbf {R} _ {j \\rightarrow i}, \\mathbf {t} _ {j \\rightarrow i} a _ {k}, m _ {k i}\\right)&: k \\in K (j)\\end{array}\\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.174, + 0.143, + 0.186 + ], + "angle": 0, + "content": "We solve" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.183, + 0.469, + 0.224 + ], + "angle": 0, + "content": "\\[\n\\underset {s _ {i} \\in \\mathbb {R} > 0} {\\arg \\max } \\sum_ {k \\in K (i)} \\mathbf {1} \\left[ \\frac {1}{\\lambda} < \\frac {d _ {k}}{s _ {i} \\cdot d _ {k} ^ {\\prime}} < \\lambda \\right] \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.23, + 0.469, + 0.365 + ], + "angle": 0, + "content": "to recover the translation magnitude, where the hyperparameter \\(\\lambda = 1.05\\). In layman terms, eq. 9 seeks what translation magnitude would align the most triangulated points to the existing 3D map? We can obtain a reasonably good solution to eq 9 by brute-force checking \\(s_i = d_k / d_k'\\) \\(\\forall k\\in K(i)\\), similar to RANSAC. If the number of inliers is too small, we retry on a new candidate pair. Conversely, the scale-difference is \\(s_i / s_j\\), where \\(s_j\\) is estimated the same way, but for anchor points in \\(K(j)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.366, + 0.469, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\underset {s _ {j} \\in \\mathbb {R} > 0} {\\arg \\max } \\sum_ {k \\in K (j)} \\mathbf {1} \\left[ \\frac {1}{\\lambda} < \\frac {d _ {k}}{s _ {j} \\cdot d _ {k} ^ {\\prime}} < \\lambda \\right] \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.42, + 0.469, + 0.464 + ], + "angle": 0, + "content": "Given the rotation \\(\\mathbf{R}_{j\\rightarrow i}\\) and translation \\(\\mathbf{t}_{j\\rightarrow i}\\) from the two-view solver layers, the resulting transformation can be used to align the two trajectories:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.469, + 0.469, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {j \\rightarrow i} = \\left[\\begin{array}{c c}\\frac {s _ {i}}{s _ {j}} \\mathbf {R}&s _ {j} \\mathbf {t}\\\\0&1\\end{array}\\right] \\in \\operatorname {S i m} (3) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.507, + 0.452, + 0.522 + ], + "angle": 0, + "content": "The graphs are then merged by concatenating all buffers." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.531, + 0.182, + 0.547 + ], + "angle": 0, + "content": "3.6. Training" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.554, + 0.469, + 0.6 + ], + "angle": 0, + "content": "We train our backbone separately for VO and for two-view pose. It is trained using a matching loss and a pose loss. The pose loss for the two-view training" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.611, + 0.469, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p o s e 2 V}} = \\sum_ {t = 1} ^ {1 2} \\cos^ {- 1} \\left(\\bar {\\mathbf {t}} ^ {t} \\cdot \\bar {\\mathbf {t}} ^ {\\bar {\\mathbf {g}}} ^ {t}\\right) + \\alpha \\left\\| \\left[ \\left(\\mathbf {R} ^ {t}\\right) ^ {T} \\mathbf {R} ^ {\\mathbf {g}} ^ {t} \\right] \\right\\| _ {S O (3)} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.661, + 0.469, + 0.72 + ], + "angle": 0, + "content": "penalizes angle error for both the predicted translation direction and predicted orientation. The pose loss for the VO training is the SE3 manifold distance between the predicted and ground-truth poses:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.722, + 0.469, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p o s e V O}} = \\sum_ {t = 1} ^ {T} \\left\\| \\left[ \\left(\\mathbf {G} ^ {t}\\right) ^ {- 1} \\mathbf {G} ^ {\\mathbf {g t}} \\right] \\right\\| _ {S E (3)} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.767, + 0.202, + 0.78 + ], + "angle": 0, + "content": "The matching loss" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.781, + 0.469, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a t c h i n g}} = \\frac {1}{| N |} \\sum_ {a} ^ {N} \\sum_ {t = 1} ^ {T} \\left\\| m _ {a} ^ {t} - m _ {a} ^ {\\mathbf {g t}} \\right\\| _ {2} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.469, + 0.902 + ], + "angle": 0, + "content": "is a standard endpoint-error [24, 43], applied only on pixels which have valid depth and are verified to be visible in both images. Our update operator is applied 12 times per training example; supervision is applied after every update (See Fig. 5). The final loss is \\(\\mathcal{L} = \\mathcal{L}_{\\text{pose}} + \\beta \\cdot \\mathcal{L}_{\\text{matching}}\\)." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.091, + 0.63, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.215, + 0.628, + 0.227 + ], + "angle": 0, + "content": "Stage 1: Image Retrieval" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.092, + 0.768, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.212, + 0.737, + 0.23 + ], + "angle": 0, + "content": "Stage 2: Sim(3) Estimation" + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.098, + 0.884, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.214, + 0.888, + 0.227 + ], + "angle": 0, + "content": "Stage 3: Graph Merging" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.242, + 0.892, + 0.325 + ], + "angle": 0, + "content": "Figure 7. High-level overview of our approach to aligning disjoint trajectories. We use NetVLAD [1] to retrieve two co-visible images, one from each trajectory, and apply our two-view update operator to estimate a relative Sim(3) alignment, following the procedure detailed in Sec 3.5. We then transform the second trajectory into the reference frame of the first, and merge all buffers." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.341, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Following prior work [23], we pre-train our two-view method on synthetic homographies (without pose loss) for two epochs on the Oxford-Paris 1M Distractors dataset [29]. We then tune our model on a \\(50/50\\) mixture of Scannet [6] and Megadepth [21] for 100,000 steps and a batch size of 120, using 10 A6000 GPUs for 5 days. The VO backbone is trained using the procedure from [45] on TartanAir." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.461, + 0.756, + 0.477 + ], + "angle": 0, + "content": "3.7. Multi-Session SLAM System" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.608 + ], + "angle": 0, + "content": "Overview: To perform Multi-Session SLAM, we use our VO-trained backbone to perform visual odometry and global optimization; our two-view backbone connects disjoint trajectories following the procedure in Sec 3.5. Our full system builds the pose graph incrementally: new video frames inserted into the graph as they are received, and unjoined trajectories are aligned and merged as soon as a connection is found." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.736 + ], + "angle": 0, + "content": "VO Frontend: Our VO frontend extracts features and estimates camera poses and depth for incoming frames, operating on a sliding window covering the most recently observed 20 keyframes. Keyframing occurs retroactively on the \\(4^{th}\\) oldest keyframe if it is found to be redundant. New poses are initialized using a linear-motion model, and depth is copied from the previous frame. Like DPVO [45], we initialize after 8 frames with significant motion are observed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.892, + 0.834 + ], + "angle": 0, + "content": "Global Optimization: We perform global optimization by introducing proximity factors using the existing pose and depth measurements, following the approach from DROID-SLAM [44], and update the matches and poses/depth using the update operator. We only run the backend periodically, and after joining trajectories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Trajectory Joining: We query the database of NetVLAD descriptors to find co-visible frames. If several pairs are retrieved, we compute relative poses in a mini-batch and use the estimate with the highest inlier-ratio." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19631" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.24 + ], + "angle": 0, + "content": "
Overall ApproachPose error AUC [%] ↑
@5°@10°@20°
Superglue [34]Matching16.2(17.7)33.8(35.6)51.8(54.2)
LightGlue [23]16.5(19.4)33.4(36.9)50.1(53.5)
LoFTR [41]↓22.1(25.7)40.1(45.0)47.6(61.4)
MatchFormer [47]RANSAC [15]24.3(27.3)43.9(47.6)61.4(64.9)
ASpanFormer [5](LO-RANSAC [18])25.6(28.4)46.0(48.8)63.3(65.8)
Roessle&Nießner [31]Matching→W8PA20.741.661.7
Roessle&Nießner [31]Matching→W8PA→BA25.747.266.4
OursOptical-Flow\n←\nClamp ← Solver30.951.167.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.334 + ], + "angle": 0, + "content": "Table 1. Two-view results on Scannet [6]. Existing methods perform matching as pre-processing, then use a solver; our method is iterative. Results in parenthesis use the LO-RANSAC [18] estimator. \"W8PA \\(\\rightarrow\\) BA\" = weighted 8-point-algo. & bundle-adjustment. Our design outperforms prior approaches. We use the same model weights as in Tab. 2." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.344, + 0.467, + 0.494 + ], + "angle": 0, + "content": "
Overall ApproachPose error AUC [%] ↑
@5°@10°@20°
Superglue [34]Matching49.7(65.8)67.1(78.7)80.6(87.5)
LightGlue [23]49.9(66.7)67.0(79.3)80.1(87.9)
LoFTR [41]↓52.8(66.4)69.2(78.6)81.2(86.5)
MatchFormer [47]RANSAC [15]53.3(66.5)69.7(78.9)81.8(87.5)
ASpanFormer [5](LO-RANSAC [18])55.3(69.4)71.5(81.1)83.1(88.9)
Roessle&Nießner [31]Matching→W8PA46.962.876.3
Roessle&Nießner [31]Matching→W8PA→BA61.274.985.0
OursOptical-Flow60.472.781.3
Clamp ← Solver
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.628 + ], + "angle": 0, + "content": "Table 2. Two-view results on Megadepth [21]. Existing methods perform matching as pre-processing, then use a solver; our method is iterative. \"W8PA \\(\\rightarrow\\) BA\" = weighted 8-point-algo. & bundle-adjustment. Results in parenthesis use the LO-RANSAC [18] estimator. Our radically different design leads to better results indoors (Tab. 1), but is less competitive in photo-tourism settings where high-volume matching is easier and therefore RANSAC-based methods work better. We use the same model weights as in Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.639, + 0.208, + 0.656 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.469, + 0.696 + ], + "angle": 0, + "content": "We evaluate the performance of our two-view method in isolation, and our Multi-Session SLAM system as a whole." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.701, + 0.279, + 0.715 + ], + "angle": 0, + "content": "4.1. Two-View Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.469, + 0.854 + ], + "angle": 0, + "content": "We evaluate our two-view method on two popular relative pose benchmarks: The Scannet [6] 1500 and MegaDepth [21] 1500 test datasets. We compare to existing matching networks [5, 23, 34, 41, 47], which are typically in service of improving COLMAP [38]-based SfM pipelines [22, 33]. In contrast, our approach is in service of Multi-Session SLAM. We report pose error AUC, as is done in prior matching work (higher is better). We use the same configuration and model weights on both datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Most existing methods perform matching as a preprocessing step using a deep network, then estimate pose with a minimal solver using RANSAC [15, 18]. Roessle" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "and Nießner [31] replace the RANSAC optimizer with a weighted-8-point algorithm and bundle adjustment. In contrast, our two-view method couples the prediction of pose and matches, refining the predictions over many iterations. We ablate over our two-view model in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.171, + 0.892, + 0.23 + ], + "angle": 0, + "content": "Scannet: We report two-view results on Scannet [6] in Tab. 1. Our approach outperforms existing methods on Scannet, which contains fewer salient keypoints and many texture-less surfaces and motion blur." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.892, + 0.34 + ], + "angle": 0, + "content": "MegaDepth: We report two-view results on Megadepth [21] in Tab. 2. On Megadepth, the matching problem is substantially easier since photo-tourism images contain many salient keypoints, resulting in better performance across all methods; our approach is on-par with existing approaches using the OpenCV RANSAC pose estimator [15], but not the LO-RANSAC [18] estimator." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.346, + 0.782, + 0.36 + ], + "angle": 0, + "content": "4.2. Multi-Session SLAM evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.532 + ], + "angle": 0, + "content": "We evaluate our approach on Multi-Session SLAM on the EuRoC-MAV [2] and ETH3D [39] datasets, since they provide camera poses under a single global reference. Following the evaluation for single-video SLAM, we align the final predictions to the ground truth by computing a global 7-DOF alignment to account for the guage freedoms. We report RMSE ATE[m] in metric units. We sample 96 anchors in each video frame using a mix of random and Superpoint [8] keypoints. Our two-view method uses those same anchors to join pairs of trajectories. We use the same weights and configuration on both datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.793 + ], + "angle": 0, + "content": "EuRoC-MAV: In Tab. 9 and Fig. 10, we report results on the EuRoC MAV dataset [2]. Following the evaluation from [4], we compare on four groups of disjoint trajectories across the three distinct environments: 3 in the Machine Hall, 5 in the Machine Hall, 3 in Vicon 1, and 3 in Vicon 2. Ground-truth poses are obtained using a laser tracker and Vicon cameras. These video sequences are long, with the entirety of the Machine-Hall sequence being 11.5 minutes of video (13.7K frames) spread across 5 sequences. Our approach achieves significantly lower error than ORBSLAM3 on all groups, including 12x lower error on the Vicon 2 sequences (0.022 vs 0.284). Video is recorded at 20-FPS; all methods run in real-time. We also compare to CCM-SLAM [36], which only provides results on the on MH01-03 sequence, the mono-inertial ORB-SLAM3, and the mono-inertial VINS [28]. We outperform the other approaches on all trajectory groups." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "ETH3D: In Tab. 4, we report results on the training set from ETH3D [39]. We compare across 5 unique scenes composed of multiple trajectories: 1-4 from Sofa, 3&4 from Table, 1-3 from Plant Scene, 1&2 from Einstein, and 2&3 from Planar. Sequences are excluded if they belong to the official ETH3D test set (with no ground-truth) or are from a different scene altogether. The sequences are trimmed" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19632" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.891, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.893, + 0.298 + ], + "angle": 0, + "content": "Figure 8. Qualitative results on Scannet. Our two-view variant is able to estimate accurate relative poses across wide camera baselines. It initializes all matches with uniform depth and identity relative pose. Progressive applications of our update operator lead to more accurate matches and higher predicted confidence. Some image pairs (row 2) take more iterations to converge than others (row 1)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.31, + 0.468, + 0.469 + ], + "angle": 0, + "content": "
Scene nameMH01-03MH01-05V101-103V201-203Cam.
# Disjoint Trajectories3533hz
Ours0.0220.0360.0300.02220
Mono-Visual
CCM-SLAM [36]0.077---38
Mono-Visual
ORB-SLAM3 [4]0.0300.0580.0580.28420
Mono-Visual
VINS [28]-0.210---
Mono-Inertial
ORB-SLAM30.0370.0650.0400.04820
Mono-Inertial
" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.478, + 0.47, + 0.547 + ], + "angle": 0, + "content": "Figure 9. Multi-Session SLAM evaluation on EuRoC MAV [2] using RMSE ATE[m] \\(\\downarrow\\). We perform similarly to ORB-SLAM3 (within \\(2mm\\)) in the Machine Hall, but perform best in the Vicon 1 and Vicon 2 rooms. Results for baseline approaches were obtained from [4]." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.555, + 0.457, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.79, + 0.47, + 0.817 + ], + "angle": 0, + "content": "Figure 10. Our prediction on the Vicon1 [2] sequences. Our method is accurate and robust to chaotic motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "from a single, longer video; to make joining the trajectories non-trivial, we reverse every other sequence, ensuring that there is a sufficient disconnect between subsequent videos. This is a novel benchmark, so [28, 36] do not report results." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.311, + 0.89, + 0.435 + ], + "angle": 0, + "content": "
Ablation ExperimentPose error AUC @10°[%] ↑
Megadepth ValScannet Val
BaselineFull model53.333.9
ArchitectureShaller correlation pyramid44.327.1
No attention in update-op48.630.5
No ReLU-attn in feat-extractor47.429.7
SolverNo pre-conditioning35.520.6
No SED solver23.813.1
No clamping step21.910.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.443, + 0.893, + 0.512 + ], + "angle": 0, + "content": "Table 3. Two-view ablation experiments on two validation sets. Our proposed RNN and feature extractor changes from DPVO improve the result. Removing components from our proposed solver design decreases performance, as does removing the reprojection step." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.527, + 0.89, + 0.615 + ], + "angle": 0, + "content": "
Scene nameSofaTablePlant SceneEinsteinPlanar
# Disjoint Trajectories42322
ORB-SLAM3 [4] \nMono-VisualFAIL \n(no init)0.018FAIL \n(init→lost)FAIL \n(init→lost)0.010
Ours \nMono-Visual0.0130.0100.0090.0050.035
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.623, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Table 4. Multi-Session SLAM evaluation on ETH3D [39] using RMSE ATE[m] \\( \\downarrow \\) . Our method is robust to catastrophic failures." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.666, + 0.892, + 0.743 + ], + "angle": 0, + "content": "ORB-SLAM3 fails on Sofa, Plant Scene, and Einstein for various reasons (couldn't initialize, or lost the feature tracks). Our approach outperforms ORB-SLAM3 on 4/5 groups, and does not fail on any. ORB-SLAM3 only succeeds on 2/5 groups." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.76, + 0.619, + 0.775 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce a new method for mono-visual Multi-Session SLAM. Our system utilizes a novel backbone which can estimate two-view pose and perform visual odometry. We leverage a novel differentiable solver which minimizes the symmetric epipolar distance. We compare against existing approaches and show strong performance across several datasets. This work was partially supported by IARPA and the National Science Foundation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19633" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pa-jdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5297-5307, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[2] Michael Burri, Janosch Nikolic, Pascal Gohl, Thomas Schneider, Joern Rehder, Sammy Omari, Markus W Achtelik, and Roland Siegwart. The euroc micro aerial vehicle datasets. The International Journal of Robotics Research, 35 (10):1157-1163, 2016. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.471, + 0.327 + ], + "angle": 0, + "content": "[3] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.471, + 0.398 + ], + "angle": 0, + "content": "[4] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.471, + 0.47 + ], + "angle": 0, + "content": "[5] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming-min Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.471, + 0.542 + ], + "angle": 0, + "content": "[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.543, + 0.471, + 0.598 + ], + "angle": 0, + "content": "[7] Hayyan Afeef Daoud, Aznul Qalid Md. Sabri, Chu Kiong Loo, and Ali Mohammed Mansoor. Slamm: Visual monocular slam with continuous mapping using multiple maps. PloS one, 13(4):e0195878, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.601, + 0.471, + 0.669 + ], + "angle": 0, + "content": "[8] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.672, + 0.471, + 0.724 + ], + "angle": 0, + "content": "[9] Jakob Engel, Thomas Schöps, and Daniel Cremers. Lsd-slam: Large-scale direct monocular slam. In European conference on computer vision, pages 834-849. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[10] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. IEEE transactions on pattern analysis and machine intelligence, 40(3):611-625, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[11] Mohammed E Fathy, Ashraf S Hussein, and Mohammed F Tolba. Fundamental matrix estimation: A study of error criteria. Pattern Recognition Letters, 32(2):383-391, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[12] Dorian Gálvez-López and Juan D Tardos. Bags of binary words for fast place recognition in image sequences. IEEE Transactions on Robotics, 28(5):1188-1197, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, New York, NY, USA, 2 edition, 2003. 4, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[14] Richard I Hartley. In defense of the eight-point algorithm. IEEE Transactions on pattern analysis and machine intelligence, 19(6):580-593, 1997. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[15] Itseez. Open source computer vision library. https://github.com/itseez/opencv, 2015.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[16] Mathieu Labbe and François Michaud. Rtab-map as an opensource lidar and visual simultaneous localization and mapping library for large-scale and long-term online operation. Journal of field robotics, 36(2):416-446, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[17] Mathieu Labbe and François Michaud. Multi-session visual slam for illumination-invariant re-localization in indoor environments. Frontiers in Robotics and AI, 9:801886, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[18] Viktor Larsson and contributors. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[19] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[20] Stefan Leutenegger, Paul Furgale, Vincent Rabaud, Margarita Chli, Kurt Konolige, and Roland Siegwart. Keyframe-based visual-inertial slam using nonlinear optimization. Proceedings of Robotics Science and Systems (RSS) 2013, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[21] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Computer Vision and Pattern Recognition (CVPR), 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.894, + 0.546 + ], + "angle": 0, + "content": "[22] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987-5997, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[23] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In ICCV, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[24] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[25] Xiao Xin Lu. A review of solutions for perspective-n-point problem in camera pose estimation. In Journal of Physics: Conference Series, page 052009. IOP Publishing, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[26] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[27] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[28] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics, 34(4):1004-1020, 2018. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] F. Radenović, A. Iscen, G. Tolias, Y. Avrithis, and O. Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In CVPR, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19634" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[30] Rene Ranftl and Vladlen Koltun. Deep fundamental matrix estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[31] Barbara Roessle and Matthias Nießner. End2end multi-view feature matching with differentiable pose optimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 477-487, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.247 + ], + "angle": 0, + "content": "[32] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.47, + 0.319 + ], + "angle": 0, + "content": "[33] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.471, + 0.39 + ], + "angle": 0, + "content": "[34] 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. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[35] Patrik Schmuck and Margarita Chli. Multi-uav collaborative monocular slam. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3863-3870. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[36] Patrik Schmuck and Margarita Chli. Ccm-slam: Robust and efficient centralized collaborative monocular simultaneous localization and mapping for robotic teams. Journal of Field Robotics, 36(4):763-781, 2019. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.574 + ], + "angle": 0, + "content": "[37] Thomas Schneider, Marcin Dymczyk, Marius Fehr, Kevin Egger, Simon Lynen, Igor Gilitschenski, and Roland Siegwart. maplab: An open framework for research in visual-inertial mapping and localization. IEEE Robotics and Automation Letters, 3(3):1418-1425, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.47, + 0.618 + ], + "angle": 0, + "content": "[38] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[39] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[40] Hauke Strasdat, J Montiel, and Andrew J Davison. Scale drift-aware large scale monocular slam. Robotics: science and Systems VI, 2(3):7, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[41] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[42] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[43] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 1, 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[44] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.192 + ], + "angle": 0, + "content": "[45] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.893, + 0.247 + ], + "angle": 0, + "content": "[46] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.893, + 0.317 + ], + "angle": 0, + "content": "[47] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, pages 2746-2762, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.319, + 0.893, + 0.402 + ], + "angle": 0, + "content": "[48] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19635" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/04406ae4-0193-431d-9be9-c73afd625413_origin.pdf b/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/04406ae4-0193-431d-9be9-c73afd625413_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4250be66cbb6eb3e7dd74363ec5611dc049e9f93 --- /dev/null +++ b/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/04406ae4-0193-431d-9be9-c73afd625413_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edb7b7730c4a9a2f6d1da548de1d898d6a837d241e62f7ad64540ad104f67934 +size 4617402 diff --git a/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/full.md b/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b356b6e8c5e5151044ca8477b68118d45af90517 --- /dev/null +++ b/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/full.md @@ -0,0 +1,379 @@ +# Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization + +Lahav Lipson + +Princeton University + +llipson@princeton.edu + +Jia Deng + +Princeton University + +jiadeng@princeton.edu + +# Abstract + +We introduce a new system for Multi-Session SLAM, which tracks camera motion across multiple disjoint videos under a single global reference. Our approach couples the prediction of optical flow with solver layers to estimate camera pose. The backbone is trained end-to-end using a novel differentiable solver for wide-baseline two-view pose. The full system can connect disjoint sequences, perform visual odometry, and global optimization. Compared to existing approaches, our design is accurate and robust to catastrophic failures. Code is available at https://github.com/princeton-v1/MultiSlam_DiffPose + +# 1. Introduction + +Simultaneous Localization and Mapping (SLAM) is the task of estimating camera motion and a 3D map from video. The standard setup assumes a single continuous video. However, video data in the wild often consists of not a single continuous stream, but rather multiple disjoint sessions, either deliberately such as in collaborative mapping when multiple robots perform joint rapid 3D reconstruction, or inadvertently due to visual discontinuities in the video stream which can result from camera failures, extreme parallax, rapid turns, auto-exposure lag, dark areas, or extreme occlusion by dynamic objects. Handling such disjoint videos is important for many applications in AR and robotics, and gives rise to the task of Multi-Session SLAM. + +In Multi-Session SLAM, the input consists of multiple disjoint video sequences and the goal is to estimate camera poses for all video frames under a single global reference. This is in contrast to "single-video SLAM", whose input is a single continuous video. In this work, we focus on the monocular, RGB-only Multi-Session SLAM setting. + +Several approaches have been proposed to deal with Multi-Session SLAM, however existing solutions typically require additional sensor data in order to remove gauge freedoms and make tracking easier [16, 28, 37]. Only a small number of methods, notably CCM-SLAM [36] and ORBSLAM3 [4], support Multi-Session SLAM from monocular video alone, due to the difficulty of aligning disjoint + +![](images/4891576479077935d55e89ddf5bb9d1bf95a249293d965daacfe5934c762b57b.jpg) +Figure 1. Our method estimates camera pose from multiple disconnected video streams. + +sequences under the 7-DOF gauge freedoms in monocular video. However, these approaches are based on classical feature descriptors, making them less accurate on average compared to recent designs based on deep networks. + +In the standard SLAM setting, Teed and Deng [44] proposed to use a deep optical flow network (RAFT [43]) to track 2D motion, while jointly updating camera poses with a bundle adjustment layer. The method, DROID-SLAM, is accurate and avoids tracking failures, but the design assumes a continuous video stream and is not capable of the wide-baseline matching and non-local optimization necessary for Multi-Session SLAM. Deep Patch Visual Odometry [45] (DPVO) introduced a sparse visual-odometry-only analog of DROID-SLAM which achieves similar accuracy on single-video SLAM, but at much lower cost. However, DPVO also does not support Multi-Session SLAM for the same reasons as DROID-SLAM. + +We propose a method for Multi-Session SLAM, capable of both wide-baseline relative pose and visual odometry using a single backbone architecture inspired by [45]. We introduce a differentiable solver layer which minimizes the symmetric epipolar distance (SED) from bi-directional optical flow. From this we construct a method for two-view pose which is capable of matching from far-apart views. This same design can be repurposed for visual odometry by swapping out the solver for bundle adjustment. By employing a unified backbone architecture for both tasks, we enable a simple approach to Multi-Session SLAM. + +We evaluate our approach on challenging real-world datasets: EuRoC-MAV [2] and ETH3D [39]. Our sys + +tem is more accurate than prior approaches, and is robust to catastrophic failures. We also evaluate our two-view pose method in isolation on the Scannet and Megadeph datasets, and show that it is competitive with transformer-based matching networks. For pairs of far-apart views, our method is capable estimating accurate relative pose. + +Our backbone predicts iterative updates to optical flow coupled with a differentiable solver layer for estimating camera pose. This framework, based on RAFT [43], has worked exceptionally well for Visual Odometry [45] and SLAM [44]. By leveraging this idea for wide-baseline matching, we can extend these methods to the Multi-Session setting without introducing substantial complexity. + +# 2. Related Work + +Most prior works treat Visual Simultaneous Localization and Mapping (Visual SLAM) as an optimization problem solving for a 3D scene model and camera trajectory which best explains the visual measurements [3]. + +Indirect approaches to SLAM perform keypoint matching as a pre-processing step, then estimate a 3D point cloud and camera poses by optimizing the 2D reprojection error over all matches [4, 20, 26, 27, 33, 38]. Reprojection-error is easier to optimize than photometric alignment, making indirect methods robust to lower camera $\text{hz}$ [44]. Keypoint matching also enables a straightforward approach to estimating two-view relative pose, a necessary step for Visual Multi-Session SLAM [14, 19, 25, 33, 46]. However, keypoint-based SLAM is less robust to low-texture environments compared to those which use photometric alignment [9, 10] or optical flow [44, 45]. + +Semi-Indirect approaches similarly optimize 2D reprojection error like indirect methods, but without requiring matches as input [42, 44, 45]. Instead, these approaches alternate between predicting optical flow residuals and performing bundle adjustment. Semi-indirect methods do not require repeatable keypoints across images, making them robust to low-texture settings while retaining the easier reprojection-error objective. + +Our approach is most similar to Deep Patch Visual Odometry [45] (DPVO), which is a sparse analog of DROID-SLAM. DPVO predicts sparse optical flow instead of dense, performing similarly to DROID-SLAM while running faster and using half the memory. + +Differentiable solver layers for camera pose estimation have been used in order to learn outlier rejection with data-driven training. For two-view relative pose, Ranftl and Koltun [30] used a deep network to learn an iteratively reweighted least-squares algorithm, which solved a weighted variant of the 8-point-algorithm [14] using confidences predicted by the network. [30] required matched points as input, whereas our approach can work from images alone. Roessle and Nießner [31] proposed an + +end-to-end architecture which used the weighted 8-point-algorithm from [30] on top of matches produced using Superglue [34], and supervised directly on the predicted pose. [31] works well, but is unable to outperform existing methods that use minimal solvers with LO-RANSAC [18] instead of differentiable solvers. In contrast, our approach is not built on top of an existing SOTA 2-view matcher; our design is also iterative, applying a recurrent module and solver layers multiple times to refine the prediction. + +Multi-Session SLAM is the task of performing SLAM on multiple trajectories of the same scene. Like SLAM, Multi-Session SLAM is an online task where camera motion is estimated from a stream of images. However, in the Multi-Session setting, there are known breaks in the data-stream over which the small-baseline assumption no longer holds. While local optimization is sufficient for motion tracking from video, estimating wide-baseline camera pose is often non-convex [40] and more challenging due to large viewpoint changes. In monocular visual SLAM, the scale of each sequence is also ambiguous; aligning two trajectories requires estimating 7 degrees-of-freedom (translation, rotation, scale) for all sequences, excluding the first which can be considered the reference. + +ORB-SLAM3 [4] performs Multi-Session SLAM by matching between ORB [32] descriptors. While tracking camera motion over each sequence, ORB-SLAM3 continually updates an Atlas of the scene - a lookup table between keypoints and their 3D map point in the reference frame of their original sequence. To align disjoint sequences, candidate cross-sequence image pairs are identified using image-retrieval [12], the 3D map points are obtained from the Atlas, and the relative transformation is found using the Umeyama algorithm [46]. Several methods perform Multi-Session Visual Inertial SLAM [17, 28], however we focus on the visual-only monocular setting where the scale of each session is ambiguous and must be estimated when joining sequences. + +CCM-SLAM [35, 36] and SLAMM [7] perform Multi-Session visual SLAM and are built on top of ORBSLAM [27] as well, but are optimized for limited bandwidth and distributed processing, while ORB-SLAM3 performs better and is optimized for accuracy and speed. + +# 3. Approach + +Overview: We propose a backbone for matching between $\geq 2$ views, and then construct a method for multi-session SLAM upon it. Our backbone approaches matching as optical flow; it borrows several ideas from RAFT [43], such as iteratively predicting flow residuals using a recurrent network, and the correlation feature pyramid. Our method maintains a running estimate of both camera pose and bi-directional optical flow, and uses the update operator to refine them both. We provide an overview in Fig. 2. An + +![](images/b2f41baebe36e1b30f6fab21abdcb137a14a168bc840a97406c1622d93fdf413.jpg) +Figure 2. Overview of our backbone. Given a group of $\geq 2$ frames, our method jointly estimates bi-directional optical flow and camera poses. This module is applicable to (1) wide-baseline two-view matching and (2) visual odometry. We use our backbone for Multi-Session SLAM, which requires the ability to perform both (1) and (2). For each image, we select a set of 2D anchor points and initialize their depth and the camera poses trivially. Our approach then iteratively refines the matches for each anchor, similar to RAFT [43], while updating the camera poses. We alternate between matching and pose updates, where each one informs the update to the other. This entire procedure is repeated several times until convergence. + +invariant of our backbone is that the matches are always clamped to plausible values given the pose estimates and an assumption of rigid scene geometry, therefore WLOG the matches can be considered a depth estimate. + +At initialization, our backbone produces dense correlation feature pyramids for each image and context features which remain fixed throughout the forward pass. The update operator uses an RNN to predict an update to the matches, and a differentiable solver to update poses. Our backbone treats the two-view and multi-view settings differently: In the two-view setting, the solver updates the poses to minimize the symmetric epipolar distance (SED). In the multi-view setting, the solver updates both poses and depth to minimize the reprojection error. After the solver layer, the matches are adjusted to agree with the poses/depth. + +# 3.1. Initialization + +Feature Extraction & Feature Pyramid: Similar to RAFT, our method separately extracts context and correlation feature for each image. The context features are provided as input to the recurrent update operator, whereas the correlation features are used to evaluate the visual similarity between any two pixels using a dot product. The context feature maps are produced at $1/8$ resolution using a residual network. We then apply a linear-self-attention residual for long-range feature sharing. The correlation feature maps are also produced using a residual network, but with several exit ramps to produce features at $1/2$ , $1/4$ , and $1/8$ of the input image resolution. We also average-pool the last one three additional times to produce a correlation feature pyramid with 6 levels. We depict their architecture in the Supplement. + +Anchor-Point Selection: Our method predicts sparse optical flow, where matches have one end anchored and the other end free to move in $\mathbb{R}^2$ . We use a mix of detector [8]-chosen and randomly-chosen anchor points. Each point is assigned an initial match in the other image(s) at the same + +pixel coordinate. For the rest of the paper, we will refer to the $k^{th}$ anchor point as $a_{k}$ , and its match in image $j$ as $m_{kj}$ . + +We index the context feature map at each $a_{k}$ to produce a unique context feature vector $ctx_{k} \in \mathbb{R}^{384}$ , which is used by the update operator. Beyond this point, the full context feature map is not used and is discarded to save memory. + +Correlation: Similar to other RAFT-based methods, we use correlation features to assess the visual alignment/similarity given by the current matching estimate. For each $(a_{k},m_{kj})$ pair, we bilinearly sample both feature pyramids for their respective frames at locations $a_{k}$ and $m_{kj}$ and take their inner product at each level, producing + +$$ +\langle f _ {k} ^ {i}, f _ {k j} ^ {i} \rangle \in \mathbb {R}: i = 1 \dots 6 \tag {1} +$$ + +To provide additional spatial context, we also perturb both $a_{k}$ and $m_{kj}$ in $3 \times 3$ and $7 \times 7$ grids, respectively, and calculate eq. 1 for all pairs. The resulting feature vector $\mathbf{C}_{kj} \in \mathbb{R}^{(6 \times 3 \times 3 \times 7 \times 7)}$ is then passed to the update operator. + +# 3.2. Update Operator + +The update operator produces a revision to all $m_{kj}$ and the estimated relative pose. It consists of three stages, depicted in Fig. 3: (1) An RNN predicts a 2D update to $m_{kj}$ and an associated confidence weight $w_{kj}$ . (2) We solve for a pose estimate which is consistent with the newly predicted matches and confidence. (3) We adjust the matches to be physically plausible assuming rigid scene geometry. In the two-view setting, we clamp the matches to the epipolar lines, and in the multi-view setting we reproject the anchor points using the depth and poses from the solver. + +Through the recurrent iterations, our method maintains a running state for each anchor-match $(a_{k},m_{kj})$ pair, consisting of a hidden state vector $h_{kj}$ , a confidence weight $w_{kj}$ and a match location $m_{kj}$ in frame $j$ . This state is continually updated, during which $m_{kj}$ should ideally approach the true match location of $a_{k}$ in $j$ , and $w_{kj}$ should approach 1. This is what we observe empirically. + +![](images/ab4da5dbd964d6bd8253334e6fbcea6680d6f1c2050d25785bf408754cd779a8.jpg) +Figure 3. A single update iteration (two-view). For each anchor-match pair, we predict an update to the matches using the RNN. We then solve for an update to the camera poses which minimizes the symmetric epipolar distance (SED), producing a new set of epipolar lines. Finally, we clamp the matches back to the best-fit epipolar lines, and repeat the whole process again. In Fig. 8 and the Supplement, we + +visualize these iterations on real-world images. + +RNN Module: The learnable component of the update operator is the recurrent network which predicts updates to $h_{kj}$ , $m_{kj}$ and $w_{kj}$ for all anchor points. We initialize this + +operator in Fig. 4. + +The input to the RNN are the context features, the previous hidden state and the correlation features generated from the current matching estimate, all of which are added together and normalized with layer norm. A self-attention residual is also applied to edges with the same source and destination frame, followed by three gated-residual-units, whose architecture is depicted in the Supplement. The output is an updated hidden state, from which we predict an updated match and confidence using the flow head and confidence head, respectively, which are implemented as two-layer MLPs. The confidence head includes a softmax to + +$\text{estriel} w_{kj}\in (0,1)$ + +![](images/7f5f718f01f4a5b00c7f12616c4796de2611b4357652e20d602376f9f7d7eeda.jpg) +Correlation Pyramids + +![](images/115da04c3ca4fafc617bdfec51ba3c455883c48ed58f3f880cfbeeee325f1cf8.jpg) +Figure 4. The RNN Module. For each anchor-match pair $(a_{k},m_{kj})$ , it predicts an update to $m_{kj}$ and an associated confidence $w_{kj}$ . It also updates a hidden state $h_{kj} \in \mathbb{R}^{384}$ . Internally, the RNN shares features via attention between updates with the same incoming and outgoing frame. + +![](images/1e973e8a48aedfc16da1ad981d8296568bf46dc4fa8e1f90400f327c3395b790.jpg) + +# 3.3. Two-View Solver and Pre-Conditioning + +Our two-view solver aims to align the relative pose between two frames to the predicted matches by minimizing the distance between each match and its epipolar line. This strategy is accurate, however the objective function is nonconvex, meaning it will only converge to the global minimum if it is initialized close to it. In contrast, the 8-point-algorithm, a common approach to this task, has the opposite problem; it does not suffer from local minimum since it is a + +homogeneous least squares, but the solution is sub-optic. In Fig. 6, we visualize how our solver either converges within a fraction of a degree, or not at all, meanwhile + +8-point-algorithm is more robust but less accurate. + +The approach our method uses is to pre-condition the pose estimate using a weighted dense variant of the 8-point algorithm [30], and then run our solver layer to refine the pose. This combined strategy obtains the best of both worlds, since the pre-conditioning will typically initialize the pose within the basin of convergence of our solver, which then refines the prediction further. We visualize this basin on the TartanAir [48] dataset using groundtruth matches in Fig. 6. + +Pre-conditioning: We solve the homogeneous least squares problem [14] + +$$ +\arg \min _ {\mathbf {F}} \overline {{| | d i a g (\vec {w}) \mathbf {M} \vec {\mathbf {F}} | | ^ {2}}} \quad s. t. | | \mathbf {F} | | ^ {2} = 1 \tag {2} +$$ + +where $\mathbf{M}$ and $\vec{w}$ are constructed from the anchor-match pairs and confidence predictions. The points are normalized to $[-1, 1]$ beforehand. Afterwards, we reshape and uncalibrate $\mathbf{F}$ to obtain the essential matrix. + +We obtain the four relative pose candidates following the procedure detailed in [13] and in the supplement. During training, we select the pose candidate closest to the ground-truth on the $SE(3)$ manifold. During inference, we select the candidate by testing for chirality [13]. + +SED Solver Layer: The solver layer seeks to minimize the distance between each match $m_{kj}$ , and the epipolar line induced by the current pose estimate and its respective anchor point $a_k$ . The solver objective is bi-directional, since every image contains a unique set of anchor points; the objective is also known as the titular symmetric epipolar distance (SED) [11]. We parameterize the pose update as small rotations to the translation direction and orientation: $\xi_{\mathbf{t}}, \xi_{\mathbf{R}} \in \mathfrak{so}(3)$ . These are the free variables in eq. 3. Let $K(i)$ be the set of anchor indices for frame $i$ , $\mathbf{R}$ and $\mathbf{t}$ be the + +![](images/61ddc7591b5c3aa86415c52e218a9eb77ac13fef93b52cef8712fe977e19148d.jpg) +Figure 5. The flow of gradients through our solver. The updated matches from the RNN are supervised directly, and then detached from the gradient tape. The solver output is supervised with a pose loss. In the backward pass, gradients from the pose loss are used to supervise the confidence head in order to learn outlier rejection. + +output of the pre-conditioning stage, and $epi(x,R,t)$ compute the epipolar line for a point $x$ given the relative pose $(R,t)$ . Formally, the minimization objective in our solver is + +$$ +\begin{array}{l} l _ {k j} = e p i \left(a _ {k}, \left(e ^ {\xi_ {\mathbf {R}}} \mathbf {R}\right) _ {i \rightarrow j}, \left(e ^ {\xi_ {\mathbf {t}}} \mathbf {t}\right) _ {i \rightarrow j}\right): k \in K (i) \\ E _ {i \rightarrow j} = \sum_ {k \in K (i)} w _ {k j} \cdot \| \operatorname {e r r} \left(m _ {k j}, l _ {k j}\right) \| _ {2} ^ {2} \tag {3} \\ \stackrel {\smile}{S E D} = \operatorname *{arg min}_{\xi_{\mathbf{R}},\xi_{\mathbf{t}}}\left(E_{i\to j} + E_{j\to i}\right) \\ \end{array} +$$ + +where $err$ computes the 2D point-to-line error. The predicted weights $w_{kj}$ are included to allow the network to down-weight the contribution of any match which it deems unreliable. This solver layer is implemented in Pytorch, enabling gradients to propagate backward from the pose loss to the confidence head using autograd. We depict the gradient flow in Fig. 5. + +To minimize eq. 3, we employ the Levenberg-Marquardt algorithm. This requires computing the Jacobian for all terms. We provide the full derivations in the Supplement, but define the residual function $err$ in its entirety here. + +Let $l$ be the epipolar line produced by $(a, e^{\xi_{\mathbf{R}}}\mathbf{R}, e^{\xi_{\mathbf{t}}}\mathbf{t})$ . We can express each term of eq. 3 as: + +$$ +\operatorname {e r r} (m, l) = \left(\frac {l _ {x} m _ {x} + l _ {y} m _ {y} + l _ {z}}{l _ {x} ^ {2} + l _ {y} ^ {2}}\right) \left[ \begin{array}{l} l _ {x} \\ l _ {y} \end{array} \right] \tag {4} +$$ + +and the epipolar line as + +$$ +\begin{array}{l} \mathbf {E} = \left(e ^ {\xi_ {\mathrm {R}}} \mathbf {R}\right) ^ {\top} \left[ e ^ {\xi_ {\mathrm {t}}} \mathbf {t} \right] _ {\times} \\ l = \mathbf {F} \left[ \begin{array}{l} a _ {x} \\ a _ {y} \\ 1 \end{array} \right] = \left[ \left(K _ {2} ^ {\top}\right) ^ {- 1} \mathbf {E} K _ {1} ^ {\top} \right] \left[ \begin{array}{l} a _ {x} \\ a _ {y} \\ 1 \end{array} \right] \tag {5} \\ \end{array} +$$ + +where $\mathbf{F}$ and $\mathbf{E}$ are the fundamental and essential matrices. + +# 3.4. Adapting our backbone to Visual Odometry + +To adapt our backbone to VO, we make several changes: + +![](images/27405bfbdf6641ceb6b0276b35feefc787b9b8ea6be9bd0659c4ca2c0881034f.jpg) +Figure 6. Convergence of our solver, given ground-truth matches (from TartanAir [48]). Our SED solver is accurate, but only converges to the global minimum when initialized close to it. We initialize the pose using the 8-pt-algorithm to ensure it is within the SED convergence basin. + +Multi-view Solver (BA): Our local optimizer minimizes reprojection error and treats depth as a separate variable. This is identical to the bundle adjustment from DPVO. Formally, let $F$ be the set of connected frames, $G$ be the global poses, $d_{k}$ be the depth estimate for anchor $k$ , and $\Pi(\cdot)$ be the $3D \rightarrow 2D$ projection function. The bundle adjustment objective is: + +$$ +\underset {G, \mathbf {d}} {\arg \min } \sum_ {(i, j) \in F} \sum_ {k \in K (i)} w _ {k j} \cdot \left\| \Pi \left[ G _ {j} ^ {- 1} G _ {i} \Pi^ {- 1} \left(a _ {k}, d _ {k}\right) \right] - m _ {k j} \right\| ^ {2} \tag {6} +$$ + +The preconditioning stage is not necessary as we can linearly-extrapolate the camera poses from previous frames to achieve good initialization. + +Clamping: Since epipolar lines do not make sense in the multi-view setting, we reset the optical flow using the pose and depth from the bundle adjustment. + +RNN changes: We include mechanisms in the RNN to share latent features between updates if they stem from the same anchor point. Specifically, we use the message passing and temporal convolutions from DPVO [45]. + +# 3.5. Trajectory Alignment + +In the Multi-Session SLAM problem, we estimate a relative Sim(3) between pairs of disjoint trajectories in order to align them. All anchor points already have a depth estimate $d_{k}$ as a result of the bundle adjustment (eq. 6) from the VO system. We use our two-view method to estimate the relative rotation and translation direction, and compute the translation magnitude and relative scaling by comparing the $d_{k}$ 's to the depth from the triangulated two-view matches. + +1) Estimate relative rotation, translation direction, and matches. We first retrieve a candidate image pair $(i,j)$ using NetVLAD [1], one from each trajectory, and apply our two-view model to estimate their relative rotation and translation direction. +2) Align the depth from the two-view and SLAM operators. We previously defined $d_{k}$ as the depth output of the VO + +system. Let $d_k'$ be the triangulated depth obtained from our two-view matches: + +$$ +d _ {k} ^ {\prime} = \left\{\begin{array}{l l}\operatorname {t r i a n g} \left(\mathbf {R} _ {i \rightarrow j}, \mathbf {t} _ {i \rightarrow j}, a _ {k}, m _ {k j}\right)&: k \in K (i)\\\operatorname {t r i a n g} \left(\mathbf {R} _ {j \rightarrow i}, \mathbf {t} _ {j \rightarrow i} a _ {k}, m _ {k i}\right)&: k \in K (j)\end{array}\right. \tag {7} +$$ + +We solve + +$$ +\underset {s _ {i} \in \mathbb {R} > 0} {\arg \max } \sum_ {k \in K (i)} \mathbf {1} \left[ \frac {1}{\lambda} < \frac {d _ {k}}{s _ {i} \cdot d _ {k} ^ {\prime}} < \lambda \right] \tag {8} +$$ + +to recover the translation magnitude, where the hyperparameter $\lambda = 1.05$ . In layman terms, eq. 9 seeks what translation magnitude would align the most triangulated points to the existing 3D map? We can obtain a reasonably good solution to eq 9 by brute-force checking $s_i = d_k / d_k'$ $\forall k\in K(i)$ , similar to RANSAC. If the number of inliers is too small, we retry on a new candidate pair. Conversely, the scale-difference is $s_i / s_j$ , where $s_j$ is estimated the same way, but for anchor points in $K(j)$ : + +$$ +\underset {s _ {j} \in \mathbb {R} > 0} {\arg \max } \sum_ {k \in K (j)} \mathbf {1} \left[ \frac {1}{\lambda} < \frac {d _ {k}}{s _ {j} \cdot d _ {k} ^ {\prime}} < \lambda \right] \tag {9} +$$ + +Given the rotation $\mathbf{R}_{j\rightarrow i}$ and translation $\mathbf{t}_{j\rightarrow i}$ from the two-view solver layers, the resulting transformation can be used to align the two trajectories: + +$$ +\mathcal {S} _ {j \rightarrow i} = \left[\begin{array}{c c}\frac {s _ {i}}{s _ {j}} \mathbf {R}&s _ {j} \mathbf {t}\\0&1\end{array}\right] \in \operatorname {S i m} (3) \tag {10} +$$ + +The graphs are then merged by concatenating all buffers. + +# 3.6. Training + +We train our backbone separately for VO and for two-view pose. It is trained using a matching loss and a pose loss. The pose loss for the two-view training + +$$ +\mathcal {L} _ {\text {p o s e 2 V}} = \sum_ {t = 1} ^ {1 2} \cos^ {- 1} \left(\bar {\mathbf {t}} ^ {t} \cdot \bar {\mathbf {t}} ^ {\bar {\mathbf {g}}} ^ {t}\right) + \alpha \left\| \left[ \left(\mathbf {R} ^ {t}\right) ^ {T} \mathbf {R} ^ {\mathbf {g}} ^ {t} \right] \right\| _ {S O (3)} \tag {11} +$$ + +penalizes angle error for both the predicted translation direction and predicted orientation. The pose loss for the VO training is the SE3 manifold distance between the predicted and ground-truth poses: + +$$ +\mathcal {L} _ {\text {p o s e V O}} = \sum_ {t = 1} ^ {T} \left\| \left[ \left(\mathbf {G} ^ {t}\right) ^ {- 1} \mathbf {G} ^ {\mathbf {g t}} \right] \right\| _ {S E (3)} \tag {12} +$$ + +The matching loss + +$$ +\mathcal {L} _ {\text {m a t c h i n g}} = \frac {1}{| N |} \sum_ {a} ^ {N} \sum_ {t = 1} ^ {T} \left\| m _ {a} ^ {t} - m _ {a} ^ {\mathbf {g t}} \right\| _ {2} \tag {13} +$$ + +is a standard endpoint-error [24, 43], applied only on pixels which have valid depth and are verified to be visible in both images. Our update operator is applied 12 times per training example; supervision is applied after every update (See Fig. 5). The final loss is $\mathcal{L} = \mathcal{L}_{\text{pose}} + \beta \cdot \mathcal{L}_{\text{matching}}$ . + +![](images/29bd9d0c7eec35b33a57b9d37784db0e9d3d963bd66c78ce6335dfb35a6bad10.jpg) +Stage 1: Image Retrieval + +![](images/667dacec7e81c5a3476d64d62d609f784f7d21e2d0471acfa6c8a9a155509477.jpg) +Stage 2: Sim(3) Estimation +Stage 3: Graph Merging + +![](images/89462d1f56b69e3ee1367046a5087e4dec77904aab2a5d46fae3b8ab21e837ad.jpg) +Figure 7. High-level overview of our approach to aligning disjoint trajectories. We use NetVLAD [1] to retrieve two co-visible images, one from each trajectory, and apply our two-view update operator to estimate a relative Sim(3) alignment, following the procedure detailed in Sec 3.5. We then transform the second trajectory into the reference frame of the first, and merge all buffers. + +Following prior work [23], we pre-train our two-view method on synthetic homographies (without pose loss) for two epochs on the Oxford-Paris 1M Distractors dataset [29]. We then tune our model on a $50/50$ mixture of Scannet [6] and Megadepth [21] for 100,000 steps and a batch size of 120, using 10 A6000 GPUs for 5 days. The VO backbone is trained using the procedure from [45] on TartanAir. + +# 3.7. Multi-Session SLAM System + +Overview: To perform Multi-Session SLAM, we use our VO-trained backbone to perform visual odometry and global optimization; our two-view backbone connects disjoint trajectories following the procedure in Sec 3.5. Our full system builds the pose graph incrementally: new video frames inserted into the graph as they are received, and unjoined trajectories are aligned and merged as soon as a connection is found. + +VO Frontend: Our VO frontend extracts features and estimates camera poses and depth for incoming frames, operating on a sliding window covering the most recently observed 20 keyframes. Keyframing occurs retroactively on the $4^{th}$ oldest keyframe if it is found to be redundant. New poses are initialized using a linear-motion model, and depth is copied from the previous frame. Like DPVO [45], we initialize after 8 frames with significant motion are observed. + +Global Optimization: We perform global optimization by introducing proximity factors using the existing pose and depth measurements, following the approach from DROID-SLAM [44], and update the matches and poses/depth using the update operator. We only run the backend periodically, and after joining trajectories. + +Trajectory Joining: We query the database of NetVLAD descriptors to find co-visible frames. If several pairs are retrieved, we compute relative poses in a mini-batch and use the estimate with the highest inlier-ratio. + +
Overall ApproachPose error AUC [%] ↑
@5°@10°@20°
Superglue [34]Matching16.2(17.7)33.8(35.6)51.8(54.2)
LightGlue [23]16.5(19.4)33.4(36.9)50.1(53.5)
LoFTR [41]↓22.1(25.7)40.1(45.0)47.6(61.4)
MatchFormer [47]RANSAC [15]24.3(27.3)43.9(47.6)61.4(64.9)
ASpanFormer [5](LO-RANSAC [18])25.6(28.4)46.0(48.8)63.3(65.8)
Roessle&Nießner [31]Matching→W8PA20.741.661.7
Roessle&Nießner [31]Matching→W8PA→BA25.747.266.4
OursOptical-Flow +← +Clamp ← Solver30.951.167.6
+ +Table 1. Two-view results on Scannet [6]. Existing methods perform matching as pre-processing, then use a solver; our method is iterative. Results in parenthesis use the LO-RANSAC [18] estimator. "W8PA $\rightarrow$ BA" = weighted 8-point-algo. & bundle-adjustment. Our design outperforms prior approaches. We use the same model weights as in Tab. 2. + +
Overall ApproachPose error AUC [%] ↑
@5°@10°@20°
Superglue [34]Matching49.7(65.8)67.1(78.7)80.6(87.5)
LightGlue [23]49.9(66.7)67.0(79.3)80.1(87.9)
LoFTR [41]↓52.8(66.4)69.2(78.6)81.2(86.5)
MatchFormer [47]RANSAC [15]53.3(66.5)69.7(78.9)81.8(87.5)
ASpanFormer [5](LO-RANSAC [18])55.3(69.4)71.5(81.1)83.1(88.9)
Roessle&Nießner [31]Matching→W8PA46.962.876.3
Roessle&Nießner [31]Matching→W8PA→BA61.274.985.0
OursOptical-Flow60.472.781.3
Clamp ← Solver
+ +Table 2. Two-view results on Megadepth [21]. Existing methods perform matching as pre-processing, then use a solver; our method is iterative. "W8PA $\rightarrow$ BA" = weighted 8-point-algo. & bundle-adjustment. Results in parenthesis use the LO-RANSAC [18] estimator. Our radically different design leads to better results indoors (Tab. 1), but is less competitive in photo-tourism settings where high-volume matching is easier and therefore RANSAC-based methods work better. We use the same model weights as in Tab. 1. + +# 4. Experiments + +We evaluate the performance of our two-view method in isolation, and our Multi-Session SLAM system as a whole. + +# 4.1. Two-View Evaluation + +We evaluate our two-view method on two popular relative pose benchmarks: The Scannet [6] 1500 and MegaDepth [21] 1500 test datasets. We compare to existing matching networks [5, 23, 34, 41, 47], which are typically in service of improving COLMAP [38]-based SfM pipelines [22, 33]. In contrast, our approach is in service of Multi-Session SLAM. We report pose error AUC, as is done in prior matching work (higher is better). We use the same configuration and model weights on both datasets. + +Most existing methods perform matching as a preprocessing step using a deep network, then estimate pose with a minimal solver using RANSAC [15, 18]. Roessle + +and Nießner [31] replace the RANSAC optimizer with a weighted-8-point algorithm and bundle adjustment. In contrast, our two-view method couples the prediction of pose and matches, refining the predictions over many iterations. We ablate over our two-view model in Tab. 3. + +Scannet: We report two-view results on Scannet [6] in Tab. 1. Our approach outperforms existing methods on Scannet, which contains fewer salient keypoints and many texture-less surfaces and motion blur. + +MegaDepth: We report two-view results on Megadepth [21] in Tab. 2. On Megadepth, the matching problem is substantially easier since photo-tourism images contain many salient keypoints, resulting in better performance across all methods; our approach is on-par with existing approaches using the OpenCV RANSAC pose estimator [15], but not the LO-RANSAC [18] estimator. + +# 4.2. Multi-Session SLAM evaluation + +We evaluate our approach on Multi-Session SLAM on the EuRoC-MAV [2] and ETH3D [39] datasets, since they provide camera poses under a single global reference. Following the evaluation for single-video SLAM, we align the final predictions to the ground truth by computing a global 7-DOF alignment to account for the guage freedoms. We report RMSE ATE[m] in metric units. We sample 96 anchors in each video frame using a mix of random and Superpoint [8] keypoints. Our two-view method uses those same anchors to join pairs of trajectories. We use the same weights and configuration on both datasets. + +EuRoC-MAV: In Tab. 9 and Fig. 10, we report results on the EuRoC MAV dataset [2]. Following the evaluation from [4], we compare on four groups of disjoint trajectories across the three distinct environments: 3 in the Machine Hall, 5 in the Machine Hall, 3 in Vicon 1, and 3 in Vicon 2. Ground-truth poses are obtained using a laser tracker and Vicon cameras. These video sequences are long, with the entirety of the Machine-Hall sequence being 11.5 minutes of video (13.7K frames) spread across 5 sequences. Our approach achieves significantly lower error than ORBSLAM3 on all groups, including 12x lower error on the Vicon 2 sequences (0.022 vs 0.284). Video is recorded at 20-FPS; all methods run in real-time. We also compare to CCM-SLAM [36], which only provides results on the on MH01-03 sequence, the mono-inertial ORB-SLAM3, and the mono-inertial VINS [28]. We outperform the other approaches on all trajectory groups. + +ETH3D: In Tab. 4, we report results on the training set from ETH3D [39]. We compare across 5 unique scenes composed of multiple trajectories: 1-4 from Sofa, 3&4 from Table, 1-3 from Plant Scene, 1&2 from Einstein, and 2&3 from Planar. Sequences are excluded if they belong to the official ETH3D test set (with no ground-truth) or are from a different scene altogether. The sequences are trimmed + +![](images/39ee103b7e694f32df19e18b1a2b150cf3540eb70b49e77258168d64285f4f64.jpg) +Figure 8. Qualitative results on Scannet. Our two-view variant is able to estimate accurate relative poses across wide camera baselines. It initializes all matches with uniform depth and identity relative pose. Progressive applications of our update operator lead to more accurate matches and higher predicted confidence. Some image pairs (row 2) take more iterations to converge than others (row 1). + +
Scene nameMH01-03MH01-05V101-103V201-203Cam.
# Disjoint Trajectories3533hz
Ours0.0220.0360.0300.02220
Mono-Visual
CCM-SLAM [36]0.077---38
Mono-Visual
ORB-SLAM3 [4]0.0300.0580.0580.28420
Mono-Visual
VINS [28]-0.210---
Mono-Inertial
ORB-SLAM30.0370.0650.0400.04820
Mono-Inertial
+ +![](images/2055a6290fe878c490de84d2166b9784dbc2a86d85c363ac87fb554965774010.jpg) +Figure 9. Multi-Session SLAM evaluation on EuRoC MAV [2] using RMSE ATE[m] $\downarrow$ . We perform similarly to ORB-SLAM3 (within $2mm$ ) in the Machine Hall, but perform best in the Vicon 1 and Vicon 2 rooms. Results for baseline approaches were obtained from [4]. +Figure 10. Our prediction on the Vicon1 [2] sequences. Our method is accurate and robust to chaotic motion. + +from a single, longer video; to make joining the trajectories non-trivial, we reverse every other sequence, ensuring that there is a sufficient disconnect between subsequent videos. This is a novel benchmark, so [28, 36] do not report results. + +
Ablation ExperimentPose error AUC @10°[%] ↑
Megadepth ValScannet Val
BaselineFull model53.333.9
ArchitectureShaller correlation pyramid44.327.1
No attention in update-op48.630.5
No ReLU-attn in feat-extractor47.429.7
SolverNo pre-conditioning35.520.6
No SED solver23.813.1
No clamping step21.910.7
+ +Table 3. Two-view ablation experiments on two validation sets. Our proposed RNN and feature extractor changes from DPVO improve the result. Removing components from our proposed solver design decreases performance, as does removing the reprojection step. + +
Scene nameSofaTablePlant SceneEinsteinPlanar
# Disjoint Trajectories42322
ORB-SLAM3 [4] +Mono-VisualFAIL +(no init)0.018FAIL +(init→lost)FAIL +(init→lost)0.010
Ours +Mono-Visual0.0130.0100.0090.0050.035
+ +Table 4. Multi-Session SLAM evaluation on ETH3D [39] using RMSE ATE[m] $\downarrow$ . Our method is robust to catastrophic failures. + +ORB-SLAM3 fails on Sofa, Plant Scene, and Einstein for various reasons (couldn't initialize, or lost the feature tracks). Our approach outperforms ORB-SLAM3 on 4/5 groups, and does not fail on any. ORB-SLAM3 only succeeds on 2/5 groups. + +# 5. Conclusion + +We introduce a new method for mono-visual Multi-Session SLAM. Our system utilizes a novel backbone which can estimate two-view pose and perform visual odometry. We leverage a novel differentiable solver which minimizes the symmetric epipolar distance. We compare against existing approaches and show strong performance across several datasets. This work was partially supported by IARPA and the National Science Foundation. + +# References + +[1] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pa-jdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5297-5307, 2016. 5, 6 +[2] Michael Burri, Janosch Nikolic, Pascal Gohl, Thomas Schneider, Joern Rehder, Sammy Omari, Markus W Achtelik, and Roland Siegwart. The euroc micro aerial vehicle datasets. The International Journal of Robotics Research, 35 (10):1157-1163, 2016. 1, 7, 8 +[3] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 2 +[4] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 1, 2, 7, 8 +[5] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming-min Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 7 +[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6, 7 +[7] Hayyan Afeef Daoud, Aznul Qalid Md. Sabri, Chu Kiong Loo, and Ali Mohammed Mansoor. Slamm: Visual monocular slam with continuous mapping using multiple maps. PloS one, 13(4):e0195878, 2018. 2 +[8] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 3, 7 +[9] Jakob Engel, Thomas Schöps, and Daniel Cremers. Lsd-slam: Large-scale direct monocular slam. In European conference on computer vision, pages 834-849. Springer, 2014. 2 +[10] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. IEEE transactions on pattern analysis and machine intelligence, 40(3):611-625, 2017. 2 +[11] Mohammed E Fathy, Ashraf S Hussein, and Mohammed F Tolba. Fundamental matrix estimation: A study of error criteria. Pattern Recognition Letters, 32(2):383-391, 2011. 4 +[12] Dorian Gálvez-López and Juan D Tardos. Bags of binary words for fast place recognition in image sequences. IEEE Transactions on Robotics, 28(5):1188-1197, 2012. 2 +[13] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, New York, NY, USA, 2 edition, 2003. 4, 2 + +[14] Richard I Hartley. In defense of the eight-point algorithm. IEEE Transactions on pattern analysis and machine intelligence, 19(6):580-593, 1997. 2, 4 +[15] Itseez. Open source computer vision library. https://github.com/itseez/opencv, 2015.7 +[16] Mathieu Labbe and François Michaud. Rtab-map as an opensource lidar and visual simultaneous localization and mapping library for large-scale and long-term online operation. Journal of field robotics, 36(2):416-446, 2019. 1 +[17] Mathieu Labbe and François Michaud. Multi-session visual slam for illumination-invariant re-localization in indoor environments. Frontiers in Robotics and AI, 9:801886, 2022. 2 +[18] Viktor Larsson and contributors. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 2, 7 +[19] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 2 +[20] Stefan Leutenegger, Paul Furgale, Vincent Rabaud, Margarita Chli, Kurt Konolige, and Roland Siegwart. Keyframe-based visual-inertial slam using nonlinear optimization. Proceedings of Robotics Science and Systems (RSS) 2013, 2013. 2 +[21] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Computer Vision and Pattern Recognition (CVPR), 2018. 6, 7 +[22] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987-5997, 2021. 7 +[23] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In ICCV, 2023. 6, 7 +[24] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 6 +[25] Xiao Xin Lu. A review of solutions for perspective-n-point problem in camera pose estimation. In Journal of Physics: Conference Series, page 052009. IOP Publishing, 2018. 2 +[26] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 2 +[27] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 2 +[28] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics, 34(4):1004-1020, 2018. 1, 2, 7, 8 +[29] F. Radenović, A. Iscen, G. Tolias, Y. Avrithis, and O. Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In CVPR, 2018. 6 + +[30] Rene Ranftl and Vladlen Koltun. Deep fundamental matrix estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2, 4 +[31] Barbara Roessle and Matthias Nießner. End2end multi-view feature matching with differentiable pose optimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 477-487, 2023. 2, 7 +[32] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011. 2 +[33] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019. 2, 7 +[34] 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. 2, 7 +[35] Patrik Schmuck and Margarita Chli. Multi-uav collaborative monocular slam. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3863-3870. IEEE, 2017. 2 +[36] Patrik Schmuck and Margarita Chli. Ccm-slam: Robust and efficient centralized collaborative monocular simultaneous localization and mapping for robotic teams. Journal of Field Robotics, 36(4):763-781, 2019. 1, 2, 7, 8 +[37] Thomas Schneider, Marcin Dymczyk, Marius Fehr, Kevin Egger, Simon Lynen, Igor Gilitschenski, and Roland Siegwart. maplab: An open framework for research in visual-inertial mapping and localization. IEEE Robotics and Automation Letters, 3(3):1418-1425, 2018. 1 +[38] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 7 +[39] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 7, 8 +[40] Hauke Strasdat, J Montiel, and Andrew J Davison. Scale drift-aware large scale monocular slam. Robotics: science and Systems VI, 2(3):7, 2010. 2 +[41] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 7 +[42] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 2 +[43] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 1, 2, 3, 6 + +[44] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1, 2, 6 +[45] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 5, 6 +[46] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 2 +[47] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, pages 2746-2762, 2022. 7 +[48] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 4, 5 \ No newline at end of file diff --git a/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/images.zip b/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3ea12dab24ca0d632155c596a3a7a9a5b854038b --- /dev/null +++ b/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7516de79b4451d680b8f13f7592329edf931821325ffef8eeb1d014c4ed15af0 +size 643868 diff --git a/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/layout.json b/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..522c7dc91db1f3fc2362385354efee644032f34d --- /dev/null +++ b/2024/Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization/layout.json @@ -0,0 +1,9305 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "text", + "content": "Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 182, + 144, + 252, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 144, + 252, + 157 + ], + "spans": [ + { + "bbox": [ + 182, + 144, + 252, + 157 + ], + "type": "text", + "content": "Lahav Lipson" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 159, + 268, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 159, + 268, + 171 + ], + "spans": [ + { + "bbox": [ + 167, + 159, + 268, + 171 + ], + "type": "text", + "content": "Princeton University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 161, + 174, + 275, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 174, + 275, + 184 + ], + "spans": [ + { + "bbox": [ + 161, + 174, + 275, + 184 + ], + "type": "text", + "content": "llipson@princeton.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 351, + 144, + 396, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 144, + 396, + 157 + ], + "spans": [ + { + "bbox": [ + 351, + 144, + 396, + 157 + ], + "type": "text", + "content": "Jia Deng" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 159, + 424, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 159, + 424, + 171 + ], + "spans": [ + { + "bbox": [ + 323, + 159, + 424, + 171 + ], + "type": "text", + "content": "Princeton University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 174, + 430, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 430, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 430, + 185 + ], + "type": "text", + "content": "jiadeng@princeton.edu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 239, + 290, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 290, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 290, + 370 + ], + "type": "text", + "content": "We introduce a new system for Multi-Session SLAM, which tracks camera motion across multiple disjoint videos under a single global reference. Our approach couples the prediction of optical flow with solver layers to estimate camera pose. The backbone is trained end-to-end using a novel differentiable solver for wide-baseline two-view pose. The full system can connect disjoint sequences, perform visual odometry, and global optimization. Compared to existing approaches, our design is accurate and robust to catastrophic failures. Code is available at https://github.com/princeton-v1/MultiSlam_DiffPose" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 380, + 128, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 128, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 128, + 392 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 401, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 556 + ], + "type": "text", + "content": "Simultaneous Localization and Mapping (SLAM) is the task of estimating camera motion and a 3D map from video. The standard setup assumes a single continuous video. However, video data in the wild often consists of not a single continuous stream, but rather multiple disjoint sessions, either deliberately such as in collaborative mapping when multiple robots perform joint rapid 3D reconstruction, or inadvertently due to visual discontinuities in the video stream which can result from camera failures, extreme parallax, rapid turns, auto-exposure lag, dark areas, or extreme occlusion by dynamic objects. Handling such disjoint videos is important for many applications in AR and robotics, and gives rise to the task of Multi-Session SLAM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 557, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 629 + ], + "type": "text", + "content": "In Multi-Session SLAM, the input consists of multiple disjoint video sequences and the goal is to estimate camera poses for all video frames under a single global reference. This is in contrast to \"single-video SLAM\", whose input is a single continuous video. In this work, we focus on the monocular, RGB-only Multi-Session SLAM setting." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "Several approaches have been proposed to deal with Multi-Session SLAM, however existing solutions typically require additional sensor data in order to remove gauge freedoms and make tracking easier [16, 28, 37]. Only a small number of methods, notably CCM-SLAM [36] and ORBSLAM3 [4], support Multi-Session SLAM from monocular video alone, due to the difficulty of aligning disjoint" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 309, + 213, + 547, + 299 + ], + "blocks": [ + { + "bbox": [ + 309, + 213, + 547, + 299 + ], + "lines": [ + { + "bbox": [ + 309, + 213, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 309, + 213, + 547, + 299 + ], + "type": "image", + "image_path": "4891576479077935d55e89ddf5bb9d1bf95a249293d965daacfe5934c762b57b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 308, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 305, + 308, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 545, + 329 + ], + "type": "text", + "content": "Figure 1. Our method estimates camera pose from multiple disconnected video streams." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 352, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 400 + ], + "type": "text", + "content": "sequences under the 7-DOF gauge freedoms in monocular video. However, these approaches are based on classical feature descriptors, making them less accurate on average compared to recent designs based on deep networks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 401, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 546, + 555 + ], + "type": "text", + "content": "In the standard SLAM setting, Teed and Deng [44] proposed to use a deep optical flow network (RAFT [43]) to track 2D motion, while jointly updating camera poses with a bundle adjustment layer. The method, DROID-SLAM, is accurate and avoids tracking failures, but the design assumes a continuous video stream and is not capable of the wide-baseline matching and non-local optimization necessary for Multi-Session SLAM. Deep Patch Visual Odometry [45] (DPVO) introduced a sparse visual-odometry-only analog of DROID-SLAM which achieves similar accuracy on single-video SLAM, but at much lower cost. However, DPVO also does not support Multi-Session SLAM for the same reasons as DROID-SLAM." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "text", + "content": "We propose a method for Multi-Session SLAM, capable of both wide-baseline relative pose and visual odometry using a single backbone architecture inspired by [45]. We introduce a differentiable solver layer which minimizes the symmetric epipolar distance (SED) from bi-directional optical flow. From this we construct a method for two-view pose which is capable of matching from far-apart views. This same design can be repurposed for visual odometry by swapping out the solver for bundle adjustment. By employing a unified backbone architecture for both tasks, we enable a simple approach to Multi-Session SLAM." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We evaluate our approach on challenging real-world datasets: EuRoC-MAV [2] and ETH3D [39]. Our sys" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "19626" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "tem is more accurate than prior approaches, and is robust to catastrophic failures. We also evaluate our two-view pose method in isolation on the Scannet and Megadeph datasets, and show that it is competitive with transformer-based matching networks. For pairs of far-apart views, our method is capable estimating accurate relative pose." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 286, + 228 + ], + "type": "text", + "content": "Our backbone predicts iterative updates to optical flow coupled with a differentiable solver layer for estimating camera pose. This framework, based on RAFT [43], has worked exceptionally well for Visual Odometry [45] and SLAM [44]. By leveraging this idea for wide-baseline matching, we can extend these methods to the Multi-Session setting without introducing substantial complexity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 237, + 134, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 134, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 134, + 249 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 286, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 286, + 300 + ], + "type": "text", + "content": "Most prior works treat Visual Simultaneous Localization and Mapping (Visual SLAM) as an optimization problem solving for a 3D scene model and camera trajectory which best explains the visual measurements [3]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 302, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 286, + 445 + ], + "type": "text", + "content": "Indirect approaches to SLAM perform keypoint matching as a pre-processing step, then estimate a 3D point cloud and camera poses by optimizing the 2D reprojection error over all matches [4, 20, 26, 27, 33, 38]. Reprojection-error is easier to optimize than photometric alignment, making indirect methods robust to lower camera " + }, + { + "bbox": [ + 46, + 302, + 286, + 445 + ], + "type": "inline_equation", + "content": "\\text{hz}" + }, + { + "bbox": [ + 46, + 302, + 286, + 445 + ], + "type": "text", + "content": " [44]. Keypoint matching also enables a straightforward approach to estimating two-view relative pose, a necessary step for Visual Multi-Session SLAM [14, 19, 25, 33, 46]. However, keypoint-based SLAM is less robust to low-texture environments compared to those which use photometric alignment [9, 10] or optical flow [44, 45]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 448, + 286, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 286, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 286, + 543 + ], + "type": "text", + "content": "Semi-Indirect approaches similarly optimize 2D reprojection error like indirect methods, but without requiring matches as input [42, 44, 45]. Instead, these approaches alternate between predicting optical flow residuals and performing bundle adjustment. Semi-indirect methods do not require repeatable keypoints across images, making them robust to low-texture settings while retaining the easier reprojection-error objective." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 544, + 286, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 286, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 286, + 602 + ], + "type": "text", + "content": "Our approach is most similar to Deep Patch Visual Odometry [45] (DPVO), which is a sparse analog of DROID-SLAM. DPVO predicts sparse optical flow instead of dense, performing similarly to DROID-SLAM while running faster and using half the memory." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "Differentiable solver layers for camera pose estimation have been used in order to learn outlier rejection with data-driven training. For two-view relative pose, Ranftl and Koltun [30] used a deep network to learn an iteratively reweighted least-squares algorithm, which solved a weighted variant of the 8-point-algorithm [14] using confidences predicted by the network. [30] required matched points as input, whereas our approach can work from images alone. Roessle and Nießner [31] proposed an" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "end-to-end architecture which used the weighted 8-point-algorithm from [30] on top of matches produced using Superglue [34], and supervised directly on the predicted pose. [31] works well, but is unable to outperform existing methods that use minimal solvers with LO-RANSAC [18] instead of differentiable solvers. In contrast, our approach is not built on top of an existing SOTA 2-view matcher; our design is also iterative, applying a recurrent module and solver layers multiple times to refine the prediction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 183, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 545, + 350 + ], + "type": "text", + "content": "Multi-Session SLAM is the task of performing SLAM on multiple trajectories of the same scene. Like SLAM, Multi-Session SLAM is an online task where camera motion is estimated from a stream of images. However, in the Multi-Session setting, there are known breaks in the data-stream over which the small-baseline assumption no longer holds. While local optimization is sufficient for motion tracking from video, estimating wide-baseline camera pose is often non-convex [40] and more challenging due to large viewpoint changes. In monocular visual SLAM, the scale of each sequence is also ambiguous; aligning two trajectories requires estimating 7 degrees-of-freedom (translation, rotation, scale) for all sequences, excluding the first which can be considered the reference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 351, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 518 + ], + "type": "text", + "content": "ORB-SLAM3 [4] performs Multi-Session SLAM by matching between ORB [32] descriptors. While tracking camera motion over each sequence, ORB-SLAM3 continually updates an Atlas of the scene - a lookup table between keypoints and their 3D map point in the reference frame of their original sequence. To align disjoint sequences, candidate cross-sequence image pairs are identified using image-retrieval [12], the 3D map points are obtained from the Atlas, and the relative transformation is found using the Umeyama algorithm [46]. Several methods perform Multi-Session Visual Inertial SLAM [17, 28], however we focus on the visual-only monocular setting where the scale of each session is ambiguous and must be estimated when joining sequences." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 520, + 545, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 580 + ], + "type": "text", + "content": "CCM-SLAM [35, 36] and SLAMM [7] perform Multi-Session visual SLAM and are built on top of ORBSLAM [27] as well, but are optimized for limited bandwidth and distributed processing, while ORB-SLAM3 performs better and is optimized for accuracy and speed." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 585, + 372, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 585, + 372, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 372, + 597 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Overview: We propose a backbone for matching between " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\geq 2" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " views, and then construct a method for multi-session SLAM upon it. Our backbone approaches matching as optical flow; it borrows several ideas from RAFT [43], such as iteratively predicting flow residuals using a recurrent network, and the correlation feature pyramid. Our method maintains a running estimate of both camera pose and bi-directional optical flow, and uses the update operator to refine them both. We provide an overview in Fig. 2. An" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19627" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 69, + 532, + 173 + ], + "blocks": [ + { + "bbox": [ + 62, + 69, + 532, + 173 + ], + "lines": [ + { + "bbox": [ + 62, + 69, + 532, + 173 + ], + "spans": [ + { + "bbox": [ + 62, + 69, + 532, + 173 + ], + "type": "image", + "image_path": "b2f41baebe36e1b30f6fab21abdcb137a14a168bc840a97406c1622d93fdf413.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 184, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 184, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 184, + 547, + 251 + ], + "type": "text", + "content": "Figure 2. Overview of our backbone. Given a group of " + }, + { + "bbox": [ + 46, + 184, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\geq 2" + }, + { + "bbox": [ + 46, + 184, + 547, + 251 + ], + "type": "text", + "content": " frames, our method jointly estimates bi-directional optical flow and camera poses. This module is applicable to (1) wide-baseline two-view matching and (2) visual odometry. We use our backbone for Multi-Session SLAM, which requires the ability to perform both (1) and (2). For each image, we select a set of 2D anchor points and initialize their depth and the camera poses trivially. Our approach then iteratively refines the matches for each anchor, similar to RAFT [43], while updating the camera poses. We alternate between matching and pose updates, where each one informs the update to the other. This entire procedure is repeated several times until convergence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 262, + 287, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 310 + ], + "type": "text", + "content": "invariant of our backbone is that the matches are always clamped to plausible values given the pose estimates and an assumption of rigid scene geometry, therefore WLOG the matches can be considered a depth estimate." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 312, + 288, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 443 + ], + "type": "text", + "content": "At initialization, our backbone produces dense correlation feature pyramids for each image and context features which remain fixed throughout the forward pass. The update operator uses an RNN to predict an update to the matches, and a differentiable solver to update poses. Our backbone treats the two-view and multi-view settings differently: In the two-view setting, the solver updates the poses to minimize the symmetric epipolar distance (SED). In the multi-view setting, the solver updates both poses and depth to minimize the reprojection error. After the solver layer, the matches are adjusted to agree with the poses/depth." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 451, + 130, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 451, + 130, + 463 + ], + "spans": [ + { + "bbox": [ + 47, + 451, + 130, + 463 + ], + "type": "text", + "content": "3.1. Initialization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": "Feature Extraction & Feature Pyramid: Similar to RAFT, our method separately extracts context and correlation feature for each image. The context features are provided as input to the recurrent update operator, whereas the correlation features are used to evaluate the visual similarity between any two pixels using a dot product. The context feature maps are produced at " + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": " resolution using a residual network. We then apply a linear-self-attention residual for long-range feature sharing. The correlation feature maps are also produced using a residual network, but with several exit ramps to produce features at " + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": " of the input image resolution. We also average-pool the last one three additional times to produce a correlation feature pyramid with 6 levels. We depict their architecture in the Supplement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Anchor-Point Selection: Our method predicts sparse optical flow, where matches have one end anchored and the other end free to move in " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ". We use a mix of detector [8]-chosen and randomly-chosen anchor points. Each point is assigned an initial match in the other image(s) at the same" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 262, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 287 + ], + "type": "text", + "content": "pixel coordinate. For the rest of the paper, we will refer to the " + }, + { + "bbox": [ + 304, + 262, + 545, + 287 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 304, + 262, + 545, + 287 + ], + "type": "text", + "content": " anchor point as " + }, + { + "bbox": [ + 304, + 262, + 545, + 287 + ], + "type": "inline_equation", + "content": "a_{k}" + }, + { + "bbox": [ + 304, + 262, + 545, + 287 + ], + "type": "text", + "content": ", and its match in image " + }, + { + "bbox": [ + 304, + 262, + 545, + 287 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 262, + 545, + 287 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 262, + 545, + 287 + ], + "type": "inline_equation", + "content": "m_{kj}" + }, + { + "bbox": [ + 304, + 262, + 545, + 287 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "text", + "content": "We index the context feature map at each " + }, + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "inline_equation", + "content": "a_{k}" + }, + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "text", + "content": " to produce a unique context feature vector " + }, + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "inline_equation", + "content": "ctx_{k} \\in \\mathbb{R}^{384}" + }, + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "text", + "content": ", which is used by the update operator. Beyond this point, the full context feature map is not used and is discarded to save memory." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 339, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 411 + ], + "type": "text", + "content": "Correlation: Similar to other RAFT-based methods, we use correlation features to assess the visual alignment/similarity given by the current matching estimate. For each " + }, + { + "bbox": [ + 304, + 339, + 545, + 411 + ], + "type": "inline_equation", + "content": "(a_{k},m_{kj})" + }, + { + "bbox": [ + 304, + 339, + 545, + 411 + ], + "type": "text", + "content": " pair, we bilinearly sample both feature pyramids for their respective frames at locations " + }, + { + "bbox": [ + 304, + 339, + 545, + 411 + ], + "type": "inline_equation", + "content": "a_{k}" + }, + { + "bbox": [ + 304, + 339, + 545, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 339, + 545, + 411 + ], + "type": "inline_equation", + "content": "m_{kj}" + }, + { + "bbox": [ + 304, + 339, + 545, + 411 + ], + "type": "text", + "content": " and take their inner product at each level, producing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 375, + 415, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 415, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 375, + 415, + 545, + 430 + ], + "type": "interline_equation", + "content": "\\langle f _ {k} ^ {i}, f _ {k j} ^ {i} \\rangle \\in \\mathbb {R}: i = 1 \\dots 6 \\tag {1}", + "image_path": "62a027d493e90c4c8c2bf0fcee027b0196ea2c97348b8fc117abdb46b85a3e89.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 434, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 483 + ], + "type": "text", + "content": "To provide additional spatial context, we also perturb both " + }, + { + "bbox": [ + 304, + 434, + 545, + 483 + ], + "type": "inline_equation", + "content": "a_{k}" + }, + { + "bbox": [ + 304, + 434, + 545, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 434, + 545, + 483 + ], + "type": "inline_equation", + "content": "m_{kj}" + }, + { + "bbox": [ + 304, + 434, + 545, + 483 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 434, + 545, + 483 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 434, + 545, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 434, + 545, + 483 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 304, + 434, + 545, + 483 + ], + "type": "text", + "content": " grids, respectively, and calculate eq. 1 for all pairs. The resulting feature vector " + }, + { + "bbox": [ + 304, + 434, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{kj} \\in \\mathbb{R}^{(6 \\times 3 \\times 3 \\times 7 \\times 7)}" + }, + { + "bbox": [ + 304, + 434, + 545, + 483 + ], + "type": "text", + "content": " is then passed to the update operator." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 491, + 409, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 409, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 409, + 503 + ], + "type": "text", + "content": "3.2. Update Operator" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "text", + "content": "The update operator produces a revision to all " + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "inline_equation", + "content": "m_{kj}" + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "text", + "content": " and the estimated relative pose. It consists of three stages, depicted in Fig. 3: (1) An RNN predicts a 2D update to " + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "inline_equation", + "content": "m_{kj}" + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "text", + "content": " and an associated confidence weight " + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "inline_equation", + "content": "w_{kj}" + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "text", + "content": ". (2) We solve for a pose estimate which is consistent with the newly predicted matches and confidence. (3) We adjust the matches to be physically plausible assuming rigid scene geometry. In the two-view setting, we clamp the matches to the epipolar lines, and in the multi-view setting we reproject the anchor points using the depth and poses from the solver." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Through the recurrent iterations, our method maintains a running state for each anchor-match " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "(a_{k},m_{kj})" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " pair, consisting of a hidden state vector " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "h_{kj}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ", a confidence weight " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "w_{kj}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and a match location " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "m_{kj}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". This state is continually updated, during which " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "m_{kj}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " should ideally approach the true match location of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "a_{k}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "w_{kj}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " should approach 1. This is what we observe empirically." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19628" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 532, + 205 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 532, + 205 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 532, + 205 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 532, + 205 + ], + "type": "image", + "image_path": "ab4da5dbd964d6bd8253334e6fbcea6680d6f1c2050d25785bf408754cd779a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 215, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 215, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 547, + 247 + ], + "type": "text", + "content": "Figure 3. A single update iteration (two-view). For each anchor-match pair, we predict an update to the matches using the RNN. We then solve for an update to the camera poses which minimizes the symmetric epipolar distance (SED), producing a new set of epipolar lines. Finally, we clamp the matches back to the best-fit epipolar lines, and repeat the whole process again. In Fig. 8 and the Supplement, we" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 248, + 218, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 218, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 218, + 268 + ], + "type": "text", + "content": "visualize these iterations on real-world images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 269, + 294, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 294, + 305 + ], + "type": "text", + "content": "RNN Module: The learnable component of the update operator is the recurrent network which predicts updates to " + }, + { + "bbox": [ + 47, + 269, + 294, + 305 + ], + "type": "inline_equation", + "content": "h_{kj}" + }, + { + "bbox": [ + 47, + 269, + 294, + 305 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 269, + 294, + 305 + ], + "type": "inline_equation", + "content": "m_{kj}" + }, + { + "bbox": [ + 47, + 269, + 294, + 305 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 269, + 294, + 305 + ], + "type": "inline_equation", + "content": "w_{kj}" + }, + { + "bbox": [ + 47, + 269, + 294, + 305 + ], + "type": "text", + "content": " for all anchor points. We initialize this" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 305, + 122, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 305, + 122, + 317 + ], + "spans": [ + { + "bbox": [ + 47, + 305, + 122, + 317 + ], + "type": "text", + "content": "operator in Fig. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 317, + 288, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 317, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 288, + 448 + ], + "type": "text", + "content": "The input to the RNN are the context features, the previous hidden state and the correlation features generated from the current matching estimate, all of which are added together and normalized with layer norm. A self-attention residual is also applied to edges with the same source and destination frame, followed by three gated-residual-units, whose architecture is depicted in the Supplement. The output is an updated hidden state, from which we predict an updated match and confidence using the flow head and confidence head, respectively, which are implemented as two-layer MLPs. The confidence head includes a softmax to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 448, + 133, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 448, + 133, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 448, + 133, + 460 + ], + "type": "inline_equation", + "content": "\\text{estriel} w_{kj}\\in (0,1)" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 23, + 445, + 49, + 482 + ], + "blocks": [ + { + "bbox": [ + 23, + 445, + 49, + 482 + ], + "lines": [ + { + "bbox": [ + 23, + 445, + 49, + 482 + ], + "spans": [ + { + "bbox": [ + 23, + 445, + 49, + 482 + ], + "type": "image", + "image_path": "7f5f718f01f4a5b00c7f12616c4796de2611b4357652e20d602376f9f7d7eeda.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 468, + 97, + 473 + ], + "lines": [ + { + "bbox": [ + 50, + 468, + 97, + 473 + ], + "spans": [ + { + "bbox": [ + 50, + 468, + 97, + 473 + ], + "type": "text", + "content": "Correlation Pyramids" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 25, + 492, + 49, + 519 + ], + "blocks": [ + { + "bbox": [ + 25, + 492, + 49, + 519 + ], + "lines": [ + { + "bbox": [ + 25, + 492, + 49, + 519 + ], + "spans": [ + { + "bbox": [ + 25, + 492, + 49, + 519 + ], + "type": "image", + "image_path": "115da04c3ca4fafc617bdfec51ba3c455883c48ed58f3f880cfbeeee325f1cf8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 532, + 287, + 587 + ], + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 587 + ], + "type": "text", + "content": "Figure 4. The RNN Module. For each anchor-match pair " + }, + { + "bbox": [ + 46, + 532, + 287, + 587 + ], + "type": "inline_equation", + "content": "(a_{k},m_{kj})" + }, + { + "bbox": [ + 46, + 532, + 287, + 587 + ], + "type": "text", + "content": ", it predicts an update to " + }, + { + "bbox": [ + 46, + 532, + 287, + 587 + ], + "type": "inline_equation", + "content": "m_{kj}" + }, + { + "bbox": [ + 46, + 532, + 287, + 587 + ], + "type": "text", + "content": " and an associated confidence " + }, + { + "bbox": [ + 46, + 532, + 287, + 587 + ], + "type": "inline_equation", + "content": "w_{kj}" + }, + { + "bbox": [ + 46, + 532, + 287, + 587 + ], + "type": "text", + "content": ". It also updates a hidden state " + }, + { + "bbox": [ + 46, + 532, + 287, + 587 + ], + "type": "inline_equation", + "content": "h_{kj} \\in \\mathbb{R}^{384}" + }, + { + "bbox": [ + 46, + 532, + 287, + 587 + ], + "type": "text", + "content": ". Internally, the RNN shares features via attention between updates with the same incoming and outgoing frame." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 50, + 474, + 284, + 521 + ], + "blocks": [ + { + "bbox": [ + 50, + 474, + 284, + 521 + ], + "lines": [ + { + "bbox": [ + 50, + 474, + 284, + 521 + ], + "spans": [ + { + "bbox": [ + 50, + 474, + 284, + 521 + ], + "type": "image", + "image_path": "1e973e8a48aedfc16da1ad981d8296568bf46dc4fa8e1f90400f327c3395b790.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 599, + 254, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 254, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 254, + 612 + ], + "type": "text", + "content": "3.3. Two-View Solver and Pre-Conditioning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "Our two-view solver aims to align the relative pose between two frames to the predicted matches by minimizing the distance between each match and its epipolar line. This strategy is accurate, however the objective function is nonconvex, meaning it will only converge to the global minimum if it is initialized close to it. In contrast, the 8-point-algorithm, a common approach to this task, has the opposite problem; it does not suffer from local minimum since it is a" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 247, + 528, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 247, + 528, + 304 + ], + "spans": [ + { + "bbox": [ + 294, + 247, + 528, + 304 + ], + "type": "text", + "content": "homogeneous least squares, but the solution is sub-optic. In Fig. 6, we visualize how our solver either converges within a fraction of a degree, or not at all, meanwhile" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 305, + 509, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 305, + 509, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 305, + 509, + 316 + ], + "type": "text", + "content": "8-point-algorithm is more robust but less accurate." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 318, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 318, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 294, + 318, + 545, + 425 + ], + "type": "text", + "content": "The approach our method uses is to pre-condition the pose estimate using a weighted dense variant of the 8-point algorithm [30], and then run our solver layer to refine the pose. This combined strategy obtains the best of both worlds, since the pre-conditioning will typically initialize the pose within the basin of convergence of our solver, which then refines the prediction further. We visualize this basin on the TartanAir [48] dataset using groundtruth matches in Fig. 6." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 429, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 429, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 429, + 545, + 453 + ], + "type": "text", + "content": "Pre-conditioning: We solve the homogeneous least squares problem [14]" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 340, + 458, + 545, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 458, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 340, + 458, + 545, + 479 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\mathbf {F}} \\overline {{| | d i a g (\\vec {w}) \\mathbf {M} \\vec {\\mathbf {F}} | | ^ {2}}} \\quad s. t. | | \\mathbf {F} | | ^ {2} = 1 \\tag {2}", + "image_path": "ce9c8adaa957748ab6ea89d0acfd26f4436e6433800071ed8d0e7322e3454ec3.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 482, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 482, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 482, + 545, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 482, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\vec{w}" + }, + { + "bbox": [ + 304, + 482, + 545, + 529 + ], + "type": "text", + "content": " are constructed from the anchor-match pairs and confidence predictions. The points are normalized to " + }, + { + "bbox": [ + 304, + 482, + 545, + 529 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 304, + 482, + 545, + 529 + ], + "type": "text", + "content": " beforehand. Afterwards, we reshape and uncalibrate " + }, + { + "bbox": [ + 304, + 482, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 482, + 545, + 529 + ], + "type": "text", + "content": " to obtain the essential matrix." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 530, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 590 + ], + "type": "text", + "content": "We obtain the four relative pose candidates following the procedure detailed in [13] and in the supplement. During training, we select the pose candidate closest to the ground-truth on the " + }, + { + "bbox": [ + 304, + 530, + 545, + 590 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 304, + 530, + 545, + 590 + ], + "type": "text", + "content": " manifold. During inference, we select the candidate by testing for chirality [13]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "SED Solver Layer: The solver layer seeks to minimize the distance between each match " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "m_{kj}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ", and the epipolar line induced by the current pose estimate and its respective anchor point " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "a_k" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ". The solver objective is bi-directional, since every image contains a unique set of anchor points; the objective is also known as the titular symmetric epipolar distance (SED) [11]. We parameterize the pose update as small rotations to the translation direction and orientation: " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\xi_{\\mathbf{t}}, \\xi_{\\mathbf{R}} \\in \\mathfrak{so}(3)" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ". These are the free variables in eq. 3. Let " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "K(i)" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " be the set of anchor indices for frame " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " be the" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19629" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 280, + 174 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 280, + 174 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 280, + 174 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 280, + 174 + ], + "type": "image", + "image_path": "61ddc7591b5c3aa86415c52e218a9eb77ac13fef93b52cef8712fe977e19148d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 187, + 289, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 289, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 289, + 243 + ], + "type": "text", + "content": "Figure 5. The flow of gradients through our solver. The updated matches from the RNN are supervised directly, and then detached from the gradient tape. The solver output is supervised with a pose loss. In the backward pass, gradients from the pose loss are used to supervise the confidence head in order to learn outlier rejection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 255, + 287, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 287, + 291 + ], + "type": "text", + "content": "output of the pre-conditioning stage, and " + }, + { + "bbox": [ + 47, + 255, + 287, + 291 + ], + "type": "inline_equation", + "content": "epi(x,R,t)" + }, + { + "bbox": [ + 47, + 255, + 287, + 291 + ], + "type": "text", + "content": " compute the epipolar line for a point " + }, + { + "bbox": [ + 47, + 255, + 287, + 291 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 255, + 287, + 291 + ], + "type": "text", + "content": " given the relative pose " + }, + { + "bbox": [ + 47, + 255, + 287, + 291 + ], + "type": "inline_equation", + "content": "(R,t)" + }, + { + "bbox": [ + 47, + 255, + 287, + 291 + ], + "type": "text", + "content": ". Formally, the minimization objective in our solver is" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 299, + 287, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 299, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 61, + 299, + 287, + 369 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} l _ {k j} = e p i \\left(a _ {k}, \\left(e ^ {\\xi_ {\\mathbf {R}}} \\mathbf {R}\\right) _ {i \\rightarrow j}, \\left(e ^ {\\xi_ {\\mathbf {t}}} \\mathbf {t}\\right) _ {i \\rightarrow j}\\right): k \\in K (i) \\\\ E _ {i \\rightarrow j} = \\sum_ {k \\in K (i)} w _ {k j} \\cdot \\| \\operatorname {e r r} \\left(m _ {k j}, l _ {k j}\\right) \\| _ {2} ^ {2} \\tag {3} \\\\ \\stackrel {\\smile}{S E D} = \\operatorname *{arg min}_{\\xi_{\\mathbf{R}},\\xi_{\\mathbf{t}}}\\left(E_{i\\to j} + E_{j\\to i}\\right) \\\\ \\end{array}", + "image_path": "65b7f059d6a88db765c4b1494aebc77addd48632ae179e26af8c6316e8dab16a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "inline_equation", + "content": "err" + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": " computes the 2D point-to-line error. The predicted weights " + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "inline_equation", + "content": "w_{kj}" + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": " are included to allow the network to down-weight the contribution of any match which it deems unreliable. This solver layer is implemented in Pytorch, enabling gradients to propagate backward from the pose loss to the confidence head using autograd. We depict the gradient flow in Fig. 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 508 + ], + "type": "text", + "content": "To minimize eq. 3, we employ the Levenberg-Marquardt algorithm. This requires computing the Jacobian for all terms. We provide the full derivations in the Supplement, but define the residual function " + }, + { + "bbox": [ + 46, + 460, + 287, + 508 + ], + "type": "inline_equation", + "content": "err" + }, + { + "bbox": [ + 46, + 460, + 287, + 508 + ], + "type": "text", + "content": " in its entirety here." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 535 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 510, + 287, + 535 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 510, + 287, + 535 + ], + "type": "text", + "content": " be the epipolar line produced by " + }, + { + "bbox": [ + 46, + 510, + 287, + 535 + ], + "type": "inline_equation", + "content": "(a, e^{\\xi_{\\mathbf{R}}}\\mathbf{R}, e^{\\xi_{\\mathbf{t}}}\\mathbf{t})" + }, + { + "bbox": [ + 46, + 510, + 287, + 535 + ], + "type": "text", + "content": ". We can express each term of eq. 3 as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 543, + 287, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 543, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 82, + 543, + 287, + 575 + ], + "type": "interline_equation", + "content": "\\operatorname {e r r} (m, l) = \\left(\\frac {l _ {x} m _ {x} + l _ {y} m _ {y} + l _ {z}}{l _ {x} ^ {2} + l _ {y} ^ {2}}\\right) \\left[ \\begin{array}{l} l _ {x} \\\\ l _ {y} \\end{array} \\right] \\tag {4}", + "image_path": "c69cd604055997b5d738e81f4a1a80c9422519b4782d77bf94873d1fe7280af1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 582, + 143, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 143, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 143, + 594 + ], + "type": "text", + "content": "and the epipolar line as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 87, + 601, + 287, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 601, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 87, + 601, + 287, + 657 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {E} = \\left(e ^ {\\xi_ {\\mathrm {R}}} \\mathbf {R}\\right) ^ {\\top} \\left[ e ^ {\\xi_ {\\mathrm {t}}} \\mathbf {t} \\right] _ {\\times} \\\\ l = \\mathbf {F} \\left[ \\begin{array}{l} a _ {x} \\\\ a _ {y} \\\\ 1 \\end{array} \\right] = \\left[ \\left(K _ {2} ^ {\\top}\\right) ^ {- 1} \\mathbf {E} K _ {1} ^ {\\top} \\right] \\left[ \\begin{array}{l} a _ {x} \\\\ a _ {y} \\\\ 1 \\end{array} \\right] \\tag {5} \\\\ \\end{array}", + "image_path": "f105632e74d4cbeaffdde1f65639f511b18384967428f7b42aaf586bbc7c72c4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 664, + 285, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 285, + 675 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 285, + 675 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 664, + 285, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 47, + 664, + 285, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 664, + 285, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 47, + 664, + 285, + 675 + ], + "type": "text", + "content": " are the fundamental and essential matrices." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 683, + 275, + 696 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 275, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 275, + 696 + ], + "type": "text", + "content": "3.4. Adapting our backbone to Visual Odometry" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 701, + 276, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 276, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 276, + 713 + ], + "type": "text", + "content": "To adapt our backbone to VO, we make several changes:" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 73, + 537, + 160 + ], + "blocks": [ + { + "bbox": [ + 316, + 73, + 537, + 160 + ], + "lines": [ + { + "bbox": [ + 316, + 73, + 537, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 537, + 160 + ], + "type": "image", + "image_path": "27405bfbdf6641ceb6b0276b35feefc787b9b8ea6be9bd0659c4ca2c0881034f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 172, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 305, + 172, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 172, + 545, + 228 + ], + "type": "text", + "content": "Figure 6. Convergence of our solver, given ground-truth matches (from TartanAir [48]). Our SED solver is accurate, but only converges to the global minimum when initialized close to it. We initialize the pose using the 8-pt-algorithm to ensure it is within the SED convergence basin." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "type": "text", + "content": "Multi-view Solver (BA): Our local optimizer minimizes reprojection error and treats depth as a separate variable. This is identical to the bundle adjustment from DPVO. Formally, let " + }, + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "type": "text", + "content": " be the set of connected frames, " + }, + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "type": "text", + "content": " be the global poses, " + }, + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "type": "text", + "content": " be the depth estimate for anchor " + }, + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "type": "inline_equation", + "content": "\\Pi(\\cdot)" + }, + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "type": "text", + "content": " be the " + }, + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "type": "inline_equation", + "content": "3D \\rightarrow 2D" + }, + { + "bbox": [ + 305, + 241, + 545, + 324 + ], + "type": "text", + "content": " projection function. The bundle adjustment objective is:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 334, + 555, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 334, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 555, + 372 + ], + "type": "interline_equation", + "content": "\\underset {G, \\mathbf {d}} {\\arg \\min } \\sum_ {(i, j) \\in F} \\sum_ {k \\in K (i)} w _ {k j} \\cdot \\left\\| \\Pi \\left[ G _ {j} ^ {- 1} G _ {i} \\Pi^ {- 1} \\left(a _ {k}, d _ {k}\\right) \\right] - m _ {k j} \\right\\| ^ {2} \\tag {6}", + "image_path": "bdf4a1df381f7f72fe3cab3c1080791d57151cd6711acc618e9e1fbb74e43949.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 373, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 408 + ], + "type": "text", + "content": "The preconditioning stage is not necessary as we can linearly-extrapolate the camera poses from previous frames to achieve good initialization." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "text", + "content": "Clamping: Since epipolar lines do not make sense in the multi-view setting, we reset the optical flow using the pose and depth from the bundle adjustment." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 452, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 452, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 305, + 452, + 545, + 500 + ], + "type": "text", + "content": "RNN changes: We include mechanisms in the RNN to share latent features between updates if they stem from the same anchor point. Specifically, we use the message passing and temporal convolutions from DPVO [45]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 508, + 430, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 508, + 430, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 508, + 430, + 521 + ], + "type": "text", + "content": "3.5. Trajectory Alignment" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 527, + 545, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 527, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 305, + 527, + 545, + 623 + ], + "type": "text", + "content": "In the Multi-Session SLAM problem, we estimate a relative Sim(3) between pairs of disjoint trajectories in order to align them. All anchor points already have a depth estimate " + }, + { + "bbox": [ + 305, + 527, + 545, + 623 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 305, + 527, + 545, + 623 + ], + "type": "text", + "content": " as a result of the bundle adjustment (eq. 6) from the VO system. We use our two-view method to estimate the relative rotation and translation direction, and compute the translation magnitude and relative scaling by comparing the " + }, + { + "bbox": [ + 305, + 527, + 545, + 623 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 305, + 527, + 545, + 623 + ], + "type": "text", + "content": "'s to the depth from the triangulated two-view matches." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 626, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 306, + 626, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 626, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 626, + 545, + 685 + ], + "type": "text", + "content": "1) Estimate relative rotation, translation direction, and matches. We first retrieve a candidate image pair " + }, + { + "bbox": [ + 306, + 626, + 545, + 685 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 306, + 626, + 545, + 685 + ], + "type": "text", + "content": " using NetVLAD [1], one from each trajectory, and apply our two-view model to estimate their relative rotation and translation direction." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "2) Align the depth from the two-view and SLAM operators. We previously defined " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " as the depth output of the VO" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19630" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "system. Let " + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "d_k'" + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": " be the triangulated depth obtained from our two-view matches:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 98, + 287, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 98, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 98, + 287, + 129 + ], + "type": "interline_equation", + "content": "d _ {k} ^ {\\prime} = \\left\\{\\begin{array}{l l}\\operatorname {t r i a n g} \\left(\\mathbf {R} _ {i \\rightarrow j}, \\mathbf {t} _ {i \\rightarrow j}, a _ {k}, m _ {k j}\\right)&: k \\in K (i)\\\\\\operatorname {t r i a n g} \\left(\\mathbf {R} _ {j \\rightarrow i}, \\mathbf {t} _ {j \\rightarrow i} a _ {k}, m _ {k i}\\right)&: k \\in K (j)\\end{array}\\right. \\tag {7}", + "image_path": "3b1b469411816f316cb514eb8c0ea5b3539eba16c711bbddd997c8cd42811d88.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 137, + 87, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 137, + 87, + 147 + ], + "spans": [ + { + "bbox": [ + 47, + 137, + 87, + 147 + ], + "type": "text", + "content": "We solve" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 144, + 287, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 144, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 90, + 144, + 287, + 177 + ], + "type": "interline_equation", + "content": "\\underset {s _ {i} \\in \\mathbb {R} > 0} {\\arg \\max } \\sum_ {k \\in K (i)} \\mathbf {1} \\left[ \\frac {1}{\\lambda} < \\frac {d _ {k}}{s _ {i} \\cdot d _ {k} ^ {\\prime}} < \\lambda \\right] \\tag {8}", + "image_path": "1da862fc134e056ce8563e425929917fd0fe636764b78923b24d5c20c4d07795.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 182, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 287, + 289 + ], + "type": "text", + "content": "to recover the translation magnitude, where the hyperparameter " + }, + { + "bbox": [ + 47, + 182, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\lambda = 1.05" + }, + { + "bbox": [ + 47, + 182, + 287, + 289 + ], + "type": "text", + "content": ". In layman terms, eq. 9 seeks what translation magnitude would align the most triangulated points to the existing 3D map? We can obtain a reasonably good solution to eq 9 by brute-force checking " + }, + { + "bbox": [ + 47, + 182, + 287, + 289 + ], + "type": "inline_equation", + "content": "s_i = d_k / d_k'" + }, + { + "bbox": [ + 47, + 182, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\forall k\\in K(i)" + }, + { + "bbox": [ + 47, + 182, + 287, + 289 + ], + "type": "text", + "content": ", similar to RANSAC. If the number of inliers is too small, we retry on a new candidate pair. Conversely, the scale-difference is " + }, + { + "bbox": [ + 47, + 182, + 287, + 289 + ], + "type": "inline_equation", + "content": "s_i / s_j" + }, + { + "bbox": [ + 47, + 182, + 287, + 289 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 182, + 287, + 289 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 47, + 182, + 287, + 289 + ], + "type": "text", + "content": " is estimated the same way, but for anchor points in " + }, + { + "bbox": [ + 47, + 182, + 287, + 289 + ], + "type": "inline_equation", + "content": "K(j)" + }, + { + "bbox": [ + 47, + 182, + 287, + 289 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 289, + 287, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 289, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 89, + 289, + 287, + 323 + ], + "type": "interline_equation", + "content": "\\underset {s _ {j} \\in \\mathbb {R} > 0} {\\arg \\max } \\sum_ {k \\in K (j)} \\mathbf {1} \\left[ \\frac {1}{\\lambda} < \\frac {d _ {k}}{s _ {j} \\cdot d _ {k} ^ {\\prime}} < \\lambda \\right] \\tag {9}", + "image_path": "646c2a855d6e8a0b664a98b544804c06144ced0189f7493e6e6759fcd5601c05.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 332, + 287, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 332, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 332, + 287, + 367 + ], + "type": "text", + "content": "Given the rotation " + }, + { + "bbox": [ + 47, + 332, + 287, + 367 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{j\\rightarrow i}" + }, + { + "bbox": [ + 47, + 332, + 287, + 367 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 47, + 332, + 287, + 367 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{j\\rightarrow i}" + }, + { + "bbox": [ + 47, + 332, + 287, + 367 + ], + "type": "text", + "content": " from the two-view solver layers, the resulting transformation can be used to align the two trajectories:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 101, + 371, + 287, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 371, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 101, + 371, + 287, + 397 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {j \\rightarrow i} = \\left[\\begin{array}{c c}\\frac {s _ {i}}{s _ {j}} \\mathbf {R}&s _ {j} \\mathbf {t}\\\\0&1\\end{array}\\right] \\in \\operatorname {S i m} (3) \\tag {10}", + "image_path": "2f37191d3b341c0ccb0aa31366e092320cf32ba33abf644366f16d59405bec37.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 401, + 276, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 401, + 276, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 401, + 276, + 413 + ], + "type": "text", + "content": "The graphs are then merged by concatenating all buffers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 420, + 111, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 111, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 111, + 433 + ], + "type": "text", + "content": "3.6. Training" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "type": "text", + "content": "We train our backbone separately for VO and for two-view pose. It is trained using a matching loss and a pose loss. The pose loss for the two-view training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 483, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 483, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 53, + 483, + 287, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p o s e 2 V}} = \\sum_ {t = 1} ^ {1 2} \\cos^ {- 1} \\left(\\bar {\\mathbf {t}} ^ {t} \\cdot \\bar {\\mathbf {t}} ^ {\\bar {\\mathbf {g}}} ^ {t}\\right) + \\alpha \\left\\| \\left[ \\left(\\mathbf {R} ^ {t}\\right) ^ {T} \\mathbf {R} ^ {\\mathbf {g}} ^ {t} \\right] \\right\\| _ {S O (3)} \\tag {11}", + "image_path": "55e36ab656d2b477e8c38b1b7b5cd0fe427e3cac98790bf7b218342e33424dca.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 523, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 287, + 570 + ], + "type": "text", + "content": "penalizes angle error for both the predicted translation direction and predicted orientation. The pose loss for the VO training is the SE3 manifold distance between the predicted and ground-truth poses:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 91, + 571, + 287, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 571, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 91, + 571, + 287, + 604 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p o s e V O}} = \\sum_ {t = 1} ^ {T} \\left\\| \\left[ \\left(\\mathbf {G} ^ {t}\\right) ^ {- 1} \\mathbf {G} ^ {\\mathbf {g t}} \\right] \\right\\| _ {S E (3)} \\tag {12}", + "image_path": "ce7b676b8ea66dd0d4f069bdea65ccec927263cbb8222b52aa38f061a000d25e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 607, + 123, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 607, + 123, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 607, + 123, + 617 + ], + "type": "text", + "content": "The matching loss" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 84, + 618, + 287, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 618, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 84, + 618, + 287, + 651 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a t c h i n g}} = \\frac {1}{| N |} \\sum_ {a} ^ {N} \\sum_ {t = 1} ^ {T} \\left\\| m _ {a} ^ {t} - m _ {a} ^ {\\mathbf {g t}} \\right\\| _ {2} \\tag {13}", + "image_path": "3a998f84efe3df611582663162cc478d63ea22e134620454e1f358b5fcde6285.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "content": "is a standard endpoint-error [24, 43], applied only on pixels which have valid depth and are verified to be visible in both images. Our update operator is applied 12 times per training example; supervision is applied after every update (See Fig. 5). The final loss is " + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\text{pose}} + \\beta \\cdot \\mathcal{L}_{\\text{matching}}" + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 385, + 167 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 385, + 167 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 385, + 167 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 385, + 167 + ], + "type": "image", + "image_path": "29bd9d0c7eec35b33a57b9d37784db0e9d3d963bd66c78ce6335dfb35a6bad10.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 170, + 384, + 179 + ], + "lines": [ + { + "bbox": [ + 313, + 170, + 384, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 384, + 179 + ], + "type": "text", + "content": "Stage 1: Image Retrieval" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 389, + 72, + 470, + 163 + ], + "blocks": [ + { + "bbox": [ + 389, + 72, + 470, + 163 + ], + "lines": [ + { + "bbox": [ + 389, + 72, + 470, + 163 + ], + "spans": [ + { + "bbox": [ + 389, + 72, + 470, + 163 + ], + "type": "image", + "image_path": "667dacec7e81c5a3476d64d62d609f784f7d21e2d0471acfa6c8a9a155509477.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 167, + 451, + 182 + ], + "lines": [ + { + "bbox": [ + 404, + 167, + 451, + 182 + ], + "spans": [ + { + "bbox": [ + 404, + 167, + 451, + 182 + ], + "type": "text", + "content": "Stage 2: Sim(3) Estimation" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 169, + 543, + 179 + ], + "lines": [ + { + "bbox": [ + 474, + 169, + 543, + 179 + ], + "spans": [ + { + "bbox": [ + 474, + 169, + 543, + 179 + ], + "type": "text", + "content": "Stage 3: Graph Merging" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 474, + 77, + 541, + 161 + ], + "blocks": [ + { + "bbox": [ + 474, + 77, + 541, + 161 + ], + "lines": [ + { + "bbox": [ + 474, + 77, + 541, + 161 + ], + "spans": [ + { + "bbox": [ + 474, + 77, + 541, + 161 + ], + "type": "image", + "image_path": "89462d1f56b69e3ee1367046a5087e4dec77904aab2a5d46fae3b8ab21e837ad.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 191, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 304, + 191, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 545, + 257 + ], + "type": "text", + "content": "Figure 7. High-level overview of our approach to aligning disjoint trajectories. We use NetVLAD [1] to retrieve two co-visible images, one from each trajectory, and apply our two-view update operator to estimate a relative Sim(3) alignment, following the procedure detailed in Sec 3.5. We then transform the second trajectory into the reference frame of the first, and merge all buffers." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 270, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 545, + 353 + ], + "type": "text", + "content": "Following prior work [23], we pre-train our two-view method on synthetic homographies (without pose loss) for two epochs on the Oxford-Paris 1M Distractors dataset [29]. We then tune our model on a " + }, + { + "bbox": [ + 305, + 270, + 545, + 353 + ], + "type": "inline_equation", + "content": "50/50" + }, + { + "bbox": [ + 305, + 270, + 545, + 353 + ], + "type": "text", + "content": " mixture of Scannet [6] and Megadepth [21] for 100,000 steps and a batch size of 120, using 10 A6000 GPUs for 5 days. The VO backbone is trained using the procedure from [45] on TartanAir." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 365, + 462, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 462, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 462, + 377 + ], + "type": "text", + "content": "3.7. Multi-Session SLAM System" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 385, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 481 + ], + "type": "text", + "content": "Overview: To perform Multi-Session SLAM, we use our VO-trained backbone to perform visual odometry and global optimization; our two-view backbone connects disjoint trajectories following the procedure in Sec 3.5. Our full system builds the pose graph incrementally: new video frames inserted into the graph as they are received, and unjoined trajectories are aligned and merged as soon as a connection is found." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 487, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 582 + ], + "type": "text", + "content": "VO Frontend: Our VO frontend extracts features and estimates camera poses and depth for incoming frames, operating on a sliding window covering the most recently observed 20 keyframes. Keyframing occurs retroactively on the " + }, + { + "bbox": [ + 304, + 487, + 545, + 582 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 304, + 487, + 545, + 582 + ], + "type": "text", + "content": " oldest keyframe if it is found to be redundant. New poses are initialized using a linear-motion model, and depth is copied from the previous frame. Like DPVO [45], we initialize after 8 frames with significant motion are observed." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 588, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 545, + 660 + ], + "type": "text", + "content": "Global Optimization: We perform global optimization by introducing proximity factors using the existing pose and depth measurements, following the approach from DROID-SLAM [44], and update the matches and poses/depth using the update operator. We only run the backend periodically, and after joining trajectories." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Trajectory Joining: We query the database of NetVLAD descriptors to find co-visible frames. If several pairs are retrieved, we compute relative poses in a mini-batch and use the estimate with the highest inlier-ratio." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19631" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 286, + 190 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 190 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 190 + ], + "type": "table", + "html": "
Overall ApproachPose error AUC [%] ↑
@5°@10°@20°
Superglue [34]Matching16.2(17.7)33.8(35.6)51.8(54.2)
LightGlue [23]16.5(19.4)33.4(36.9)50.1(53.5)
LoFTR [41]↓22.1(25.7)40.1(45.0)47.6(61.4)
MatchFormer [47]RANSAC [15]24.3(27.3)43.9(47.6)61.4(64.9)
ASpanFormer [5](LO-RANSAC [18])25.6(28.4)46.0(48.8)63.3(65.8)
Roessle&Nießner [31]Matching→W8PA20.741.661.7
Roessle&Nießner [31]Matching→W8PA→BA25.747.266.4
OursOptical-Flow\n←\nClamp ← Solver30.951.167.6
", + "image_path": "60b32ef2eaa78fe043510dcc2636baa972208dbe0e9d87d8a5040b4b47413719.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 272, + 285, + 391 + ], + "blocks": [ + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "type": "text", + "content": "Table 1. Two-view results on Scannet [6]. Existing methods perform matching as pre-processing, then use a solver; our method is iterative. Results in parenthesis use the LO-RANSAC [18] estimator. \"W8PA " + }, + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 198, + 287, + 264 + ], + "type": "text", + "content": " BA\" = weighted 8-point-algo. & bundle-adjustment. Our design outperforms prior approaches. We use the same model weights as in Tab. 2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 272, + 285, + 391 + ], + "lines": [ + { + "bbox": [ + 50, + 272, + 285, + 391 + ], + "spans": [ + { + "bbox": [ + 50, + 272, + 285, + 391 + ], + "type": "table", + "html": "
Overall ApproachPose error AUC [%] ↑
@5°@10°@20°
Superglue [34]Matching49.7(65.8)67.1(78.7)80.6(87.5)
LightGlue [23]49.9(66.7)67.0(79.3)80.1(87.9)
LoFTR [41]↓52.8(66.4)69.2(78.6)81.2(86.5)
MatchFormer [47]RANSAC [15]53.3(66.5)69.7(78.9)81.8(87.5)
ASpanFormer [5](LO-RANSAC [18])55.3(69.4)71.5(81.1)83.1(88.9)
Roessle&Nießner [31]Matching→W8PA46.962.876.3
Roessle&Nießner [31]Matching→W8PA→BA61.274.985.0
OursOptical-Flow60.472.781.3
Clamp ← Solver
", + "image_path": "a8635de030a474c883167004848236e7df002e5eeda0b1190445814da91f04b9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 287, + 497 + ], + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 497 + ], + "type": "text", + "content": "Table 2. Two-view results on Megadepth [21]. Existing methods perform matching as pre-processing, then use a solver; our method is iterative. \"W8PA " + }, + { + "bbox": [ + 46, + 399, + 287, + 497 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 399, + 287, + 497 + ], + "type": "text", + "content": " BA\" = weighted 8-point-algo. & bundle-adjustment. Results in parenthesis use the LO-RANSAC [18] estimator. Our radically different design leads to better results indoors (Tab. 1), but is less competitive in photo-tourism settings where high-volume matching is easier and therefore RANSAC-based methods work better. We use the same model weights as in Tab. 1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 506, + 127, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 127, + 519 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 127, + 519 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 526, + 287, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 287, + 551 + ], + "type": "text", + "content": "We evaluate the performance of our two-view method in isolation, and our Multi-Session SLAM system as a whole." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 555, + 170, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 170, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 170, + 566 + ], + "type": "text", + "content": "4.1. Two-View Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 569, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 676 + ], + "type": "text", + "content": "We evaluate our two-view method on two popular relative pose benchmarks: The Scannet [6] 1500 and MegaDepth [21] 1500 test datasets. We compare to existing matching networks [5, 23, 34, 41, 47], which are typically in service of improving COLMAP [38]-based SfM pipelines [22, 33]. In contrast, our approach is in service of Multi-Session SLAM. We report pose error AUC, as is done in prior matching work (higher is better). We use the same configuration and model weights on both datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Most existing methods perform matching as a preprocessing step using a deep network, then estimate pose with a minimal solver using RANSAC [15, 18]. Roessle" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "and Nießner [31] replace the RANSAC optimizer with a weighted-8-point algorithm and bundle adjustment. In contrast, our two-view method couples the prediction of pose and matches, refining the predictions over many iterations. We ablate over our two-view model in Tab. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 135, + 545, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 135, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 135, + 545, + 182 + ], + "type": "text", + "content": "Scannet: We report two-view results on Scannet [6] in Tab. 1. Our approach outperforms existing methods on Scannet, which contains fewer salient keypoints and many texture-less surfaces and motion blur." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 185, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 269 + ], + "type": "text", + "content": "MegaDepth: We report two-view results on Megadepth [21] in Tab. 2. On Megadepth, the matching problem is substantially easier since photo-tourism images contain many salient keypoints, resulting in better performance across all methods; our approach is on-par with existing approaches using the OpenCV RANSAC pose estimator [15], but not the LO-RANSAC [18] estimator." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 274, + 478, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 274, + 478, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 274, + 478, + 285 + ], + "type": "text", + "content": "4.2. Multi-Session SLAM evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 289, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 421 + ], + "type": "text", + "content": "We evaluate our approach on Multi-Session SLAM on the EuRoC-MAV [2] and ETH3D [39] datasets, since they provide camera poses under a single global reference. Following the evaluation for single-video SLAM, we align the final predictions to the ground truth by computing a global 7-DOF alignment to account for the guage freedoms. We report RMSE ATE[m] in metric units. We sample 96 anchors in each video frame using a mix of random and Superpoint [8] keypoints. Our two-view method uses those same anchors to join pairs of trajectories. We use the same weights and configuration on both datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 423, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 628 + ], + "type": "text", + "content": "EuRoC-MAV: In Tab. 9 and Fig. 10, we report results on the EuRoC MAV dataset [2]. Following the evaluation from [4], we compare on four groups of disjoint trajectories across the three distinct environments: 3 in the Machine Hall, 5 in the Machine Hall, 3 in Vicon 1, and 3 in Vicon 2. Ground-truth poses are obtained using a laser tracker and Vicon cameras. These video sequences are long, with the entirety of the Machine-Hall sequence being 11.5 minutes of video (13.7K frames) spread across 5 sequences. Our approach achieves significantly lower error than ORBSLAM3 on all groups, including 12x lower error on the Vicon 2 sequences (0.022 vs 0.284). Video is recorded at 20-FPS; all methods run in real-time. We also compare to CCM-SLAM [36], which only provides results on the on MH01-03 sequence, the mono-inertial ORB-SLAM3, and the mono-inertial VINS [28]. We outperform the other approaches on all trajectory groups." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "ETH3D: In Tab. 4, we report results on the training set from ETH3D [39]. We compare across 5 unique scenes composed of multiple trajectories: 1-4 from Sofa, 3&4 from Table, 1-3 from Plant Scene, 1&2 from Einstein, and 2&3 from Planar. Sequences are excluded if they belong to the official ETH3D test set (with no ground-truth) or are from a different scene altogether. The sequences are trimmed" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19632" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 545, + 190 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 545, + 190 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 545, + 190 + ], + "type": "image", + "image_path": "39ee103b7e694f32df19e18b1a2b150cf3540eb70b49e77258168d64285f4f64.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 546, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 546, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 546, + 236 + ], + "type": "text", + "content": "Figure 8. Qualitative results on Scannet. Our two-view variant is able to estimate accurate relative poses across wide camera baselines. It initializes all matches with uniform depth and identity relative pose. Progressive applications of our update operator lead to more accurate matches and higher predicted confidence. Some image pairs (row 2) take more iterations to converge than others (row 1)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 245, + 286, + 371 + ], + "blocks": [ + { + "bbox": [ + 50, + 245, + 286, + 371 + ], + "lines": [ + { + "bbox": [ + 50, + 245, + 286, + 371 + ], + "spans": [ + { + "bbox": [ + 50, + 245, + 286, + 371 + ], + "type": "table", + "html": "
Scene nameMH01-03MH01-05V101-103V201-203Cam.
# Disjoint Trajectories3533hz
Ours0.0220.0360.0300.02220
Mono-Visual
CCM-SLAM [36]0.077---38
Mono-Visual
ORB-SLAM3 [4]0.0300.0580.0580.28420
Mono-Visual
VINS [28]-0.210---
Mono-Inertial
ORB-SLAM30.0370.0650.0400.04820
Mono-Inertial
", + "image_path": "a6f21b60a13fa1e893b8034709b4fe09dd45c4f114444ffc158dfe4ae53f913a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 53, + 439, + 279, + 622 + ], + "blocks": [ + { + "bbox": [ + 46, + 378, + 287, + 433 + ], + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 433 + ], + "type": "text", + "content": "Figure 9. Multi-Session SLAM evaluation on EuRoC MAV [2] using RMSE ATE[m] " + }, + { + "bbox": [ + 46, + 378, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 378, + 287, + 433 + ], + "type": "text", + "content": ". We perform similarly to ORB-SLAM3 (within " + }, + { + "bbox": [ + 46, + 378, + 287, + 433 + ], + "type": "inline_equation", + "content": "2mm" + }, + { + "bbox": [ + 46, + 378, + 287, + 433 + ], + "type": "text", + "content": ") in the Machine Hall, but perform best in the Vicon 1 and Vicon 2 rooms. Results for baseline approaches were obtained from [4]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 439, + 279, + 622 + ], + "lines": [ + { + "bbox": [ + 53, + 439, + 279, + 622 + ], + "spans": [ + { + "bbox": [ + 53, + 439, + 279, + 622 + ], + "type": "image", + "image_path": "2055a6290fe878c490de84d2166b9784dbc2a86d85c363ac87fb554965774010.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 625, + 287, + 647 + ], + "lines": [ + { + "bbox": [ + 46, + 625, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 287, + 647 + ], + "type": "text", + "content": "Figure 10. Our prediction on the Vicon1 [2] sequences. Our method is accurate and robust to chaotic motion." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "from a single, longer video; to make joining the trajectories non-trivial, we reverse every other sequence, ensuring that there is a sufficient disconnect between subsequent videos. This is a novel benchmark, so [28, 36] do not report results." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 246, + 544, + 344 + ], + "blocks": [ + { + "bbox": [ + 308, + 246, + 544, + 344 + ], + "lines": [ + { + "bbox": [ + 308, + 246, + 544, + 344 + ], + "spans": [ + { + "bbox": [ + 308, + 246, + 544, + 344 + ], + "type": "table", + "html": "
Ablation ExperimentPose error AUC @10°[%] ↑
Megadepth ValScannet Val
BaselineFull model53.333.9
ArchitectureShaller correlation pyramid44.327.1
No attention in update-op48.630.5
No ReLU-attn in feat-extractor47.429.7
SolverNo pre-conditioning35.520.6
No SED solver23.813.1
No clamping step21.910.7
", + "image_path": "b1e87adcb4b98685c5612a2112637968e489724b56ce20a867656af0cd1ff604.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 417, + 544, + 487 + ], + "blocks": [ + { + "bbox": [ + 305, + 350, + 546, + 405 + ], + "lines": [ + { + "bbox": [ + 305, + 350, + 546, + 405 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 546, + 405 + ], + "type": "text", + "content": "Table 3. Two-view ablation experiments on two validation sets. Our proposed RNN and feature extractor changes from DPVO improve the result. Removing components from our proposed solver design decreases performance, as does removing the reprojection step." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 417, + 544, + 487 + ], + "lines": [ + { + "bbox": [ + 308, + 417, + 544, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 417, + 544, + 487 + ], + "type": "table", + "html": "
Scene nameSofaTablePlant SceneEinsteinPlanar
# Disjoint Trajectories42322
ORB-SLAM3 [4] \nMono-VisualFAIL \n(no init)0.018FAIL \n(init→lost)FAIL \n(init→lost)0.010
Ours \nMono-Visual0.0130.0100.0090.0050.035
", + "image_path": "562569f4080a5ff80200bb038c3de4177d6c4654a1e6efc55ca75371dd1f0c97.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 493, + 545, + 515 + ], + "lines": [ + { + "bbox": [ + 305, + 493, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 493, + 545, + 515 + ], + "type": "text", + "content": "Table 4. Multi-Session SLAM evaluation on ETH3D [39] using RMSE ATE[m] " + }, + { + "bbox": [ + 305, + 493, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 305, + 493, + 545, + 515 + ], + "type": "text", + "content": " . Our method is robust to catastrophic failures." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 527, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 527, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 305, + 527, + 545, + 588 + ], + "type": "text", + "content": "ORB-SLAM3 fails on Sofa, Plant Scene, and Einstein for various reasons (couldn't initialize, or lost the feature tracks). Our approach outperforms ORB-SLAM3 on 4/5 groups, and does not fail on any. ORB-SLAM3 only succeeds on 2/5 groups." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 601, + 378, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 601, + 378, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 601, + 378, + 613 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We introduce a new method for mono-visual Multi-Session SLAM. Our system utilizes a novel backbone which can estimate two-view pose and perform visual odometry. We leverage a novel differentiable solver which minimizes the symmetric epipolar distance. We compare against existing approaches and show strong performance across several datasets. This work was partially supported by IARPA and the National Science Foundation." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19633" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pa-jdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5297-5307, 2016. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "type": "text", + "content": "[2] Michael Burri, Janosch Nikolic, Pascal Gohl, Thomas Schneider, Joern Rehder, Sammy Omari, Markus W Achtelik, and Roland Siegwart. The euroc micro aerial vehicle datasets. The International Journal of Robotics Research, 35 (10):1157-1163, 2016. 1, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 258 + ], + "type": "text", + "content": "[3] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 260, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 288, + 315 + ], + "type": "text", + "content": "[4] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 1, 2, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 317, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 288, + 372 + ], + "type": "text", + "content": "[5] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming-min Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 373, + 288, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 288, + 429 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 288, + 429 + ], + "type": "text", + "content": "[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 430, + 288, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 288, + 473 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 288, + 473 + ], + "type": "text", + "content": "[7] Hayyan Afeef Daoud, Aznul Qalid Md. Sabri, Chu Kiong Loo, and Ali Mohammed Mansoor. Slamm: Visual monocular slam with continuous mapping using multiple maps. PloS one, 13(4):e0195878, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 475, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 288, + 529 + ], + "type": "text", + "content": "[8] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 532, + 288, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 532, + 288, + 573 + ], + "spans": [ + { + "bbox": [ + 53, + 532, + 288, + 573 + ], + "type": "text", + "content": "[9] Jakob Engel, Thomas Schöps, and Daniel Cremers. Lsd-slam: Large-scale direct monocular slam. In European conference on computer vision, pages 834-849. Springer, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "type": "text", + "content": "[10] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. IEEE transactions on pattern analysis and machine intelligence, 40(3):611-625, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "type": "text", + "content": "[11] Mohammed E Fathy, Ashraf S Hussein, and Mohammed F Tolba. Fundamental matrix estimation: A study of error criteria. Pattern Recognition Letters, 32(2):383-391, 2011. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "text", + "content": "[12] Dorian Gálvez-López and Juan D Tardos. Bags of binary words for fast place recognition in image sequences. IEEE Transactions on Robotics, 28(5):1188-1197, 2012. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[13] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, New York, NY, USA, 2 edition, 2003. 4, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Richard I Hartley. In defense of the eight-point algorithm. IEEE Transactions on pattern analysis and machine intelligence, 19(6):580-593, 1997. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "type": "text", + "content": "[15] Itseez. Open source computer vision library. https://github.com/itseez/opencv, 2015.7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "type": "text", + "content": "[16] Mathieu Labbe and François Michaud. Rtab-map as an opensource lidar and visual simultaneous localization and mapping library for large-scale and long-term online operation. Journal of field robotics, 36(2):416-446, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[17] Mathieu Labbe and François Michaud. Multi-session visual slam for illumination-invariant re-localization in indoor environments. Frontiers in Robotics and AI, 9:801886, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 220, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 241 + ], + "type": "text", + "content": "[18] Viktor Larsson and contributors. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "text", + "content": "[19] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 287, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 342 + ], + "type": "text", + "content": "[20] Stefan Leutenegger, Paul Furgale, Vincent Rabaud, Margarita Chli, Kurt Konolige, and Roland Siegwart. Keyframe-based visual-inertial slam using nonlinear optimization. Proceedings of Robotics Science and Systems (RSS) 2013, 2013. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "type": "text", + "content": "[21] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Computer Vision and Pattern Recognition (CVPR), 2018. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 377, + 547, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 547, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 547, + 432 + ], + "type": "text", + "content": "[22] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987-5997, 2021. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 434, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 465 + ], + "type": "text", + "content": "[23] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In ICCV, 2023. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "type": "text", + "content": "[24] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "text", + "content": "[25] Xiao Xin Lu. A review of solutions for perspective-n-point problem in camera pose estimation. In Journal of Physics: Conference Series, page 052009. IOP Publishing, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "type": "text", + "content": "[26] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "text", + "content": "[27] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[28] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics, 34(4):1004-1020, 2018. 1, 2, 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[29] F. Radenović, A. Iscen, G. Tolias, Y. Avrithis, and O. Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In CVPR, 2018. 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19634" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[30] Rene Ranftl and Vladlen Koltun. Deep fundamental matrix estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "text", + "content": "[31] Barbara Roessle and Matthias Nießner. End2end multi-view feature matching with differentiable pose optimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 477-487, 2023. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "text", + "content": "[32] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 252 + ], + "type": "text", + "content": "[33] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 288, + 308 + ], + "type": "text", + "content": "[34] 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. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "type": "text", + "content": "[35] Patrik Schmuck and Margarita Chli. Multi-uav collaborative monocular slam. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3863-3870. IEEE, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "text", + "content": "[36] Patrik Schmuck and Margarita Chli. Ccm-slam: Robust and efficient centralized collaborative monocular simultaneous localization and mapping for robotic teams. Journal of Field Robotics, 36(4):763-781, 2019. 1, 2, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "type": "text", + "content": "[37] Thomas Schneider, Marcin Dymczyk, Marius Fehr, Kevin Egger, Simon Lynen, Igor Gilitschenski, and Roland Siegwart. maplab: An open framework for research in visual-inertial mapping and localization. IEEE Robotics and Automation Letters, 3(3):1418-1425, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 489 + ], + "type": "text", + "content": "[38] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "text", + "content": "[39] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 534, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 567 + ], + "type": "text", + "content": "[40] Hauke Strasdat, J Montiel, and Andrew J Davison. Scale drift-aware large scale monocular slam. Robotics: science and Systems VI, 2(3):7, 2010. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "text", + "content": "[41] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[42] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[43] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 318 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "text", + "content": "[44] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1, 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 546, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 152 + ], + "type": "text", + "content": "[45] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 546, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 546, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 546, + 195 + ], + "type": "text", + "content": "[46] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 546, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 546, + 251 + ], + "type": "text", + "content": "[47] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, pages 2746-2762, 2022. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 252, + 546, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 252, + 546, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 252, + 546, + 318 + ], + "type": "text", + "content": "[48] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 4, 5" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19635" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/b0f5aa24-c6e7-4f07-a0f8-7432b1bda5de_content_list.json b/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/b0f5aa24-c6e7-4f07-a0f8-7432b1bda5de_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1641cf7f8463cb153db831c344519f2251caf5d7 --- /dev/null +++ b/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/b0f5aa24-c6e7-4f07-a0f8-7432b1bda5de_content_list.json @@ -0,0 +1,1734 @@ +[ + { + "type": "text", + "text": "Multi-Space Alignments Towards Universal LiDAR Segmentation", + "text_level": 1, + "bbox": [ + 153, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youquan Liu\\*,1 Lingdong Kong\\*,2,3 Xiaoyang Wu4 Runnan Chen4 Xin Li5 Liang Pan2 Ziwei Liu6 Yuexin Ma1", + "bbox": [ + 158, + 179, + 825, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ShanghaiTech University, $^{2}$ Shanghai AI Laboratory, $^{3}$ National University of Singapore $^{4}$ University of Hong Kong, $^{5}$ East China Normal University, $^{6}$ S-Lab, Nanyang Technological University https://github.com/youquanl/M3Net", + "bbox": [ + 88, + 215, + 897, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 303, + 313, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A unified and versatile LiDAR segmentation model with strong robustness and generalizability is desirable for safe autonomous driving perception. This work presents M3Net, a one-of-a-kind framework for fulfilling multi-task, multi-dataset, multi-modality LiDAR segmentation in a universal manner using just a single set of parameters. To better exploit data volume and diversity, we first combine large-scale driving datasets acquired by different types of sensors from diverse scenes and then conduct alignments in three spaces, namely data, feature, and label spaces, during the training. As a result, M3Net is capable of taming heterogeneous data for training state-of-the-art LiDAR segmentation models. Extensive experiments on twelve LiDAR segmentation datasets verify our effectiveness. Notably, using a shared set of parameters, M3Net achieves $75.1\\%$ , $83.1\\%$ , and $72.4\\%$ mIoU scores, respectively, on the official benchmarks of SemanticKITTI, nuScenes, and Waymo Open.", + "bbox": [ + 75, + 335, + 473, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 623, + 209, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dense and structural 3D surrounding scene understanding provides crucial information for autonomous vehicles to make proper decisions [70]. With the recent advancements in sensing technologies, especially the Light Detection and Ranging (LiDAR) sensor, a holistic scene perception can be achieved by segmenting the acquired sensor data [28, 83].", + "bbox": [ + 75, + 648, + 468, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing LiDAR segmentation models [1, 36, 111, 121, 128] are trained and tested in a single-task, single-dataset, single-modality manner. Despite achieving commendable results in the single domain, there is a significant performance drop when transitioning to new domains [40, 48]. The limited generalization capability hinders their facilitation of real-world applications [47, 50, 87]. In reality, LiDAR datasets are marred by significant variances, encompassing variations in data patterns due to different", + "bbox": [ + 75, + 739, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/610d391e645481a25a4707f26a4e4b53a8d20eca1327984146383805462fea08.jpg", + "image_caption": [ + "Figure 1. Performance comparisons among M3Net [●], Single-Dataset Training [●], and Naïve Joint Training [●] across twelve LiDAR segmentation datasets. For better comparisons, the radius is normalized based on M3Net's scores. The larger the area coverage, the higher the overall performance. Best viewed in colors." + ], + "image_footnote": [], + "bbox": [ + 501, + 300, + 893, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sensor types and weather conditions, diverse class distributions arising from varying capture scenarios, and distinct label spaces shaped by specific annotation protocols. These factors collectively pose a formidable challenge in harmonizing disparate LiDAR point clouds and jointly optimizing model parameters to effectively address multiple tasks across a range of sensor modalities [89, 118]. Empirical evidence in Fig. 3 further reveals that naively combining heterogeneous data to train a LiDAR segmentation model – without strategic alignments – often leads to sub-opt results.", + "bbox": [ + 496, + 671, + 890, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent works [6, 39, 48, 79, 87, 94, 106] resort to unsupervised domain adaptation (UDA) for utilizing training data from both source and target domains to optimize one parameter set. Nevertheless, they either focus on only the sharing mapping between two domains (by ignoring dis", + "bbox": [ + 496, + 824, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* The first two authors contributed equally to this work.", + "bbox": [ + 99, + 886, + 395, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14648", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "joint classes) or directly merge source domain labels to align with the target domain [40, 113]. The overlook of the performance degradation on the source dataset and the destruction of original label mappings inevitably constrains such a learning paradigm. Furthermore, there have been efforts [85, 96, 104, 118] to employ multi-dataset learning strategies to bolster the generalization prowess of 3D perception models. However, they either necessitate dataset-specific fine-tuning, deviating from a truly universal learning approach, or converge label spaces to a coarser set, resulting in the dilution of fine-grained segmentation capabilities across diverse semantic categories.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we define a novel paradigm towards leveraging LiDAR point clouds from different datasets to tame a single set of parameters for multi-task LiDAR segmentation. Sibling to image segmentation communities [43, 53, 124], we call this paradigm universal LiDAR segmentation. The ultimate goal of such a synergistic way of learning is to build a powerful segmentation model that can absorb rich cross-domain knowledge and, in return, achieve strong resilience and generalizability for practical usage. Given the substantial differences among datasets in terms of data characteristics, feature distributions, and labeling conventions, we introduce a comprehensive multi-space alignment approach that encompasses data-, feature-, and label-level alignments, to effectively pave the path for efficient and universally applicable LiDAR segmentation. In particular, the multi-modal data, including images and texts, is fully exploited to assist the alignment process with the guidance of more general knowledge. Through aforementioned processes, we propose M3Net to learn common knowledge across datasets, modalities, and tasks, thereby significantly enhancing its applicability in practical scenarios.", + "bbox": [ + 75, + 272, + 472, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To substantiate the efficacy of M3Net and the utility of each module developed, we have carried out a series of thorough comparative and ablation studies across an extensive array of driving datasets, as shown in Fig. 1. Notably, our best model achieves state-of-the-art LiDAR segmentation performance with $75.1\\%$ , $83.1\\%$ , $72.4\\%$ mIoU scores on SemanticKITTI [3], nuScenes [27], Waymo Open [90], respectively, using a shared set of parameters. Moreover, our approach also performs well for direct knowledge transfer and out-of-distribution adaptations, further underscoring its robust capability for effective knowledge transfer.", + "bbox": [ + 75, + 589, + 472, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 768, + 218, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LiDAR Segmentation. A holistic perception of 3D scenes is crucial for safe autonomous driving [4, 7, 33, 51, 61]. Various LiDAR segmentation models have been proposed, with distinct focuses on aspects include LiDAR representations [20, 74, 91-93, 103, 121, 128], model architectures [1, 17, 24, 35, 46, 52, 80, 112], sensor fusion [18, 62, 64, 113, 129], post-processing [111, 123], data aug", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mentations [75, 84, 105], etc. Most recently, researchers started to explore data efficiency [49, 56], annotation efficiency [57, 63, 65, 86, 97], annotation-free learning [10, 11, 122], zero-shot learning [12, 69], domain adaptation [6, 39, 48, 54, 73, 79, 106], and robustness [47] in LiDAR segmentation, shedding lights for practitioners. Existing pursuits, however, learn separate parameter sets for each dataset, impeding the scalability. This motivates us to explore LiDAR segmentation in a multi-task, multi-dataset, multi-modality manner with just a single set of parameters.", + "bbox": [ + 496, + 90, + 890, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Task Learning. A proper pipeline design could enable the model to generate suitable predictions to fulfill multiple tasks simultaneously [16, 31]. The current research endeavors mainly focus on building image or video segmentation models to handle semantic, instance, and panoptic segmentation tasks [38, 59, 98, 100, 119, 120, 130]. Recently, several attempts have been made to enable multitask segmentation on LiDAR point clouds. MaskRange [30] and MaskPLS [71] extend the mask classification paradigm [15] for joint semantic and panoptic LiDAR segmentation. LidarMultiNet [117] uses global context pooling and task-specific heads to handle LiDAR-based detection and segmentation. P3Former [108] proposed a specialized positional embedding to handle the geometry ambiguity in panoptic LiDAR segmentation. Our framework also supports multi-task learning. Different from existing approaches, the proposed M3Net stands out by combining knowledge from different sensor data across multiple data sources, which achieves superior performance on each task.", + "bbox": [ + 496, + 246, + 892, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Dataset Learning. Leveraging data samples from different sources for training has been proven effective in enhancing robustness and generalizability [72]. Various approaches have been proposed to merge image datasets for object detection [14, 58, 60, 99, 125, 126], image segmentation [29, 42, 43, 53, 124], depth estimation [13, 82], etc. Due to large domain gaps, the image-based methods are often hard to be transferred to 3D. To combine multiple Li-DAR datasets for 3D object detection, MDT3D [89] defines a coarse label set to handle the label space conflicts in different point cloud datasets. MS3D++ [95, 96] ensembles pre-trained detectors from different source datasets for multi-domain adaptation. Uni3D [118] resorts to dataset-specific detection heads and feature re-coupling for training a unified 3D object detector. Recently, PPT [104] proposed to pre-train a point cloud segmentation network using data from multiple datasets. However, the pre-trained weights are then fine-tuned on each specific dataset, which breaks the universal learning manner. The closest work to us is COLA [85], which trains a single model across multiple sources by converting dataset-specific labels to a common coarse set. Such a conversion, however, leads to the loss of fine-grained segmentation across the various semantic categories. Differently, our M3Net is tailored to tame a single", + "bbox": [ + 496, + 537, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14649", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/766b15b8b801135d7eabe8e60a7bdaa7e165b07eadaae8904a1d98a4a475b11e.jpg", + "image_caption": [ + "Figure 2. Statistical analysis of six sharing semantic classes in the nuScenes [●], SemanticKITTI [●], and Waymo Open [●] datasets. Each violin plot shows the class distribution across LiDAR scenes spanning 50 meters, centered around the ego-vehicle. Best viewed in colors." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 200, + 270 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1e97aa12f4e903a23eebbbb9b226519a3af01def186fa0199b251ba872c246ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 85, + 339, + 270 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e7e1ea742b504d35b9462e6605d179dc2f0f66e4d05dcf6d627975e1f3bce0f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 85, + 475, + 270 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0290866900cf110d7921e2ad02cc7ebd6f1081777035dd7d72036204f086a3b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 85, + 616, + 270 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9052bb2739c8011afdc180a8fec8b872ccf983ab9bfdd0b01f6d58ee31a9d630.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 85, + 754, + 270 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cfba0f78cfc727b73924884070d78c5dcc2d5d0cc32428cf22407bc12d91bddc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 85, + 893, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "parameter set to fulfill multi-task prediction across multiple datasets while still maintaining the original label mappings. Multi-Modality Learning. Recent trend favors synergistic learning from data of different modalities, such as vision, language, and speech [2, 8, 19, 26, 76, 81, 101]. For LiDAR segmentation, several works [9, 39, 40, 67, 115] explored the distillation of image features to point clouds. Recently, OpenScene [78] and CLIP2Scene [11] proposed to leverage point clouds along with multi-view images and language for open-vocabulary learning. PPKT [66], SLidR [86], and Seal [65] form cross-sensor contrastive learning objectives to pre-train the LiDAR segmentation models. The advantages of sensor fusion have been consistently proven. In this work, to pursue universal LiDAR segmentation, we propose to align multi-space point clouds via images and texts.", + "bbox": [ + 75, + 327, + 472, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 569, + 187, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our study serves as an early attempt at combining multi-task, multi-dataset, multi-modality knowledge into a single set of parameters to fulfill universal LiDAR segmentation. We start with a pilot study to unveil the difficulties in merging heterogeneous LiDAR point clouds (cf. Sec. 3.1). We then present M3Net, a versatile LiDAR segmentation network tailored to pursue i) statistical consistency in the data space (cf. Sec. 3.2), ii) cross-modality-assisted alignment in the feature space (cf. Sec. 3.3), and iii) language-guided unification in the label space (cf. Sec. 3.4).", + "bbox": [ + 75, + 594, + 468, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pilot Study", + "text_level": 1, + "bbox": [ + 76, + 755, + 200, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The current de facto of training a LiDAR segmentation network adopts a task-by-task and dataset-by-dataset pipeline. Despite the superior performance achieved under such standalone settings, the trained parameter sets cannot be shared to satisfy out-of-domain requirements and, therefore, limits their use cases for practical applications.", + "bbox": [ + 75, + 779, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Naïve Joint Training. A natural alternative to breaking the above constraint is to jointly train a network across multiple", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "datasets for better generalizability. However, as depicted in Fig. 2, it is often non-trivial to naively combine heterogeneous data with large data distribution gaps to train a universal LiDAR segmentation model without proper alignments. To testify this, we conducted a pilot study using the prior art MinkUNet [20] for both standalone and joint training on three large-scale datasets [3, 27, 90]. As shown in Fig. 3 (a) and (d), a brutal combination undermines the segmentation performance. Due to large discrepancies in aspects like sensor configurations, data acquisitions, label mappings, and domain shifts, the jointly trained representations tend to be disruptive instead of being more general.", + "bbox": [ + 496, + 327, + 893, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LiDAR Sensor Discrepancy. To understand the root cause of performance degradation, we conducted another study that controls point cloud density discrepancies when merging datasets. As shown in Fig. 3 (b) and (c), joint training on data collected by sensors with different beam numbers tends to suffer more severely than merging less density variant data. We hypothesize that this is mainly caused by the data statistical variations. In light of these observations, we propose a bag of suitable operations in the following sections to alleviate the large domain gaps among different LiDAR segmentation datasets [3, 4, 7, 27, 90].", + "bbox": [ + 496, + 510, + 893, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Data-Space Alignment", + "text_level": 1, + "bbox": [ + 500, + 691, + 710, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a total of $S$ datasets $D^{s} = \\{(x^{s},y^{s})|1\\leq s\\leq S\\}$ where $(x^s,y^s)$ denotes the data-label pairs constituting a dataset. For the LiDAR segmentation task, $x^{s}$ often encompasses the LiDAR point cloud $P^{s} = \\{p_{x},p_{y},p_{z}\\}^{s}\\in$ $\\mathbb{R}^{N\\times 3}$ and synchronized multi-view camera images $V^{s} =$ $\\{\\mathcal{I}_1,\\dots,\\mathcal{I}_l\\} |l = 1,\\dots,L\\}$ , where $\\mathcal{I}_t\\in \\mathbb{R}^{H\\times W\\times 3}$ $N$ is the number of points, $L$ denotes the number of camera sensors, $H$ and $W$ are the height and width of the image, respectively. $y^{s}\\in \\mathbb{R}^{N}$ denotes point cloud labels in the label space $\\mathbb{Y}^s$ , we unify the label space as $\\mathbb{Y}^u = \\mathbb{Y}^1\\cup \\mathbb{Y}^2\\dots \\cup \\mathbb{Y}^S$", + "bbox": [ + 496, + 715, + 893, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-Modality Data Alignment. As a multi-sensing system, the information encoded in $P_{i}^{s}$ and $V_{i}^{s}$ are intuitively", + "bbox": [ + 496, + 869, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14650", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/034652c7fd889434320ee744bf594cb70749a3e09166a44978313464ba3b0cc1.jpg", + "image_caption": [ + "Standalone Training" + ], + "image_footnote": [], + "bbox": [ + 76, + 107, + 243, + 237 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/63746aa1134b4c098642f2b87f6880f4c4065873d812c8c78112bf5e0a0e0f12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 248, + 191, + 267, + 226 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/454289ff174e478ba3a0b50cde2c33065935cf04c313dd8e8997e04683da73b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 191, + 294, + 226 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/be2644d1baa9563208d57cde624574e1a758ff44124c902dfe803732d4eb44b7.jpg", + "image_caption": [ + "Joint Training (Same # Beam)" + ], + "image_footnote": [], + "bbox": [ + 303, + 107, + 467, + 237 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/df956c51dd541326673b31a545bd870fb2314fa94dfa8589d60ccbc66164f7a6.jpg", + "image_caption": [ + "Figure 3. A pilot study of naively merging different datasets for training the MinkUNet [20] model. Compared to the standalone training in (a), either jointly training with (b) the same, (c) different, or (d) all sensor-acquired data will cause severe degradation." + ], + "image_footnote": [], + "bbox": [ + 76, + 244, + 243, + 335 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bf2472168858d61efd730da86ff7b5690e9670a0a78d2e07d7c3049012a6279f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 246, + 467, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "complementary to each other [3, 7, 90]. To leverage such an advantage, we resort to the correspondences embedded in camera calibration matrices to bridge the LiDAR points and camera image pixels. Specifically, for each point $\\mathbf{p} = (p^x,p^y,p^z)$ in $P^s$ , the corresponding pixel $(u,v)$ can be found by the following transformations:", + "bbox": [ + 75, + 417, + 468, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ u, v, 1 ] ^ {\\mathrm {T}} = \\frac {1}{p ^ {z}} \\cdot T _ {s} \\cdot T \\cdot [ p ^ {x}, p ^ {y}, 1 ] ^ {\\mathrm {T}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 513, + 468, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T \\in \\mathbb{R}^{4 \\times 4}$ is the camera extrinsic matrix that consists of a rotation matrix and a translation matrix, and $T_{s} \\in \\mathbb{R}^{3 \\times 4}$ is the camera intrinsic matrix. As we will show in the following sections, such a cross-sensor data alignment serves as the foundation for alignments in other spaces.", + "bbox": [ + 75, + 551, + 468, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-Sensor Statistical Alignment. To mitigate the discrepancies in sensor installations across different datasets, we incorporate a point coordinate alignment operation. Specifically, drawing upon insights from prior domain adaptation approaches [102, 116], we adjust the coordinate origins of point clouds from different datasets by introducing an offset $\\sigma \\in \\mathbb{R}^{1\\times 3}$ to the ground plane. We find empirically that such an alignment can largely reduce the degradation caused by the variations in different sensor setups.", + "bbox": [ + 75, + 628, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset-Specific Rasterization. It is conventional to rasterize LiDAR point clouds $P^s$ using unified rasterization parameters, e.g., voxel size [91, 128] or horizontal range view resolution [74, 111]. However, the point clouds acquired in different LiDAR datasets naturally differ in density, range, intensity, etc., which tends to favor different rasterization parameters. To meet such a requirement, we select dataset-specific parameters for rasterization on each dataset through empirical experiments and analyses.", + "bbox": [ + 75, + 765, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decoupled BN. Another challenge in training across multiple datasets is the presence of domain gaps, which can result in significant statistical shifts of feature learning among datasets. Such shifts can hinder the convergence and affect the model's ability to generalize well across diverse datasets. We adopt a decoupled batch norm (BN) for point cloud features in each dataset. Instead of using the traditional BN, which calculates mean and variance across all samples in a mini-batch, the decoupled BN tends to adapt each dataset's specific characteristics independently.", + "bbox": [ + 496, + 90, + 890, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Feature-Space Alignment", + "text_level": 1, + "bbox": [ + 500, + 253, + 732, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to acquire a generalized feature representation for downstream tasks. Compared to point clouds, images contribute stronger visual, textural, and semantic information. Thus, the collaboration between pixels and points could enrich the overall representation. Previous research [11, 60, 64] has consistently demonstrated that such a combination results often leads to improved performance.", + "bbox": [ + 496, + 276, + 890, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-Modality Assisted Alignment. In the context of multi-dataset joint training, our objective is to establish a unified feature space by leveraging image features to assist point cloud features. Acknowledging that images used in training lack ground truth labels [3, 27], we utilize image features from a pre-trained model as an alternative, facilitating a more universally applicable representation. We feed camera images $V^{s}$ into a pre-trained DeepLab [98] and a vision-language model (VLM) and visualize the output image features by t-SNE [25]. As shown in Fig. 4, we observe that image features from DeepLab appear disorderly and lack semantics. In contrast, features from VLM share a more unified feature space. Motivated by this, we propose a cross-modality assisted alignment that uses VLM to help align the feature space. Specifically, the camera images $V^{s}$ are fed to the frozen image encoder from VLM to obtain image features $F_{v} = \\{\\mathcal{F}_{v}^{1},\\mathcal{F}_{v}^{2},\\dots,\\mathcal{F}_{v}^{s}\\}$ , where $\\mathcal{F}_v^s\\in \\mathbb{R}^{c\\times h\\times w}$ . The LiDAR point clouds $P^{s}$ , on the other hand, are fed to the point encoder followed by a projection layer to generate the point features $F_{p} = \\{\\mathcal{F}_{p}^{1},\\mathcal{F}_{p}^{2},\\dots,\\mathcal{F}_{p}^{s}\\}$ , where $\\mathcal{F}_p^s\\in \\mathbb{R}^{m\\times c}$ ; $m$ denotes the number of non-empty grids. We then leverage the paired image features $\\hat{F}_v\\in \\mathbb{R}^{m_p\\times c}$ and point feature $\\hat{F}_p\\in \\mathbb{R}^{m_p\\times c}$ for alignment, where $m_{p}$ is the number of point-pixel pairs. After obtaining $\\hat{F}_p$ and $\\hat{F}_v$ , the cross-modality alignment is expressed as follows:", + "bbox": [ + 496, + 383, + 892, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c m a}} \\left(\\hat {F} _ {v}, \\hat {F} _ {p}\\right) = 1 - \\frac {\\hat {F} _ {v} \\cdot \\hat {F} _ {p}}{\\| \\hat {F} _ {v} \\| \\cdot \\| \\hat {F} _ {p} \\|}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 775, + 890, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Domain-Aware Cross-Modality Alignment. With cross-modality alignment, we transfer the knowledge of VLM to the point encoder, enabling the point features to gain a more comprehensive representation. However, during the execution of the above alignment, we have narrowed", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14651", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/310b97e835a536564a0cd79f100b852668275897818711eb5272fc58eddeae88.jpg", + "image_caption": [ + "Figure 4. The t-SNE plots of learned features before and after the feature-space alignment in merging the nuScenes [●], SemanticKITTI [●], and Waymo Open [●] datasets. We show image features from (a) standalone networks; (b) SAM [44], and point cloud features (c) before and (d) after the feature-space alignment." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 472, + 342 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b42b377a5d75e26c03b1a296f884ac1428fe40fec203eb95355c8c5ba01c8f72.jpg", + "image_caption": [ + "Figure 5. Feature-space alignment in M3Net. We leverage both image features $F_{v}$ and LiDAR point cloud features $F_{p}$ extracted from image encoder $E_{img}$ and point encoder $E_{pcd}$ to employ the regularization via V2P loss and achieve feature-space alignment." + ], + "image_footnote": [], + "bbox": [ + 76, + 441, + 472, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "it exclusively to image and point features from the same dataset. In this mode, point features solely learn from matching image features, restricting their knowledge acquisition. Ideally, we aim to ensure that image features encompass not only scenes identical to those represented in point clouds but also scenes from other datasets. To address this, we propose a domain-aware cross-modality guided alignment, as illustrated in Fig. 5. Specifically, we first extract, for each dataset, $F_{v}$ and $F_{p}$ from the same image encoder $E_{img}$ and point encoder $E_{pcd}$ during the cross", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "modality assisted alignment. The sets of features from all datasets are concatenated along the channel dimension to form $\\widetilde{F}_v\\in \\mathbb{R}^{c_v\\times h\\times w}$ . Subsequently, we sequentially feed $\\widetilde{F}_v$ through a branch that consists of a global average pooling and an MLP. Simultaneously, $\\widetilde{F}_v$ is fed to an auxiliary branch that undergoes the same processing flow and generates an output after the softmax function $\\mathcal{G}(\\cdot)$ . The outputs from both branches are multiplied to obtain $F_{m}\\in \\mathbb{R}^{c_{v}\\times 1\\times 1}$ . The overall process can be described as follows:", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {m} = M L P \\left(P o o l \\left(\\widetilde {F} _ {v}\\right)\\right) \\cdot \\mathcal {G} \\left(M L P \\left(P o o l \\left(\\widetilde {F} _ {v}\\right)\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 236, + 890, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we forward $F_{m}$ to a sigmoid activation function $\\mathcal{H}(\\cdot)$ and multiply it with input image features $\\widetilde{F}_v$ . The resulting output is added to $\\widetilde{F}_v$ and passed through the MLP layers to obtain the final image features $F_{vf} \\in \\mathbb{R}^{c \\times h \\times w}$ . The forward process of this operation is depicted as follows:", + "bbox": [ + 496, + 265, + 890, + 342 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {v f} = M L P \\left(\\left(\\mathcal {H} \\left(F _ {m}\\right) \\cdot \\widetilde {F} _ {v}\\right) + \\widetilde {F} _ {v}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 351, + 890, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we leverage the cross-modality data alignment to acquire paired image features $\\hat{F}_{vf} \\in \\mathbb{R}^{m_p \\times c}$ and paired point feature $\\hat{F}_p$ . The overall objective function is:", + "bbox": [ + 496, + 380, + 890, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v 2 p} \\left(\\hat {F} _ {v f}, \\hat {F} _ {p}\\right) = 1 - \\frac {\\hat {F} _ {v f} \\cdot \\hat {F} _ {p}}{\\| \\hat {F} _ {v f} \\| \\cdot \\| \\hat {F} _ {p} \\|}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 438, + 890, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Label-Space Alignment", + "text_level": 1, + "bbox": [ + 500, + 484, + 718, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Label Conflict. In multi-dataset joint training settings, label conflicts emerge as a significant challenge. This often refers to the inconsistencies in class labels across different datasets involved in the training process. The discrepancy can arise due to variations in annotation conventions, labeling errors, or even differences in the underlying semantics of classes between datasets. In our baseline, we unionize the different label spaces across datasets into $\\mathbb{Y}^u$ , where all datasets share a single LiDAR segmentation head. However, this may introduce several potential drawbacks:", + "bbox": [ + 496, + 508, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Loss of granularity: Unified label spaces could lose semantic granularity, particularly when dealing with subtle category differences in between different datasets.", + "- Information loss: During label space consolidation, details unique to each dataset may be obscured or lost, especially for those related to domain-specific categories.", + "- Increased complexity: Handling a unified label space may necessitate more complex model architectures or training strategies, thereby increasing overall complexity." + ], + "bbox": [ + 500, + 659, + 890, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address these issues, we introduce a language-guided label-space alignment to facilitate a more holistic semantic correlation across datasets. Given the natural correspondence between images and texts and the strong correlation between images and point clouds, we aim to strategically utilize the image modality as a bridge to establish language-guided alignments. Such a process consists of a text-driven", + "bbox": [ + 496, + 796, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14652", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/284d94ef22a16ef0fb1f4d2d8981c135be38ff15bb7ed5154900f94613a35560.jpg", + "image_caption": [ + "Figure 6. Label-space alignment in M3Net. We leverage image features $F_{v}$ , point cloud features $F_{p}$ , and text embedding $F_{t}$ extracted from $E_{img}$ , $E_{pcd}$ , and $E_{txt}$ , respectively, for regularization via the I2P, P2T, and V2T losses in the label-space alignment." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 472, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "point alignment, a text-driven image alignment, and a cross-modality-assisted label alignment.", + "bbox": [ + 75, + 380, + 468, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-Driven Alignments. As depicted in Fig. 6, images $V^s$ are fed into the frozen image encoder $E_{img}$ to extract the image features $F_v$ . Concurrently, the LiDAR point clouds $P^s$ are processed by the point encoder $E_{pcd}$ to generate the point features $F_p$ . Additionally, given the text input $T^s$ , text embedding features $F_t \\in \\mathbb{R}^{Q \\times c}$ are obtained from a frozen text encoder $E_{txt}$ , where $Q$ represents the number of categories across datasets. The text is composed of class names from unified label space $\\mathbb{Y}_u$ placed into pre-defined templates, and the text embedding captures semantic information of the corresponding classes. Subsequently, pixel-text pairs $\\{v_k, t_k\\}_{k=1}^M$ and point-text pairs $\\{p_k, t_k\\}_{k=1}^M$ are generated, where $M$ represents the number of pairs. Leveraging the semantic information contained in the text, we selectively choose positive and negative samples for both images and points for contrastive learning. It is noteworthy that negative samples are confined to the specific dataset category space. The overall objective of the text-driven point alignment function is shown as follows:", + "bbox": [ + 75, + 410, + 468, + 698 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p 2 t} = - \\sum_ {q = 1} ^ {Q} \\log \\left(\\frac {\\sum_ {t _ {k} \\in q , p _ {k}} \\exp \\left(< t _ {k} , p _ {k} > / \\tau\\right)}{\\sum_ {t _ {k} \\in q , t _ {k} \\notin q , p _ {j}} \\exp \\left(< t _ {k} , p _ {k} > / \\tau\\right)}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 705, + 468, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $t_k \\in q$ indicates that $t_k$ is generated by the $q$ -th classes name, and $Q$ is the number of classes. Symbol $<, >$ denotes the scalar product operation and $\\tau$ is a temperature term $(\\tau > 0)$ . Similarly, the objective of the text-driven image alignment function is illustrated as follows:", + "bbox": [ + 75, + 761, + 468, + 838 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v 2 t} = - \\sum_ {q = 1} ^ {Q} \\log \\left(\\frac {\\sum_ {t _ {k} \\in q , v _ {k}} \\exp \\left(< t _ {k} , v _ {k} > / \\tau\\right)}{\\sum_ {t _ {k} \\in q , t _ {k} \\notin q , v _ {j}} \\exp \\left(< t _ {k} , v _ {k} > / \\tau\\right)}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 844, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-Modality-Assisted Label Alignment. After text-driven alignments, the subsequent crucial step entails aligning the point and image modalities within the label space. We first obtain image logits $F_{vl} = \\{\\mathcal{F}_{vl}^{1},\\mathcal{F}_{vl}^{2},\\dots,\\mathcal{F}_{vl}^{s}\\}$ and point logits $F_{pl} = \\{\\mathcal{F}_{pl}^{1},\\mathcal{F}_{pl}^{2},\\dots,\\mathcal{F}_{pl}^{s}\\}$ from text-driven alignments, where $\\mathcal{F}_{vl}^{s}\\in \\mathbb{R}^{Q\\times H\\times W}$ , $\\mathcal{F}_{pl}^{s}\\in \\mathbb{R}^{N\\times Q}$ . Subsequently, we conduct cross-modality alignment to obtain paired image logits $\\hat{F}_{vl}\\in \\mathbb{R}^{m_p\\times Q}$ and paired point logits $\\hat{F}_{pl}\\in \\mathbb{R}^{m_p\\times Q}$ . Formally, the cross-modality-assisted alignment in the label space is formulated as follows:", + "bbox": [ + 496, + 90, + 890, + 244 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i 2 p} \\left(\\hat {F} _ {v l}, \\hat {F} _ {p l}\\right) = 1 - \\frac {\\hat {F} _ {v l} \\cdot \\hat {F} _ {p l}}{\\| \\hat {F} _ {v l} \\| \\cdot \\| \\hat {F} _ {p l} \\|}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 253, + 890, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, the complete objective function for the language-guided label-space alignment is expressed as follows:", + "bbox": [ + 498, + 301, + 890, + 330 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l a b e l}} = \\mathcal {L} _ {p 2 t} + \\mathcal {L} _ {i 2 p} + \\mathcal {L} _ {v 2 t}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 342, + 890, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Universal LiDAR Segmentation", + "text_level": 1, + "bbox": [ + 500, + 369, + 777, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We enhance the versatility of M3Net via multi-tasking learning. This integration involves an instance extractor to enable joint semantic and panoptic LiDAR segmentation.", + "bbox": [ + 496, + 392, + 890, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Panoptic LiDAR Segmentation. Motivated by DSNet [32, 33], our instance extractor comprises an instance head and a clustering step. The instance head encompasses several MLPs designed to predict the offsets between instance centers. The clustering step uses semantic predictions to filter out stuff points, thereby retaining only those associated with thing points. The remaining points undergo a mean-shift clustering [21], utilizing features from the instance head to discern distinct instances. Lastly, we employ the L1 loss $\\mathcal{L}_{l1}$ to optimize the thing point regression process.", + "bbox": [ + 496, + 438, + 890, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall Objectives. Putting everything together, the overall objective of M3Net is to minimize the following losses:", + "bbox": [ + 498, + 589, + 890, + 619 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {v 2 p} + \\mathcal {L} _ {\\text {l a b e l}} + \\mathcal {L} _ {c e} + \\mathcal {L} _ {\\text {l o v a s z}} + \\mathcal {L} _ {l 1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 630, + 890, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{ce}$ and $\\mathcal{L}_{lovasz}$ denote the cross-entropy loss and the Lovasz-softmax [5] loss, respectively.", + "bbox": [ + 498, + 657, + 890, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 700, + 633, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 500, + 726, + 696, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Our M3Net framework and baselines are trained on a combination of nuScenes [27], SemanticKITTI [3], and Waymo Open [90]. Meanwhile, we resort to another five LiDAR-based perception datasets [41, 45, 77, 106, 107] and two 3D robustness evaluation datasets [47] to verify the strong generalizability of M3Net. Due to space limits, additional details regarding the datasets are in the Appendix.", + "bbox": [ + 496, + 750, + 890, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. M3Net is implemented based on Pointcept [23] and MMDetection3D [22]. We use two backbones in our experiments, i.e., MinkUNet [20] and", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14653", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a0cd17d042a20307c403a6134be9e53e6c24f0b495a910bf2eda3bdec4ec3827.jpg", + "table_caption": [ + "Table 1. Ablation study on the M3Net alignments happen in the Data, Feature, and Label spaces, respectively, when combining the SemanticKITTI [3], nuScenes [27], and Waymo Open [90] datasets. The mAcc and mIoU scores are in percentage. Best scores are in bold." + ], + "table_footnote": [], + "table_body": "
-ConfigurationsMinkUNet [20]PTv2+ [103]
SemKITTInuScenesWaymoSemKITTInuScenesWaymo
mAccmIoUmAccmIoUmAccmIoUmAccmIoUmAccmIoUmAccmIoU
BaselineNaïve Joint Training62.4354.0365.0559.8473.7665.3967.9661.5976.5369.6575.6867.00
M3Net (Ours)DataFeatureLabel------
✓73.8269.0183.6676.8977.8869.3778.5569.9586.2279.1380.9672.15
✓✓74.3669.6485.1778.8878.3169.7079.4370.8787.1080.2680.7472.33
✓✓73.8569.3485.2078.9078.0469.5580.3071.1387.4480.4580.6972.30
✓✓✓74.4069.8585.3079.0078.6670.1580.0072.0087.9180.9081.1172.40
", + "bbox": [ + 80, + 125, + 893, + 253 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4e5bfd0dee79c4ea430f5e670000ad1f9512f532f4c7ceee4e4c651825a23020.jpg", + "table_caption": [ + "Table 2. Panoptic LiDAR segmentation results on the val sets of the Panoptic-SemanticKITTI [3] and Panoptic-nuScenes [27] datasets. All scores are given in percentage. The best and second-best scores are highlighted in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
MethodConfigurationsPanoptic-SemanticKITTIPanoptic-nuScenes
PQ\\( PQ^† \\)RQSQmIoUPQ\\( PQ^† \\)RQSQmIoU
Panoptic-TrackNet [37]40.0-48.373.053.851.456.263.380.258.0
Panoptic-PolarNet [127]59.164.170.278.364.563.467.275.383.966.9
EfficientLPS [88]Single-Dataset Training59.265.169.875.064.959.262.882.970.769.4
DSNet [32]61.465.272.779.069.664.767.676.183.576.3
Panoptic-PHNet [55]61.7---65.774.777.784.288.279.7
BaselineNaïve Joint Training56.0359.6465.7873.7261.5956.6760.6166.7583.4969.65
M3Net (Ours)DataFeatureLabel--
✓62.3465.1772.6074.6769.9568.4971.1179.1385.4979.13
✓✓62.9165.7373.3275.4770.8771.4773.8681.5386.7180.26
✓✓63.2367.8973.6181.6671.1371.5373.9181.8086.9280.45
✓✓✓63.8768.6673.1082.3572.0071.7074.0182.2086.4780.90
", + "bbox": [ + 81, + 306, + 890, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PTv2+ [103]. We trained M3Net on four A100 GPUs for 50 epochs with a batch size of 6 for each GPU. The initial learning rate is set to 0.002. We adopt the AdamW optimizer [68] with a weight decay of 0.005 and cosine decay learning rate scheduler. For the dataset-specific rasterization, we set voxel sizes to $0.05\\mathrm{m}$ , $0.1\\mathrm{m}$ , and $0.05\\mathrm{m}$ for SemanticKITTI [3], nuScenes [27], and Waymo Open [90], respectively. For the data augmentation, we employ random flipping, jittering, scaling, rotation, and Mix3D [75]. Due to space limits, kindly refer to Appendix for additional details. Evaluation Metrics. We adopt conventional reportings of mAcc and mIoU for LiDAR semantic segmentation, $PQ$ , $PQ^{\\dagger}$ , $SQ$ , and $RQ$ for panoptic segmentation, and mCE and mRR for 3D robustness evaluation. Due to the space limit, kindly refer to our Appendix for more detailed definitions.", + "bbox": [ + 75, + 518, + 468, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 756, + 230, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-Space Alignments. The effectiveness of three proposed alignments over the joint training baselines is shown in Tab. 1. We observe that the data-space alignment plays the most crucial role in improving the universal LiDAR segmentation performance. Without proper data alignments, joint training with either MinkUNet [20] or the stronger PTv2+ [103] will suffer severe degradation, especially on sparser point clouds [27]. On top of the data-space align-", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ment, the combinations of multi-view images at the feature space and the language-guided knowledge at the label space further enhance the learned feature representations. The results show that they work synergistically in merging knowledge from heterogeneous domains during joint training.", + "bbox": [ + 496, + 518, + 890, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Panoptic LiDAR Segmentation. In Tab. 2, we present another ablation study focusing on panoptic LiDAR segmentation. All three alignments incorporated in M3Net demonstrate significant improvements over the baselines. This highlights the pronounced efficacy of our multi-space alignments. Moreover, our approach outperforms the single-dataset state-of-the-art method Panoptic-PHNet [55] by a notable $2.17\\%$ PQ on Panoptic-SemanticKITTI [3] and achieves compelling results on Panoptic-nuScenes [27].", + "bbox": [ + 496, + 595, + 892, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Feature Alignments. We conduct a qualitative analysis of the learned visual feature distributions in the form of t-SNE [25]. Fig. 4 (a) and (b) represent the distributions of learned visual features among three datasets from DeepLab and VLM backbones, respectively. The features obtained by the latter exhibit more distinct semantics in feature space. The concentrated distribution space is advantageous for achieving feature alignments across multiple datasets. Additionally, Fig. 4 (c) and (d) illustrate the distribution of point cloud features before and after feature-space alignment. As can be seen, the feature distribution distances", + "bbox": [ + 496, + 734, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14654", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e59fa22168883494ffb8b5ac2fbdb0cabbdb456f81353bc4137d77c06dfe2695.jpg", + "table_caption": [ + "Table 3. Knowledge transfer and generalization analyses across five LiDAR segmentation datasets and two 3D robustness evaluation datasets. All scores are given in percentage. The best and second-best scores are highlighted in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
MethodRELLIS-3DSemanticPOSSSemanticSTFSynLiDARDAPS-3DSemKITTI-CnuScenes-C
1%10%HalfFullHalfFull1%10%HalfFullmCEmRRmCEmRR
PPKT [66]49.7154.3350.1856.0050.9254.6937.5746.4878.9084.00--105.6476.06
SLidR [86]49.7554.5751.5655.3652.0154.3542.0547.8481.0085.40--106.0875.99
Seal [65]51.0955.0353.2656.8953.4655.3643.5849.2681.8885.90--92.6383.08
Naïve Joint37.7750.2342.1952.3146.7048.0018.5642.3773.9177.89113.6584.73128.9781.45
Single-Dataset40.1754.2547.6955.0050.3351.1923.1745.0875.1080.8795.1184.9599.6379.06
M3Net (Ours)51.2755.0553.6057.1753.7855.4244.1049.9382.0886.0086.4385.7791.0379.15
", + "bbox": [ + 80, + 125, + 893, + 242 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6d1d973134a2cb1f10b6c754a90a7525d4d0506d10b278210317a4bed7731d76.jpg", + "table_caption": [ + "Table 4. LiDAR semantic segmentation results on the val and test sets of SemanticKITTI [3] and nuScenes [27], and the val set of Waymo Open [90]. All scores are in percentage. The best and second-best scores are highlighted in bold and underline." + ], + "table_footnote": [], + "table_body": "
MethodSemKITTInuScenesWaymo
ValTestValTestmIoUmAcc
RangeNet++ [74]-52.2-65.5--
PolarNet [121]57.254.371.069.8--
SalsaNext [24]-59.5-72.2--
RangeViT [1]60.764.075.2---
MinkUNet [20]63.863.773.3-65.976.6
SPVNAS [91]64.766.4-77.467.4-
AMVNet [62]65.265.377.277.3--
RPVNet [112]65.570.377.6---
(AF)2-S3Net [18]-69.7-78.3--
Cylinder3D [128]65.967.876.177.966.0-
PVKD [34]66.471.276.0---
WaffleIron [80]66.870.879.1---
RangeFormer [46]67.673.378.180.1--
SphereFormer [52]67.874.878.481.969.9-
FRNet [114]68.773.379.082.5--
PTv2+ [103]70.370.680.282.670.680.2
LidarMultiNet [117]---81.473.8-
M3Net (Ours)72.075.180.983.172.481.1
", + "bbox": [ + 80, + 325, + 468, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "between the three datasets have been largely reduced, providing evidence of the alignment effectiveness.", + "bbox": [ + 76, + 589, + 468, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Comparative Study", + "text_level": 1, + "bbox": [ + 76, + 632, + 264, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparisons to State of the Arts. In Tab. 4, we compare M3Net with current best-performing models on the benchmarks of SemanticKITTI [3], nuScenes [27], and Waymo Open [90]. Remarkably, M3Net consistently outperforms existing approaches across all three datasets. Specifically, on SemanticKITTI [3], M3Net achieves a $72.0\\%$ mIoU on the validation set, surpassing the closest method by a notable margin of $1.7\\%$ mIoU. Similarly, on nuScenes [27], M3Net achieves $80.9\\%$ mIoU and $83.1\\%$ mIoU on the validation and test sets, demonstrating its robustness and generalization capabilities. Additionally, the performance of M3Net on Waymo Open [90] is competitive with prior arts. We achieve a mIoU of $72.4\\%$ and a mAcc of $81.1\\%$ . These results highlight again the superiority of M3Net in handling complex diverse LiDAR segmentation tasks.", + "bbox": [ + 75, + 657, + 468, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Direct Knowledge Transfer. To further validate the strong", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "knowledge transfer capability of M3Net, we conduct extensive experiments on five different LiDAR-based perception datasets [41, 45, 77, 106, 107]. These datasets have unique data collection protocols and data distributions. As shown in Fig. 1 and the first ten columns in Tab. 3, our framework constantly outperforms the prior arts, the naive joint training, and the single-dataset baselines across all five datasets. This concretely supports the strong knowledge transfer efficacy brought by multi-space alignments in M3Net.", + "bbox": [ + 496, + 263, + 890, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Out-of-Distribution Generalization. Evaluating the generalization ability of models on out-of-training-distribution data is crucial, particularly in safety-critical fields like autonomous driving [50, 109, 110]. In this context, we resort to the two corruption datasets from the Robo3D [47] benchmark, i.e., SemanticKITTI-C and nuScenes-C, to conduct our assessment. From the last four columns of Tab. 3, we observe that M3Net achieves better results than the naive joint training and other single-dataset approaches, proving the strong generalizability of the learned representations.", + "bbox": [ + 496, + 401, + 892, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 566, + 617, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we presented M3Net, a universal framework capable of fulfilling multi-task, multi-dataset, multimodality LiDAR segmentation using a single set of parameters. Through extensive analyses, we validated the effectiveness of applying data-, feature-, and label-space alignments to handle such a challenging task. In addition, our comprehensive analysis and discourse have delved into the fundamental challenges of acquiring the general knowledge for scalable 3D perception, which holds substantial potential to propel further research in this domain. Our future strides focus on combining more data resources to further enhance the alignments and adaptations in our framework.", + "bbox": [ + 496, + 590, + 890, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was partially supported by NSFC (No.62206173) and MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University). This work was also supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14655", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Angelika Ando, Spyros Gidaris, Andrei Bursuc, Gilles Puy, Alexandre Boulch, and Renaud Marlet. Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5240-5250, 2023.", + "[2] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312, 2022.", + "[3] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019.", + "[4] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Jürgen Gall, and Cyrill Stachniss. Towards 3d lidar-based semantic scene understanding of 3d point cloud sequences: The semantickitti dataset. International Journal of Robotics Research, 40:959-96, 2021.", + "[5] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: a tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4413-4421, 2018.", + "[6] Alexandre Boulch, Corentin Sautier, Björn Michele, Gilles Puy, and Renaud Marlet. Also: Automotive lidar self-supervision by occupancy estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13455-13465, 2023.", + "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020.", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021.", + "[9] Jun Cen, Shiwei Zhang, Yixuan Pei, Kun Li, Hang Zheng, Maochun Luo, Yingya Zhang, and Qifeng Chen. Cmdfusion: Bidirectional fusion network with cross-modality knowledge distillation for lidar semantic segmentation. arXiv preprint arXiv:2307.04091, 2023.", + "[10] Runnan Chen, Youquan Liu, Lingdong Kong, Nenglun Chen, Xinge Zhu, Yuexin Ma, Tongliang Liu, and Wenping Wang. Towards label-free scene understanding by vision foundation models. In Advances in Neural Information Processing Systems, 2023.", + "[11] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023." + ], + "bbox": [ + 86, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Runnan Chen, Xinge Zhu, Nenglun Chen, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Bridging language and geometric primitives for zero-shot point cloud segmentation. In ACM International Conference on Multimedia, pages 5380-5388, 2023.", + "[13] Tian Chen, Shijie An, Yuan Zhang, Chongyang Ma, Huayan Wang, Xiaoyan Guo, and Wen Zheng. Improving monocular depth estimation by leveraging structural awareness and complementary datasets. In European Conference on Computer Vision, pages 90-108, 2020.", + "[14] Yanbei Chen, Manchen Wang, Abhay Mittal, Zhenlin Xu, Paolo Favaro, Joseph Tighe, and Davide Modolo. Scaledet: A scalable multi-dataset object detector. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7288-7297, 2023.", + "[15] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. In Advances in Neural Information Processing Systems, pages 17864-17875, 2021.", + "[16] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022.", + "[17] Huixian Cheng, Xianfeng Han, and Guoqiang Xiao. Cenet: Toward concise and efficient lidar semantic segmentation for autonomous driving. In IEEE International Conference on Multimedia and Expo, pages 1-6, 2022.", + "[18] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. Af2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021.", + "[19] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022.", + "[20] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019." + ], + "bbox": [ + 509, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14656", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Dorin Comaniciu and Peter Meer. Mean shift: A robust approach toward feature space analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(5):603-619, 2005.", + "[22] MMDetection3D Contributors. MMDetection3D: Open-MMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.", + "[23] Pointcept Contributors. Pointcept: A codebase for point cloud perception research. https://github.com/Pointcept/Pointcept, 2023.", + "[24] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In International Symposium on Visual Computing, pages 207-222, 2020.", + "[25] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (11), 2008.", + "[26] Christoph Feichtenhofer, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In Advances in Neural Information Processing Systems, 2022.", + "[27] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters, 7:3795-3802, 2022.", + "[28] Biao Gao, Yancheng Pan, Chengkun Li, Sibo Geng, and Huijing Zhao. Are we hungry for 3d lidar data for semantic segmentation? a survey of datasets and methods. IEEE Transactions on Intelligent Transportation Systems, 2021.", + "[29] Xiuye Gu, Yin Cui, Jonathan Huang, Abdullah Rashwan, Xuan Yang, Xingyi Zhou, Golnaz Ghiasi, Weicheng Kuo, Huizhong Chen, Liang-Chieh Chen, and David A. Ross. Dataseg: Taming a universal multi-dataset multi-task segmentation model. arXiv preprint arXiv:2306.01736, 2023.", + "[30] Yi Gu, Yuming Huang, Chengzhong Xu, and Hui Kong. Maskrange: A mask-classification model for range-view based lidar segmentation. arXiv preprint arXiv:2206.12073, 2022.", + "[31] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In IEEE/CVF International Conference on Computer Vision, pages 2961-2969, 2017.", + "[32] Fangzhou Hong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Lidar-based panoptic segmentation via dynamic shifting network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13090-13099, 2021.", + "[33] Fangzhou Hong, Lingdong Kong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Unified 3d and 4d panoptic segmentation via dynamic shifting networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2024.", + "[34] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-voxel knowledge distillation for lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8479–8488, 2022.", + "[35] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham." + ], + "bbox": [ + 86, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Randla-net: Efficient semantic segmentation of large-scale point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11108-11117, 2020.", + "[36] Qingyong Hu, Bo Yang, Sheikh Khalid, Wen Xiao, Niki Trigoni, and Andrew Markham. Towards semantic segmentation of urban-scale 3d point clouds: A dataset, benchmarks and challenges. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4977-4987, 2021.", + "[37] Juana Valeria Hurtado, Rohit Mohan, Wolfram Burgard, and Abhinav Valada. Mopt: Multi-object panoptic tracking. arXiv preprint arXiv:2004.08189, 2020.", + "[38] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023.", + "[39] Maximilian Jaritz, Tuan-Hung Vu, Raoul de Charette, Emilie Wirbel, and Patrick Pérez. xmuda: Cross-modal unsupervised domain adaptation for 3d semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12605-12614, 2020.", + "[40] Maximilian Jaritz, Tuan-Hung Vu, Raoul De Charette, Emilie Wirbel, and Patrick Pérez. Cross-modal learning for domain adaptation in 3d semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1533-1544, 2023.", + "[41] Peng Jiang, Philip Osteen, Maggie Wigness, and Srikanth Saripallig. Rellis-3d dataset: Data, benchmarks and analysis. In IEEE International Conference on Robotics and Automation, pages 1110–1116, 2021.", + "[42] Tarun Kalluri, Girish Varma, Manmohan Chandraker, and C. V. Jawahar. Universal semi-supervised semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 5259-5270, 2019.", + "[43] Dongwan Kim, Yi-Hsuan Tsai, Yumin Suh, Masoud Faraki, Sparsh Garg, Manmohan Chandraker, and Bohyung Han. Learning semantic segmentation from multiple datasets with label shifts. In European Conference on Computer Vision, pages 20-36, 2022.", + "[44] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023.", + "[45] Alexey Klokov, Di Un Pak, Aleksandr Khorin, Dmitry Yudin, Leon Kochiev, Vladimir Luchinskiy, and Vitaly Bezuglyj. Daps3d: Domain adaptive projective segmentation of 3d lidar point clouds. IEEE Access, 11:79341-79356, 2023.", + "[46] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Zwei Liu. Rethinking range view representation for lidar segmentation. In IEEE/CVF International Conference on Computer Vision, pages 228-240, 2023.", + "[47] Lingdong Kong, Youquan Liu, Xin Li, Runnan Chen, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14657", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Liu. Robo3d: Towards robust and reliable 3d perception against corruptions. In IEEE/CVF International Conference on Computer Vision, pages 19994-20006, 2023.", + "[48] Lingdong Kong, Niamul Quader, and Venice Erin Liong. Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In IEEE International Conference on Robotics and Automation, pages 9338-9345, 2023.", + "[49] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21705-21715, 2023.", + "[50] Lingdong Kong, Shaoyuan Xie, Hanjiang Hu, Lai Xing Ng, Benoit R. Cottereau, and Wei Tsang Ooi. Robodepth: Robust out-of-distribution depth estimation under corruptions. In Advances in Neural Information Processing Systems, 2023.", + "[51] Lingdong Kong, Xiang Xu, Jun Cen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Calib3d: Calibrating model preferences for reliable 3d scene understanding. arXiv preprint arXiv:2403.17010, 2024.", + "[52] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17545-17555, 2023.", + "[53] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. Mseg: A composite dataset for multidomain semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2879-2888, 2020.", + "[54] Guangrui Li, Guoliang Kang, Xiaohan Wang, Yunchao Wei, and Yi Yang. Adversarily masking synthetic to mimic real: Adaptive noise injection for point cloud segmentation adaptation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20464-20474, 2023.", + "[55] Jinke Li, Xiao He, Yang Wen, Yuan Gao, Xiaogiang Cheng, and Dan Zhang. Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11809-11818, 2022.", + "[56] Li Li, Hubert PH Shum, and Toby P. Breckon. Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9361-9371, 2023.", + "[57] Rong Li, Raoul de Charette, and C. A. O. Anh-Quan. Coarse3d: Class-prototypes for contrastive learning in weakly-supervised 3d point cloud segmentation. In *British Machine Vision Conference*, 2022.", + "[58] Xin Li, Botian Shi, Yuenan Hou, Xingjiao Wu, Tianlong Ma, Yikang Li, and Liang He. Homogeneous multi-modal feature fusion and interaction for 3d object detection. In European Conference on Computer Vision, pages 691-707. Springer, 2022.", + "[59] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy." + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Video k-net: A simple, strong, and unified baseline for video segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18847-18857, 2022.", + "[60] Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17524-17534, 2023.", + "[61] Ye Li, Lingdong Kong, Hanjiang Hu, Xiaohao Xu, and Xiaonan Huang. Optimizing lidar placements for robust driving perception in adverse conditions. arXiv preprint arXiv:2403.17009, 2024.", + "[62] Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja, Dhananjai Sharma, and Zhuang Jie Chong. Amvnet: Assertion-based multi-view fusion network for lidar semantic segmentation. arXiv preprint arXiv:2012.04934, 2020.", + "[63] Minghua Liu, Yin Zhou, Charles R. Qi, Boqing Gong, Hao Su, and Dragomir Anguelov. Less: Label-efficient semantic segmentation for lidar point clouds. In European Conference on Computer Vision, pages 70-89, 2022.", + "[64] Youquan Liu, Runnan Chen, Xin Li, Lingdong Kong, Yuchen Yang, Zhaoyang Xia, Yeqi Bai, Xinge Zhu, Yuexin Ma, Yikang Li, Yu Qiao, and Yuenan Hou. Uniseg: A unified multi-modal lidar segmentation network and the openpcseg codebase. In IEEE/CVF International Conference on Computer Vision, pages 21662-21673, 2023.", + "[65] Youquan Liu, Lingdong Kong, Jun Cen, Runnan Chen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Segment any point cloud sequences by distilling vision foundation models. In Advances in Neural Information Processing Systems, 2023.", + "[66] Yueh-Cheng Liu, Yu-Kai Huang, Hung-Yueh Chiang, Hung-Ting Su, Zhe-Yu Liu, Chin-Tang Chen, Ching-Yu Tseng, and Winston H. Hsu. Learning from 2d: Contrastive pixel-to-point knowledge transfer for 3d pretraining. arXiv preprint arXiv:2104.0468, 2021.", + "[67] Zhijian Liu, Alexander Amini Haotian Tang, Xinyu Yang, Huizi Mao, Daniela L. Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In IEEE International Conference on Robotics and Automation, pages 2774-2781, 2023.", + "[68] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019.", + "[69] Yuhang Lu, Qi Jiang, Runnan Chen, Yuenan Hou, Xinge Zhu, and Yuexin Ma. See more and know more: Zero-shot point cloud segmentation via multi-modal visual data. In IEEE/CVF International Conference on Computer Vision, pages 21674-21684, 2023.", + "[70] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A comprehensive survey. International Journal of Computer Vision, 2023.", + "[71] Rodrigo Marcuzzi, Lucas Nunes, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Mask-based panoptic lidar" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14658", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "segmentation for autonomous driving. IEEE Robotics and Automation Letters, 8(2):1141-1148, 2023.", + "[72] Panagiotis Meletis and Gijs Dubbelman. Training of convolutional networks on multiple heterogeneous datasets for street scene semantic segmentation. In IEEE Intelligent Vehicles Symposium, pages 1045-1050, 2018.", + "[73] Björn Michele, Alexandre Boulch, Gilles Puy, Tuan-Hung Vu, Renaud Marlet, and Nicolas Courty. Saluda: Surface-based automotive lidar unsupervised domain adaptation. arXiv preprint arXiv:2304.03251, 2023.", + "[74] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4213-4220, 2019.", + "[75] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3d: Out-of-context data augmentation for 3d scenes. In International Conference on 3D Vision, pages 116-125, 2021.", + "[76] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023.", + "[77] Yancheng Pan, Biao Gao, Jilin Mei, Sibo Geng, Chengkun Li, and Huijing Zhao. Semanticposs: A point cloud dataset with large quantity of dynamic instances. In IEEE Intelligent Vehicles Symposium, pages 687-693, 2020.", + "[78] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. Openscene: 3d scene understanding with open vocabularies. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-824, 2023.", + "[79] Xidong Peng, Runnan Chen, Feng Qiao, Lingdong Kong, Youquan Liu, Tai Wang, Xinge Zhu, and Yuexin Ma. Sam-guided unsupervised domain adaptation for 3d segmentation. arXiv preprint arXiv:2310.08820, 2023.", + "[80] Gilles Puy, Alexandre Boulch, and Renaud Marlet. Using a waffle iron for automotive point cloud semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 3379-3389, 2023.", + "[81] 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 International Conference on Machine Learning, pages 8748-8763, 2021.", + "[82] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1623-1637, 2020.", + "[83] Giulia Rizzoli, Francesco Barbato, and Pietro Zanuttigh. Multimodal semantic segmentation in autonomous driving:" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "A review of current approaches and future perspectives. Technologies, 10(4), 2022.", + "[84] Cristiano Saltori, Fabio Galasso, Giuseppe Fiameni, Nicu Sebe, Elisa Ricci, and Fabio Poiesi. Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In European Conference on Computer Vision, pages 586-602, 2022.", + "[85] Jules Sanchez, Jean-Emmanuel Deschaud, and François Goulette. Cola: Coarse-label multi-source lidar semantic segmentation for autonomous driving. arXiv preprint arXiv:2311.03017, 2023.", + "[86] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022.", + "[87] Alvari Seppänen, Risto Ojala, and Kari Tammi. 4denoiseNET: Adverse weather denoising from adjacent point clouds. IEEE Robotics and Automation Letters, 8:456-463, 2022.", + "[88] Kshitij Sirohi, Rohit Mohan, Daniel Buscher, Wolfram Burgard, and Abhinav Valada. Efficientlps: Efficient lidar panoptic segmentation. IEEE Transactions on Robotics, 2021.", + "[89] Louis Soum-Fontez, Jean-Emmanuel Deschaud, and François Goulette. Mdt3d: Multi-dataset training for lidar 3d object detection generalization. arXiv preprint arXiv:2308.01000, 2023.", + "[90] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2446-2454, 2020.", + "[91] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, pages 685–702, 2020.", + "[92] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In IEEE/CVF International Conference on Computer Vision, pages 6411–6420, 2019.", + "[93] Larissa T Triess, David Peter, Christoph B Rist, and J Marius Zöllner. Scan-based semantic segmentation of lidar point clouds: An experimental study. In IEEE Intelligent Vehicles Symposium, pages 1116-1121, 2020.", + "[94] Larissa T. Triess, Mariella Dreissig, Christoph B. Rist, and J. Marius Zöllner. A survey on deep domain adaptation for lidar perception. In IEEE Intelligent Vehicles Symposium Workshop, pages 350-357, 2021.", + "[95] Darren Tsai, Julie Stephany Berrio, Mao Shan, Eduardo Nebot, and Stewart Worrall. Ms3d: Leveraging multiple" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14659", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "detectors for unsupervised domain adaptation in 3d object detection. arXiv preprint arXiv:2304.02431, 2023.", + "[96] Darren Tsai, Julie Stephany Berrio, Mao Shan, Eduardo Nebot, and Stewart Worrall. Ms3d++: Ensemble of experts for multi-source unsupervised domain adaption in 3d object detection. arXiv preprint arXiv:2308.05988, 2023.", + "[97] Ozan Unal, Dengxin Dai, and Luc Van Gool. Scribblesupervised lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2697-2707, 2022.", + "[98] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021.", + "[99] Xudong Wang, Zhaowei Cai, Dashan Gao, and Nuno Vasconcelos. Towards universal object detection by domain attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7289-7298, 2019.", + "[100] Xudong Wang, Shufan Li, Konstantinos Kallidromitis, Yusuke Kato, Kazuki Kozuka, and Trevor Darrell. Hierarchical open-vocabulary universal image segmentation. In Advances in Neural Information Processing Systems, 2023.", + "[101] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023.", + "[102] Yi Wei, Zibu Wei, Yongming Rao, Jiaxin Li, Jie Zhou, and Jiwen Lu. Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In European Conference on Computer Vision, pages 179-195. Springer, 2022.", + "[103] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In Advances in Neural Information Processing Systems, 2022.", + "[104] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards large-scale 3d representation learning with multi-dataset point prompt training. arXiv preprint arXiv:2308.09718, 2023.", + "[105] Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, and Ling Shao. Polarmix: A general data augmentation technique for lidar point clouds. In Advances in Neural Information Processing Systems, pages 11035-11048, 2022.", + "[106] Aoran Xiao, Jiaxing Huang, Dayan Guan, Fangneng Zhan, and Shijian Lu. Transfer learning from synthetic to real lidar point cloud for semantic segmentation. In AAAI Conference on Artificial Intelligence, pages 2795-2803, 2022.", + "[107] Aoran Xiao, Jiaxing Huang, Weihao Xuan, Ruijie Ren, Kangcheng Liu, Dayan Guan, Abdulmotaleb El Saddik, Shijian Lu, and Eric Xing. 3d semantic segmentation in the wild: Learning generalized models for adverse-condition point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9382-9392, 2023.", + "[108] Zeqi Xiao, Wenwei Zhang, Tai Wang, Chen Change Loy, Dahua Lin, and Jiangmiao Pang. Position-guided point cloud panoptic segmentation transformer. arXiv preprint arXiv:2303.13509, 2023." + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[109] Shaoyuan Xie, Lingdong Kong, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Robobev: Towards robust bird's eye view perception under corruptions. arXiv preprint arXiv:2304.06719, 2023.", + "[110] Shaoyuan Xie, Lingdong Kong, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Benchmarking and analyzing bird's eye view perception robustness to corruptions. Preprint, 2023.", + "[111] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In European Conference on Computer Vision, pages 1-19, 2020.", + "[112] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In IEEE/CVF International Conference on Computer Vision, pages 16024-16033, 2021.", + "[113] Jingyi Xu, Weidong Yang, Lingdong Kong, Youquan Liu, Rui Zhang, Qingyuan Zhou, and Ben Fei. Visual foundation models boost cross-modal unsupervised domain adaptation for 3d semantic segmentation. arXiv preprint arXiv:2403.10001, 2024.", + "[114] Xiang Xu, Lingdong Kong, Hui Shuai, and Qingshan Liu. Frnet: Frustum-range networks for scalable lidar segmentation. arXiv preprint arXiv:2312.04484, 2023.", + "[115] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European Conference on Computer Vision, pages 677-695, 2022.", + "[116] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d: Self-training for unsupervised domain adaptation on 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10368-10378, 2021.", + "[117] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: Towards a unified multi-task network for lidar perception. In AAAI Conference on Artificial Intelligence, pages 3231-3240, 2023.", + "[118] Bo Zhang, Jiakang Yuan, Botian Shi, Tao Chen, Yikang Li, and Yu Qiao. Uni3d: A unified baseline for multi-dataset 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9253-9262, 2023.", + "[119] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianfeng Gao, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. arXiv preprint arXiv:2303.08131, 2023.", + "[120] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. In Advances in Neural Information Processing Systems, pages 10326-10338, 2021.", + "[121] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In IEEE/CVF Conference on Com" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "14660", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "puter Vision and Pattern Recognition, pages 9601-9610, 2020.", + "[122] Zihui Zhang, Bo Yang, Bing Wang, and Bo Li. Growsp: Unsupervised semantic segmentation of 3d point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17619-17629, 2023.", + "[123] Yiming Zhao, Lin Bai, and Xinming Huang. Fidnet: Lidar point cloud semantic segmentation with fully interpolation decoding. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4453-4458, 2021.", + "[124] Qiang Zhou, Yuang Liu, Chaohui Yu, Jingliang Li, Zhibin Wang, and Fan Wang. Lmseg: Language-guided multi-dataset segmentation. In International Conference on Learning Representations, 2022.", + "[125] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368, 2022.", + "[126] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7571-7580, 2022.", + "[127] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-polararnet: Proposal-free lidar point cloud panoptic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13194-13203, 2021.", + "[128] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021.", + "[129] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multisensor fusion for 3d lidar semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 16280-16290, 2021.", + "[130] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. In Advances in Neural Information Processing Systems, 2023." + ], + "bbox": [ + 78, + 92, + 467, + 656 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14661", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/b0f5aa24-c6e7-4f07-a0f8-7432b1bda5de_model.json b/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/b0f5aa24-c6e7-4f07-a0f8-7432b1bda5de_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c4a41e9a7b2d221ecbb3178ff8fb96ae05bff69a --- /dev/null +++ b/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/b0f5aa24-c6e7-4f07-a0f8-7432b1bda5de_model.json @@ -0,0 +1,3176 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.154, + 0.131, + 0.819, + 0.154 + ], + "angle": 0, + "content": "Multi-Space Alignments Towards Universal LiDAR Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.18, + 0.826, + 0.216 + ], + "angle": 0, + "content": "Youquan Liu\\*,1 Lingdong Kong\\*,2,3 Xiaoyang Wu4 Runnan Chen4 Xin Li5 Liang Pan2 Ziwei Liu6 Yuexin Ma1" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.217, + 0.898, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\)ShanghaiTech University, \\(^{2}\\)Shanghai AI Laboratory, \\(^{3}\\)National University of Singapore \\(^{4}\\)University of Hong Kong, \\(^{5}\\)East China Normal University, \\(^{6}\\)S-Lab, Nanyang Technological University https://github.com/youquanl/M3Net" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.304, + 0.314, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.474, + 0.594 + ], + "angle": 0, + "content": "A unified and versatile LiDAR segmentation model with strong robustness and generalizability is desirable for safe autonomous driving perception. This work presents M3Net, a one-of-a-kind framework for fulfilling multi-task, multi-dataset, multi-modality LiDAR segmentation in a universal manner using just a single set of parameters. To better exploit data volume and diversity, we first combine large-scale driving datasets acquired by different types of sensors from diverse scenes and then conduct alignments in three spaces, namely data, feature, and label spaces, during the training. As a result, M3Net is capable of taming heterogeneous data for training state-of-the-art LiDAR segmentation models. Extensive experiments on twelve LiDAR segmentation datasets verify our effectiveness. Notably, using a shared set of parameters, M3Net achieves \\(75.1\\%\\), \\(83.1\\%\\), and \\(72.4\\%\\) mIoU scores, respectively, on the official benchmarks of SemanticKITTI, nuScenes, and Waymo Open." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.624, + 0.21, + 0.64 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.47, + 0.741 + ], + "angle": 0, + "content": "Dense and structural 3D surrounding scene understanding provides crucial information for autonomous vehicles to make proper decisions [70]. With the recent advancements in sensing technologies, especially the Light Detection and Ranging (LiDAR) sensor, a holistic scene perception can be achieved by segmenting the acquired sensor data [28, 83]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.741, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Most existing LiDAR segmentation models [1, 36, 111, 121, 128] are trained and tested in a single-task, single-dataset, single-modality manner. Despite achieving commendable results in the single domain, there is a significant performance drop when transitioning to new domains [40, 48]. The limited generalization capability hinders their facilitation of real-world applications [47, 50, 87]. In reality, LiDAR datasets are marred by significant variances, encompassing variations in data patterns due to different" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.301, + 0.895, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.577, + 0.894, + 0.646 + ], + "angle": 0, + "content": "Figure 1. Performance comparisons among M3Net [●], Single-Dataset Training [●], and Naïve Joint Training [●] across twelve LiDAR segmentation datasets. For better comparisons, the radius is normalized based on M3Net's scores. The larger the area coverage, the higher the overall performance. Best viewed in colors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.824 + ], + "angle": 0, + "content": "sensor types and weather conditions, diverse class distributions arising from varying capture scenarios, and distinct label spaces shaped by specific annotation protocols. These factors collectively pose a formidable challenge in harmonizing disparate LiDAR point clouds and jointly optimizing model parameters to effectively address multiple tasks across a range of sensor modalities [89, 118]. Empirical evidence in Fig. 3 further reveals that naively combining heterogeneous data to train a LiDAR segmentation model – without strategic alignments – often leads to sub-opt results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Recent works [6, 39, 48, 79, 87, 94, 106] resort to unsupervised domain adaptation (UDA) for utilizing training data from both source and target domains to optimize one parameter set. Nevertheless, they either focus on only the sharing mapping between two domains (by ignoring dis" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.396, + 0.9 + ], + "angle": 0, + "content": "* The first two authors contributed equally to this work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14648" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "joint classes) or directly merge source domain labels to align with the target domain [40, 113]. The overlook of the performance degradation on the source dataset and the destruction of original label mappings inevitably constrains such a learning paradigm. Furthermore, there have been efforts [85, 96, 104, 118] to employ multi-dataset learning strategies to bolster the generalization prowess of 3D perception models. However, they either necessitate dataset-specific fine-tuning, deviating from a truly universal learning approach, or converge label spaces to a coarser set, resulting in the dilution of fine-grained segmentation capabilities across diverse semantic categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.59 + ], + "angle": 0, + "content": "In this work, we define a novel paradigm towards leveraging LiDAR point clouds from different datasets to tame a single set of parameters for multi-task LiDAR segmentation. Sibling to image segmentation communities [43, 53, 124], we call this paradigm universal LiDAR segmentation. The ultimate goal of such a synergistic way of learning is to build a powerful segmentation model that can absorb rich cross-domain knowledge and, in return, achieve strong resilience and generalizability for practical usage. Given the substantial differences among datasets in terms of data characteristics, feature distributions, and labeling conventions, we introduce a comprehensive multi-space alignment approach that encompasses data-, feature-, and label-level alignments, to effectively pave the path for efficient and universally applicable LiDAR segmentation. In particular, the multi-modal data, including images and texts, is fully exploited to assist the alignment process with the guidance of more general knowledge. Through aforementioned processes, we propose M3Net to learn common knowledge across datasets, modalities, and tasks, thereby significantly enhancing its applicability in practical scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.473, + 0.756 + ], + "angle": 0, + "content": "To substantiate the efficacy of M3Net and the utility of each module developed, we have carried out a series of thorough comparative and ablation studies across an extensive array of driving datasets, as shown in Fig. 1. Notably, our best model achieves state-of-the-art LiDAR segmentation performance with \\(75.1\\%\\), \\(83.1\\%\\), \\(72.4\\%\\) mIoU scores on SemanticKITTI [3], nuScenes [27], Waymo Open [90], respectively, using a shared set of parameters. Moreover, our approach also performs well for direct knowledge transfer and out-of-distribution adaptations, further underscoring its robust capability for effective knowledge transfer." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.77, + 0.22, + 0.785 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "LiDAR Segmentation. A holistic perception of 3D scenes is crucial for safe autonomous driving [4, 7, 33, 51, 61]. Various LiDAR segmentation models have been proposed, with distinct focuses on aspects include LiDAR representations [20, 74, 91-93, 103, 121, 128], model architectures [1, 17, 24, 35, 46, 52, 80, 112], sensor fusion [18, 62, 64, 113, 129], post-processing [111, 123], data aug" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.244 + ], + "angle": 0, + "content": "mentations [75, 84, 105], etc. Most recently, researchers started to explore data efficiency [49, 56], annotation efficiency [57, 63, 65, 86, 97], annotation-free learning [10, 11, 122], zero-shot learning [12, 69], domain adaptation [6, 39, 48, 54, 73, 79, 106], and robustness [47] in LiDAR segmentation, shedding lights for practitioners. Existing pursuits, however, learn separate parameter sets for each dataset, impeding the scalability. This motivates us to explore LiDAR segmentation in a multi-task, multi-dataset, multi-modality manner with just a single set of parameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.247, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Multi-Task Learning. A proper pipeline design could enable the model to generate suitable predictions to fulfill multiple tasks simultaneously [16, 31]. The current research endeavors mainly focus on building image or video segmentation models to handle semantic, instance, and panoptic segmentation tasks [38, 59, 98, 100, 119, 120, 130]. Recently, several attempts have been made to enable multitask segmentation on LiDAR point clouds. MaskRange [30] and MaskPLS [71] extend the mask classification paradigm [15] for joint semantic and panoptic LiDAR segmentation. LidarMultiNet [117] uses global context pooling and task-specific heads to handle LiDAR-based detection and segmentation. P3Former [108] proposed a specialized positional embedding to handle the geometry ambiguity in panoptic LiDAR segmentation. Our framework also supports multi-task learning. Different from existing approaches, the proposed M3Net stands out by combining knowledge from different sensor data across multiple data sources, which achieves superior performance on each task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Multi-Dataset Learning. Leveraging data samples from different sources for training has been proven effective in enhancing robustness and generalizability [72]. Various approaches have been proposed to merge image datasets for object detection [14, 58, 60, 99, 125, 126], image segmentation [29, 42, 43, 53, 124], depth estimation [13, 82], etc. Due to large domain gaps, the image-based methods are often hard to be transferred to 3D. To combine multiple Li-DAR datasets for 3D object detection, MDT3D [89] defines a coarse label set to handle the label space conflicts in different point cloud datasets. MS3D++ [95, 96] ensembles pre-trained detectors from different source datasets for multi-domain adaptation. Uni3D [118] resorts to dataset-specific detection heads and feature re-coupling for training a unified 3D object detector. Recently, PPT [104] proposed to pre-train a point cloud segmentation network using data from multiple datasets. However, the pre-trained weights are then fine-tuned on each specific dataset, which breaks the universal learning manner. The closest work to us is COLA [85], which trains a single model across multiple sources by converting dataset-specific labels to a common coarse set. Such a conversion, however, leads to the loss of fine-grained segmentation across the various semantic categories. Differently, our M3Net is tailored to tame a single" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14649" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.201, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.087, + 0.34, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.087, + 0.477, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.087, + 0.617, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.087, + 0.755, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.087, + 0.895, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.279, + 0.893, + 0.308 + ], + "angle": 0, + "content": "Figure 2. Statistical analysis of six sharing semantic classes in the nuScenes [●], SemanticKITTI [●], and Waymo Open [●] datasets. Each violin plot shows the class distribution across LiDAR scenes spanning 50 meters, centered around the ego-vehicle. Best viewed in colors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.473, + 0.556 + ], + "angle": 0, + "content": "parameter set to fulfill multi-task prediction across multiple datasets while still maintaining the original label mappings. Multi-Modality Learning. Recent trend favors synergistic learning from data of different modalities, such as vision, language, and speech [2, 8, 19, 26, 76, 81, 101]. For LiDAR segmentation, several works [9, 39, 40, 67, 115] explored the distillation of image features to point clouds. Recently, OpenScene [78] and CLIP2Scene [11] proposed to leverage point clouds along with multi-view images and language for open-vocabulary learning. PPKT [66], SLidR [86], and Seal [65] form cross-sensor contrastive learning objectives to pre-train the LiDAR segmentation models. The advantages of sensor fusion have been consistently proven. In this work, to pursue universal LiDAR segmentation, we propose to align multi-space point clouds via images and texts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.57, + 0.188, + 0.587 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.746 + ], + "angle": 0, + "content": "Our study serves as an early attempt at combining multi-task, multi-dataset, multi-modality knowledge into a single set of parameters to fulfill universal LiDAR segmentation. We start with a pilot study to unveil the difficulties in merging heterogeneous LiDAR point clouds (cf. Sec. 3.1). We then present M3Net, a versatile LiDAR segmentation network tailored to pursue i) statistical consistency in the data space (cf. Sec. 3.2), ii) cross-modality-assisted alignment in the feature space (cf. Sec. 3.3), and iii) language-guided unification in the label space (cf. Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.201, + 0.772 + ], + "angle": 0, + "content": "3.1. Pilot Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.87 + ], + "angle": 0, + "content": "The current de facto of training a LiDAR segmentation network adopts a task-by-task and dataset-by-dataset pipeline. Despite the superior performance achieved under such standalone settings, the trained parameter sets cannot be shared to satisfy out-of-domain requirements and, therefore, limits their use cases for practical applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Naïve Joint Training. A natural alternative to breaking the above constraint is to jointly train a network across multiple" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.895, + 0.51 + ], + "angle": 0, + "content": "datasets for better generalizability. However, as depicted in Fig. 2, it is often non-trivial to naively combine heterogeneous data with large data distribution gaps to train a universal LiDAR segmentation model without proper alignments. To testify this, we conducted a pilot study using the prior art MinkUNet [20] for both standalone and joint training on three large-scale datasets [3, 27, 90]. As shown in Fig. 3 (a) and (d), a brutal combination undermines the segmentation performance. Due to large discrepancies in aspects like sensor configurations, data acquisitions, label mappings, and domain shifts, the jointly trained representations tend to be disruptive instead of being more general." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.895, + 0.679 + ], + "angle": 0, + "content": "LiDAR Sensor Discrepancy. To understand the root cause of performance degradation, we conducted another study that controls point cloud density discrepancies when merging datasets. As shown in Fig. 3 (b) and (c), joint training on data collected by sensors with different beam numbers tends to suffer more severely than merging less density variant data. We hypothesize that this is mainly caused by the data statistical variations. In light of these observations, we propose a bag of suitable operations in the following sections to alleviate the large domain gaps among different LiDAR segmentation datasets [3, 4, 7, 27, 90]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.692, + 0.712, + 0.709 + ], + "angle": 0, + "content": "3.2. Data-Space Alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.894, + 0.869 + ], + "angle": 0, + "content": "Given a total of \\(S\\) datasets \\(D^{s} = \\{(x^{s},y^{s})|1\\leq s\\leq S\\}\\) where \\((x^s,y^s)\\) denotes the data-label pairs constituting a dataset. For the LiDAR segmentation task, \\(x^{s}\\) often encompasses the LiDAR point cloud \\(P^{s} = \\{p_{x},p_{y},p_{z}\\}^{s}\\in\\) \\(\\mathbb{R}^{N\\times 3}\\) and synchronized multi-view camera images \\(V^{s} =\\) \\(\\{\\mathcal{I}_1,\\dots,\\mathcal{I}_l\\} |l = 1,\\dots,L\\}\\) , where \\(\\mathcal{I}_t\\in \\mathbb{R}^{H\\times W\\times 3}\\) \\(N\\) is the number of points, \\(L\\) denotes the number of camera sensors, \\(H\\) and \\(W\\) are the height and width of the image, respectively. \\(y^{s}\\in \\mathbb{R}^{N}\\) denotes point cloud labels in the label space \\(\\mathbb{Y}^s\\) , we unify the label space as \\(\\mathbb{Y}^u = \\mathbb{Y}^1\\cup \\mathbb{Y}^2\\dots \\cup \\mathbb{Y}^S\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Cross-Modality Data Alignment. As a multi-sensing system, the information encoded in \\( P_{i}^{s} \\) and \\( V_{i}^{s} \\) are intuitively" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14650" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.088, + 0.191, + 0.102 + ], + "angle": 0, + "content": "Standalone Training" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.108, + 0.245, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.193, + 0.268, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.193, + 0.295, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.089, + 0.465, + 0.102 + ], + "angle": 0, + "content": "Joint Training (Same # Beam)" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.108, + 0.468, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.246, + 0.244, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.247, + 0.468, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.345, + 0.47, + 0.4 + ], + "angle": 0, + "content": "Figure 3. A pilot study of naively merging different datasets for training the MinkUNet [20] model. Compared to the standalone training in (a), either jointly training with (b) the same, (c) different, or (d) all sensor-acquired data will cause severe degradation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.509 + ], + "angle": 0, + "content": "complementary to each other [3, 7, 90]. To leverage such an advantage, we resort to the correspondences embedded in camera calibration matrices to bridge the LiDAR points and camera image pixels. Specifically, for each point \\(\\mathbf{p} = (p^x,p^y,p^z)\\) in \\(P^s\\), the corresponding pixel \\((u,v)\\) can be found by the following transformations:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.515, + 0.469, + 0.547 + ], + "angle": 0, + "content": "\\[\n[ u, v, 1 ] ^ {\\mathrm {T}} = \\frac {1}{p ^ {z}} \\cdot T _ {s} \\cdot T \\cdot [ p ^ {x}, p ^ {y}, 1 ] ^ {\\mathrm {T}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.469, + 0.629 + ], + "angle": 0, + "content": "where \\( T \\in \\mathbb{R}^{4 \\times 4} \\) is the camera extrinsic matrix that consists of a rotation matrix and a translation matrix, and \\( T_{s} \\in \\mathbb{R}^{3 \\times 4} \\) is the camera intrinsic matrix. As we will show in the following sections, such a cross-sensor data alignment serves as the foundation for alignments in other spaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.469, + 0.765 + ], + "angle": 0, + "content": "Cross-Sensor Statistical Alignment. To mitigate the discrepancies in sensor installations across different datasets, we incorporate a point coordinate alignment operation. Specifically, drawing upon insights from prior domain adaptation approaches [102, 116], we adjust the coordinate origins of point clouds from different datasets by introducing an offset \\(\\sigma \\in \\mathbb{R}^{1\\times 3}\\) to the ground plane. We find empirically that such an alignment can largely reduce the degradation caused by the variations in different sensor setups." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Dataset-Specific Rasterization. It is conventional to rasterize LiDAR point clouds \\( P^s \\) using unified rasterization parameters, e.g., voxel size [91, 128] or horizontal range view resolution [74, 111]. However, the point clouds acquired in different LiDAR datasets naturally differ in density, range, intensity, etc., which tends to favor different rasterization parameters. To meet such a requirement, we select dataset-specific parameters for rasterization on each dataset through empirical experiments and analyses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.244 + ], + "angle": 0, + "content": "Decoupled BN. Another challenge in training across multiple datasets is the presence of domain gaps, which can result in significant statistical shifts of feature learning among datasets. Such shifts can hinder the convergence and affect the model's ability to generalize well across diverse datasets. We adopt a decoupled batch norm (BN) for point cloud features in each dataset. Instead of using the traditional BN, which calculates mean and variance across all samples in a mini-batch, the decoupled BN tends to adapt each dataset's specific characteristics independently." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.254, + 0.733, + 0.27 + ], + "angle": 0, + "content": "3.3. Feature-Space Alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.383 + ], + "angle": 0, + "content": "We aim to acquire a generalized feature representation for downstream tasks. Compared to point clouds, images contribute stronger visual, textural, and semantic information. Thus, the collaboration between pixels and points could enrich the overall representation. Previous research [11, 60, 64] has consistently demonstrated that such a combination results often leads to improved performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.893, + 0.765 + ], + "angle": 0, + "content": "Cross-Modality Assisted Alignment. In the context of multi-dataset joint training, our objective is to establish a unified feature space by leveraging image features to assist point cloud features. Acknowledging that images used in training lack ground truth labels [3, 27], we utilize image features from a pre-trained model as an alternative, facilitating a more universally applicable representation. We feed camera images \\( V^{s} \\) into a pre-trained DeepLab [98] and a vision-language model (VLM) and visualize the output image features by t-SNE [25]. As shown in Fig. 4, we observe that image features from DeepLab appear disorderly and lack semantics. In contrast, features from VLM share a more unified feature space. Motivated by this, we propose a cross-modality assisted alignment that uses VLM to help align the feature space. Specifically, the camera images \\( V^{s} \\) are fed to the frozen image encoder from VLM to obtain image features \\( F_{v} = \\{\\mathcal{F}_{v}^{1},\\mathcal{F}_{v}^{2},\\dots,\\mathcal{F}_{v}^{s}\\} \\), where \\( \\mathcal{F}_v^s\\in \\mathbb{R}^{c\\times h\\times w} \\). The LiDAR point clouds \\( P^{s} \\), on the other hand, are fed to the point encoder followed by a projection layer to generate the point features \\( F_{p} = \\{\\mathcal{F}_{p}^{1},\\mathcal{F}_{p}^{2},\\dots,\\mathcal{F}_{p}^{s}\\} \\), where \\( \\mathcal{F}_p^s\\in \\mathbb{R}^{m\\times c} \\); \\( m \\) denotes the number of non-empty grids. We then leverage the paired image features \\( \\hat{F}_v\\in \\mathbb{R}^{m_p\\times c} \\) and point feature \\( \\hat{F}_p\\in \\mathbb{R}^{m_p\\times c} \\) for alignment, where \\( m_{p} \\) is the number of point-pixel pairs. After obtaining \\( \\hat{F}_p \\) and \\( \\hat{F}_v \\), the cross-modality alignment is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.776, + 0.892, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c m a}} \\left(\\hat {F} _ {v}, \\hat {F} _ {p}\\right) = 1 - \\frac {\\hat {F} _ {v} \\cdot \\hat {F} _ {p}}{\\| \\hat {F} _ {v} \\| \\cdot \\| \\hat {F} _ {p} \\|}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Domain-Aware Cross-Modality Alignment. With cross-modality alignment, we transfer the knowledge of VLM to the point encoder, enabling the point features to gain a more comprehensive representation. However, during the execution of the above alignment, we have narrowed" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14651" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.473, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.348, + 0.471, + 0.417 + ], + "angle": 0, + "content": "Figure 4. The t-SNE plots of learned features before and after the feature-space alignment in merging the nuScenes [●], SemanticKITTI [●], and Waymo Open [●] datasets. We show image features from (a) standalone networks; (b) SAM [44], and point cloud features (c) before and (d) after the feature-space alignment." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.442, + 0.473, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.666, + 0.47, + 0.722 + ], + "angle": 0, + "content": "Figure 5. Feature-space alignment in M3Net. We leverage both image features \\( F_{v} \\) and LiDAR point cloud features \\( F_{p} \\) extracted from image encoder \\( E_{img} \\) and point encoder \\( E_{pcd} \\) to employ the regularization via V2P loss and achieve feature-space alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "it exclusively to image and point features from the same dataset. In this mode, point features solely learn from matching image features, restricting their knowledge acquisition. Ideally, we aim to ensure that image features encompass not only scenes identical to those represented in point clouds but also scenes from other datasets. To address this, we propose a domain-aware cross-modality guided alignment, as illustrated in Fig. 5. Specifically, we first extract, for each dataset, \\( F_{v} \\) and \\( F_{p} \\) from the same image encoder \\( E_{img} \\) and point encoder \\( E_{pcd} \\) during the cross" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "modality assisted alignment. The sets of features from all datasets are concatenated along the channel dimension to form \\(\\widetilde{F}_v\\in \\mathbb{R}^{c_v\\times h\\times w}\\). Subsequently, we sequentially feed \\(\\widetilde{F}_v\\) through a branch that consists of a global average pooling and an MLP. Simultaneously, \\(\\widetilde{F}_v\\) is fed to an auxiliary branch that undergoes the same processing flow and generates an output after the softmax function \\(\\mathcal{G}(\\cdot)\\). The outputs from both branches are multiplied to obtain \\(F_{m}\\in \\mathbb{R}^{c_{v}\\times 1\\times 1}\\). The overall process can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.237, + 0.892, + 0.256 + ], + "angle": 0, + "content": "\\[\nF _ {m} = M L P \\left(P o o l \\left(\\widetilde {F} _ {v}\\right)\\right) \\cdot \\mathcal {G} \\left(M L P \\left(P o o l \\left(\\widetilde {F} _ {v}\\right)\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.343 + ], + "angle": 0, + "content": "Next, we forward \\(F_{m}\\) to a sigmoid activation function \\(\\mathcal{H}(\\cdot)\\) and multiply it with input image features \\(\\widetilde{F}_v\\). The resulting output is added to \\(\\widetilde{F}_v\\) and passed through the MLP layers to obtain the final image features \\(F_{vf} \\in \\mathbb{R}^{c \\times h \\times w}\\). The forward process of this operation is depicted as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.352, + 0.892, + 0.371 + ], + "angle": 0, + "content": "\\[\nF _ {v f} = M L P \\left(\\left(\\mathcal {H} \\left(F _ {m}\\right) \\cdot \\widetilde {F} _ {v}\\right) + \\widetilde {F} _ {v}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.429 + ], + "angle": 0, + "content": "Finally, we leverage the cross-modality data alignment to acquire paired image features \\(\\hat{F}_{vf} \\in \\mathbb{R}^{m_p \\times c}\\) and paired point feature \\(\\hat{F}_p\\). The overall objective function is:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.439, + 0.892, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v 2 p} \\left(\\hat {F} _ {v f}, \\hat {F} _ {p}\\right) = 1 - \\frac {\\hat {F} _ {v f} \\cdot \\hat {F} _ {p}}{\\| \\hat {F} _ {v f} \\| \\cdot \\| \\hat {F} _ {p} \\|}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.485, + 0.719, + 0.502 + ], + "angle": 0, + "content": "3.4. Label-Space Alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.659 + ], + "angle": 0, + "content": "Label Conflict. In multi-dataset joint training settings, label conflicts emerge as a significant challenge. This often refers to the inconsistencies in class labels across different datasets involved in the training process. The discrepancy can arise due to variations in annotation conventions, labeling errors, or even differences in the underlying semantics of classes between datasets. In our baseline, we unionize the different label spaces across datasets into \\(\\mathbb{Y}^u\\), where all datasets share a single LiDAR segmentation head. However, this may introduce several potential drawbacks:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.66, + 0.892, + 0.704 + ], + "angle": 0, + "content": "- Loss of granularity: Unified label spaces could lose semantic granularity, particularly when dealing with subtle category differences in between different datasets." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.705, + 0.892, + 0.75 + ], + "angle": 0, + "content": "- Information loss: During label space consolidation, details unique to each dataset may be obscured or lost, especially for those related to domain-specific categories." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.751, + 0.892, + 0.796 + ], + "angle": 0, + "content": "- Increased complexity: Handling a unified label space may necessitate more complex model architectures or training strategies, thereby increasing overall complexity." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.66, + 0.892, + 0.796 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.797, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To address these issues, we introduce a language-guided label-space alignment to facilitate a more holistic semantic correlation across datasets. Given the natural correspondence between images and texts and the strong correlation between images and point clouds, we aim to strategically utilize the image modality as a bridge to establish language-guided alignments. Such a process consists of a text-driven" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14652" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.473, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.47, + 0.365 + ], + "angle": 0, + "content": "Figure 6. Label-space alignment in M3Net. We leverage image features \\( F_{v} \\), point cloud features \\( F_{p} \\), and text embedding \\( F_{t} \\) extracted from \\( E_{img} \\), \\( E_{pcd} \\), and \\( E_{txt} \\), respectively, for regularization via the I2P, P2T, and V2T losses in the label-space alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.469, + 0.411 + ], + "angle": 0, + "content": "point alignment, a text-driven image alignment, and a cross-modality-assisted label alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.47, + 0.699 + ], + "angle": 0, + "content": "Text-Driven Alignments. As depicted in Fig. 6, images \\( V^s \\) are fed into the frozen image encoder \\( E_{img} \\) to extract the image features \\( F_v \\). Concurrently, the LiDAR point clouds \\( P^s \\) are processed by the point encoder \\( E_{pcd} \\) to generate the point features \\( F_p \\). Additionally, given the text input \\( T^s \\), text embedding features \\( F_t \\in \\mathbb{R}^{Q \\times c} \\) are obtained from a frozen text encoder \\( E_{txt} \\), where \\( Q \\) represents the number of categories across datasets. The text is composed of class names from unified label space \\( \\mathbb{Y}_u \\) placed into pre-defined templates, and the text embedding captures semantic information of the corresponding classes. Subsequently, pixel-text pairs \\( \\{v_k, t_k\\}_{k=1}^M \\) and point-text pairs \\( \\{p_k, t_k\\}_{k=1}^M \\) are generated, where \\( M \\) represents the number of pairs. Leveraging the semantic information contained in the text, we selectively choose positive and negative samples for both images and points for contrastive learning. It is noteworthy that negative samples are confined to the specific dataset category space. The overall objective of the text-driven point alignment function is shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.706, + 0.469, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p 2 t} = - \\sum_ {q = 1} ^ {Q} \\log \\left(\\frac {\\sum_ {t _ {k} \\in q , p _ {k}} \\exp \\left(< t _ {k} , p _ {k} > / \\tau\\right)}{\\sum_ {t _ {k} \\in q , t _ {k} \\notin q , p _ {j}} \\exp \\left(< t _ {k} , p _ {k} > / \\tau\\right)}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.469, + 0.839 + ], + "angle": 0, + "content": "where \\( t_k \\in q \\) indicates that \\( t_k \\) is generated by the \\( q \\)-th classes name, and \\( Q \\) is the number of classes. Symbol \\( <, > \\) denotes the scalar product operation and \\( \\tau \\) is a temperature term \\( (\\tau > 0) \\). Similarly, the objective of the text-driven image alignment function is illustrated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.845, + 0.469, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v 2 t} = - \\sum_ {q = 1} ^ {Q} \\log \\left(\\frac {\\sum_ {t _ {k} \\in q , v _ {k}} \\exp \\left(< t _ {k} , v _ {k} > / \\tau\\right)}{\\sum_ {t _ {k} \\in q , t _ {k} \\notin q , v _ {j}} \\exp \\left(< t _ {k} , v _ {k} > / \\tau\\right)}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.245 + ], + "angle": 0, + "content": "Cross-Modality-Assisted Label Alignment. After text-driven alignments, the subsequent crucial step entails aligning the point and image modalities within the label space. We first obtain image logits \\( F_{vl} = \\{\\mathcal{F}_{vl}^{1},\\mathcal{F}_{vl}^{2},\\dots,\\mathcal{F}_{vl}^{s}\\} \\) and point logits \\( F_{pl} = \\{\\mathcal{F}_{pl}^{1},\\mathcal{F}_{pl}^{2},\\dots,\\mathcal{F}_{pl}^{s}\\} \\) from text-driven alignments, where \\( \\mathcal{F}_{vl}^{s}\\in \\mathbb{R}^{Q\\times H\\times W} \\), \\( \\mathcal{F}_{pl}^{s}\\in \\mathbb{R}^{N\\times Q} \\). Subsequently, we conduct cross-modality alignment to obtain paired image logits \\( \\hat{F}_{vl}\\in \\mathbb{R}^{m_p\\times Q} \\) and paired point logits \\( \\hat{F}_{pl}\\in \\mathbb{R}^{m_p\\times Q} \\). Formally, the cross-modality-assisted alignment in the label space is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.255, + 0.892, + 0.294 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i 2 p} \\left(\\hat {F} _ {v l}, \\hat {F} _ {p l}\\right) = 1 - \\frac {\\hat {F} _ {v l} \\cdot \\hat {F} _ {p l}}{\\| \\hat {F} _ {v l} \\| \\cdot \\| \\hat {F} _ {p l} \\|}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.302, + 0.892, + 0.332 + ], + "angle": 0, + "content": "Finally, the complete objective function for the language-guided label-space alignment is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.343, + 0.892, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l a b e l}} = \\mathcal {L} _ {p 2 t} + \\mathcal {L} _ {i 2 p} + \\mathcal {L} _ {v 2 t}. \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.37, + 0.779, + 0.385 + ], + "angle": 0, + "content": "3.5. Universal LiDAR Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.438 + ], + "angle": 0, + "content": "We enhance the versatility of M3Net via multi-tasking learning. This integration involves an instance extractor to enable joint semantic and panoptic LiDAR segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Panoptic LiDAR Segmentation. Motivated by DSNet [32, 33], our instance extractor comprises an instance head and a clustering step. The instance head encompasses several MLPs designed to predict the offsets between instance centers. The clustering step uses semantic predictions to filter out stuff points, thereby retaining only those associated with thing points. The remaining points undergo a mean-shift clustering [21], utilizing features from the instance head to discern distinct instances. Lastly, we employ the L1 loss \\(\\mathcal{L}_{l1}\\) to optimize the thing point regression process." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.59, + 0.891, + 0.62 + ], + "angle": 0, + "content": "Overall Objectives. Putting everything together, the overall objective of M3Net is to minimize the following losses:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.631, + 0.892, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {v 2 p} + \\mathcal {L} _ {\\text {l a b e l}} + \\mathcal {L} _ {c e} + \\mathcal {L} _ {\\text {l o v a s z}} + \\mathcal {L} _ {l 1}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.658, + 0.892, + 0.689 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{ce}\\) and \\(\\mathcal{L}_{lovasz}\\) denote the cross-entropy loss and the Lovasz-softmax [5] loss, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.702, + 0.634, + 0.719 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.697, + 0.743 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Datasets. Our M3Net framework and baselines are trained on a combination of nuScenes [27], SemanticKITTI [3], and Waymo Open [90]. Meanwhile, we resort to another five LiDAR-based perception datasets [41, 45, 77, 106, 107] and two 3D robustness evaluation datasets [47] to verify the strong generalizability of M3Net. Due to space limits, additional details regarding the datasets are in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. M3Net is implemented based on Pointcept [23] and MMDetection3D [22]. We use two backbones in our experiments, i.e., MinkUNet [20] and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14653" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 1. Ablation study on the M3Net alignments happen in the Data, Feature, and Label spaces, respectively, when combining the SemanticKITTI [3], nuScenes [27], and Waymo Open [90] datasets. The mAcc and mIoU scores are in percentage. Best scores are in bold." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.125, + 0.895, + 0.255 + ], + "angle": 0, + "content": "
-ConfigurationsMinkUNet [20]PTv2+ [103]
SemKITTInuScenesWaymoSemKITTInuScenesWaymo
mAccmIoUmAccmIoUmAccmIoUmAccmIoUmAccmIoUmAccmIoU
BaselineNaïve Joint Training62.4354.0365.0559.8473.7665.3967.9661.5976.5369.6575.6867.00
M3Net (Ours)DataFeatureLabel------
✓73.8269.0183.6676.8977.8869.3778.5569.9586.2279.1380.9672.15
✓✓74.3669.6485.1778.8878.3169.7079.4370.8787.1080.2680.7472.33
✓✓73.8569.3485.2078.9078.0469.5580.3071.1387.4480.4580.6972.30
✓✓✓74.4069.8585.3079.0078.6670.1580.0072.0087.9180.9081.1172.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.272, + 0.89, + 0.302 + ], + "angle": 0, + "content": "Table 2. Panoptic LiDAR segmentation results on the val sets of the Panoptic-SemanticKITTI [3] and Panoptic-nuScenes [27] datasets. All scores are given in percentage. The best and second-best scores are highlighted in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.308, + 0.892, + 0.49 + ], + "angle": 0, + "content": "
MethodConfigurationsPanoptic-SemanticKITTIPanoptic-nuScenes
PQ\\( PQ^† \\)RQSQmIoUPQ\\( PQ^† \\)RQSQmIoU
Panoptic-TrackNet [37]40.0-48.373.053.851.456.263.380.258.0
Panoptic-PolarNet [127]59.164.170.278.364.563.467.275.383.966.9
EfficientLPS [88]Single-Dataset Training59.265.169.875.064.959.262.882.970.769.4
DSNet [32]61.465.272.779.069.664.767.676.183.576.3
Panoptic-PHNet [55]61.7---65.774.777.784.288.279.7
BaselineNaïve Joint Training56.0359.6465.7873.7261.5956.6760.6166.7583.4969.65
M3Net (Ours)DataFeatureLabel--
✓62.3465.1772.6074.6769.9568.4971.1179.1385.4979.13
✓✓62.9165.7373.3275.4770.8771.4773.8681.5386.7180.26
✓✓63.2367.8973.6181.6671.1371.5373.9181.8086.9280.45
✓✓✓63.8768.6673.1082.3572.0071.7074.0182.2086.4780.90
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.747 + ], + "angle": 0, + "content": "PTv2+ [103]. We trained M3Net on four A100 GPUs for 50 epochs with a batch size of 6 for each GPU. The initial learning rate is set to 0.002. We adopt the AdamW optimizer [68] with a weight decay of 0.005 and cosine decay learning rate scheduler. For the dataset-specific rasterization, we set voxel sizes to \\(0.05\\mathrm{m}\\), \\(0.1\\mathrm{m}\\), and \\(0.05\\mathrm{m}\\) for SemanticKITTI [3], nuScenes [27], and Waymo Open [90], respectively. For the data augmentation, we employ random flipping, jittering, scaling, rotation, and Mix3D [75]. Due to space limits, kindly refer to Appendix for additional details. Evaluation Metrics. We adopt conventional reportings of mAcc and mIoU for LiDAR semantic segmentation, \\(PQ\\), \\(PQ^{\\dagger}\\), \\(SQ\\), and \\(RQ\\) for panoptic segmentation, and mCE and mRR for 3D robustness evaluation. Due to the space limit, kindly refer to our Appendix for more detailed definitions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.231, + 0.773 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Multi-Space Alignments. The effectiveness of three proposed alignments over the joint training baselines is shown in Tab. 1. We observe that the data-space alignment plays the most crucial role in improving the universal LiDAR segmentation performance. Without proper data alignments, joint training with either MinkUNet [20] or the stronger PTv2+ [103] will suffer severe degradation, especially on sparser point clouds [27]. On top of the data-space align-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.596 + ], + "angle": 0, + "content": "ment, the combinations of multi-view images at the feature space and the language-guided knowledge at the label space further enhance the learned feature representations. The results show that they work synergistically in merging knowledge from heterogeneous domains during joint training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.893, + 0.734 + ], + "angle": 0, + "content": "Panoptic LiDAR Segmentation. In Tab. 2, we present another ablation study focusing on panoptic LiDAR segmentation. All three alignments incorporated in M3Net demonstrate significant improvements over the baselines. This highlights the pronounced efficacy of our multi-space alignments. Moreover, our approach outperforms the single-dataset state-of-the-art method Panoptic-PHNet [55] by a notable \\(2.17\\%\\) PQ on Panoptic-SemanticKITTI [3] and achieves compelling results on Panoptic-nuScenes [27]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Visual Feature Alignments. We conduct a qualitative analysis of the learned visual feature distributions in the form of t-SNE [25]. Fig. 4 (a) and (b) represent the distributions of learned visual features among three datasets from DeepLab and VLM backbones, respectively. The features obtained by the latter exhibit more distinct semantics in feature space. The concentrated distribution space is advantageous for achieving feature alignments across multiple datasets. Additionally, Fig. 4 (c) and (d) illustrate the distribution of point cloud features before and after feature-space alignment. As can be seen, the feature distribution distances" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14654" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 3. Knowledge transfer and generalization analyses across five LiDAR segmentation datasets and two 3D robustness evaluation datasets. All scores are given in percentage. The best and second-best scores are highlighted in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.125, + 0.894, + 0.243 + ], + "angle": 0, + "content": "
MethodRELLIS-3DSemanticPOSSSemanticSTFSynLiDARDAPS-3DSemKITTI-CnuScenes-C
1%10%HalfFullHalfFull1%10%HalfFullmCEmRRmCEmRR
PPKT [66]49.7154.3350.1856.0050.9254.6937.5746.4878.9084.00--105.6476.06
SLidR [86]49.7554.5751.5655.3652.0154.3542.0547.8481.0085.40--106.0875.99
Seal [65]51.0955.0353.2656.8953.4655.3643.5849.2681.8885.90--92.6383.08
Naïve Joint37.7750.2342.1952.3146.7048.0018.5642.3773.9177.89113.6584.73128.9781.45
Single-Dataset40.1754.2547.6955.0050.3351.1923.1745.0875.1080.8795.1184.9599.6379.06
M3Net (Ours)51.2755.0553.6057.1753.7855.4244.1049.9382.0886.0086.4385.7791.0379.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.263, + 0.472, + 0.319 + ], + "angle": 0, + "content": "Table 4. LiDAR semantic segmentation results on the val and test sets of SemanticKITTI [3] and nuScenes [27], and the val set of Waymo Open [90]. All scores are in percentage. The best and second-best scores are highlighted in bold and underline." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.327, + 0.47, + 0.564 + ], + "angle": 0, + "content": "
MethodSemKITTInuScenesWaymo
ValTestValTestmIoUmAcc
RangeNet++ [74]-52.2-65.5--
PolarNet [121]57.254.371.069.8--
SalsaNext [24]-59.5-72.2--
RangeViT [1]60.764.075.2---
MinkUNet [20]63.863.773.3-65.976.6
SPVNAS [91]64.766.4-77.467.4-
AMVNet [62]65.265.377.277.3--
RPVNet [112]65.570.377.6---
(AF)2-S3Net [18]-69.7-78.3--
Cylinder3D [128]65.967.876.177.966.0-
PVKD [34]66.471.276.0---
WaffleIron [80]66.870.879.1---
RangeFormer [46]67.673.378.180.1--
SphereFormer [52]67.874.878.481.969.9-
FRNet [114]68.773.379.082.5--
PTv2+ [103]70.370.680.282.670.680.2
LidarMultiNet [117]---81.473.8-
M3Net (Ours)72.075.180.983.172.481.1
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.59, + 0.47, + 0.62 + ], + "angle": 0, + "content": "between the three datasets have been largely reduced, providing evidence of the alignment effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.633, + 0.265, + 0.65 + ], + "angle": 0, + "content": "4.3. Comparative Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Comparisons to State of the Arts. In Tab. 4, we compare M3Net with current best-performing models on the benchmarks of SemanticKITTI [3], nuScenes [27], and Waymo Open [90]. Remarkably, M3Net consistently outperforms existing approaches across all three datasets. Specifically, on SemanticKITTI [3], M3Net achieves a \\(72.0\\%\\) mIoU on the validation set, surpassing the closest method by a notable margin of \\(1.7\\%\\) mIoU. Similarly, on nuScenes [27], M3Net achieves \\(80.9\\%\\) mIoU and \\(83.1\\%\\) mIoU on the validation and test sets, demonstrating its robustness and generalization capabilities. Additionally, the performance of M3Net on Waymo Open [90] is competitive with prior arts. We achieve a mIoU of \\(72.4\\%\\) and a mAcc of \\(81.1\\%\\). These results highlight again the superiority of M3Net in handling complex diverse LiDAR segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Direct Knowledge Transfer. To further validate the strong" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.892, + 0.401 + ], + "angle": 0, + "content": "knowledge transfer capability of M3Net, we conduct extensive experiments on five different LiDAR-based perception datasets [41, 45, 77, 106, 107]. These datasets have unique data collection protocols and data distributions. As shown in Fig. 1 and the first ten columns in Tab. 3, our framework constantly outperforms the prior arts, the naive joint training, and the single-dataset baselines across all five datasets. This concretely supports the strong knowledge transfer efficacy brought by multi-space alignments in M3Net." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.893, + 0.552 + ], + "angle": 0, + "content": "Out-of-Distribution Generalization. Evaluating the generalization ability of models on out-of-training-distribution data is crucial, particularly in safety-critical fields like autonomous driving [50, 109, 110]. In this context, we resort to the two corruption datasets from the Robo3D [47] benchmark, i.e., SemanticKITTI-C and nuScenes-C, to conduct our assessment. From the last four columns of Tab. 3, we observe that M3Net achieves better results than the naive joint training and other single-dataset approaches, proving the strong generalizability of the learned representations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.567, + 0.619, + 0.582 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.773 + ], + "angle": 0, + "content": "In this work, we presented M3Net, a universal framework capable of fulfilling multi-task, multi-dataset, multimodality LiDAR segmentation using a single set of parameters. Through extensive analyses, we validated the effectiveness of applying data-, feature-, and label-space alignments to handle such a challenging task. In addition, our comprehensive analysis and discourse have delved into the fundamental challenges of acquiring the general knowledge for scalable 3D perception, which holds substantial potential to propel further research in this domain. Our future strides focus on combining more data resources to further enhance the alignments and adaptations in our framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was partially supported by NSFC (No.62206173) and MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University). This work was also supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14655" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Angelika Ando, Spyros Gidaris, Andrei Bursuc, Gilles Puy, Alexandre Boulch, and Renaud Marlet. Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5240-5250, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.186, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.256, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[3] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.326, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[4] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Jürgen Gall, and Cyrill Stachniss. Towards 3d lidar-based semantic scene understanding of 3d point cloud sequences: The semantickitti dataset. International Journal of Robotics Research, 40:959-96, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.397, + 0.471, + 0.466 + ], + "angle": 0, + "content": "[5] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: a tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4413-4421, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.467, + 0.47, + 0.535 + ], + "angle": 0, + "content": "[6] Alexandre Boulch, Corentin Sautier, Björn Michele, Gilles Puy, and Renaud Marlet. Also: Automotive lidar self-supervision by occupancy estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13455-13465, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.537, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.621, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.691, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[9] Jun Cen, Shiwei Zhang, Yixuan Pei, Kun Li, Hang Zheng, Maochun Luo, Yingya Zhang, and Qifeng Chen. Cmdfusion: Bidirectional fusion network with cross-modality knowledge distillation for lidar semantic segmentation. arXiv preprint arXiv:2307.04091, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.761, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[10] Runnan Chen, Youquan Liu, Lingdong Kong, Nenglun Chen, Xinge Zhu, Yuexin Ma, Tongliang Liu, and Wenping Wang. Towards label-free scene understanding by vision foundation models. In Advances in Neural Information Processing Systems, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.832, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[11] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[12] Runnan Chen, Xinge Zhu, Nenglun Chen, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Bridging language and geometric primitives for zero-shot point cloud segmentation. In ACM International Conference on Multimedia, pages 5380-5388, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[13] Tian Chen, Shijie An, Yuan Zhang, Chongyang Ma, Huayan Wang, Xiaoyan Guo, and Wen Zheng. Improving monocular depth estimation by leveraging structural awareness and complementary datasets. In European Conference on Computer Vision, pages 90-108, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.233, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[14] Yanbei Chen, Manchen Wang, Abhay Mittal, Zhenlin Xu, Paolo Favaro, Joseph Tighe, and Davide Modolo. Scaledet: A scalable multi-dataset object detector. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7288-7297, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.302, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[15] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. In Advances in Neural Information Processing Systems, pages 17864-17875, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.359, + 0.892, + 0.428 + ], + "angle": 0, + "content": "[16] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.429, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[17] Huixian Cheng, Xianfeng Han, and Guoqiang Xiao. Cenet: Toward concise and efficient lidar semantic segmentation for autonomous driving. In IEEE International Conference on Multimedia and Expo, pages 1-6, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.485, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[18] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. Af2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.555, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[19] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.846, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[20] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14656" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[21] Dorin Comaniciu and Peter Meer. Mean shift: A robust approach toward feature space analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(5):603-619, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[22] MMDetection3D Contributors. MMDetection3D: Open-MMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.205, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[23] Pointcept Contributors. Pointcept: A codebase for point cloud perception research. https://github.com/Pointcept/Pointcept, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.247, + 0.47, + 0.301 + ], + "angle": 0, + "content": "[24] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In International Symposium on Visual Computing, pages 207-222, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.303, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[25] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (11), 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.47, + 0.385 + ], + "angle": 0, + "content": "[26] Christoph Feichtenhofer, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[27] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters, 7:3795-3802, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.456, + 0.47, + 0.511 + ], + "angle": 0, + "content": "[28] Biao Gao, Yancheng Pan, Chengkun Li, Sibo Geng, and Huijing Zhao. Are we hungry for 3d lidar data for semantic segmentation? a survey of datasets and methods. IEEE Transactions on Intelligent Transportation Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.512, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[29] Xiuye Gu, Yin Cui, Jonathan Huang, Abdullah Rashwan, Xuan Yang, Xingyi Zhou, Golnaz Ghiasi, Weicheng Kuo, Huizhong Chen, Liang-Chieh Chen, and David A. Ross. Dataseg: Taming a universal multi-dataset multi-task segmentation model. arXiv preprint arXiv:2306.01736, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.581, + 0.47, + 0.635 + ], + "angle": 0, + "content": "[30] Yi Gu, Yuming Huang, Chengzhong Xu, and Hui Kong. Maskrange: A mask-classification model for range-view based lidar segmentation. arXiv preprint arXiv:2206.12073, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.637, + 0.47, + 0.678 + ], + "angle": 0, + "content": "[31] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In IEEE/CVF International Conference on Computer Vision, pages 2961-2969, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.679, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[32] Fangzhou Hong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Lidar-based panoptic segmentation via dynamic shifting network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13090-13099, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.735, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[33] Fangzhou Hong, Lingdong Kong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Unified 3d and 4d panoptic segmentation via dynamic shifting networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.805, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[34] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-voxel knowledge distillation for lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8479–8488, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[35] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham." + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Randla-net: Efficient semantic segmentation of large-scale point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11108-11117, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.136, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[36] Qingyong Hu, Bo Yang, Sheikh Khalid, Wen Xiao, Niki Trigoni, and Andrew Markham. Towards semantic segmentation of urban-scale 3d point clouds: A dataset, benchmarks and challenges. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4977-4987, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.219, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[37] Juana Valeria Hurtado, Rohit Mohan, Wolfram Burgard, and Abhinav Valada. Mopt: Multi-object panoptic tracking. arXiv preprint arXiv:2004.08189, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[38] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.33, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[39] Maximilian Jaritz, Tuan-Hung Vu, Raoul de Charette, Emilie Wirbel, and Patrick Pérez. xmuda: Cross-modal unsupervised domain adaptation for 3d semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12605-12614, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.401, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[40] Maximilian Jaritz, Tuan-Hung Vu, Raoul De Charette, Emilie Wirbel, and Patrick Pérez. Cross-modal learning for domain adaptation in 3d semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1533-1544, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.47, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[41] Peng Jiang, Philip Osteen, Maggie Wigness, and Srikanth Saripallig. Rellis-3d dataset: Data, benchmarks and analysis. In IEEE International Conference on Robotics and Automation, pages 1110–1116, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.526, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[42] Tarun Kalluri, Girish Varma, Manmohan Chandraker, and C. V. Jawahar. Universal semi-supervised semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 5259-5270, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[43] Dongwan Kim, Yi-Hsuan Tsai, Yumin Suh, Masoud Faraki, Sparsh Garg, Manmohan Chandraker, and Bohyung Han. Learning semantic segmentation from multiple datasets with label shifts. In European Conference on Computer Vision, pages 20-36, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.651, + 0.892, + 0.732 + ], + "angle": 0, + "content": "[44] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.734, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[45] Alexey Klokov, Di Un Pak, Aleksandr Khorin, Dmitry Yudin, Leon Kochiev, Vladimir Luchinskiy, and Vitaly Bezuglyj. Daps3d: Domain adaptive projective segmentation of 3d lidar point clouds. IEEE Access, 11:79341-79356, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[46] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Zwei Liu. Rethinking range view representation for lidar segmentation. In IEEE/CVF International Conference on Computer Vision, pages 228-240, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[47] Lingdong Kong, Youquan Liu, Xin Li, Runnan Chen, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14657" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Liu. Robo3d: Towards robust and reliable 3d perception against corruptions. In IEEE/CVF International Conference on Computer Vision, pages 19994-20006, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[48] Lingdong Kong, Niamul Quader, and Venice Erin Liong. Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In IEEE International Conference on Robotics and Automation, pages 9338-9345, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.207, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[49] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21705-21715, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.264, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[50] Lingdong Kong, Shaoyuan Xie, Hanjiang Hu, Lai Xing Ng, Benoit R. Cottereau, and Wei Tsang Ooi. Robodepth: Robust out-of-distribution depth estimation under corruptions. In Advances in Neural Information Processing Systems, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.334, + 0.47, + 0.39 + ], + "angle": 0, + "content": "[51] Lingdong Kong, Xiang Xu, Jun Cen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Calib3d: Calibrating model preferences for reliable 3d scene understanding. arXiv preprint arXiv:2403.17010, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.392, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[52] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17545-17555, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.449, + 0.47, + 0.516 + ], + "angle": 0, + "content": "[53] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. Mseg: A composite dataset for multidomain semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2879-2888, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.519, + 0.47, + 0.601 + ], + "angle": 0, + "content": "[54] Guangrui Li, Guoliang Kang, Xiaohan Wang, Yunchao Wei, and Yi Yang. Adversarily masking synthetic to mimic real: Adaptive noise injection for point cloud segmentation adaptation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20464-20474, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.604, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[55] Jinke Li, Xiao He, Yang Wen, Yuan Gao, Xiaogiang Cheng, and Dan Zhang. Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11809-11818, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.675, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[56] Li Li, Hubert PH Shum, and Toby P. Breckon. Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9361-9371, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.746, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[57] Rong Li, Raoul de Charette, and C. A. O. Anh-Quan. Coarse3d: Class-prototypes for contrastive learning in weakly-supervised 3d point cloud segmentation. In *British Machine Vision Conference*, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.803, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[58] Xin Li, Botian Shi, Yuenan Hou, Xingjiao Wu, Tianlong Ma, Yikang Li, and Liang He. Homogeneous multi-modal feature fusion and interaction for 3d object detection. In European Conference on Computer Vision, pages 691-707. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[59] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Video k-net: A simple, strong, and unified baseline for video segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18847-18857, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.15, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[60] Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17524-17534, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.234, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[61] Ye Li, Lingdong Kong, Hanjiang Hu, Xiaohao Xu, and Xiaonan Huang. Optimizing lidar placements for robust driving perception in adverse conditions. arXiv preprint arXiv:2403.17009, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[62] Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja, Dhananjai Sharma, and Zhuang Jie Chong. Amvnet: Assertion-based multi-view fusion network for lidar semantic segmentation. arXiv preprint arXiv:2012.04934, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.349, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[63] Minghua Liu, Yin Zhou, Charles R. Qi, Boqing Gong, Hao Su, and Dragomir Anguelov. Less: Label-efficient semantic segmentation for lidar point clouds. In European Conference on Computer Vision, pages 70-89, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.406, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[64] Youquan Liu, Runnan Chen, Xin Li, Lingdong Kong, Yuchen Yang, Zhaoyang Xia, Yeqi Bai, Xinge Zhu, Yuexin Ma, Yikang Li, Yu Qiao, and Yuenan Hou. Uniseg: A unified multi-modal lidar segmentation network and the openpcseg codebase. In IEEE/CVF International Conference on Computer Vision, pages 21662-21673, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.49, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[65] Youquan Liu, Lingdong Kong, Jun Cen, Runnan Chen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Segment any point cloud sequences by distilling vision foundation models. In Advances in Neural Information Processing Systems, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.561, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[66] Yueh-Cheng Liu, Yu-Kai Huang, Hung-Yueh Chiang, Hung-Ting Su, Zhe-Yu Liu, Chin-Tang Chen, Ching-Yu Tseng, and Winston H. Hsu. Learning from 2d: Contrastive pixel-to-point knowledge transfer for 3d pretraining. arXiv preprint arXiv:2104.0468, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.632, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[67] Zhijian Liu, Alexander Amini Haotian Tang, Xinyu Yang, Huizi Mao, Daniela L. Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In IEEE International Conference on Robotics and Automation, pages 2774-2781, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.703, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[68] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.746, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[69] Yuhang Lu, Qi Jiang, Runnan Chen, Yuenan Hou, Xinge Zhu, and Yuexin Ma. See more and know more: Zero-shot point cloud segmentation via multi-modal visual data. In IEEE/CVF International Conference on Computer Vision, pages 21674-21684, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[70] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A comprehensive survey. International Journal of Computer Vision, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[71] Rodrigo Marcuzzi, Lucas Nunes, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Mask-based panoptic lidar" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14658" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "segmentation for autonomous driving. IEEE Robotics and Automation Letters, 8(2):1141-1148, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[72] Panagiotis Meletis and Gijs Dubbelman. Training of convolutional networks on multiple heterogeneous datasets for street scene semantic segmentation. In IEEE Intelligent Vehicles Symposium, pages 1045-1050, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.178, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[73] Björn Michele, Alexandre Boulch, Gilles Puy, Tuan-Hung Vu, Renaud Marlet, and Nicolas Courty. Saluda: Surface-based automotive lidar unsupervised domain adaptation. arXiv preprint arXiv:2304.03251, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.233, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[74] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4213-4220, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.289, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[75] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3d: Out-of-context data augmentation for 3d scenes. In International Conference on 3D Vision, pages 116-125, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.345, + 0.47, + 0.481 + ], + "angle": 0, + "content": "[76] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.484, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[77] Yancheng Pan, Biao Gao, Jilin Mei, Sibo Geng, Chengkun Li, and Huijing Zhao. Semanticposs: A point cloud dataset with large quantity of dynamic instances. In IEEE Intelligent Vehicles Symposium, pages 687-693, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.539, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[78] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. Openscene: 3d scene understanding with open vocabularies. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-824, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.609, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[79] Xidong Peng, Runnan Chen, Feng Qiao, Lingdong Kong, Youquan Liu, Tai Wang, Xinge Zhu, and Yuexin Ma. Sam-guided unsupervised domain adaptation for 3d segmentation. arXiv preprint arXiv:2310.08820, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.665, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[80] Gilles Puy, Alexandre Boulch, and Renaud Marlet. Using a waffle iron for automotive point cloud semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 3379-3389, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.721, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[81] 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 International Conference on Machine Learning, pages 8748-8763, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.805, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[82] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1623-1637, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[83] Giulia Rizzoli, Francesco Barbato, and Pietro Zanuttigh. Multimodal semantic segmentation in autonomous driving:" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "A review of current approaches and future perspectives. Technologies, 10(4), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[84] Cristiano Saltori, Fabio Galasso, Giuseppe Fiameni, Nicu Sebe, Elisa Ricci, and Fabio Poiesi. Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In European Conference on Computer Vision, pages 586-602, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[85] Jules Sanchez, Jean-Emmanuel Deschaud, and François Goulette. Cola: Coarse-label multi-source lidar semantic segmentation for autonomous driving. arXiv preprint arXiv:2311.03017, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.247, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[86] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.317, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[87] Alvari Seppänen, Risto Ojala, and Kari Tammi. 4denoiseNET: Adverse weather denoising from adjacent point clouds. IEEE Robotics and Automation Letters, 8:456-463, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.372, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[88] Kshitij Sirohi, Rohit Mohan, Daniel Buscher, Wolfram Burgard, and Abhinav Valada. Efficientlps: Efficient lidar panoptic segmentation. IEEE Transactions on Robotics, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[89] Louis Soum-Fontez, Jean-Emmanuel Deschaud, and François Goulette. Mdt3d: Multi-dataset training for lidar 3d object detection generalization. arXiv preprint arXiv:2308.01000, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.484, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[90] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2446-2454, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.623, + 0.892, + 0.69 + ], + "angle": 0, + "content": "[91] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, pages 685–702, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.692, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[92] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In IEEE/CVF International Conference on Computer Vision, pages 6411–6420, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.762, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[93] Larissa T Triess, David Peter, Christoph B Rist, and J Marius Zöllner. Scan-based semantic segmentation of lidar point clouds: An experimental study. In IEEE Intelligent Vehicles Symposium, pages 1116-1121, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[94] Larissa T. Triess, Mariella Dreissig, Christoph B. Rist, and J. Marius Zöllner. A survey on deep domain adaptation for lidar perception. In IEEE Intelligent Vehicles Symposium Workshop, pages 350-357, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[95] Darren Tsai, Julie Stephany Berrio, Mao Shan, Eduardo Nebot, and Stewart Worrall. Ms3d: Leveraging multiple" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14659" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "detectors for unsupervised domain adaptation in 3d object detection. arXiv preprint arXiv:2304.02431, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.121, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[96] Darren Tsai, Julie Stephany Berrio, Mao Shan, Eduardo Nebot, and Stewart Worrall. Ms3d++: Ensemble of experts for multi-source unsupervised domain adaption in 3d object detection. arXiv preprint arXiv:2308.05988, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.177, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[97] Ozan Unal, Dengxin Dai, and Luc Van Gool. Scribblesupervised lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2697-2707, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.233, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[98] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.302, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[99] Xudong Wang, Zhaowei Cai, Dashan Gao, and Nuno Vasconcelos. Towards universal object detection by domain attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7289-7298, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[100] Xudong Wang, Shufan Li, Konstantinos Kallidromitis, Yusuke Kato, Kazuki Kozuka, and Trevor Darrell. Hierarchical open-vocabulary universal image segmentation. In Advances in Neural Information Processing Systems, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[101] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[102] Yi Wei, Zibu Wei, Yongming Rao, Jiaxin Li, Jie Zhou, and Jiwen Lu. Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In European Conference on Computer Vision, pages 179-195. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[103] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.636 + ], + "angle": 0, + "content": "[104] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards large-scale 3d representation learning with multi-dataset point prompt training. arXiv preprint arXiv:2308.09718, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[105] Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, and Ling Shao. Polarmix: A general data augmentation technique for lidar point clouds. In Advances in Neural Information Processing Systems, pages 11035-11048, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.762 + ], + "angle": 0, + "content": "[106] Aoran Xiao, Jiaxing Huang, Dayan Guan, Fangneng Zhan, and Shijian Lu. Transfer learning from synthetic to real lidar point cloud for semantic segmentation. In AAAI Conference on Artificial Intelligence, pages 2795-2803, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[107] Aoran Xiao, Jiaxing Huang, Weihao Xuan, Ruijie Ren, Kangcheng Liu, Dayan Guan, Abdulmotaleb El Saddik, Shijian Lu, and Eric Xing. 3d semantic segmentation in the wild: Learning generalized models for adverse-condition point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9382-9392, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[108] Zeqi Xiao, Wenwei Zhang, Tai Wang, Chen Change Loy, Dahua Lin, and Jiangmiao Pang. Position-guided point cloud panoptic segmentation transformer. arXiv preprint arXiv:2303.13509, 2023." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[109] Shaoyuan Xie, Lingdong Kong, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Robobev: Towards robust bird's eye view perception under corruptions. arXiv preprint arXiv:2304.06719, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[110] Shaoyuan Xie, Lingdong Kong, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Benchmarking and analyzing bird's eye view perception robustness to corruptions. Preprint, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[111] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In European Conference on Computer Vision, pages 1-19, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[112] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In IEEE/CVF International Conference on Computer Vision, pages 16024-16033, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[113] Jingyi Xu, Weidong Yang, Lingdong Kong, Youquan Liu, Rui Zhang, Qingyuan Zhou, and Ben Fei. Visual foundation models boost cross-modal unsupervised domain adaptation for 3d semantic segmentation. arXiv preprint arXiv:2403.10001, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[114] Xiang Xu, Lingdong Kong, Hui Shuai, and Qingshan Liu. Frnet: Frustum-range networks for scalable lidar segmentation. arXiv preprint arXiv:2312.04484, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[115] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European Conference on Computer Vision, pages 677-695, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[116] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d: Self-training for unsupervised domain adaptation on 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10368-10378, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[117] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: Towards a unified multi-task network for lidar perception. In AAAI Conference on Artificial Intelligence, pages 3231-3240, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[118] Bo Zhang, Jiakang Yuan, Botian Shi, Tao Chen, Yikang Li, and Yu Qiao. Uni3d: A unified baseline for multi-dataset 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9253-9262, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[119] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianfeng Gao, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. arXiv preprint arXiv:2303.08131, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[120] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. In Advances in Neural Information Processing Systems, pages 10326-10338, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[121] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In IEEE/CVF Conference on Com" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14660" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.468, + 0.119 + ], + "angle": 0, + "content": "puter Vision and Pattern Recognition, pages 9601-9610, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[122] Zihui Zhang, Bo Yang, Bing Wang, and Bo Li. Growsp: Unsupervised semantic segmentation of 3d point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17619-17629, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[123] Yiming Zhao, Lin Bai, and Xinming Huang. Fidnet: Lidar point cloud semantic segmentation with fully interpolation decoding. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4453-4458, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.29 + ], + "angle": 0, + "content": "[124] Qiang Zhou, Yuang Liu, Chaohui Yu, Jingliang Li, Zhibin Wang, and Fan Wang. Lmseg: Language-guided multi-dataset segmentation. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.468, + 0.347 + ], + "angle": 0, + "content": "[125] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[126] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7571-7580, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.468, + 0.46 + ], + "angle": 0, + "content": "[127] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-polararnet: Proposal-free lidar point cloud panoptic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13194-13203, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.462, + 0.468, + 0.53 + ], + "angle": 0, + "content": "[128] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.532, + 0.468, + 0.6 + ], + "angle": 0, + "content": "[129] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multisensor fusion for 3d lidar semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 16280-16290, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.602, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[130] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. In Advances in Neural Information Processing Systems, 2023." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14661" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/b0f5aa24-c6e7-4f07-a0f8-7432b1bda5de_origin.pdf b/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/b0f5aa24-c6e7-4f07-a0f8-7432b1bda5de_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e80ee368b5fef85f34c90bc8e69a066a6960158b --- /dev/null +++ b/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/b0f5aa24-c6e7-4f07-a0f8-7432b1bda5de_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f50d4cca07a1d0dc6f1c5981cfe0ee0667dcc1436bdfa689ff4a1f5b3a064d1 +size 2404113 diff --git a/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/full.md b/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6d131c557d77e385bf5922ff3dc30288dcd6c4a3 --- /dev/null +++ b/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/full.md @@ -0,0 +1,397 @@ +# Multi-Space Alignments Towards Universal LiDAR Segmentation + +Youquan Liu\*,1 Lingdong Kong\*,2,3 Xiaoyang Wu4 Runnan Chen4 Xin Li5 Liang Pan2 Ziwei Liu6 Yuexin Ma1 + +$^{1}$ ShanghaiTech University, $^{2}$ Shanghai AI Laboratory, $^{3}$ National University of Singapore $^{4}$ University of Hong Kong, $^{5}$ East China Normal University, $^{6}$ S-Lab, Nanyang Technological University https://github.com/youquanl/M3Net + +# Abstract + +A unified and versatile LiDAR segmentation model with strong robustness and generalizability is desirable for safe autonomous driving perception. This work presents M3Net, a one-of-a-kind framework for fulfilling multi-task, multi-dataset, multi-modality LiDAR segmentation in a universal manner using just a single set of parameters. To better exploit data volume and diversity, we first combine large-scale driving datasets acquired by different types of sensors from diverse scenes and then conduct alignments in three spaces, namely data, feature, and label spaces, during the training. As a result, M3Net is capable of taming heterogeneous data for training state-of-the-art LiDAR segmentation models. Extensive experiments on twelve LiDAR segmentation datasets verify our effectiveness. Notably, using a shared set of parameters, M3Net achieves $75.1\%$ , $83.1\%$ , and $72.4\%$ mIoU scores, respectively, on the official benchmarks of SemanticKITTI, nuScenes, and Waymo Open. + +# 1. Introduction + +Dense and structural 3D surrounding scene understanding provides crucial information for autonomous vehicles to make proper decisions [70]. With the recent advancements in sensing technologies, especially the Light Detection and Ranging (LiDAR) sensor, a holistic scene perception can be achieved by segmenting the acquired sensor data [28, 83]. + +Most existing LiDAR segmentation models [1, 36, 111, 121, 128] are trained and tested in a single-task, single-dataset, single-modality manner. Despite achieving commendable results in the single domain, there is a significant performance drop when transitioning to new domains [40, 48]. The limited generalization capability hinders their facilitation of real-world applications [47, 50, 87]. In reality, LiDAR datasets are marred by significant variances, encompassing variations in data patterns due to different + +![](images/610d391e645481a25a4707f26a4e4b53a8d20eca1327984146383805462fea08.jpg) +Figure 1. Performance comparisons among M3Net [●], Single-Dataset Training [●], and Naïve Joint Training [●] across twelve LiDAR segmentation datasets. For better comparisons, the radius is normalized based on M3Net's scores. The larger the area coverage, the higher the overall performance. Best viewed in colors. + +sensor types and weather conditions, diverse class distributions arising from varying capture scenarios, and distinct label spaces shaped by specific annotation protocols. These factors collectively pose a formidable challenge in harmonizing disparate LiDAR point clouds and jointly optimizing model parameters to effectively address multiple tasks across a range of sensor modalities [89, 118]. Empirical evidence in Fig. 3 further reveals that naively combining heterogeneous data to train a LiDAR segmentation model – without strategic alignments – often leads to sub-opt results. + +Recent works [6, 39, 48, 79, 87, 94, 106] resort to unsupervised domain adaptation (UDA) for utilizing training data from both source and target domains to optimize one parameter set. Nevertheless, they either focus on only the sharing mapping between two domains (by ignoring dis + +joint classes) or directly merge source domain labels to align with the target domain [40, 113]. The overlook of the performance degradation on the source dataset and the destruction of original label mappings inevitably constrains such a learning paradigm. Furthermore, there have been efforts [85, 96, 104, 118] to employ multi-dataset learning strategies to bolster the generalization prowess of 3D perception models. However, they either necessitate dataset-specific fine-tuning, deviating from a truly universal learning approach, or converge label spaces to a coarser set, resulting in the dilution of fine-grained segmentation capabilities across diverse semantic categories. + +In this work, we define a novel paradigm towards leveraging LiDAR point clouds from different datasets to tame a single set of parameters for multi-task LiDAR segmentation. Sibling to image segmentation communities [43, 53, 124], we call this paradigm universal LiDAR segmentation. The ultimate goal of such a synergistic way of learning is to build a powerful segmentation model that can absorb rich cross-domain knowledge and, in return, achieve strong resilience and generalizability for practical usage. Given the substantial differences among datasets in terms of data characteristics, feature distributions, and labeling conventions, we introduce a comprehensive multi-space alignment approach that encompasses data-, feature-, and label-level alignments, to effectively pave the path for efficient and universally applicable LiDAR segmentation. In particular, the multi-modal data, including images and texts, is fully exploited to assist the alignment process with the guidance of more general knowledge. Through aforementioned processes, we propose M3Net to learn common knowledge across datasets, modalities, and tasks, thereby significantly enhancing its applicability in practical scenarios. + +To substantiate the efficacy of M3Net and the utility of each module developed, we have carried out a series of thorough comparative and ablation studies across an extensive array of driving datasets, as shown in Fig. 1. Notably, our best model achieves state-of-the-art LiDAR segmentation performance with $75.1\%$ , $83.1\%$ , $72.4\%$ mIoU scores on SemanticKITTI [3], nuScenes [27], Waymo Open [90], respectively, using a shared set of parameters. Moreover, our approach also performs well for direct knowledge transfer and out-of-distribution adaptations, further underscoring its robust capability for effective knowledge transfer. + +# 2. Related Work + +LiDAR Segmentation. A holistic perception of 3D scenes is crucial for safe autonomous driving [4, 7, 33, 51, 61]. Various LiDAR segmentation models have been proposed, with distinct focuses on aspects include LiDAR representations [20, 74, 91-93, 103, 121, 128], model architectures [1, 17, 24, 35, 46, 52, 80, 112], sensor fusion [18, 62, 64, 113, 129], post-processing [111, 123], data aug + +mentations [75, 84, 105], etc. Most recently, researchers started to explore data efficiency [49, 56], annotation efficiency [57, 63, 65, 86, 97], annotation-free learning [10, 11, 122], zero-shot learning [12, 69], domain adaptation [6, 39, 48, 54, 73, 79, 106], and robustness [47] in LiDAR segmentation, shedding lights for practitioners. Existing pursuits, however, learn separate parameter sets for each dataset, impeding the scalability. This motivates us to explore LiDAR segmentation in a multi-task, multi-dataset, multi-modality manner with just a single set of parameters. + +Multi-Task Learning. A proper pipeline design could enable the model to generate suitable predictions to fulfill multiple tasks simultaneously [16, 31]. The current research endeavors mainly focus on building image or video segmentation models to handle semantic, instance, and panoptic segmentation tasks [38, 59, 98, 100, 119, 120, 130]. Recently, several attempts have been made to enable multitask segmentation on LiDAR point clouds. MaskRange [30] and MaskPLS [71] extend the mask classification paradigm [15] for joint semantic and panoptic LiDAR segmentation. LidarMultiNet [117] uses global context pooling and task-specific heads to handle LiDAR-based detection and segmentation. P3Former [108] proposed a specialized positional embedding to handle the geometry ambiguity in panoptic LiDAR segmentation. Our framework also supports multi-task learning. Different from existing approaches, the proposed M3Net stands out by combining knowledge from different sensor data across multiple data sources, which achieves superior performance on each task. + +Multi-Dataset Learning. Leveraging data samples from different sources for training has been proven effective in enhancing robustness and generalizability [72]. Various approaches have been proposed to merge image datasets for object detection [14, 58, 60, 99, 125, 126], image segmentation [29, 42, 43, 53, 124], depth estimation [13, 82], etc. Due to large domain gaps, the image-based methods are often hard to be transferred to 3D. To combine multiple Li-DAR datasets for 3D object detection, MDT3D [89] defines a coarse label set to handle the label space conflicts in different point cloud datasets. MS3D++ [95, 96] ensembles pre-trained detectors from different source datasets for multi-domain adaptation. Uni3D [118] resorts to dataset-specific detection heads and feature re-coupling for training a unified 3D object detector. Recently, PPT [104] proposed to pre-train a point cloud segmentation network using data from multiple datasets. However, the pre-trained weights are then fine-tuned on each specific dataset, which breaks the universal learning manner. The closest work to us is COLA [85], which trains a single model across multiple sources by converting dataset-specific labels to a common coarse set. Such a conversion, however, leads to the loss of fine-grained segmentation across the various semantic categories. Differently, our M3Net is tailored to tame a single + +![](images/766b15b8b801135d7eabe8e60a7bdaa7e165b07eadaae8904a1d98a4a475b11e.jpg) +Figure 2. Statistical analysis of six sharing semantic classes in the nuScenes [●], SemanticKITTI [●], and Waymo Open [●] datasets. Each violin plot shows the class distribution across LiDAR scenes spanning 50 meters, centered around the ego-vehicle. Best viewed in colors. + +![](images/1e97aa12f4e903a23eebbbb9b226519a3af01def186fa0199b251ba872c246ec.jpg) + +![](images/e7e1ea742b504d35b9462e6605d179dc2f0f66e4d05dcf6d627975e1f3bce0f9.jpg) + +![](images/0290866900cf110d7921e2ad02cc7ebd6f1081777035dd7d72036204f086a3b2.jpg) + +![](images/9052bb2739c8011afdc180a8fec8b872ccf983ab9bfdd0b01f6d58ee31a9d630.jpg) + +![](images/cfba0f78cfc727b73924884070d78c5dcc2d5d0cc32428cf22407bc12d91bddc.jpg) + +parameter set to fulfill multi-task prediction across multiple datasets while still maintaining the original label mappings. Multi-Modality Learning. Recent trend favors synergistic learning from data of different modalities, such as vision, language, and speech [2, 8, 19, 26, 76, 81, 101]. For LiDAR segmentation, several works [9, 39, 40, 67, 115] explored the distillation of image features to point clouds. Recently, OpenScene [78] and CLIP2Scene [11] proposed to leverage point clouds along with multi-view images and language for open-vocabulary learning. PPKT [66], SLidR [86], and Seal [65] form cross-sensor contrastive learning objectives to pre-train the LiDAR segmentation models. The advantages of sensor fusion have been consistently proven. In this work, to pursue universal LiDAR segmentation, we propose to align multi-space point clouds via images and texts. + +# 3. Approach + +Our study serves as an early attempt at combining multi-task, multi-dataset, multi-modality knowledge into a single set of parameters to fulfill universal LiDAR segmentation. We start with a pilot study to unveil the difficulties in merging heterogeneous LiDAR point clouds (cf. Sec. 3.1). We then present M3Net, a versatile LiDAR segmentation network tailored to pursue i) statistical consistency in the data space (cf. Sec. 3.2), ii) cross-modality-assisted alignment in the feature space (cf. Sec. 3.3), and iii) language-guided unification in the label space (cf. Sec. 3.4). + +# 3.1. Pilot Study + +The current de facto of training a LiDAR segmentation network adopts a task-by-task and dataset-by-dataset pipeline. Despite the superior performance achieved under such standalone settings, the trained parameter sets cannot be shared to satisfy out-of-domain requirements and, therefore, limits their use cases for practical applications. + +Naïve Joint Training. A natural alternative to breaking the above constraint is to jointly train a network across multiple + +datasets for better generalizability. However, as depicted in Fig. 2, it is often non-trivial to naively combine heterogeneous data with large data distribution gaps to train a universal LiDAR segmentation model without proper alignments. To testify this, we conducted a pilot study using the prior art MinkUNet [20] for both standalone and joint training on three large-scale datasets [3, 27, 90]. As shown in Fig. 3 (a) and (d), a brutal combination undermines the segmentation performance. Due to large discrepancies in aspects like sensor configurations, data acquisitions, label mappings, and domain shifts, the jointly trained representations tend to be disruptive instead of being more general. + +LiDAR Sensor Discrepancy. To understand the root cause of performance degradation, we conducted another study that controls point cloud density discrepancies when merging datasets. As shown in Fig. 3 (b) and (c), joint training on data collected by sensors with different beam numbers tends to suffer more severely than merging less density variant data. We hypothesize that this is mainly caused by the data statistical variations. In light of these observations, we propose a bag of suitable operations in the following sections to alleviate the large domain gaps among different LiDAR segmentation datasets [3, 4, 7, 27, 90]. + +# 3.2. Data-Space Alignment + +Given a total of $S$ datasets $D^{s} = \{(x^{s},y^{s})|1\leq s\leq S\}$ where $(x^s,y^s)$ denotes the data-label pairs constituting a dataset. For the LiDAR segmentation task, $x^{s}$ often encompasses the LiDAR point cloud $P^{s} = \{p_{x},p_{y},p_{z}\}^{s}\in$ $\mathbb{R}^{N\times 3}$ and synchronized multi-view camera images $V^{s} =$ $\{\mathcal{I}_1,\dots,\mathcal{I}_l\} |l = 1,\dots,L\}$ , where $\mathcal{I}_t\in \mathbb{R}^{H\times W\times 3}$ $N$ is the number of points, $L$ denotes the number of camera sensors, $H$ and $W$ are the height and width of the image, respectively. $y^{s}\in \mathbb{R}^{N}$ denotes point cloud labels in the label space $\mathbb{Y}^s$ , we unify the label space as $\mathbb{Y}^u = \mathbb{Y}^1\cup \mathbb{Y}^2\dots \cup \mathbb{Y}^S$ + +Cross-Modality Data Alignment. As a multi-sensing system, the information encoded in $P_{i}^{s}$ and $V_{i}^{s}$ are intuitively + +![](images/034652c7fd889434320ee744bf594cb70749a3e09166a44978313464ba3b0cc1.jpg) +Standalone Training + +![](images/63746aa1134b4c098642f2b87f6880f4c4065873d812c8c78112bf5e0a0e0f12.jpg) + +![](images/454289ff174e478ba3a0b50cde2c33065935cf04c313dd8e8997e04683da73b4.jpg) + +![](images/be2644d1baa9563208d57cde624574e1a758ff44124c902dfe803732d4eb44b7.jpg) +Joint Training (Same # Beam) + +![](images/df956c51dd541326673b31a545bd870fb2314fa94dfa8589d60ccbc66164f7a6.jpg) +Figure 3. A pilot study of naively merging different datasets for training the MinkUNet [20] model. Compared to the standalone training in (a), either jointly training with (b) the same, (c) different, or (d) all sensor-acquired data will cause severe degradation. + +![](images/bf2472168858d61efd730da86ff7b5690e9670a0a78d2e07d7c3049012a6279f.jpg) + +complementary to each other [3, 7, 90]. To leverage such an advantage, we resort to the correspondences embedded in camera calibration matrices to bridge the LiDAR points and camera image pixels. Specifically, for each point $\mathbf{p} = (p^x,p^y,p^z)$ in $P^s$ , the corresponding pixel $(u,v)$ can be found by the following transformations: + +$$ +[ u, v, 1 ] ^ {\mathrm {T}} = \frac {1}{p ^ {z}} \cdot T _ {s} \cdot T \cdot [ p ^ {x}, p ^ {y}, 1 ] ^ {\mathrm {T}}, \tag {1} +$$ + +where $T \in \mathbb{R}^{4 \times 4}$ is the camera extrinsic matrix that consists of a rotation matrix and a translation matrix, and $T_{s} \in \mathbb{R}^{3 \times 4}$ is the camera intrinsic matrix. As we will show in the following sections, such a cross-sensor data alignment serves as the foundation for alignments in other spaces. + +Cross-Sensor Statistical Alignment. To mitigate the discrepancies in sensor installations across different datasets, we incorporate a point coordinate alignment operation. Specifically, drawing upon insights from prior domain adaptation approaches [102, 116], we adjust the coordinate origins of point clouds from different datasets by introducing an offset $\sigma \in \mathbb{R}^{1\times 3}$ to the ground plane. We find empirically that such an alignment can largely reduce the degradation caused by the variations in different sensor setups. + +Dataset-Specific Rasterization. It is conventional to rasterize LiDAR point clouds $P^s$ using unified rasterization parameters, e.g., voxel size [91, 128] or horizontal range view resolution [74, 111]. However, the point clouds acquired in different LiDAR datasets naturally differ in density, range, intensity, etc., which tends to favor different rasterization parameters. To meet such a requirement, we select dataset-specific parameters for rasterization on each dataset through empirical experiments and analyses. + +Decoupled BN. Another challenge in training across multiple datasets is the presence of domain gaps, which can result in significant statistical shifts of feature learning among datasets. Such shifts can hinder the convergence and affect the model's ability to generalize well across diverse datasets. We adopt a decoupled batch norm (BN) for point cloud features in each dataset. Instead of using the traditional BN, which calculates mean and variance across all samples in a mini-batch, the decoupled BN tends to adapt each dataset's specific characteristics independently. + +# 3.3. Feature-Space Alignment + +We aim to acquire a generalized feature representation for downstream tasks. Compared to point clouds, images contribute stronger visual, textural, and semantic information. Thus, the collaboration between pixels and points could enrich the overall representation. Previous research [11, 60, 64] has consistently demonstrated that such a combination results often leads to improved performance. + +Cross-Modality Assisted Alignment. In the context of multi-dataset joint training, our objective is to establish a unified feature space by leveraging image features to assist point cloud features. Acknowledging that images used in training lack ground truth labels [3, 27], we utilize image features from a pre-trained model as an alternative, facilitating a more universally applicable representation. We feed camera images $V^{s}$ into a pre-trained DeepLab [98] and a vision-language model (VLM) and visualize the output image features by t-SNE [25]. As shown in Fig. 4, we observe that image features from DeepLab appear disorderly and lack semantics. In contrast, features from VLM share a more unified feature space. Motivated by this, we propose a cross-modality assisted alignment that uses VLM to help align the feature space. Specifically, the camera images $V^{s}$ are fed to the frozen image encoder from VLM to obtain image features $F_{v} = \{\mathcal{F}_{v}^{1},\mathcal{F}_{v}^{2},\dots,\mathcal{F}_{v}^{s}\}$ , where $\mathcal{F}_v^s\in \mathbb{R}^{c\times h\times w}$ . The LiDAR point clouds $P^{s}$ , on the other hand, are fed to the point encoder followed by a projection layer to generate the point features $F_{p} = \{\mathcal{F}_{p}^{1},\mathcal{F}_{p}^{2},\dots,\mathcal{F}_{p}^{s}\}$ , where $\mathcal{F}_p^s\in \mathbb{R}^{m\times c}$ ; $m$ denotes the number of non-empty grids. We then leverage the paired image features $\hat{F}_v\in \mathbb{R}^{m_p\times c}$ and point feature $\hat{F}_p\in \mathbb{R}^{m_p\times c}$ for alignment, where $m_{p}$ is the number of point-pixel pairs. After obtaining $\hat{F}_p$ and $\hat{F}_v$ , the cross-modality alignment is expressed as follows: + +$$ +\mathcal {L} _ {\mathrm {c m a}} \left(\hat {F} _ {v}, \hat {F} _ {p}\right) = 1 - \frac {\hat {F} _ {v} \cdot \hat {F} _ {p}}{\| \hat {F} _ {v} \| \cdot \| \hat {F} _ {p} \|}. \tag {2} +$$ + +Domain-Aware Cross-Modality Alignment. With cross-modality alignment, we transfer the knowledge of VLM to the point encoder, enabling the point features to gain a more comprehensive representation. However, during the execution of the above alignment, we have narrowed + +![](images/310b97e835a536564a0cd79f100b852668275897818711eb5272fc58eddeae88.jpg) +Figure 4. The t-SNE plots of learned features before and after the feature-space alignment in merging the nuScenes [●], SemanticKITTI [●], and Waymo Open [●] datasets. We show image features from (a) standalone networks; (b) SAM [44], and point cloud features (c) before and (d) after the feature-space alignment. + +![](images/b42b377a5d75e26c03b1a296f884ac1428fe40fec203eb95355c8c5ba01c8f72.jpg) +Figure 5. Feature-space alignment in M3Net. We leverage both image features $F_{v}$ and LiDAR point cloud features $F_{p}$ extracted from image encoder $E_{img}$ and point encoder $E_{pcd}$ to employ the regularization via V2P loss and achieve feature-space alignment. + +it exclusively to image and point features from the same dataset. In this mode, point features solely learn from matching image features, restricting their knowledge acquisition. Ideally, we aim to ensure that image features encompass not only scenes identical to those represented in point clouds but also scenes from other datasets. To address this, we propose a domain-aware cross-modality guided alignment, as illustrated in Fig. 5. Specifically, we first extract, for each dataset, $F_{v}$ and $F_{p}$ from the same image encoder $E_{img}$ and point encoder $E_{pcd}$ during the cross + +modality assisted alignment. The sets of features from all datasets are concatenated along the channel dimension to form $\widetilde{F}_v\in \mathbb{R}^{c_v\times h\times w}$ . Subsequently, we sequentially feed $\widetilde{F}_v$ through a branch that consists of a global average pooling and an MLP. Simultaneously, $\widetilde{F}_v$ is fed to an auxiliary branch that undergoes the same processing flow and generates an output after the softmax function $\mathcal{G}(\cdot)$ . The outputs from both branches are multiplied to obtain $F_{m}\in \mathbb{R}^{c_{v}\times 1\times 1}$ . The overall process can be described as follows: + +$$ +F _ {m} = M L P \left(P o o l \left(\widetilde {F} _ {v}\right)\right) \cdot \mathcal {G} \left(M L P \left(P o o l \left(\widetilde {F} _ {v}\right)\right)\right). \tag {3} +$$ + +Next, we forward $F_{m}$ to a sigmoid activation function $\mathcal{H}(\cdot)$ and multiply it with input image features $\widetilde{F}_v$ . The resulting output is added to $\widetilde{F}_v$ and passed through the MLP layers to obtain the final image features $F_{vf} \in \mathbb{R}^{c \times h \times w}$ . The forward process of this operation is depicted as follows: + +$$ +F _ {v f} = M L P \left(\left(\mathcal {H} \left(F _ {m}\right) \cdot \widetilde {F} _ {v}\right) + \widetilde {F} _ {v}\right). \tag {4} +$$ + +Finally, we leverage the cross-modality data alignment to acquire paired image features $\hat{F}_{vf} \in \mathbb{R}^{m_p \times c}$ and paired point feature $\hat{F}_p$ . The overall objective function is: + +$$ +\mathcal {L} _ {v 2 p} \left(\hat {F} _ {v f}, \hat {F} _ {p}\right) = 1 - \frac {\hat {F} _ {v f} \cdot \hat {F} _ {p}}{\| \hat {F} _ {v f} \| \cdot \| \hat {F} _ {p} \|}. \tag {5} +$$ + +# 3.4. Label-Space Alignment + +Label Conflict. In multi-dataset joint training settings, label conflicts emerge as a significant challenge. This often refers to the inconsistencies in class labels across different datasets involved in the training process. The discrepancy can arise due to variations in annotation conventions, labeling errors, or even differences in the underlying semantics of classes between datasets. In our baseline, we unionize the different label spaces across datasets into $\mathbb{Y}^u$ , where all datasets share a single LiDAR segmentation head. However, this may introduce several potential drawbacks: + +- Loss of granularity: Unified label spaces could lose semantic granularity, particularly when dealing with subtle category differences in between different datasets. +- Information loss: During label space consolidation, details unique to each dataset may be obscured or lost, especially for those related to domain-specific categories. +- Increased complexity: Handling a unified label space may necessitate more complex model architectures or training strategies, thereby increasing overall complexity. + +To address these issues, we introduce a language-guided label-space alignment to facilitate a more holistic semantic correlation across datasets. Given the natural correspondence between images and texts and the strong correlation between images and point clouds, we aim to strategically utilize the image modality as a bridge to establish language-guided alignments. Such a process consists of a text-driven + +![](images/284d94ef22a16ef0fb1f4d2d8981c135be38ff15bb7ed5154900f94613a35560.jpg) +Figure 6. Label-space alignment in M3Net. We leverage image features $F_{v}$ , point cloud features $F_{p}$ , and text embedding $F_{t}$ extracted from $E_{img}$ , $E_{pcd}$ , and $E_{txt}$ , respectively, for regularization via the I2P, P2T, and V2T losses in the label-space alignment. + +point alignment, a text-driven image alignment, and a cross-modality-assisted label alignment. + +Text-Driven Alignments. As depicted in Fig. 6, images $V^s$ are fed into the frozen image encoder $E_{img}$ to extract the image features $F_v$ . Concurrently, the LiDAR point clouds $P^s$ are processed by the point encoder $E_{pcd}$ to generate the point features $F_p$ . Additionally, given the text input $T^s$ , text embedding features $F_t \in \mathbb{R}^{Q \times c}$ are obtained from a frozen text encoder $E_{txt}$ , where $Q$ represents the number of categories across datasets. The text is composed of class names from unified label space $\mathbb{Y}_u$ placed into pre-defined templates, and the text embedding captures semantic information of the corresponding classes. Subsequently, pixel-text pairs $\{v_k, t_k\}_{k=1}^M$ and point-text pairs $\{p_k, t_k\}_{k=1}^M$ are generated, where $M$ represents the number of pairs. Leveraging the semantic information contained in the text, we selectively choose positive and negative samples for both images and points for contrastive learning. It is noteworthy that negative samples are confined to the specific dataset category space. The overall objective of the text-driven point alignment function is shown as follows: + +$$ +\mathcal {L} _ {p 2 t} = - \sum_ {q = 1} ^ {Q} \log \left(\frac {\sum_ {t _ {k} \in q , p _ {k}} \exp \left(< t _ {k} , p _ {k} > / \tau\right)}{\sum_ {t _ {k} \in q , t _ {k} \notin q , p _ {j}} \exp \left(< t _ {k} , p _ {k} > / \tau\right)}\right), \tag {6} +$$ + +where $t_k \in q$ indicates that $t_k$ is generated by the $q$ -th classes name, and $Q$ is the number of classes. Symbol $<, >$ denotes the scalar product operation and $\tau$ is a temperature term $(\tau > 0)$ . Similarly, the objective of the text-driven image alignment function is illustrated as follows: + +$$ +\mathcal {L} _ {v 2 t} = - \sum_ {q = 1} ^ {Q} \log \left(\frac {\sum_ {t _ {k} \in q , v _ {k}} \exp \left(< t _ {k} , v _ {k} > / \tau\right)}{\sum_ {t _ {k} \in q , t _ {k} \notin q , v _ {j}} \exp \left(< t _ {k} , v _ {k} > / \tau\right)}\right). \tag {7} +$$ + +Cross-Modality-Assisted Label Alignment. After text-driven alignments, the subsequent crucial step entails aligning the point and image modalities within the label space. We first obtain image logits $F_{vl} = \{\mathcal{F}_{vl}^{1},\mathcal{F}_{vl}^{2},\dots,\mathcal{F}_{vl}^{s}\}$ and point logits $F_{pl} = \{\mathcal{F}_{pl}^{1},\mathcal{F}_{pl}^{2},\dots,\mathcal{F}_{pl}^{s}\}$ from text-driven alignments, where $\mathcal{F}_{vl}^{s}\in \mathbb{R}^{Q\times H\times W}$ , $\mathcal{F}_{pl}^{s}\in \mathbb{R}^{N\times Q}$ . Subsequently, we conduct cross-modality alignment to obtain paired image logits $\hat{F}_{vl}\in \mathbb{R}^{m_p\times Q}$ and paired point logits $\hat{F}_{pl}\in \mathbb{R}^{m_p\times Q}$ . Formally, the cross-modality-assisted alignment in the label space is formulated as follows: + +$$ +\mathcal {L} _ {i 2 p} \left(\hat {F} _ {v l}, \hat {F} _ {p l}\right) = 1 - \frac {\hat {F} _ {v l} \cdot \hat {F} _ {p l}}{\| \hat {F} _ {v l} \| \cdot \| \hat {F} _ {p l} \|}. \tag {8} +$$ + +Finally, the complete objective function for the language-guided label-space alignment is expressed as follows: + +$$ +\mathcal {L} _ {\text {l a b e l}} = \mathcal {L} _ {p 2 t} + \mathcal {L} _ {i 2 p} + \mathcal {L} _ {v 2 t}. \tag {9} +$$ + +# 3.5. Universal LiDAR Segmentation + +We enhance the versatility of M3Net via multi-tasking learning. This integration involves an instance extractor to enable joint semantic and panoptic LiDAR segmentation. + +Panoptic LiDAR Segmentation. Motivated by DSNet [32, 33], our instance extractor comprises an instance head and a clustering step. The instance head encompasses several MLPs designed to predict the offsets between instance centers. The clustering step uses semantic predictions to filter out stuff points, thereby retaining only those associated with thing points. The remaining points undergo a mean-shift clustering [21], utilizing features from the instance head to discern distinct instances. Lastly, we employ the L1 loss $\mathcal{L}_{l1}$ to optimize the thing point regression process. + +Overall Objectives. Putting everything together, the overall objective of M3Net is to minimize the following losses: + +$$ +\mathcal {L} = \mathcal {L} _ {v 2 p} + \mathcal {L} _ {\text {l a b e l}} + \mathcal {L} _ {c e} + \mathcal {L} _ {\text {l o v a s z}} + \mathcal {L} _ {l 1}, \tag {10} +$$ + +where $\mathcal{L}_{ce}$ and $\mathcal{L}_{lovasz}$ denote the cross-entropy loss and the Lovasz-softmax [5] loss, respectively. + +# 4. Experiments + +# 4.1. Experimental Setups + +Datasets. Our M3Net framework and baselines are trained on a combination of nuScenes [27], SemanticKITTI [3], and Waymo Open [90]. Meanwhile, we resort to another five LiDAR-based perception datasets [41, 45, 77, 106, 107] and two 3D robustness evaluation datasets [47] to verify the strong generalizability of M3Net. Due to space limits, additional details regarding the datasets are in the Appendix. + +Implementation Details. M3Net is implemented based on Pointcept [23] and MMDetection3D [22]. We use two backbones in our experiments, i.e., MinkUNet [20] and + +Table 1. Ablation study on the M3Net alignments happen in the Data, Feature, and Label spaces, respectively, when combining the SemanticKITTI [3], nuScenes [27], and Waymo Open [90] datasets. The mAcc and mIoU scores are in percentage. Best scores are in bold. + +
-ConfigurationsMinkUNet [20]PTv2+ [103]
SemKITTInuScenesWaymoSemKITTInuScenesWaymo
mAccmIoUmAccmIoUmAccmIoUmAccmIoUmAccmIoUmAccmIoU
BaselineNaïve Joint Training62.4354.0365.0559.8473.7665.3967.9661.5976.5369.6575.6867.00
M3Net (Ours)DataFeatureLabel------
✓73.8269.0183.6676.8977.8869.3778.5569.9586.2279.1380.9672.15
✓✓74.3669.6485.1778.8878.3169.7079.4370.8787.1080.2680.7472.33
✓✓73.8569.3485.2078.9078.0469.5580.3071.1387.4480.4580.6972.30
✓✓✓74.4069.8585.3079.0078.6670.1580.0072.0087.9180.9081.1172.40
+ +Table 2. Panoptic LiDAR segmentation results on the val sets of the Panoptic-SemanticKITTI [3] and Panoptic-nuScenes [27] datasets. All scores are given in percentage. The best and second-best scores are highlighted in bold and underline, respectively. + +
MethodConfigurationsPanoptic-SemanticKITTIPanoptic-nuScenes
PQ\( PQ^† \)RQSQmIoUPQ\( PQ^† \)RQSQmIoU
Panoptic-TrackNet [37]40.0-48.373.053.851.456.263.380.258.0
Panoptic-PolarNet [127]59.164.170.278.364.563.467.275.383.966.9
EfficientLPS [88]Single-Dataset Training59.265.169.875.064.959.262.882.970.769.4
DSNet [32]61.465.272.779.069.664.767.676.183.576.3
Panoptic-PHNet [55]61.7---65.774.777.784.288.279.7
BaselineNaïve Joint Training56.0359.6465.7873.7261.5956.6760.6166.7583.4969.65
M3Net (Ours)DataFeatureLabel--
✓62.3465.1772.6074.6769.9568.4971.1179.1385.4979.13
✓✓62.9165.7373.3275.4770.8771.4773.8681.5386.7180.26
✓✓63.2367.8973.6181.6671.1371.5373.9181.8086.9280.45
✓✓✓63.8768.6673.1082.3572.0071.7074.0182.2086.4780.90
+ +PTv2+ [103]. We trained M3Net on four A100 GPUs for 50 epochs with a batch size of 6 for each GPU. The initial learning rate is set to 0.002. We adopt the AdamW optimizer [68] with a weight decay of 0.005 and cosine decay learning rate scheduler. For the dataset-specific rasterization, we set voxel sizes to $0.05\mathrm{m}$ , $0.1\mathrm{m}$ , and $0.05\mathrm{m}$ for SemanticKITTI [3], nuScenes [27], and Waymo Open [90], respectively. For the data augmentation, we employ random flipping, jittering, scaling, rotation, and Mix3D [75]. Due to space limits, kindly refer to Appendix for additional details. Evaluation Metrics. We adopt conventional reportings of mAcc and mIoU for LiDAR semantic segmentation, $PQ$ , $PQ^{\dagger}$ , $SQ$ , and $RQ$ for panoptic segmentation, and mCE and mRR for 3D robustness evaluation. Due to the space limit, kindly refer to our Appendix for more detailed definitions. + +# 4.2. Ablation Study + +Multi-Space Alignments. The effectiveness of three proposed alignments over the joint training baselines is shown in Tab. 1. We observe that the data-space alignment plays the most crucial role in improving the universal LiDAR segmentation performance. Without proper data alignments, joint training with either MinkUNet [20] or the stronger PTv2+ [103] will suffer severe degradation, especially on sparser point clouds [27]. On top of the data-space align- + +ment, the combinations of multi-view images at the feature space and the language-guided knowledge at the label space further enhance the learned feature representations. The results show that they work synergistically in merging knowledge from heterogeneous domains during joint training. + +Panoptic LiDAR Segmentation. In Tab. 2, we present another ablation study focusing on panoptic LiDAR segmentation. All three alignments incorporated in M3Net demonstrate significant improvements over the baselines. This highlights the pronounced efficacy of our multi-space alignments. Moreover, our approach outperforms the single-dataset state-of-the-art method Panoptic-PHNet [55] by a notable $2.17\%$ PQ on Panoptic-SemanticKITTI [3] and achieves compelling results on Panoptic-nuScenes [27]. + +Visual Feature Alignments. We conduct a qualitative analysis of the learned visual feature distributions in the form of t-SNE [25]. Fig. 4 (a) and (b) represent the distributions of learned visual features among three datasets from DeepLab and VLM backbones, respectively. The features obtained by the latter exhibit more distinct semantics in feature space. The concentrated distribution space is advantageous for achieving feature alignments across multiple datasets. Additionally, Fig. 4 (c) and (d) illustrate the distribution of point cloud features before and after feature-space alignment. As can be seen, the feature distribution distances + +Table 3. Knowledge transfer and generalization analyses across five LiDAR segmentation datasets and two 3D robustness evaluation datasets. All scores are given in percentage. The best and second-best scores are highlighted in bold and underline, respectively. + +
MethodRELLIS-3DSemanticPOSSSemanticSTFSynLiDARDAPS-3DSemKITTI-CnuScenes-C
1%10%HalfFullHalfFull1%10%HalfFullmCEmRRmCEmRR
PPKT [66]49.7154.3350.1856.0050.9254.6937.5746.4878.9084.00--105.6476.06
SLidR [86]49.7554.5751.5655.3652.0154.3542.0547.8481.0085.40--106.0875.99
Seal [65]51.0955.0353.2656.8953.4655.3643.5849.2681.8885.90--92.6383.08
Naïve Joint37.7750.2342.1952.3146.7048.0018.5642.3773.9177.89113.6584.73128.9781.45
Single-Dataset40.1754.2547.6955.0050.3351.1923.1745.0875.1080.8795.1184.9599.6379.06
M3Net (Ours)51.2755.0553.6057.1753.7855.4244.1049.9382.0886.0086.4385.7791.0379.15
+ +Table 4. LiDAR semantic segmentation results on the val and test sets of SemanticKITTI [3] and nuScenes [27], and the val set of Waymo Open [90]. All scores are in percentage. The best and second-best scores are highlighted in bold and underline. + +
MethodSemKITTInuScenesWaymo
ValTestValTestmIoUmAcc
RangeNet++ [74]-52.2-65.5--
PolarNet [121]57.254.371.069.8--
SalsaNext [24]-59.5-72.2--
RangeViT [1]60.764.075.2---
MinkUNet [20]63.863.773.3-65.976.6
SPVNAS [91]64.766.4-77.467.4-
AMVNet [62]65.265.377.277.3--
RPVNet [112]65.570.377.6---
(AF)2-S3Net [18]-69.7-78.3--
Cylinder3D [128]65.967.876.177.966.0-
PVKD [34]66.471.276.0---
WaffleIron [80]66.870.879.1---
RangeFormer [46]67.673.378.180.1--
SphereFormer [52]67.874.878.481.969.9-
FRNet [114]68.773.379.082.5--
PTv2+ [103]70.370.680.282.670.680.2
LidarMultiNet [117]---81.473.8-
M3Net (Ours)72.075.180.983.172.481.1
+ +between the three datasets have been largely reduced, providing evidence of the alignment effectiveness. + +# 4.3. Comparative Study + +Comparisons to State of the Arts. In Tab. 4, we compare M3Net with current best-performing models on the benchmarks of SemanticKITTI [3], nuScenes [27], and Waymo Open [90]. Remarkably, M3Net consistently outperforms existing approaches across all three datasets. Specifically, on SemanticKITTI [3], M3Net achieves a $72.0\%$ mIoU on the validation set, surpassing the closest method by a notable margin of $1.7\%$ mIoU. Similarly, on nuScenes [27], M3Net achieves $80.9\%$ mIoU and $83.1\%$ mIoU on the validation and test sets, demonstrating its robustness and generalization capabilities. Additionally, the performance of M3Net on Waymo Open [90] is competitive with prior arts. We achieve a mIoU of $72.4\%$ and a mAcc of $81.1\%$ . These results highlight again the superiority of M3Net in handling complex diverse LiDAR segmentation tasks. + +Direct Knowledge Transfer. To further validate the strong + +knowledge transfer capability of M3Net, we conduct extensive experiments on five different LiDAR-based perception datasets [41, 45, 77, 106, 107]. These datasets have unique data collection protocols and data distributions. As shown in Fig. 1 and the first ten columns in Tab. 3, our framework constantly outperforms the prior arts, the naive joint training, and the single-dataset baselines across all five datasets. This concretely supports the strong knowledge transfer efficacy brought by multi-space alignments in M3Net. + +Out-of-Distribution Generalization. Evaluating the generalization ability of models on out-of-training-distribution data is crucial, particularly in safety-critical fields like autonomous driving [50, 109, 110]. In this context, we resort to the two corruption datasets from the Robo3D [47] benchmark, i.e., SemanticKITTI-C and nuScenes-C, to conduct our assessment. From the last four columns of Tab. 3, we observe that M3Net achieves better results than the naive joint training and other single-dataset approaches, proving the strong generalizability of the learned representations. + +# 5. Conclusion + +In this work, we presented M3Net, a universal framework capable of fulfilling multi-task, multi-dataset, multimodality LiDAR segmentation using a single set of parameters. Through extensive analyses, we validated the effectiveness of applying data-, feature-, and label-space alignments to handle such a challenging task. In addition, our comprehensive analysis and discourse have delved into the fundamental challenges of acquiring the general knowledge for scalable 3D perception, which holds substantial potential to propel further research in this domain. Our future strides focus on combining more data resources to further enhance the alignments and adaptations in our framework. + +Acknowledgements. This work was partially supported by NSFC (No.62206173) and MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University). This work was also supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Angelika Ando, Spyros Gidaris, Andrei Bursuc, Gilles Puy, Alexandre Boulch, and Renaud Marlet. Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5240-5250, 2023. +[2] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312, 2022. +[3] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. +[4] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Jürgen Gall, and Cyrill Stachniss. Towards 3d lidar-based semantic scene understanding of 3d point cloud sequences: The semantickitti dataset. International Journal of Robotics Research, 40:959-96, 2021. +[5] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: a tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4413-4421, 2018. +[6] Alexandre Boulch, Corentin Sautier, Björn Michele, Gilles Puy, and Renaud Marlet. Also: Automotive lidar self-supervision by occupancy estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13455-13465, 2023. +[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. +[9] Jun Cen, Shiwei Zhang, Yixuan Pei, Kun Li, Hang Zheng, Maochun Luo, Yingya Zhang, and Qifeng Chen. Cmdfusion: Bidirectional fusion network with cross-modality knowledge distillation for lidar semantic segmentation. arXiv preprint arXiv:2307.04091, 2023. +[10] Runnan Chen, Youquan Liu, Lingdong Kong, Nenglun Chen, Xinge Zhu, Yuexin Ma, Tongliang Liu, and Wenping Wang. Towards label-free scene understanding by vision foundation models. In Advances in Neural Information Processing Systems, 2023. +[11] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023. + +[12] Runnan Chen, Xinge Zhu, Nenglun Chen, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Bridging language and geometric primitives for zero-shot point cloud segmentation. In ACM International Conference on Multimedia, pages 5380-5388, 2023. +[13] Tian Chen, Shijie An, Yuan Zhang, Chongyang Ma, Huayan Wang, Xiaoyan Guo, and Wen Zheng. Improving monocular depth estimation by leveraging structural awareness and complementary datasets. In European Conference on Computer Vision, pages 90-108, 2020. +[14] Yanbei Chen, Manchen Wang, Abhay Mittal, Zhenlin Xu, Paolo Favaro, Joseph Tighe, and Davide Modolo. Scaledet: A scalable multi-dataset object detector. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7288-7297, 2023. +[15] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. In Advances in Neural Information Processing Systems, pages 17864-17875, 2021. +[16] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. +[17] Huixian Cheng, Xianfeng Han, and Guoqiang Xiao. Cenet: Toward concise and efficient lidar semantic segmentation for autonomous driving. In IEEE International Conference on Multimedia and Expo, pages 1-6, 2022. +[18] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. Af2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021. +[19] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. +[20] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. + +[21] Dorin Comaniciu and Peter Meer. Mean shift: A robust approach toward feature space analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(5):603-619, 2005. +[22] MMDetection3D Contributors. MMDetection3D: Open-MMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020. +[23] Pointcept Contributors. Pointcept: A codebase for point cloud perception research. https://github.com/Pointcept/Pointcept, 2023. +[24] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In International Symposium on Visual Computing, pages 207-222, 2020. +[25] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (11), 2008. +[26] Christoph Feichtenhofer, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In Advances in Neural Information Processing Systems, 2022. +[27] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters, 7:3795-3802, 2022. +[28] Biao Gao, Yancheng Pan, Chengkun Li, Sibo Geng, and Huijing Zhao. Are we hungry for 3d lidar data for semantic segmentation? a survey of datasets and methods. IEEE Transactions on Intelligent Transportation Systems, 2021. +[29] Xiuye Gu, Yin Cui, Jonathan Huang, Abdullah Rashwan, Xuan Yang, Xingyi Zhou, Golnaz Ghiasi, Weicheng Kuo, Huizhong Chen, Liang-Chieh Chen, and David A. Ross. Dataseg: Taming a universal multi-dataset multi-task segmentation model. arXiv preprint arXiv:2306.01736, 2023. +[30] Yi Gu, Yuming Huang, Chengzhong Xu, and Hui Kong. Maskrange: A mask-classification model for range-view based lidar segmentation. arXiv preprint arXiv:2206.12073, 2022. +[31] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In IEEE/CVF International Conference on Computer Vision, pages 2961-2969, 2017. +[32] Fangzhou Hong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Lidar-based panoptic segmentation via dynamic shifting network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13090-13099, 2021. +[33] Fangzhou Hong, Lingdong Kong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Unified 3d and 4d panoptic segmentation via dynamic shifting networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2024. +[34] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-voxel knowledge distillation for lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8479–8488, 2022. +[35] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. + +Randla-net: Efficient semantic segmentation of large-scale point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11108-11117, 2020. +[36] Qingyong Hu, Bo Yang, Sheikh Khalid, Wen Xiao, Niki Trigoni, and Andrew Markham. Towards semantic segmentation of urban-scale 3d point clouds: A dataset, benchmarks and challenges. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4977-4987, 2021. +[37] Juana Valeria Hurtado, Rohit Mohan, Wolfram Burgard, and Abhinav Valada. Mopt: Multi-object panoptic tracking. arXiv preprint arXiv:2004.08189, 2020. +[38] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. +[39] Maximilian Jaritz, Tuan-Hung Vu, Raoul de Charette, Emilie Wirbel, and Patrick Pérez. xmuda: Cross-modal unsupervised domain adaptation for 3d semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12605-12614, 2020. +[40] Maximilian Jaritz, Tuan-Hung Vu, Raoul De Charette, Emilie Wirbel, and Patrick Pérez. Cross-modal learning for domain adaptation in 3d semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1533-1544, 2023. +[41] Peng Jiang, Philip Osteen, Maggie Wigness, and Srikanth Saripallig. Rellis-3d dataset: Data, benchmarks and analysis. In IEEE International Conference on Robotics and Automation, pages 1110–1116, 2021. +[42] Tarun Kalluri, Girish Varma, Manmohan Chandraker, and C. V. Jawahar. Universal semi-supervised semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 5259-5270, 2019. +[43] Dongwan Kim, Yi-Hsuan Tsai, Yumin Suh, Masoud Faraki, Sparsh Garg, Manmohan Chandraker, and Bohyung Han. Learning semantic segmentation from multiple datasets with label shifts. In European Conference on Computer Vision, pages 20-36, 2022. +[44] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. +[45] Alexey Klokov, Di Un Pak, Aleksandr Khorin, Dmitry Yudin, Leon Kochiev, Vladimir Luchinskiy, and Vitaly Bezuglyj. Daps3d: Domain adaptive projective segmentation of 3d lidar point clouds. IEEE Access, 11:79341-79356, 2023. +[46] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Zwei Liu. Rethinking range view representation for lidar segmentation. In IEEE/CVF International Conference on Computer Vision, pages 228-240, 2023. +[47] Lingdong Kong, Youquan Liu, Xin Li, Runnan Chen, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei + +Liu. Robo3d: Towards robust and reliable 3d perception against corruptions. In IEEE/CVF International Conference on Computer Vision, pages 19994-20006, 2023. +[48] Lingdong Kong, Niamul Quader, and Venice Erin Liong. Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In IEEE International Conference on Robotics and Automation, pages 9338-9345, 2023. +[49] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21705-21715, 2023. +[50] Lingdong Kong, Shaoyuan Xie, Hanjiang Hu, Lai Xing Ng, Benoit R. Cottereau, and Wei Tsang Ooi. Robodepth: Robust out-of-distribution depth estimation under corruptions. In Advances in Neural Information Processing Systems, 2023. +[51] Lingdong Kong, Xiang Xu, Jun Cen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Calib3d: Calibrating model preferences for reliable 3d scene understanding. arXiv preprint arXiv:2403.17010, 2024. +[52] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17545-17555, 2023. +[53] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. Mseg: A composite dataset for multidomain semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2879-2888, 2020. +[54] Guangrui Li, Guoliang Kang, Xiaohan Wang, Yunchao Wei, and Yi Yang. Adversarily masking synthetic to mimic real: Adaptive noise injection for point cloud segmentation adaptation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20464-20474, 2023. +[55] Jinke Li, Xiao He, Yang Wen, Yuan Gao, Xiaogiang Cheng, and Dan Zhang. Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11809-11818, 2022. +[56] Li Li, Hubert PH Shum, and Toby P. Breckon. Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9361-9371, 2023. +[57] Rong Li, Raoul de Charette, and C. A. O. Anh-Quan. Coarse3d: Class-prototypes for contrastive learning in weakly-supervised 3d point cloud segmentation. In *British Machine Vision Conference*, 2022. +[58] Xin Li, Botian Shi, Yuenan Hou, Xingjiao Wu, Tianlong Ma, Yikang Li, and Liang He. Homogeneous multi-modal feature fusion and interaction for 3d object detection. In European Conference on Computer Vision, pages 691-707. Springer, 2022. +[59] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. + +Video k-net: A simple, strong, and unified baseline for video segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18847-18857, 2022. +[60] Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17524-17534, 2023. +[61] Ye Li, Lingdong Kong, Hanjiang Hu, Xiaohao Xu, and Xiaonan Huang. Optimizing lidar placements for robust driving perception in adverse conditions. arXiv preprint arXiv:2403.17009, 2024. +[62] Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja, Dhananjai Sharma, and Zhuang Jie Chong. Amvnet: Assertion-based multi-view fusion network for lidar semantic segmentation. arXiv preprint arXiv:2012.04934, 2020. +[63] Minghua Liu, Yin Zhou, Charles R. Qi, Boqing Gong, Hao Su, and Dragomir Anguelov. Less: Label-efficient semantic segmentation for lidar point clouds. In European Conference on Computer Vision, pages 70-89, 2022. +[64] Youquan Liu, Runnan Chen, Xin Li, Lingdong Kong, Yuchen Yang, Zhaoyang Xia, Yeqi Bai, Xinge Zhu, Yuexin Ma, Yikang Li, Yu Qiao, and Yuenan Hou. Uniseg: A unified multi-modal lidar segmentation network and the openpcseg codebase. In IEEE/CVF International Conference on Computer Vision, pages 21662-21673, 2023. +[65] Youquan Liu, Lingdong Kong, Jun Cen, Runnan Chen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Segment any point cloud sequences by distilling vision foundation models. In Advances in Neural Information Processing Systems, 2023. +[66] Yueh-Cheng Liu, Yu-Kai Huang, Hung-Yueh Chiang, Hung-Ting Su, Zhe-Yu Liu, Chin-Tang Chen, Ching-Yu Tseng, and Winston H. Hsu. Learning from 2d: Contrastive pixel-to-point knowledge transfer for 3d pretraining. arXiv preprint arXiv:2104.0468, 2021. +[67] Zhijian Liu, Alexander Amini Haotian Tang, Xinyu Yang, Huizi Mao, Daniela L. Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In IEEE International Conference on Robotics and Automation, pages 2774-2781, 2023. +[68] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. +[69] Yuhang Lu, Qi Jiang, Runnan Chen, Yuenan Hou, Xinge Zhu, and Yuexin Ma. See more and know more: Zero-shot point cloud segmentation via multi-modal visual data. In IEEE/CVF International Conference on Computer Vision, pages 21674-21684, 2023. +[70] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A comprehensive survey. International Journal of Computer Vision, 2023. +[71] Rodrigo Marcuzzi, Lucas Nunes, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Mask-based panoptic lidar + +segmentation for autonomous driving. IEEE Robotics and Automation Letters, 8(2):1141-1148, 2023. +[72] Panagiotis Meletis and Gijs Dubbelman. Training of convolutional networks on multiple heterogeneous datasets for street scene semantic segmentation. In IEEE Intelligent Vehicles Symposium, pages 1045-1050, 2018. +[73] Björn Michele, Alexandre Boulch, Gilles Puy, Tuan-Hung Vu, Renaud Marlet, and Nicolas Courty. Saluda: Surface-based automotive lidar unsupervised domain adaptation. arXiv preprint arXiv:2304.03251, 2023. +[74] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4213-4220, 2019. +[75] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3d: Out-of-context data augmentation for 3d scenes. In International Conference on 3D Vision, pages 116-125, 2021. +[76] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. +[77] Yancheng Pan, Biao Gao, Jilin Mei, Sibo Geng, Chengkun Li, and Huijing Zhao. Semanticposs: A point cloud dataset with large quantity of dynamic instances. In IEEE Intelligent Vehicles Symposium, pages 687-693, 2020. +[78] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. Openscene: 3d scene understanding with open vocabularies. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-824, 2023. +[79] Xidong Peng, Runnan Chen, Feng Qiao, Lingdong Kong, Youquan Liu, Tai Wang, Xinge Zhu, and Yuexin Ma. Sam-guided unsupervised domain adaptation for 3d segmentation. arXiv preprint arXiv:2310.08820, 2023. +[80] Gilles Puy, Alexandre Boulch, and Renaud Marlet. Using a waffle iron for automotive point cloud semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 3379-3389, 2023. +[81] 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 International Conference on Machine Learning, pages 8748-8763, 2021. +[82] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1623-1637, 2020. +[83] Giulia Rizzoli, Francesco Barbato, and Pietro Zanuttigh. Multimodal semantic segmentation in autonomous driving: + +A review of current approaches and future perspectives. Technologies, 10(4), 2022. +[84] Cristiano Saltori, Fabio Galasso, Giuseppe Fiameni, Nicu Sebe, Elisa Ricci, and Fabio Poiesi. Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In European Conference on Computer Vision, pages 586-602, 2022. +[85] Jules Sanchez, Jean-Emmanuel Deschaud, and François Goulette. Cola: Coarse-label multi-source lidar semantic segmentation for autonomous driving. arXiv preprint arXiv:2311.03017, 2023. +[86] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022. +[87] Alvari Seppänen, Risto Ojala, and Kari Tammi. 4denoiseNET: Adverse weather denoising from adjacent point clouds. IEEE Robotics and Automation Letters, 8:456-463, 2022. +[88] Kshitij Sirohi, Rohit Mohan, Daniel Buscher, Wolfram Burgard, and Abhinav Valada. Efficientlps: Efficient lidar panoptic segmentation. IEEE Transactions on Robotics, 2021. +[89] Louis Soum-Fontez, Jean-Emmanuel Deschaud, and François Goulette. Mdt3d: Multi-dataset training for lidar 3d object detection generalization. arXiv preprint arXiv:2308.01000, 2023. +[90] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2446-2454, 2020. +[91] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, pages 685–702, 2020. +[92] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In IEEE/CVF International Conference on Computer Vision, pages 6411–6420, 2019. +[93] Larissa T Triess, David Peter, Christoph B Rist, and J Marius Zöllner. Scan-based semantic segmentation of lidar point clouds: An experimental study. In IEEE Intelligent Vehicles Symposium, pages 1116-1121, 2020. +[94] Larissa T. Triess, Mariella Dreissig, Christoph B. Rist, and J. Marius Zöllner. A survey on deep domain adaptation for lidar perception. In IEEE Intelligent Vehicles Symposium Workshop, pages 350-357, 2021. +[95] Darren Tsai, Julie Stephany Berrio, Mao Shan, Eduardo Nebot, and Stewart Worrall. Ms3d: Leveraging multiple + +detectors for unsupervised domain adaptation in 3d object detection. arXiv preprint arXiv:2304.02431, 2023. +[96] Darren Tsai, Julie Stephany Berrio, Mao Shan, Eduardo Nebot, and Stewart Worrall. Ms3d++: Ensemble of experts for multi-source unsupervised domain adaption in 3d object detection. arXiv preprint arXiv:2308.05988, 2023. +[97] Ozan Unal, Dengxin Dai, and Luc Van Gool. Scribblesupervised lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2697-2707, 2022. +[98] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021. +[99] Xudong Wang, Zhaowei Cai, Dashan Gao, and Nuno Vasconcelos. Towards universal object detection by domain attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7289-7298, 2019. +[100] Xudong Wang, Shufan Li, Konstantinos Kallidromitis, Yusuke Kato, Kazuki Kozuka, and Trevor Darrell. Hierarchical open-vocabulary universal image segmentation. In Advances in Neural Information Processing Systems, 2023. +[101] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. +[102] Yi Wei, Zibu Wei, Yongming Rao, Jiaxin Li, Jie Zhou, and Jiwen Lu. Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In European Conference on Computer Vision, pages 179-195. Springer, 2022. +[103] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In Advances in Neural Information Processing Systems, 2022. +[104] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards large-scale 3d representation learning with multi-dataset point prompt training. arXiv preprint arXiv:2308.09718, 2023. +[105] Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, and Ling Shao. Polarmix: A general data augmentation technique for lidar point clouds. In Advances in Neural Information Processing Systems, pages 11035-11048, 2022. +[106] Aoran Xiao, Jiaxing Huang, Dayan Guan, Fangneng Zhan, and Shijian Lu. Transfer learning from synthetic to real lidar point cloud for semantic segmentation. In AAAI Conference on Artificial Intelligence, pages 2795-2803, 2022. +[107] Aoran Xiao, Jiaxing Huang, Weihao Xuan, Ruijie Ren, Kangcheng Liu, Dayan Guan, Abdulmotaleb El Saddik, Shijian Lu, and Eric Xing. 3d semantic segmentation in the wild: Learning generalized models for adverse-condition point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9382-9392, 2023. +[108] Zeqi Xiao, Wenwei Zhang, Tai Wang, Chen Change Loy, Dahua Lin, and Jiangmiao Pang. Position-guided point cloud panoptic segmentation transformer. arXiv preprint arXiv:2303.13509, 2023. + +[109] Shaoyuan Xie, Lingdong Kong, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Robobev: Towards robust bird's eye view perception under corruptions. arXiv preprint arXiv:2304.06719, 2023. +[110] Shaoyuan Xie, Lingdong Kong, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Benchmarking and analyzing bird's eye view perception robustness to corruptions. Preprint, 2023. +[111] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In European Conference on Computer Vision, pages 1-19, 2020. +[112] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In IEEE/CVF International Conference on Computer Vision, pages 16024-16033, 2021. +[113] Jingyi Xu, Weidong Yang, Lingdong Kong, Youquan Liu, Rui Zhang, Qingyuan Zhou, and Ben Fei. Visual foundation models boost cross-modal unsupervised domain adaptation for 3d semantic segmentation. arXiv preprint arXiv:2403.10001, 2024. +[114] Xiang Xu, Lingdong Kong, Hui Shuai, and Qingshan Liu. Frnet: Frustum-range networks for scalable lidar segmentation. arXiv preprint arXiv:2312.04484, 2023. +[115] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European Conference on Computer Vision, pages 677-695, 2022. +[116] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d: Self-training for unsupervised domain adaptation on 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10368-10378, 2021. +[117] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: Towards a unified multi-task network for lidar perception. In AAAI Conference on Artificial Intelligence, pages 3231-3240, 2023. +[118] Bo Zhang, Jiakang Yuan, Botian Shi, Tao Chen, Yikang Li, and Yu Qiao. Uni3d: A unified baseline for multi-dataset 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9253-9262, 2023. +[119] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianfeng Gao, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. arXiv preprint arXiv:2303.08131, 2023. +[120] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. In Advances in Neural Information Processing Systems, pages 10326-10338, 2021. +[121] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In IEEE/CVF Conference on Com + +puter Vision and Pattern Recognition, pages 9601-9610, 2020. +[122] Zihui Zhang, Bo Yang, Bing Wang, and Bo Li. Growsp: Unsupervised semantic segmentation of 3d point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17619-17629, 2023. +[123] Yiming Zhao, Lin Bai, and Xinming Huang. Fidnet: Lidar point cloud semantic segmentation with fully interpolation decoding. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4453-4458, 2021. +[124] Qiang Zhou, Yuang Liu, Chaohui Yu, Jingliang Li, Zhibin Wang, and Fan Wang. Lmseg: Language-guided multi-dataset segmentation. In International Conference on Learning Representations, 2022. +[125] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368, 2022. +[126] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7571-7580, 2022. +[127] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-polararnet: Proposal-free lidar point cloud panoptic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13194-13203, 2021. +[128] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021. +[129] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multisensor fusion for 3d lidar semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 16280-16290, 2021. +[130] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. In Advances in Neural Information Processing Systems, 2023. \ No newline at end of file diff --git a/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/images.zip b/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..74bd8b95f8feb1873613085d8f0d392c2989d3e3 --- /dev/null +++ b/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c230ac3fd796de3163df4265698d675fb45813eb009852ea6a3eae73661cfbb7 +size 644770 diff --git a/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/layout.json b/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fe90dd852589f4211f801d9cfc252b39a80a504f --- /dev/null +++ b/2024/Multi-Space Alignments Towards Universal LiDAR Segmentation/layout.json @@ -0,0 +1,11994 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 501, + 121 + ], + "type": "text", + "content": "Multi-Space Alignments Towards Universal LiDAR Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 142, + 505, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 142, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 97, + 142, + 505, + 171 + ], + "type": "text", + "content": "Youquan Liu\\*,1 Lingdong Kong\\*,2,3 Xiaoyang Wu4 Runnan Chen4 Xin Li5 Liang Pan2 Ziwei Liu6 Yuexin Ma1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 171, + 549, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 171, + 549, + 213 + ], + "spans": [ + { + "bbox": [ + 54, + 171, + 549, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 54, + 171, + 549, + 213 + ], + "type": "text", + "content": "ShanghaiTech University, " + }, + { + "bbox": [ + 54, + 171, + 549, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 54, + 171, + 549, + 213 + ], + "type": "text", + "content": "Shanghai AI Laboratory, " + }, + { + "bbox": [ + 54, + 171, + 549, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 54, + 171, + 549, + 213 + ], + "type": "text", + "content": "National University of Singapore " + }, + { + "bbox": [ + 54, + 171, + 549, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 54, + 171, + 549, + 213 + ], + "type": "text", + "content": "University of Hong Kong, " + }, + { + "bbox": [ + 54, + 171, + 549, + 213 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 54, + 171, + 549, + 213 + ], + "type": "text", + "content": "East China Normal University, " + }, + { + "bbox": [ + 54, + 171, + 549, + 213 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 54, + 171, + 549, + 213 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University https://github.com/youquanl/M3Net" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 240, + 192, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 240, + 192, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 240, + 192, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 266, + 290, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 470 + ], + "type": "text", + "content": "A unified and versatile LiDAR segmentation model with strong robustness and generalizability is desirable for safe autonomous driving perception. This work presents M3Net, a one-of-a-kind framework for fulfilling multi-task, multi-dataset, multi-modality LiDAR segmentation in a universal manner using just a single set of parameters. To better exploit data volume and diversity, we first combine large-scale driving datasets acquired by different types of sensors from diverse scenes and then conduct alignments in three spaces, namely data, feature, and label spaces, during the training. As a result, M3Net is capable of taming heterogeneous data for training state-of-the-art LiDAR segmentation models. Extensive experiments on twelve LiDAR segmentation datasets verify our effectiveness. Notably, using a shared set of parameters, M3Net achieves " + }, + { + "bbox": [ + 46, + 266, + 290, + 470 + ], + "type": "inline_equation", + "content": "75.1\\%" + }, + { + "bbox": [ + 46, + 266, + 290, + 470 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 266, + 290, + 470 + ], + "type": "inline_equation", + "content": "83.1\\%" + }, + { + "bbox": [ + 46, + 266, + 290, + 470 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 266, + 290, + 470 + ], + "type": "inline_equation", + "content": "72.4\\%" + }, + { + "bbox": [ + 46, + 266, + 290, + 470 + ], + "type": "text", + "content": " mIoU scores, respectively, on the official benchmarks of SemanticKITTI, nuScenes, and Waymo Open." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 494, + 128, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 128, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 128, + 506 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 514, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 586 + ], + "type": "text", + "content": "Dense and structural 3D surrounding scene understanding provides crucial information for autonomous vehicles to make proper decisions [70]. With the recent advancements in sensing technologies, especially the Light Detection and Ranging (LiDAR) sensor, a holistic scene perception can be achieved by segmenting the acquired sensor data [28, 83]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "type": "text", + "content": "Most existing LiDAR segmentation models [1, 36, 111, 121, 128] are trained and tested in a single-task, single-dataset, single-modality manner. Despite achieving commendable results in the single domain, there is a significant performance drop when transitioning to new domains [40, 48]. The limited generalization capability hinders their facilitation of real-world applications [47, 50, 87]. In reality, LiDAR datasets are marred by significant variances, encompassing variations in data patterns due to different" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 238, + 547, + 450 + ], + "blocks": [ + { + "bbox": [ + 307, + 238, + 547, + 450 + ], + "lines": [ + { + "bbox": [ + 307, + 238, + 547, + 450 + ], + "spans": [ + { + "bbox": [ + 307, + 238, + 547, + 450 + ], + "type": "image", + "image_path": "610d391e645481a25a4707f26a4e4b53a8d20eca1327984146383805462fea08.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 456, + 547, + 511 + ], + "lines": [ + { + "bbox": [ + 305, + 456, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 305, + 456, + 547, + 511 + ], + "type": "text", + "content": "Figure 1. Performance comparisons among M3Net [●], Single-Dataset Training [●], and Naïve Joint Training [●] across twelve LiDAR segmentation datasets. For better comparisons, the radius is normalized based on M3Net's scores. The larger the area coverage, the higher the overall performance. Best viewed in colors." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 532, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 652 + ], + "type": "text", + "content": "sensor types and weather conditions, diverse class distributions arising from varying capture scenarios, and distinct label spaces shaped by specific annotation protocols. These factors collectively pose a formidable challenge in harmonizing disparate LiDAR point clouds and jointly optimizing model parameters to effectively address multiple tasks across a range of sensor modalities [89, 118]. Empirical evidence in Fig. 3 further reveals that naively combining heterogeneous data to train a LiDAR segmentation model – without strategic alignments – often leads to sub-opt results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": "Recent works [6, 39, 48, 79, 87, 94, 106] resort to unsupervised domain adaptation (UDA) for utilizing training data from both source and target domains to optimize one parameter set. Nevertheless, they either focus on only the sharing mapping between two domains (by ignoring dis" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 242, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 242, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 242, + 712 + ], + "type": "text", + "content": "* The first two authors contributed equally to this work." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14648" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "joint classes) or directly merge source domain labels to align with the target domain [40, 113]. The overlook of the performance degradation on the source dataset and the destruction of original label mappings inevitably constrains such a learning paradigm. Furthermore, there have been efforts [85, 96, 104, 118] to employ multi-dataset learning strategies to bolster the generalization prowess of 3D perception models. However, they either necessitate dataset-specific fine-tuning, deviating from a truly universal learning approach, or converge label spaces to a coarser set, resulting in the dilution of fine-grained segmentation capabilities across diverse semantic categories." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 467 + ], + "type": "text", + "content": "In this work, we define a novel paradigm towards leveraging LiDAR point clouds from different datasets to tame a single set of parameters for multi-task LiDAR segmentation. Sibling to image segmentation communities [43, 53, 124], we call this paradigm universal LiDAR segmentation. The ultimate goal of such a synergistic way of learning is to build a powerful segmentation model that can absorb rich cross-domain knowledge and, in return, achieve strong resilience and generalizability for practical usage. Given the substantial differences among datasets in terms of data characteristics, feature distributions, and labeling conventions, we introduce a comprehensive multi-space alignment approach that encompasses data-, feature-, and label-level alignments, to effectively pave the path for efficient and universally applicable LiDAR segmentation. In particular, the multi-modal data, including images and texts, is fully exploited to assist the alignment process with the guidance of more general knowledge. Through aforementioned processes, we propose M3Net to learn common knowledge across datasets, modalities, and tasks, thereby significantly enhancing its applicability in practical scenarios." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 467, + 289, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 289, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 289, + 598 + ], + "type": "text", + "content": "To substantiate the efficacy of M3Net and the utility of each module developed, we have carried out a series of thorough comparative and ablation studies across an extensive array of driving datasets, as shown in Fig. 1. Notably, our best model achieves state-of-the-art LiDAR segmentation performance with " + }, + { + "bbox": [ + 46, + 467, + 289, + 598 + ], + "type": "inline_equation", + "content": "75.1\\%" + }, + { + "bbox": [ + 46, + 467, + 289, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 467, + 289, + 598 + ], + "type": "inline_equation", + "content": "83.1\\%" + }, + { + "bbox": [ + 46, + 467, + 289, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 467, + 289, + 598 + ], + "type": "inline_equation", + "content": "72.4\\%" + }, + { + "bbox": [ + 46, + 467, + 289, + 598 + ], + "type": "text", + "content": " mIoU scores on SemanticKITTI [3], nuScenes [27], Waymo Open [90], respectively, using a shared set of parameters. Moreover, our approach also performs well for direct knowledge transfer and out-of-distribution adaptations, further underscoring its robust capability for effective knowledge transfer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 609, + 134, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 134, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 134, + 621 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "LiDAR Segmentation. A holistic perception of 3D scenes is crucial for safe autonomous driving [4, 7, 33, 51, 61]. Various LiDAR segmentation models have been proposed, with distinct focuses on aspects include LiDAR representations [20, 74, 91-93, 103, 121, 128], model architectures [1, 17, 24, 35, 46, 52, 80, 112], sensor fusion [18, 62, 64, 113, 129], post-processing [111, 123], data aug" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": "mentations [75, 84, 105], etc. Most recently, researchers started to explore data efficiency [49, 56], annotation efficiency [57, 63, 65, 86, 97], annotation-free learning [10, 11, 122], zero-shot learning [12, 69], domain adaptation [6, 39, 48, 54, 73, 79, 106], and robustness [47] in LiDAR segmentation, shedding lights for practitioners. Existing pursuits, however, learn separate parameter sets for each dataset, impeding the scalability. This motivates us to explore LiDAR segmentation in a multi-task, multi-dataset, multi-modality manner with just a single set of parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 195, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 546, + 422 + ], + "type": "text", + "content": "Multi-Task Learning. A proper pipeline design could enable the model to generate suitable predictions to fulfill multiple tasks simultaneously [16, 31]. The current research endeavors mainly focus on building image or video segmentation models to handle semantic, instance, and panoptic segmentation tasks [38, 59, 98, 100, 119, 120, 130]. Recently, several attempts have been made to enable multitask segmentation on LiDAR point clouds. MaskRange [30] and MaskPLS [71] extend the mask classification paradigm [15] for joint semantic and panoptic LiDAR segmentation. LidarMultiNet [117] uses global context pooling and task-specific heads to handle LiDAR-based detection and segmentation. P3Former [108] proposed a specialized positional embedding to handle the geometry ambiguity in panoptic LiDAR segmentation. Our framework also supports multi-task learning. Different from existing approaches, the proposed M3Net stands out by combining knowledge from different sensor data across multiple data sources, which achieves superior performance on each task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": "Multi-Dataset Learning. Leveraging data samples from different sources for training has been proven effective in enhancing robustness and generalizability [72]. Various approaches have been proposed to merge image datasets for object detection [14, 58, 60, 99, 125, 126], image segmentation [29, 42, 43, 53, 124], depth estimation [13, 82], etc. Due to large domain gaps, the image-based methods are often hard to be transferred to 3D. To combine multiple Li-DAR datasets for 3D object detection, MDT3D [89] defines a coarse label set to handle the label space conflicts in different point cloud datasets. MS3D++ [95, 96] ensembles pre-trained detectors from different source datasets for multi-domain adaptation. Uni3D [118] resorts to dataset-specific detection heads and feature re-coupling for training a unified 3D object detector. Recently, PPT [104] proposed to pre-train a point cloud segmentation network using data from multiple datasets. However, the pre-trained weights are then fine-tuned on each specific dataset, which breaks the universal learning manner. The closest work to us is COLA [85], which trains a single model across multiple sources by converting dataset-specific labels to a common coarse set. Such a conversion, however, leads to the loss of fine-grained segmentation across the various semantic categories. Differently, our M3Net is tailored to tame a single" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14649" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 123, + 214 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 123, + 214 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 123, + 214 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 123, + 214 + ], + "type": "image", + "image_path": "766b15b8b801135d7eabe8e60a7bdaa7e165b07eadaae8904a1d98a4a475b11e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 546, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 546, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 546, + 243 + ], + "type": "text", + "content": "Figure 2. Statistical analysis of six sharing semantic classes in the nuScenes [●], SemanticKITTI [●], and Waymo Open [●] datasets. Each violin plot shows the class distribution across LiDAR scenes spanning 50 meters, centered around the ego-vehicle. Best viewed in colors." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 132, + 68, + 208, + 214 + ], + "blocks": [ + { + "bbox": [ + 132, + 68, + 208, + 214 + ], + "lines": [ + { + "bbox": [ + 132, + 68, + 208, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 68, + 208, + 214 + ], + "type": "image", + "image_path": "1e97aa12f4e903a23eebbbb9b226519a3af01def186fa0199b251ba872c246ec.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 215, + 68, + 291, + 214 + ], + "blocks": [ + { + "bbox": [ + 215, + 68, + 291, + 214 + ], + "lines": [ + { + "bbox": [ + 215, + 68, + 291, + 214 + ], + "spans": [ + { + "bbox": [ + 215, + 68, + 291, + 214 + ], + "type": "image", + "image_path": "e7e1ea742b504d35b9462e6605d179dc2f0f66e4d05dcf6d627975e1f3bce0f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 301, + 68, + 377, + 214 + ], + "blocks": [ + { + "bbox": [ + 301, + 68, + 377, + 214 + ], + "lines": [ + { + "bbox": [ + 301, + 68, + 377, + 214 + ], + "spans": [ + { + "bbox": [ + 301, + 68, + 377, + 214 + ], + "type": "image", + "image_path": "0290866900cf110d7921e2ad02cc7ebd6f1081777035dd7d72036204f086a3b2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 386, + 68, + 462, + 214 + ], + "blocks": [ + { + "bbox": [ + 386, + 68, + 462, + 214 + ], + "lines": [ + { + "bbox": [ + 386, + 68, + 462, + 214 + ], + "spans": [ + { + "bbox": [ + 386, + 68, + 462, + 214 + ], + "type": "image", + "image_path": "9052bb2739c8011afdc180a8fec8b872ccf983ab9bfdd0b01f6d58ee31a9d630.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 471, + 68, + 547, + 214 + ], + "blocks": [ + { + "bbox": [ + 471, + 68, + 547, + 214 + ], + "lines": [ + { + "bbox": [ + 471, + 68, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 471, + 68, + 547, + 214 + ], + "type": "image", + "image_path": "cfba0f78cfc727b73924884070d78c5dcc2d5d0cc32428cf22407bc12d91bddc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 259, + 289, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 289, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 289, + 440 + ], + "type": "text", + "content": "parameter set to fulfill multi-task prediction across multiple datasets while still maintaining the original label mappings. Multi-Modality Learning. Recent trend favors synergistic learning from data of different modalities, such as vision, language, and speech [2, 8, 19, 26, 76, 81, 101]. For LiDAR segmentation, several works [9, 39, 40, 67, 115] explored the distillation of image features to point clouds. Recently, OpenScene [78] and CLIP2Scene [11] proposed to leverage point clouds along with multi-view images and language for open-vocabulary learning. PPKT [66], SLidR [86], and Seal [65] form cross-sensor contrastive learning objectives to pre-train the LiDAR segmentation models. The advantages of sensor fusion have been consistently proven. In this work, to pursue universal LiDAR segmentation, we propose to align multi-space point clouds via images and texts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 451, + 115, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 451, + 115, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 451, + 115, + 464 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 471, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 590 + ], + "type": "text", + "content": "Our study serves as an early attempt at combining multi-task, multi-dataset, multi-modality knowledge into a single set of parameters to fulfill universal LiDAR segmentation. We start with a pilot study to unveil the difficulties in merging heterogeneous LiDAR point clouds (cf. Sec. 3.1). We then present M3Net, a versatile LiDAR segmentation network tailored to pursue i) statistical consistency in the data space (cf. Sec. 3.2), ii) cross-modality-assisted alignment in the feature space (cf. Sec. 3.3), and iii) language-guided unification in the label space (cf. Sec. 3.4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 598, + 123, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 123, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 123, + 611 + ], + "type": "text", + "content": "3.1. Pilot Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 617, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 689 + ], + "type": "text", + "content": "The current de facto of training a LiDAR segmentation network adopts a task-by-task and dataset-by-dataset pipeline. Despite the superior performance achieved under such standalone settings, the trained parameter sets cannot be shared to satisfy out-of-domain requirements and, therefore, limits their use cases for practical applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Naïve Joint Training. A natural alternative to breaking the above constraint is to jointly train a network across multiple" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 259, + 547, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 547, + 403 + ], + "type": "text", + "content": "datasets for better generalizability. However, as depicted in Fig. 2, it is often non-trivial to naively combine heterogeneous data with large data distribution gaps to train a universal LiDAR segmentation model without proper alignments. To testify this, we conducted a pilot study using the prior art MinkUNet [20] for both standalone and joint training on three large-scale datasets [3, 27, 90]. As shown in Fig. 3 (a) and (d), a brutal combination undermines the segmentation performance. Due to large discrepancies in aspects like sensor configurations, data acquisitions, label mappings, and domain shifts, the jointly trained representations tend to be disruptive instead of being more general." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 404, + 547, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 547, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 547, + 537 + ], + "type": "text", + "content": "LiDAR Sensor Discrepancy. To understand the root cause of performance degradation, we conducted another study that controls point cloud density discrepancies when merging datasets. As shown in Fig. 3 (b) and (c), joint training on data collected by sensors with different beam numbers tends to suffer more severely than merging less density variant data. We hypothesize that this is mainly caused by the data statistical variations. In light of these observations, we propose a bag of suitable operations in the following sections to alleviate the large domain gaps among different LiDAR segmentation datasets [3, 4, 7, 27, 90]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 548, + 435, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 435, + 561 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 435, + 561 + ], + "type": "text", + "content": "3.2. Data-Space Alignment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": "Given a total of " + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": " datasets " + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "D^{s} = \\{(x^{s},y^{s})|1\\leq s\\leq S\\}" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "(x^s,y^s)" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": " denotes the data-label pairs constituting a dataset. For the LiDAR segmentation task, " + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "x^{s}" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": " often encompasses the LiDAR point cloud " + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "P^{s} = \\{p_{x},p_{y},p_{z}\\}^{s}\\in" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": " and synchronized multi-view camera images " + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "V^{s} =" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_1,\\dots,\\mathcal{I}_l\\} |l = 1,\\dots,L\\}" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": " is the number of points, " + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": " denotes the number of camera sensors, " + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": " are the height and width of the image, respectively. " + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "y^{s}\\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": " denotes point cloud labels in the label space " + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "\\mathbb{Y}^s" + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": " , we unify the label space as " + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "inline_equation", + "content": "\\mathbb{Y}^u = \\mathbb{Y}^1\\cup \\mathbb{Y}^2\\dots \\cup \\mathbb{Y}^S" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": "Cross-Modality Data Alignment. As a multi-sensing system, the information encoded in " + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "P_{i}^{s}" + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "V_{i}^{s}" + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": " are intuitively" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14650" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 85, + 149, + 188 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 116, + 80 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 116, + 80 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 116, + 80 + ], + "type": "text", + "content": "Standalone Training" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 85, + 149, + 188 + ], + "lines": [ + { + "bbox": [ + 47, + 85, + 149, + 188 + ], + "spans": [ + { + "bbox": [ + 47, + 85, + 149, + 188 + ], + "type": "image", + "image_path": "034652c7fd889434320ee744bf594cb70749a3e09166a44978313464ba3b0cc1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 152, + 152, + 164, + 179 + ], + "blocks": [ + { + "bbox": [ + 152, + 152, + 164, + 179 + ], + "lines": [ + { + "bbox": [ + 152, + 152, + 164, + 179 + ], + "spans": [ + { + "bbox": [ + 152, + 152, + 164, + 179 + ], + "type": "image", + "image_path": "63746aa1134b4c098642f2b87f6880f4c4065873d812c8c78112bf5e0a0e0f12.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 169, + 152, + 180, + 179 + ], + "blocks": [ + { + "bbox": [ + 169, + 152, + 180, + 179 + ], + "lines": [ + { + "bbox": [ + 169, + 152, + 180, + 179 + ], + "spans": [ + { + "bbox": [ + 169, + 152, + 180, + 179 + ], + "type": "image", + "image_path": "454289ff174e478ba3a0b50cde2c33065935cf04c313dd8e8997e04683da73b4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 186, + 85, + 286, + 188 + ], + "blocks": [ + { + "bbox": [ + 187, + 70, + 284, + 80 + ], + "lines": [ + { + "bbox": [ + 187, + 70, + 284, + 80 + ], + "spans": [ + { + "bbox": [ + 187, + 70, + 284, + 80 + ], + "type": "text", + "content": "Joint Training (Same # Beam)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 186, + 85, + 286, + 188 + ], + "lines": [ + { + "bbox": [ + 186, + 85, + 286, + 188 + ], + "spans": [ + { + "bbox": [ + 186, + 85, + 286, + 188 + ], + "type": "image", + "image_path": "be2644d1baa9563208d57cde624574e1a758ff44124c902dfe803732d4eb44b7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 47, + 194, + 149, + 266 + ], + "blocks": [ + { + "bbox": [ + 47, + 194, + 149, + 266 + ], + "lines": [ + { + "bbox": [ + 47, + 194, + 149, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 149, + 266 + ], + "type": "image", + "image_path": "df956c51dd541326673b31a545bd870fb2314fa94dfa8589d60ccbc66164f7a6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 273, + 287, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 273, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 287, + 316 + ], + "type": "text", + "content": "Figure 3. A pilot study of naively merging different datasets for training the MinkUNet [20] model. Compared to the standalone training in (a), either jointly training with (b) the same, (c) different, or (d) all sensor-acquired data will cause severe degradation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 186, + 195, + 286, + 266 + ], + "blocks": [ + { + "bbox": [ + 186, + 195, + 286, + 266 + ], + "lines": [ + { + "bbox": [ + 186, + 195, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 186, + 195, + 286, + 266 + ], + "type": "image", + "image_path": "bf2472168858d61efd730da86ff7b5690e9670a0a78d2e07d7c3049012a6279f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "text", + "content": "complementary to each other [3, 7, 90]. To leverage such an advantage, we resort to the correspondences embedded in camera calibration matrices to bridge the LiDAR points and camera image pixels. Specifically, for each point " + }, + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = (p^x,p^y,p^z)" + }, + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "inline_equation", + "content": "P^s" + }, + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "text", + "content": ", the corresponding pixel " + }, + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "text", + "content": " can be found by the following transformations:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 92, + 407, + 287, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 407, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 92, + 407, + 287, + 433 + ], + "type": "interline_equation", + "content": "[ u, v, 1 ] ^ {\\mathrm {T}} = \\frac {1}{p ^ {z}} \\cdot T _ {s} \\cdot T \\cdot [ p ^ {x}, p ^ {y}, 1 ] ^ {\\mathrm {T}}, \\tag {1}", + "image_path": "1b1ddc0b54282417c9960c48617c8aa3553269ce22b701f3bb4a02fa4ed647a3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "text", + "content": " is the camera extrinsic matrix that consists of a rotation matrix and a translation matrix, and " + }, + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "inline_equation", + "content": "T_{s} \\in \\mathbb{R}^{3 \\times 4}" + }, + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "text", + "content": " is the camera intrinsic matrix. As we will show in the following sections, such a cross-sensor data alignment serves as the foundation for alignments in other spaces." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": "Cross-Sensor Statistical Alignment. To mitigate the discrepancies in sensor installations across different datasets, we incorporate a point coordinate alignment operation. Specifically, drawing upon insights from prior domain adaptation approaches [102, 116], we adjust the coordinate origins of point clouds from different datasets by introducing an offset " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathbb{R}^{1\\times 3}" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " to the ground plane. We find empirically that such an alignment can largely reduce the degradation caused by the variations in different sensor setups." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "text", + "content": "Dataset-Specific Rasterization. It is conventional to rasterize LiDAR point clouds " + }, + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "inline_equation", + "content": "P^s" + }, + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "text", + "content": " using unified rasterization parameters, e.g., voxel size [91, 128] or horizontal range view resolution [74, 111]. However, the point clouds acquired in different LiDAR datasets naturally differ in density, range, intensity, etc., which tends to favor different rasterization parameters. To meet such a requirement, we select dataset-specific parameters for rasterization on each dataset through empirical experiments and analyses." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": "Decoupled BN. Another challenge in training across multiple datasets is the presence of domain gaps, which can result in significant statistical shifts of feature learning among datasets. Such shifts can hinder the convergence and affect the model's ability to generalize well across diverse datasets. We adopt a decoupled batch norm (BN) for point cloud features in each dataset. Instead of using the traditional BN, which calculates mean and variance across all samples in a mini-batch, the decoupled BN tends to adapt each dataset's specific characteristics independently." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 201, + 448, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 201, + 448, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 448, + 213 + ], + "type": "text", + "content": "3.3. Feature-Space Alignment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 219, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 303 + ], + "type": "text", + "content": "We aim to acquire a generalized feature representation for downstream tasks. Compared to point clouds, images contribute stronger visual, textural, and semantic information. Thus, the collaboration between pixels and points could enrich the overall representation. Previous research [11, 60, 64] has consistently demonstrated that such a combination results often leads to improved performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "content": "Cross-Modality Assisted Alignment. In the context of multi-dataset joint training, our objective is to establish a unified feature space by leveraging image features to assist point cloud features. Acknowledging that images used in training lack ground truth labels [3, 27], we utilize image features from a pre-trained model as an alternative, facilitating a more universally applicable representation. We feed camera images " + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "inline_equation", + "content": "V^{s}" + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "content": " into a pre-trained DeepLab [98] and a vision-language model (VLM) and visualize the output image features by t-SNE [25]. As shown in Fig. 4, we observe that image features from DeepLab appear disorderly and lack semantics. In contrast, features from VLM share a more unified feature space. Motivated by this, we propose a cross-modality assisted alignment that uses VLM to help align the feature space. Specifically, the camera images " + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "inline_equation", + "content": "V^{s}" + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "content": " are fed to the frozen image encoder from VLM to obtain image features " + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "inline_equation", + "content": "F_{v} = \\{\\mathcal{F}_{v}^{1},\\mathcal{F}_{v}^{2},\\dots,\\mathcal{F}_{v}^{s}\\}" + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_v^s\\in \\mathbb{R}^{c\\times h\\times w}" + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "content": ". The LiDAR point clouds " + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "inline_equation", + "content": "P^{s}" + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "content": ", on the other hand, are fed to the point encoder followed by a projection layer to generate the point features " + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "inline_equation", + "content": "F_{p} = \\{\\mathcal{F}_{p}^{1},\\mathcal{F}_{p}^{2},\\dots,\\mathcal{F}_{p}^{s}\\}" + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p^s\\in \\mathbb{R}^{m\\times c}" + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "content": " denotes the number of non-empty grids. We then leverage the paired image features " + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "inline_equation", + "content": "\\hat{F}_v\\in \\mathbb{R}^{m_p\\times c}" + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "content": " and point feature " + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "inline_equation", + "content": "\\hat{F}_p\\in \\mathbb{R}^{m_p\\times c}" + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "content": " for alignment, where " + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "inline_equation", + "content": "m_{p}" + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "content": " is the number of point-pixel pairs. After obtaining " + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "inline_equation", + "content": "\\hat{F}_p" + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "inline_equation", + "content": "\\hat{F}_v" + }, + { + "bbox": [ + 304, + 304, + 546, + 605 + ], + "type": "text", + "content": ", the cross-modality alignment is expressed as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 353, + 614, + 545, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 614, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 353, + 614, + 545, + 645 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c m a}} \\left(\\hat {F} _ {v}, \\hat {F} _ {p}\\right) = 1 - \\frac {\\hat {F} _ {v} \\cdot \\hat {F} _ {p}}{\\| \\hat {F} _ {v} \\| \\cdot \\| \\hat {F} _ {p} \\|}. \\tag {2}", + "image_path": "ca91b5a2f4ae135e481511b1c96445e95ddeee6e91269ee9ffd3ba7ffecc01f5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Domain-Aware Cross-Modality Alignment. With cross-modality alignment, we transfer the knowledge of VLM to the point encoder, enabling the point features to gain a more comprehensive representation. However, during the execution of the above alignment, we have narrowed" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14651" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 289, + 271 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 289, + 271 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 289, + 271 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 289, + 271 + ], + "type": "image", + "image_path": "310b97e835a536564a0cd79f100b852668275897818711eb5272fc58eddeae88.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 275, + 288, + 330 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 288, + 330 + ], + "type": "text", + "content": "Figure 4. The t-SNE plots of learned features before and after the feature-space alignment in merging the nuScenes [●], SemanticKITTI [●], and Waymo Open [●] datasets. We show image features from (a) standalone networks; (b) SAM [44], and point cloud features (c) before and (d) after the feature-space alignment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 350, + 289, + 522 + ], + "blocks": [ + { + "bbox": [ + 47, + 350, + 289, + 522 + ], + "lines": [ + { + "bbox": [ + 47, + 350, + 289, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 289, + 522 + ], + "type": "image", + "image_path": "b42b377a5d75e26c03b1a296f884ac1428fe40fec203eb95355c8c5ba01c8f72.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 527, + 287, + 571 + ], + "lines": [ + { + "bbox": [ + 46, + 527, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 287, + 571 + ], + "type": "text", + "content": "Figure 5. Feature-space alignment in M3Net. We leverage both image features " + }, + { + "bbox": [ + 46, + 527, + 287, + 571 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 46, + 527, + 287, + 571 + ], + "type": "text", + "content": " and LiDAR point cloud features " + }, + { + "bbox": [ + 46, + 527, + 287, + 571 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 46, + 527, + 287, + 571 + ], + "type": "text", + "content": " extracted from image encoder " + }, + { + "bbox": [ + 46, + 527, + 287, + 571 + ], + "type": "inline_equation", + "content": "E_{img}" + }, + { + "bbox": [ + 46, + 527, + 287, + 571 + ], + "type": "text", + "content": " and point encoder " + }, + { + "bbox": [ + 46, + 527, + 287, + 571 + ], + "type": "inline_equation", + "content": "E_{pcd}" + }, + { + "bbox": [ + 46, + 527, + 287, + 571 + ], + "type": "text", + "content": " to employ the regularization via V2P loss and achieve feature-space alignment." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "it exclusively to image and point features from the same dataset. In this mode, point features solely learn from matching image features, restricting their knowledge acquisition. Ideally, we aim to ensure that image features encompass not only scenes identical to those represented in point clouds but also scenes from other datasets. To address this, we propose a domain-aware cross-modality guided alignment, as illustrated in Fig. 5. Specifically, we first extract, for each dataset, " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " from the same image encoder " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "E_{img}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " and point encoder " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "E_{pcd}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " during the cross" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "modality assisted alignment. The sets of features from all datasets are concatenated along the channel dimension to form " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\widetilde{F}_v\\in \\mathbb{R}^{c_v\\times h\\times w}" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ". Subsequently, we sequentially feed " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\widetilde{F}_v" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " through a branch that consists of a global average pooling and an MLP. Simultaneously, " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\widetilde{F}_v" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " is fed to an auxiliary branch that undergoes the same processing flow and generates an output after the softmax function " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ". The outputs from both branches are multiplied to obtain " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "F_{m}\\in \\mathbb{R}^{c_{v}\\times 1\\times 1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ". The overall process can be described as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 187, + 545, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 187, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 318, + 187, + 545, + 202 + ], + "type": "interline_equation", + "content": "F _ {m} = M L P \\left(P o o l \\left(\\widetilde {F} _ {v}\\right)\\right) \\cdot \\mathcal {G} \\left(M L P \\left(P o o l \\left(\\widetilde {F} _ {v}\\right)\\right)\\right). \\tag {3}", + "image_path": "5eadaa907968c721e3b637f9fc391b274db8867da9275a20d59dd76af747dcde.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 210, + 545, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 271 + ], + "type": "text", + "content": "Next, we forward " + }, + { + "bbox": [ + 304, + 210, + 545, + 271 + ], + "type": "inline_equation", + "content": "F_{m}" + }, + { + "bbox": [ + 304, + 210, + 545, + 271 + ], + "type": "text", + "content": " to a sigmoid activation function " + }, + { + "bbox": [ + 304, + 210, + 545, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot)" + }, + { + "bbox": [ + 304, + 210, + 545, + 271 + ], + "type": "text", + "content": " and multiply it with input image features " + }, + { + "bbox": [ + 304, + 210, + 545, + 271 + ], + "type": "inline_equation", + "content": "\\widetilde{F}_v" + }, + { + "bbox": [ + 304, + 210, + 545, + 271 + ], + "type": "text", + "content": ". The resulting output is added to " + }, + { + "bbox": [ + 304, + 210, + 545, + 271 + ], + "type": "inline_equation", + "content": "\\widetilde{F}_v" + }, + { + "bbox": [ + 304, + 210, + 545, + 271 + ], + "type": "text", + "content": " and passed through the MLP layers to obtain the final image features " + }, + { + "bbox": [ + 304, + 210, + 545, + 271 + ], + "type": "inline_equation", + "content": "F_{vf} \\in \\mathbb{R}^{c \\times h \\times w}" + }, + { + "bbox": [ + 304, + 210, + 545, + 271 + ], + "type": "text", + "content": ". The forward process of this operation is depicted as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 350, + 278, + 545, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 278, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 350, + 278, + 545, + 293 + ], + "type": "interline_equation", + "content": "F _ {v f} = M L P \\left(\\left(\\mathcal {H} \\left(F _ {m}\\right) \\cdot \\widetilde {F} _ {v}\\right) + \\widetilde {F} _ {v}\\right). \\tag {4}", + "image_path": "484180b28e63d0f657972615018724193d04e2b23ef76c19eff8af1afca8cea0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 301, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 339 + ], + "type": "text", + "content": "Finally, we leverage the cross-modality data alignment to acquire paired image features " + }, + { + "bbox": [ + 304, + 301, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\hat{F}_{vf} \\in \\mathbb{R}^{m_p \\times c}" + }, + { + "bbox": [ + 304, + 301, + 545, + 339 + ], + "type": "text", + "content": " and paired point feature " + }, + { + "bbox": [ + 304, + 301, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\hat{F}_p" + }, + { + "bbox": [ + 304, + 301, + 545, + 339 + ], + "type": "text", + "content": ". The overall objective function is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 348, + 347, + 545, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 347, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 348, + 347, + 545, + 377 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v 2 p} \\left(\\hat {F} _ {v f}, \\hat {F} _ {p}\\right) = 1 - \\frac {\\hat {F} _ {v f} \\cdot \\hat {F} _ {p}}{\\| \\hat {F} _ {v f} \\| \\cdot \\| \\hat {F} _ {p} \\|}. \\tag {5}", + "image_path": "1d0cb95a5691acd9fad7e27fe762f01739f8ca66a736725d5c871575db387eb4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 384, + 440, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 440, + 397 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 440, + 397 + ], + "type": "text", + "content": "3.4. Label-Space Alignment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 403, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 521 + ], + "type": "text", + "content": "Label Conflict. In multi-dataset joint training settings, label conflicts emerge as a significant challenge. This often refers to the inconsistencies in class labels across different datasets involved in the training process. The discrepancy can arise due to variations in annotation conventions, labeling errors, or even differences in the underlying semantics of classes between datasets. In our baseline, we unionize the different label spaces across datasets into " + }, + { + "bbox": [ + 304, + 403, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\mathbb{Y}^u" + }, + { + "bbox": [ + 304, + 403, + 545, + 521 + ], + "type": "text", + "content": ", where all datasets share a single LiDAR segmentation head. However, this may introduce several potential drawbacks:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 522, + 545, + 630 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 306, + 522, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 522, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 545, + 557 + ], + "type": "text", + "content": "- Loss of granularity: Unified label spaces could lose semantic granularity, particularly when dealing with subtle category differences in between different datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 558, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 545, + 594 + ], + "type": "text", + "content": "- Information loss: During label space consolidation, details unique to each dataset may be obscured or lost, especially for those related to domain-specific categories." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 594, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 594, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 545, + 630 + ], + "type": "text", + "content": "- Increased complexity: Handling a unified label space may necessitate more complex model architectures or training strategies, thereby increasing overall complexity." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "text", + "content": "To address these issues, we introduce a language-guided label-space alignment to facilitate a more holistic semantic correlation across datasets. Given the natural correspondence between images and texts and the strong correlation between images and point clouds, we aim to strategically utilize the image modality as a bridge to establish language-guided alignments. Such a process consists of a text-driven" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14652" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 289, + 238 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 289, + 238 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 289, + 238 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 289, + 238 + ], + "type": "image", + "image_path": "284d94ef22a16ef0fb1f4d2d8981c135be38ff15bb7ed5154900f94613a35560.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "type": "text", + "content": "Figure 6. Label-space alignment in M3Net. We leverage image features " + }, + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "type": "text", + "content": ", point cloud features " + }, + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "type": "text", + "content": ", and text embedding " + }, + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "type": "text", + "content": " extracted from " + }, + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "type": "inline_equation", + "content": "E_{img}" + }, + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "type": "inline_equation", + "content": "E_{pcd}" + }, + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "type": "inline_equation", + "content": "E_{txt}" + }, + { + "bbox": [ + 46, + 243, + 287, + 289 + ], + "type": "text", + "content": ", respectively, for regularization via the I2P, P2T, and V2T losses in the label-space alignment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 301, + 287, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 325 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 325 + ], + "type": "text", + "content": "point alignment, a text-driven image alignment, and a cross-modality-assisted label alignment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": "Text-Driven Alignments. As depicted in Fig. 6, images " + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "inline_equation", + "content": "V^s" + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": " are fed into the frozen image encoder " + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "inline_equation", + "content": "E_{img}" + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": " to extract the image features " + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "inline_equation", + "content": "F_v" + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": ". Concurrently, the LiDAR point clouds " + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "inline_equation", + "content": "P^s" + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": " are processed by the point encoder " + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "inline_equation", + "content": "E_{pcd}" + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": " to generate the point features " + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "inline_equation", + "content": "F_p" + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": ". Additionally, given the text input " + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "inline_equation", + "content": "T^s" + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": ", text embedding features " + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "inline_equation", + "content": "F_t \\in \\mathbb{R}^{Q \\times c}" + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": " are obtained from a frozen text encoder " + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "inline_equation", + "content": "E_{txt}" + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": " represents the number of categories across datasets. The text is composed of class names from unified label space " + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathbb{Y}_u" + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": " placed into pre-defined templates, and the text embedding captures semantic information of the corresponding classes. Subsequently, pixel-text pairs " + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\{v_k, t_k\\}_{k=1}^M" + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": " and point-text pairs " + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\{p_k, t_k\\}_{k=1}^M" + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": " are generated, where " + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 325, + 287, + 553 + ], + "type": "text", + "content": " represents the number of pairs. Leveraging the semantic information contained in the text, we selectively choose positive and negative samples for both images and points for contrastive learning. It is noteworthy that negative samples are confined to the specific dataset category space. The overall objective of the text-driven point alignment function is shown as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 559, + 287, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 287, + 602 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p 2 t} = - \\sum_ {q = 1} ^ {Q} \\log \\left(\\frac {\\sum_ {t _ {k} \\in q , p _ {k}} \\exp \\left(< t _ {k} , p _ {k} > / \\tau\\right)}{\\sum_ {t _ {k} \\in q , t _ {k} \\notin q , p _ {j}} \\exp \\left(< t _ {k} , p _ {k} > / \\tau\\right)}\\right), \\tag {6}", + "image_path": "b018a072dac51965305b2cb585c59ac663f5d34adcc014fa58a40647375ff1f7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "inline_equation", + "content": "t_k \\in q" + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "text", + "content": " indicates that " + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "text", + "content": " is generated by the " + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "text", + "content": "-th classes name, and " + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "text", + "content": " is the number of classes. Symbol " + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "inline_equation", + "content": "<, >" + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "text", + "content": " denotes the scalar product operation and " + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "text", + "content": " is a temperature term " + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "inline_equation", + "content": "(\\tau > 0)" + }, + { + "bbox": [ + 46, + 603, + 287, + 664 + ], + "type": "text", + "content": ". Similarly, the objective of the text-driven image alignment function is illustrated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 669, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 669, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 669, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v 2 t} = - \\sum_ {q = 1} ^ {Q} \\log \\left(\\frac {\\sum_ {t _ {k} \\in q , v _ {k}} \\exp \\left(< t _ {k} , v _ {k} > / \\tau\\right)}{\\sum_ {t _ {k} \\in q , t _ {k} \\notin q , v _ {j}} \\exp \\left(< t _ {k} , v _ {k} > / \\tau\\right)}\\right). \\tag {7}", + "image_path": "7eafa9b3e5ee474c47f44b1f7b895b15b9b3e34678d7edd4d900a1803e4157da.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "text", + "content": "Cross-Modality-Assisted Label Alignment. After text-driven alignments, the subsequent crucial step entails aligning the point and image modalities within the label space. We first obtain image logits " + }, + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "inline_equation", + "content": "F_{vl} = \\{\\mathcal{F}_{vl}^{1},\\mathcal{F}_{vl}^{2},\\dots,\\mathcal{F}_{vl}^{s}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "text", + "content": " and point logits " + }, + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "inline_equation", + "content": "F_{pl} = \\{\\mathcal{F}_{pl}^{1},\\mathcal{F}_{pl}^{2},\\dots,\\mathcal{F}_{pl}^{s}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "text", + "content": " from text-driven alignments, where " + }, + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}^{s}\\in \\mathbb{R}^{Q\\times H\\times W}" + }, + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{pl}^{s}\\in \\mathbb{R}^{N\\times Q}" + }, + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "text", + "content": ". Subsequently, we conduct cross-modality alignment to obtain paired image logits " + }, + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\hat{F}_{vl}\\in \\mathbb{R}^{m_p\\times Q}" + }, + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "text", + "content": " and paired point logits " + }, + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\hat{F}_{pl}\\in \\mathbb{R}^{m_p\\times Q}" + }, + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "text", + "content": ". Formally, the cross-modality-assisted alignment in the label space is formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 348, + 201, + 545, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 201, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 348, + 201, + 545, + 232 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i 2 p} \\left(\\hat {F} _ {v l}, \\hat {F} _ {p l}\\right) = 1 - \\frac {\\hat {F} _ {v l} \\cdot \\hat {F} _ {p l}}{\\| \\hat {F} _ {v l} \\| \\cdot \\| \\hat {F} _ {p l} \\|}. \\tag {8}", + "image_path": "00e46801331131a3477bed35f499df1df384d06b8440cfe7502a5bd91bf973b3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 239, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 239, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 545, + 262 + ], + "type": "text", + "content": "Finally, the complete objective function for the language-guided label-space alignment is expressed as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 362, + 271, + 545, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 271, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 362, + 271, + 545, + 285 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l a b e l}} = \\mathcal {L} _ {p 2 t} + \\mathcal {L} _ {i 2 p} + \\mathcal {L} _ {v 2 t}. \\tag {9}", + "image_path": "ad41a825d7cfa573b369bbf05126fe8066ffc7acf06f45310994b644a067e5af.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 293, + 476, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 293, + 476, + 304 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 476, + 304 + ], + "type": "text", + "content": "3.5. Universal LiDAR Segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 311, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 346 + ], + "type": "text", + "content": "We enhance the versatility of M3Net via multi-tasking learning. This integration involves an instance extractor to enable joint semantic and panoptic LiDAR segmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 347, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 467 + ], + "type": "text", + "content": "Panoptic LiDAR Segmentation. Motivated by DSNet [32, 33], our instance extractor comprises an instance head and a clustering step. The instance head encompasses several MLPs designed to predict the offsets between instance centers. The clustering step uses semantic predictions to filter out stuff points, thereby retaining only those associated with thing points. The remaining points undergo a mean-shift clustering [21], utilizing features from the instance head to discern distinct instances. Lastly, we employ the L1 loss " + }, + { + "bbox": [ + 304, + 347, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{l1}" + }, + { + "bbox": [ + 304, + 347, + 545, + 467 + ], + "type": "text", + "content": " to optimize the thing point regression process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "type": "text", + "content": "Overall Objectives. Putting everything together, the overall objective of M3Net is to minimize the following losses:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 327, + 499, + 545, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 499, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 327, + 499, + 545, + 513 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {v 2 p} + \\mathcal {L} _ {\\text {l a b e l}} + \\mathcal {L} _ {c e} + \\mathcal {L} _ {\\text {l o v a s z}} + \\mathcal {L} _ {l 1}, \\tag {10}", + "image_path": "733186b13787bb648bab3fb63eb922d216a43c60c6e7f34fb56a924919ca1384.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 521, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 545, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 521, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 305, + 521, + 545, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 521, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lovasz}" + }, + { + "bbox": [ + 305, + 521, + 545, + 545 + ], + "type": "text", + "content": " denote the cross-entropy loss and the Lovasz-softmax [5] loss, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 555, + 388, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 555, + 388, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 555, + 388, + 569 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 575, + 426, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 426, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 426, + 588 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": "Datasets. Our M3Net framework and baselines are trained on a combination of nuScenes [27], SemanticKITTI [3], and Waymo Open [90]. Meanwhile, we resort to another five LiDAR-based perception datasets [41, 45, 77, 106, 107] and two 3D robustness evaluation datasets [47] to verify the strong generalizability of M3Net. Due to space limits, additional details regarding the datasets are in the Appendix." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. M3Net is implemented based on Pointcept [23] and MMDetection3D [22]. We use two backbones in our experiments, i.e., MinkUNet [20] and" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14653" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 99, + 547, + 201 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 1. Ablation study on the M3Net alignments happen in the Data, Feature, and Label spaces, respectively, when combining the SemanticKITTI [3], nuScenes [27], and Waymo Open [90] datasets. The mAcc and mIoU scores are in percentage. Best scores are in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 99, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 49, + 99, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 49, + 99, + 547, + 201 + ], + "type": "table", + "html": "
-ConfigurationsMinkUNet [20]PTv2+ [103]
SemKITTInuScenesWaymoSemKITTInuScenesWaymo
mAccmIoUmAccmIoUmAccmIoUmAccmIoUmAccmIoUmAccmIoU
BaselineNaïve Joint Training62.4354.0365.0559.8473.7665.3967.9661.5976.5369.6575.6867.00
M3Net (Ours)DataFeatureLabel------
✓73.8269.0183.6676.8977.8869.3778.5569.9586.2279.1380.9672.15
✓✓74.3669.6485.1778.8878.3169.7079.4370.8787.1080.2680.7472.33
✓✓73.8569.3485.2078.9078.0469.5580.3071.1387.4480.4580.6972.30
✓✓✓74.4069.8585.3079.0078.6670.1580.0072.0087.9180.9081.1172.40
", + "image_path": "a0cd17d042a20307c403a6134be9e53e6c24f0b495a910bf2eda3bdec4ec3827.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 243, + 545, + 388 + ], + "blocks": [ + { + "bbox": [ + 47, + 215, + 544, + 239 + ], + "lines": [ + { + "bbox": [ + 47, + 215, + 544, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 544, + 239 + ], + "type": "text", + "content": "Table 2. Panoptic LiDAR segmentation results on the val sets of the Panoptic-SemanticKITTI [3] and Panoptic-nuScenes [27] datasets. All scores are given in percentage. The best and second-best scores are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 243, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 50, + 243, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 50, + 243, + 545, + 388 + ], + "type": "table", + "html": "
MethodConfigurationsPanoptic-SemanticKITTIPanoptic-nuScenes
PQ\\( PQ^† \\)RQSQmIoUPQ\\( PQ^† \\)RQSQmIoU
Panoptic-TrackNet [37]40.0-48.373.053.851.456.263.380.258.0
Panoptic-PolarNet [127]59.164.170.278.364.563.467.275.383.966.9
EfficientLPS [88]Single-Dataset Training59.265.169.875.064.959.262.882.970.769.4
DSNet [32]61.465.272.779.069.664.767.676.183.576.3
Panoptic-PHNet [55]61.7---65.774.777.784.288.279.7
BaselineNaïve Joint Training56.0359.6465.7873.7261.5956.6760.6166.7583.4969.65
M3Net (Ours)DataFeatureLabel--
✓62.3465.1772.6074.6769.9568.4971.1179.1385.4979.13
✓✓62.9165.7373.3275.4770.8771.4773.8681.5386.7180.26
✓✓63.2367.8973.6181.6671.1371.5373.9181.8086.9280.45
✓✓✓63.8768.6673.1082.3572.0071.7074.0182.2086.4780.90
", + "image_path": "4e5bfd0dee79c4ea430f5e670000ad1f9512f532f4c7ceee4e4c651825a23020.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "text", + "content": "PTv2+ [103]. We trained M3Net on four A100 GPUs for 50 epochs with a batch size of 6 for each GPU. The initial learning rate is set to 0.002. We adopt the AdamW optimizer [68] with a weight decay of 0.005 and cosine decay learning rate scheduler. For the dataset-specific rasterization, we set voxel sizes to " + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "inline_equation", + "content": "0.05\\mathrm{m}" + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "inline_equation", + "content": "0.05\\mathrm{m}" + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "text", + "content": " for SemanticKITTI [3], nuScenes [27], and Waymo Open [90], respectively. For the data augmentation, we employ random flipping, jittering, scaling, rotation, and Mix3D [75]. Due to space limits, kindly refer to Appendix for additional details. Evaluation Metrics. We adopt conventional reportings of mAcc and mIoU for LiDAR semantic segmentation, " + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "inline_equation", + "content": "PQ" + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "inline_equation", + "content": "PQ^{\\dagger}" + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "inline_equation", + "content": "SQ" + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "inline_equation", + "content": "RQ" + }, + { + "bbox": [ + 46, + 411, + 287, + 591 + ], + "type": "text", + "content": " for panoptic segmentation, and mCE and mRR for 3D robustness evaluation. Due to the space limit, kindly refer to our Appendix for more detailed definitions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 599, + 141, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 141, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 141, + 612 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "Multi-Space Alignments. The effectiveness of three proposed alignments over the joint training baselines is shown in Tab. 1. We observe that the data-space alignment plays the most crucial role in improving the universal LiDAR segmentation performance. Without proper data alignments, joint training with either MinkUNet [20] or the stronger PTv2+ [103] will suffer severe degradation, especially on sparser point clouds [27]. On top of the data-space align-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "text", + "content": "ment, the combinations of multi-view images at the feature space and the language-guided knowledge at the label space further enhance the learned feature representations. The results show that they work synergistically in merging knowledge from heterogeneous domains during joint training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 472, + 546, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 581 + ], + "type": "text", + "content": "Panoptic LiDAR Segmentation. In Tab. 2, we present another ablation study focusing on panoptic LiDAR segmentation. All three alignments incorporated in M3Net demonstrate significant improvements over the baselines. This highlights the pronounced efficacy of our multi-space alignments. Moreover, our approach outperforms the single-dataset state-of-the-art method Panoptic-PHNet [55] by a notable " + }, + { + "bbox": [ + 304, + 472, + 546, + 581 + ], + "type": "inline_equation", + "content": "2.17\\%" + }, + { + "bbox": [ + 304, + 472, + 546, + 581 + ], + "type": "text", + "content": " PQ on Panoptic-SemanticKITTI [3] and achieves compelling results on Panoptic-nuScenes [27]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": "Visual Feature Alignments. We conduct a qualitative analysis of the learned visual feature distributions in the form of t-SNE [25]. Fig. 4 (a) and (b) represent the distributions of learned visual features among three datasets from DeepLab and VLM backbones, respectively. The features obtained by the latter exhibit more distinct semantics in feature space. The concentrated distribution space is advantageous for achieving feature alignments across multiple datasets. Additionally, Fig. 4 (c) and (d) illustrate the distribution of point cloud features before and after feature-space alignment. As can be seen, the feature distribution distances" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14654" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 99, + 547, + 192 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 3. Knowledge transfer and generalization analyses across five LiDAR segmentation datasets and two 3D robustness evaluation datasets. All scores are given in percentage. The best and second-best scores are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 99, + 547, + 192 + ], + "lines": [ + { + "bbox": [ + 49, + 99, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 49, + 99, + 547, + 192 + ], + "type": "table", + "html": "
MethodRELLIS-3DSemanticPOSSSemanticSTFSynLiDARDAPS-3DSemKITTI-CnuScenes-C
1%10%HalfFullHalfFull1%10%HalfFullmCEmRRmCEmRR
PPKT [66]49.7154.3350.1856.0050.9254.6937.5746.4878.9084.00--105.6476.06
SLidR [86]49.7554.5751.5655.3652.0154.3542.0547.8481.0085.40--106.0875.99
Seal [65]51.0955.0353.2656.8953.4655.3643.5849.2681.8885.90--92.6383.08
Naïve Joint37.7750.2342.1952.3146.7048.0018.5642.3773.9177.89113.6584.73128.9781.45
Single-Dataset40.1754.2547.6955.0050.3351.1923.1745.0875.1080.8795.1184.9599.6379.06
M3Net (Ours)51.2755.0553.6057.1753.7855.4244.1049.9382.0886.0086.4385.7791.0379.15
", + "image_path": "e59fa22168883494ffb8b5ac2fbdb0cabbdb456f81353bc4137d77c06dfe2695.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 258, + 287, + 446 + ], + "blocks": [ + { + "bbox": [ + 47, + 208, + 288, + 252 + ], + "lines": [ + { + "bbox": [ + 47, + 208, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 208, + 288, + 252 + ], + "type": "text", + "content": "Table 4. LiDAR semantic segmentation results on the val and test sets of SemanticKITTI [3] and nuScenes [27], and the val set of Waymo Open [90]. All scores are in percentage. The best and second-best scores are highlighted in bold and underline." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 258, + 287, + 446 + ], + "lines": [ + { + "bbox": [ + 49, + 258, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 49, + 258, + 287, + 446 + ], + "type": "table", + "html": "
MethodSemKITTInuScenesWaymo
ValTestValTestmIoUmAcc
RangeNet++ [74]-52.2-65.5--
PolarNet [121]57.254.371.069.8--
SalsaNext [24]-59.5-72.2--
RangeViT [1]60.764.075.2---
MinkUNet [20]63.863.773.3-65.976.6
SPVNAS [91]64.766.4-77.467.4-
AMVNet [62]65.265.377.277.3--
RPVNet [112]65.570.377.6---
(AF)2-S3Net [18]-69.7-78.3--
Cylinder3D [128]65.967.876.177.966.0-
PVKD [34]66.471.276.0---
WaffleIron [80]66.870.879.1---
RangeFormer [46]67.673.378.180.1--
SphereFormer [52]67.874.878.481.969.9-
FRNet [114]68.773.379.082.5--
PTv2+ [103]70.370.680.282.670.680.2
LidarMultiNet [117]---81.473.8-
M3Net (Ours)72.075.180.983.172.481.1
", + "image_path": "6d1d973134a2cb1f10b6c754a90a7525d4d0506d10b278210317a4bed7731d76.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 467, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 287, + 491 + ], + "type": "text", + "content": "between the three datasets have been largely reduced, providing evidence of the alignment effectiveness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 501, + 162, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 162, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 162, + 514 + ], + "type": "text", + "content": "4.3. Comparative Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "content": "Comparisons to State of the Arts. In Tab. 4, we compare M3Net with current best-performing models on the benchmarks of SemanticKITTI [3], nuScenes [27], and Waymo Open [90]. Remarkably, M3Net consistently outperforms existing approaches across all three datasets. Specifically, on SemanticKITTI [3], M3Net achieves a " + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "inline_equation", + "content": "72.0\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "content": " mIoU on the validation set, surpassing the closest method by a notable margin of " + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "content": " mIoU. Similarly, on nuScenes [27], M3Net achieves " + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "inline_equation", + "content": "80.9\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "content": " mIoU and " + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "inline_equation", + "content": "83.1\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "content": " mIoU on the validation and test sets, demonstrating its robustness and generalization capabilities. Additionally, the performance of M3Net on Waymo Open [90] is competitive with prior arts. We achieve a mIoU of " + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "inline_equation", + "content": "72.4\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "content": " and a mAcc of " + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "inline_equation", + "content": "81.1\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "content": ". These results highlight again the superiority of M3Net in handling complex diverse LiDAR segmentation tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Direct Knowledge Transfer. To further validate the strong" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 209, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 317 + ], + "type": "text", + "content": "knowledge transfer capability of M3Net, we conduct extensive experiments on five different LiDAR-based perception datasets [41, 45, 77, 106, 107]. These datasets have unique data collection protocols and data distributions. As shown in Fig. 1 and the first ten columns in Tab. 3, our framework constantly outperforms the prior arts, the naive joint training, and the single-dataset baselines across all five datasets. This concretely supports the strong knowledge transfer efficacy brought by multi-space alignments in M3Net." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "type": "text", + "content": "Out-of-Distribution Generalization. Evaluating the generalization ability of models on out-of-training-distribution data is crucial, particularly in safety-critical fields like autonomous driving [50, 109, 110]. In this context, we resort to the two corruption datasets from the Robo3D [47] benchmark, i.e., SemanticKITTI-C and nuScenes-C, to conduct our assessment. From the last four columns of Tab. 3, we observe that M3Net achieves better results than the naive joint training and other single-dataset approaches, proving the strong generalizability of the learned representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 449, + 378, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 378, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 378, + 460 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 468, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 612 + ], + "type": "text", + "content": "In this work, we presented M3Net, a universal framework capable of fulfilling multi-task, multi-dataset, multimodality LiDAR segmentation using a single set of parameters. Through extensive analyses, we validated the effectiveness of applying data-, feature-, and label-space alignments to handle such a challenging task. In addition, our comprehensive analysis and discourse have delved into the fundamental challenges of acquiring the general knowledge for scalable 3D perception, which holds substantial potential to propel further research in this domain. Our future strides focus on combining more data resources to further enhance the alignments and adaptations in our framework." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was partially supported by NSFC (No.62206173) and MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University). This work was also supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14655" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Angelika Ando, Spyros Gidaris, Andrei Bursuc, Gilles Puy, Alexandre Boulch, and Renaud Marlet. Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5240-5250, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 147, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 147, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 147, + 288, + 201 + ], + "type": "text", + "content": "[2] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 202, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 202, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 56, + 202, + 288, + 257 + ], + "type": "text", + "content": "[3] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 258, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 287, + 312 + ], + "type": "text", + "content": "[4] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Jürgen Gall, and Cyrill Stachniss. Towards 3d lidar-based semantic scene understanding of 3d point cloud sequences: The semantickitti dataset. International Journal of Robotics Research, 40:959-96, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 314, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 314, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 288, + 369 + ], + "type": "text", + "content": "[5] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: a tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4413-4421, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 369, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 369, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 287, + 423 + ], + "type": "text", + "content": "[6] Alexandre Boulch, Corentin Sautier, Björn Michele, Gilles Puy, and Renaud Marlet. Also: Automotive lidar self-supervision by occupancy estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13455-13465, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 425, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 425, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 425, + 287, + 491 + ], + "type": "text", + "content": "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 491, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 287, + 546 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 547, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 547, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 547, + 287, + 602 + ], + "type": "text", + "content": "[9] Jun Cen, Shiwei Zhang, Yixuan Pei, Kun Li, Hang Zheng, Maochun Luo, Yingya Zhang, and Qifeng Chen. Cmdfusion: Bidirectional fusion network with cross-modality knowledge distillation for lidar semantic segmentation. arXiv preprint arXiv:2307.04091, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 602, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 602, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 53, + 602, + 287, + 658 + ], + "type": "text", + "content": "[10] Runnan Chen, Youquan Liu, Lingdong Kong, Nenglun Chen, Xinge Zhu, Yuexin Ma, Tongliang Liu, and Wenping Wang. Towards label-free scene understanding by vision foundation models. In Advances in Neural Information Processing Systems, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 658, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 287, + 714 + ], + "type": "text", + "content": "[11] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 128 + ], + "type": "text", + "content": "[12] Runnan Chen, Xinge Zhu, Nenglun Chen, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Bridging language and geometric primitives for zero-shot point cloud segmentation. In ACM International Conference on Multimedia, pages 5380-5388, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 545, + 183 + ], + "type": "text", + "content": "[13] Tian Chen, Shijie An, Yuan Zhang, Chongyang Ma, Huayan Wang, Xiaoyan Guo, and Wen Zheng. Improving monocular depth estimation by leveraging structural awareness and complementary datasets. In European Conference on Computer Vision, pages 90-108, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 184, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 184, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 312, + 184, + 545, + 237 + ], + "type": "text", + "content": "[14] Yanbei Chen, Manchen Wang, Abhay Mittal, Zhenlin Xu, Paolo Favaro, Joseph Tighe, and Davide Modolo. Scaledet: A scalable multi-dataset object detector. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7288-7297, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 239, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 239, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 312, + 239, + 545, + 283 + ], + "type": "text", + "content": "[15] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. In Advances in Neural Information Processing Systems, pages 17864-17875, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 284, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 284, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 312, + 284, + 545, + 338 + ], + "type": "text", + "content": "[16] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 339, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 339, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 312, + 339, + 545, + 384 + ], + "type": "text", + "content": "[17] Huixian Cheng, Xianfeng Han, and Guoqiang Xiao. Cenet: Toward concise and efficient lidar semantic segmentation for autonomous driving. In IEEE International Conference on Multimedia and Expo, pages 1-6, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 384, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 384, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 312, + 384, + 545, + 438 + ], + "type": "text", + "content": "[18] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. Af2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 439, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 439, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 312, + 439, + 545, + 669 + ], + "type": "text", + "content": "[19] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 670, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 545, + 714 + ], + "type": "text", + "content": "[20] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14656" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "text", + "content": "[21] Dorin Comaniciu and Peter Meer. Mean shift: A robust approach toward feature space analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(5):603-619, 2005." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 287, + 160 + ], + "type": "text", + "content": "[22] MMDetection3D Contributors. MMDetection3D: Open-MMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 162, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 162, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 162, + 287, + 194 + ], + "type": "text", + "content": "[23] Pointcept Contributors. Pointcept: A codebase for point cloud perception research. https://github.com/Pointcept/Pointcept, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 195, + 287, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 287, + 238 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 287, + 238 + ], + "type": "text", + "content": "[24] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In International Symposium on Visual Computing, pages 207-222, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 271 + ], + "type": "text", + "content": "[25] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (11), 2008." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 287, + 304 + ], + "type": "text", + "content": "[26] Christoph Feichtenhofer, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 306, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 287, + 358 + ], + "type": "text", + "content": "[27] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters, 7:3795-3802, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 361, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 361, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 53, + 361, + 287, + 404 + ], + "type": "text", + "content": "[28] Biao Gao, Yancheng Pan, Chengkun Li, Sibo Geng, and Huijing Zhao. Are we hungry for 3d lidar data for semantic segmentation? a survey of datasets and methods. IEEE Transactions on Intelligent Transportation Systems, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 405, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 405, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 53, + 405, + 287, + 459 + ], + "type": "text", + "content": "[29] Xiuye Gu, Yin Cui, Jonathan Huang, Abdullah Rashwan, Xuan Yang, Xingyi Zhou, Golnaz Ghiasi, Weicheng Kuo, Huizhong Chen, Liang-Chieh Chen, and David A. Ross. Dataseg: Taming a universal multi-dataset multi-task segmentation model. arXiv preprint arXiv:2306.01736, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 460, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 460, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 460, + 287, + 502 + ], + "type": "text", + "content": "[30] Yi Gu, Yuming Huang, Chengzhong Xu, and Hui Kong. Maskrange: A mask-classification model for range-view based lidar segmentation. arXiv preprint arXiv:2206.12073, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 504, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 504, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 53, + 504, + 287, + 536 + ], + "type": "text", + "content": "[31] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In IEEE/CVF International Conference on Computer Vision, pages 2961-2969, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 537, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 537, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 53, + 537, + 287, + 581 + ], + "type": "text", + "content": "[32] Fangzhou Hong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Lidar-based panoptic segmentation via dynamic shifting network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13090-13099, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 582, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 582, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 582, + 287, + 635 + ], + "type": "text", + "content": "[33] Fangzhou Hong, Lingdong Kong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Unified 3d and 4d panoptic segmentation via dynamic shifting networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2024." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 637, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 637, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 637, + 287, + 689 + ], + "type": "text", + "content": "[34] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-voxel knowledge distillation for lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8479–8488, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "type": "text", + "content": "[35] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 106 + ], + "type": "text", + "content": "Randla-net: Efficient semantic segmentation of large-scale point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11108-11117, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 107, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 171 + ], + "type": "text", + "content": "[36] Qingyong Hu, Bo Yang, Sheikh Khalid, Wen Xiao, Niki Trigoni, and Andrew Markham. Towards semantic segmentation of urban-scale 3d point clouds: A dataset, benchmarks and challenges. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4977-4987, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 173, + 545, + 205 + ], + "type": "text", + "content": "[37] Juana Valeria Hurtado, Rohit Mohan, Wolfram Burgard, and Abhinav Valada. Mopt: Multi-object panoptic tracking. arXiv preprint arXiv:2004.08189, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 545, + 259 + ], + "type": "text", + "content": "[38] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 261, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 261, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 312, + 261, + 545, + 316 + ], + "type": "text", + "content": "[39] Maximilian Jaritz, Tuan-Hung Vu, Raoul de Charette, Emilie Wirbel, and Patrick Pérez. xmuda: Cross-modal unsupervised domain adaptation for 3d semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12605-12614, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 317, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 317, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 312, + 317, + 545, + 369 + ], + "type": "text", + "content": "[40] Maximilian Jaritz, Tuan-Hung Vu, Raoul De Charette, Emilie Wirbel, and Patrick Pérez. Cross-modal learning for domain adaptation in 3d semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1533-1544, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 372, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 372, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 372, + 545, + 415 + ], + "type": "text", + "content": "[41] Peng Jiang, Philip Osteen, Maggie Wigness, and Srikanth Saripallig. Rellis-3d dataset: Data, benchmarks and analysis. In IEEE International Conference on Robotics and Automation, pages 1110–1116, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 416, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 416, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 416, + 545, + 459 + ], + "type": "text", + "content": "[42] Tarun Kalluri, Girish Varma, Manmohan Chandraker, and C. V. Jawahar. Universal semi-supervised semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 5259-5270, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "type": "text", + "content": "[43] Dongwan Kim, Yi-Hsuan Tsai, Yumin Suh, Masoud Faraki, Sparsh Garg, Manmohan Chandraker, and Bohyung Han. Learning semantic segmentation from multiple datasets with label shifts. In European Conference on Computer Vision, pages 20-36, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 515, + 545, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 515, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 312, + 515, + 545, + 579 + ], + "type": "text", + "content": "[44] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 581, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 581, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 312, + 581, + 545, + 634 + ], + "type": "text", + "content": "[45] Alexey Klokov, Di Un Pak, Aleksandr Khorin, Dmitry Yudin, Leon Kochiev, Vladimir Luchinskiy, and Vitaly Bezuglyj. Daps3d: Domain adaptive projective segmentation of 3d lidar point clouds. IEEE Access, 11:79341-79356, 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 636, + 545, + 689 + ], + "type": "text", + "content": "[46] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Zwei Liu. Rethinking range view representation for lidar segmentation. In IEEE/CVF International Conference on Computer Vision, pages 228-240, 2023." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "text", + "content": "[47] Lingdong Kong, Youquan Liu, Xin Li, Runnan Chen, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14657" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "type": "text", + "content": "Liu. Robo3d: Towards robust and reliable 3d perception against corruptions. In IEEE/CVF International Conference on Computer Vision, pages 19994-20006, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 287, + 162 + ], + "type": "text", + "content": "[48] Lingdong Kong, Niamul Quader, and Venice Erin Liong. Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In IEEE International Conference on Robotics and Automation, pages 9338-9345, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 163, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 288, + 207 + ], + "type": "text", + "content": "[49] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21705-21715, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 54, + 209, + 287, + 262 + ], + "type": "text", + "content": "[50] Lingdong Kong, Shaoyuan Xie, Hanjiang Hu, Lai Xing Ng, Benoit R. Cottereau, and Wei Tsang Ooi. Robodepth: Robust out-of-distribution depth estimation under corruptions. In Advances in Neural Information Processing Systems, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "type": "text", + "content": "[51] Lingdong Kong, Xiang Xu, Jun Cen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Calib3d: Calibrating model preferences for reliable 3d scene understanding. arXiv preprint arXiv:2403.17010, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 310, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 310, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 53, + 310, + 287, + 354 + ], + "type": "text", + "content": "[52] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17545-17555, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 355, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 355, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 53, + 355, + 287, + 408 + ], + "type": "text", + "content": "[53] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. Mseg: A composite dataset for multidomain semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2879-2888, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 411, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 411, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 411, + 287, + 475 + ], + "type": "text", + "content": "[54] Guangrui Li, Guoliang Kang, Xiaohan Wang, Yunchao Wei, and Yi Yang. Adversarily masking synthetic to mimic real: Adaptive noise injection for point cloud segmentation adaptation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20464-20474, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 478, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 288, + 533 + ], + "type": "text", + "content": "[55] Jinke Li, Xiao He, Yang Wen, Yuan Gao, Xiaogiang Cheng, and Dan Zhang. Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11809-11818, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 534, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 534, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 534, + 287, + 588 + ], + "type": "text", + "content": "[56] Li Li, Hubert PH Shum, and Toby P. Breckon. Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9361-9371, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 287, + 634 + ], + "type": "text", + "content": "[57] Rong Li, Raoul de Charette, and C. A. O. Anh-Quan. Coarse3d: Class-prototypes for contrastive learning in weakly-supervised 3d point cloud segmentation. In *British Machine Vision Conference*, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 635, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 287, + 690 + ], + "type": "text", + "content": "[58] Xin Li, Botian Shi, Yuenan Hou, Xingjiao Wu, Tianlong Ma, Yikang Li, and Liang He. Homogeneous multi-modal feature fusion and interaction for 3d object detection. In European Conference on Computer Vision, pages 691-707. Springer, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "text", + "content": "[59] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 116 + ], + "type": "text", + "content": "Video k-net: A simple, strong, and unified baseline for video segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18847-18857, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 118, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 183 + ], + "type": "text", + "content": "[60] Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17524-17534, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 312, + 185, + 545, + 228 + ], + "type": "text", + "content": "[61] Ye Li, Lingdong Kong, Hanjiang Hu, Xiaohao Xu, and Xiaonan Huang. Optimizing lidar placements for robust driving perception in adverse conditions. arXiv preprint arXiv:2403.17009, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 312, + 231, + 545, + 274 + ], + "type": "text", + "content": "[62] Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja, Dhananjai Sharma, and Zhuang Jie Chong. Amvnet: Assertion-based multi-view fusion network for lidar semantic segmentation. arXiv preprint arXiv:2012.04934, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 276, + 545, + 319 + ], + "type": "text", + "content": "[63] Minghua Liu, Yin Zhou, Charles R. Qi, Boqing Gong, Hao Su, and Dragomir Anguelov. Less: Label-efficient semantic segmentation for lidar point clouds. In European Conference on Computer Vision, pages 70-89, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 321, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 321, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 312, + 321, + 545, + 387 + ], + "type": "text", + "content": "[64] Youquan Liu, Runnan Chen, Xin Li, Lingdong Kong, Yuchen Yang, Zhaoyang Xia, Yeqi Bai, Xinge Zhu, Yuexin Ma, Yikang Li, Yu Qiao, and Yuenan Hou. Uniseg: A unified multi-modal lidar segmentation network and the openpcseg codebase. In IEEE/CVF International Conference on Computer Vision, pages 21662-21673, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 388, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 388, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 312, + 388, + 545, + 442 + ], + "type": "text", + "content": "[65] Youquan Liu, Lingdong Kong, Jun Cen, Runnan Chen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Segment any point cloud sequences by distilling vision foundation models. In Advances in Neural Information Processing Systems, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 444, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 444, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 312, + 444, + 545, + 498 + ], + "type": "text", + "content": "[66] Yueh-Cheng Liu, Yu-Kai Huang, Hung-Yueh Chiang, Hung-Ting Su, Zhe-Yu Liu, Chin-Tang Chen, Ching-Yu Tseng, and Winston H. Hsu. Learning from 2d: Contrastive pixel-to-point knowledge transfer for 3d pretraining. arXiv preprint arXiv:2104.0468, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 500, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 500, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 500, + 545, + 555 + ], + "type": "text", + "content": "[67] Zhijian Liu, Alexander Amini Haotian Tang, Xinyu Yang, Huizi Mao, Daniela L. Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In IEEE International Conference on Robotics and Automation, pages 2774-2781, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 556, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 556, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 312, + 556, + 545, + 588 + ], + "type": "text", + "content": "[68] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 590, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 590, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 312, + 590, + 545, + 644 + ], + "type": "text", + "content": "[69] Yuhang Lu, Qi Jiang, Runnan Chen, Yuenan Hou, Xinge Zhu, and Yuexin Ma. See more and know more: Zero-shot point cloud segmentation via multi-modal visual data. In IEEE/CVF International Conference on Computer Vision, pages 21674-21684, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 689 + ], + "type": "text", + "content": "[70] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A comprehensive survey. International Journal of Computer Vision, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "text", + "content": "[71] Rodrigo Marcuzzi, Lucas Nunes, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Mask-based panoptic lidar" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14658" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "type": "text", + "content": "segmentation for autonomous driving. IEEE Robotics and Automation Letters, 8(2):1141-1148, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 288, + 139 + ], + "type": "text", + "content": "[72] Panagiotis Meletis and Gijs Dubbelman. Training of convolutional networks on multiple heterogeneous datasets for street scene semantic segmentation. In IEEE Intelligent Vehicles Symposium, pages 1045-1050, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 54, + 140, + 287, + 183 + ], + "type": "text", + "content": "[73] Björn Michele, Alexandre Boulch, Gilles Puy, Tuan-Hung Vu, Renaud Marlet, and Nicolas Courty. Saluda: Surface-based automotive lidar unsupervised domain adaptation. arXiv preprint arXiv:2304.03251, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 184, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 287, + 228 + ], + "type": "text", + "content": "[74] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4213-4220, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 271 + ], + "type": "text", + "content": "[75] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3d: Out-of-context data augmentation for 3d scenes. In International Conference on 3D Vision, pages 116-125, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 287, + 380 + ], + "type": "text", + "content": "[76] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 383, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 383, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 53, + 383, + 287, + 426 + ], + "type": "text", + "content": "[77] Yancheng Pan, Biao Gao, Jilin Mei, Sibo Geng, Chengkun Li, and Huijing Zhao. Semanticposs: A point cloud dataset with large quantity of dynamic instances. In IEEE Intelligent Vehicles Symposium, pages 687-693, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 426, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 287, + 480 + ], + "type": "text", + "content": "[78] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. Openscene: 3d scene understanding with open vocabularies. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-824, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 482, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 482, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 53, + 482, + 287, + 525 + ], + "type": "text", + "content": "[79] Xidong Peng, Runnan Chen, Feng Qiao, Lingdong Kong, Youquan Liu, Tai Wang, Xinge Zhu, and Yuexin Ma. Sam-guided unsupervised domain adaptation for 3d segmentation. arXiv preprint arXiv:2310.08820, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 526, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 526, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 53, + 526, + 287, + 569 + ], + "type": "text", + "content": "[80] Gilles Puy, Alexandre Boulch, and Renaud Marlet. Using a waffle iron for automotive point cloud semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 3379-3389, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 571, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 571, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 571, + 287, + 635 + ], + "type": "text", + "content": "[81] 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 International Conference on Machine Learning, pages 8748-8763, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 637, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 637, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 637, + 287, + 690 + ], + "type": "text", + "content": "[82] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1623-1637, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "text", + "content": "[83] Giulia Rizzoli, Francesco Barbato, and Pietro Zanuttigh. Multimodal semantic segmentation in autonomous driving:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 331, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 544, + 95 + ], + "type": "text", + "content": "A review of current approaches and future perspectives. Technologies, 10(4), 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 149 + ], + "type": "text", + "content": "[84] Cristiano Saltori, Fabio Galasso, Giuseppe Fiameni, Nicu Sebe, Elisa Ricci, and Fabio Poiesi. Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In European Conference on Computer Vision, pages 586-602, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 151, + 545, + 194 + ], + "type": "text", + "content": "[85] Jules Sanchez, Jean-Emmanuel Deschaud, and François Goulette. Cola: Coarse-label multi-source lidar semantic segmentation for autonomous driving. arXiv preprint arXiv:2311.03017, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 195, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 195, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 195, + 545, + 249 + ], + "type": "text", + "content": "[86] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 312, + 251, + 545, + 293 + ], + "type": "text", + "content": "[87] Alvari Seppänen, Risto Ojala, and Kari Tammi. 4denoiseNET: Adverse weather denoising from adjacent point clouds. IEEE Robotics and Automation Letters, 8:456-463, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 294, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 294, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 312, + 294, + 545, + 337 + ], + "type": "text", + "content": "[88] Kshitij Sirohi, Rohit Mohan, Daniel Buscher, Wolfram Burgard, and Abhinav Valada. Efficientlps: Efficient lidar panoptic segmentation. IEEE Transactions on Robotics, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 312, + 338, + 545, + 381 + ], + "type": "text", + "content": "[89] Louis Soum-Fontez, Jean-Emmanuel Deschaud, and François Goulette. Mdt3d: Multi-dataset training for lidar 3d object detection generalization. arXiv preprint arXiv:2308.01000, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 383, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 383, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 312, + 383, + 545, + 491 + ], + "type": "text", + "content": "[90] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2446-2454, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 493, + 545, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 493, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 312, + 493, + 545, + 546 + ], + "type": "text", + "content": "[91] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, pages 685–702, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 548, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 548, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 312, + 548, + 545, + 602 + ], + "type": "text", + "content": "[92] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In IEEE/CVF International Conference on Computer Vision, pages 6411–6420, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 603, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 603, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 312, + 603, + 545, + 647 + ], + "type": "text", + "content": "[93] Larissa T Triess, David Peter, Christoph B Rist, and J Marius Zöllner. Scan-based semantic segmentation of lidar point clouds: An experimental study. In IEEE Intelligent Vehicles Symposium, pages 1116-1121, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 690 + ], + "type": "text", + "content": "[94] Larissa T. Triess, Mariella Dreissig, Christoph B. Rist, and J. Marius Zöllner. A survey on deep domain adaptation for lidar perception. In IEEE Intelligent Vehicles Symposium Workshop, pages 350-357, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "type": "text", + "content": "[95] Darren Tsai, Julie Stephany Berrio, Mao Shan, Eduardo Nebot, and Stewart Worrall. Ms3d: Leveraging multiple" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14659" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "type": "text", + "content": "detectors for unsupervised domain adaptation in 3d object detection. arXiv preprint arXiv:2304.02431, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 95, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 95, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 54, + 95, + 287, + 139 + ], + "type": "text", + "content": "[96] Darren Tsai, Julie Stephany Berrio, Mao Shan, Eduardo Nebot, and Stewart Worrall. Ms3d++: Ensemble of experts for multi-source unsupervised domain adaption in 3d object detection. arXiv preprint arXiv:2308.05988, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 54, + 140, + 287, + 183 + ], + "type": "text", + "content": "[97] Ozan Unal, Dengxin Dai, and Luc Van Gool. Scribblesupervised lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2697-2707, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 184, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 184, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 54, + 184, + 287, + 237 + ], + "type": "text", + "content": "[98] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 282 + ], + "type": "text", + "content": "[99] Xudong Wang, Zhaowei Cai, Dashan Gao, and Nuno Vasconcelos. Towards universal object detection by domain attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7289-7298, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 327 + ], + "type": "text", + "content": "[100] Xudong Wang, Shufan Li, Konstantinos Kallidromitis, Yusuke Kato, Kazuki Kozuka, and Trevor Darrell. Hierarchical open-vocabulary universal image segmentation. In Advances in Neural Information Processing Systems, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "type": "text", + "content": "[101] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "text", + "content": "[102] Yi Wei, Zibu Wei, Yongming Rao, Jiaxin Li, Jie Zhou, and Jiwen Lu. Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In European Conference on Computer Vision, pages 179-195. Springer, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "type": "text", + "content": "[103] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 503 + ], + "type": "text", + "content": "[104] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards large-scale 3d representation learning with multi-dataset point prompt training. arXiv preprint arXiv:2308.09718, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 504, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 557 + ], + "type": "text", + "content": "[105] Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, and Ling Shao. Polarmix: A general data augmentation technique for lidar point clouds. In Advances in Neural Information Processing Systems, pages 11035-11048, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "type": "text", + "content": "[106] Aoran Xiao, Jiaxing Huang, Dayan Guan, Fangneng Zhan, and Shijian Lu. Transfer learning from synthetic to real lidar point cloud for semantic segmentation. In AAAI Conference on Artificial Intelligence, pages 2795-2803, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 604, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 287, + 669 + ], + "type": "text", + "content": "[107] Aoran Xiao, Jiaxing Huang, Weihao Xuan, Ruijie Ren, Kangcheng Liu, Dayan Guan, Abdulmotaleb El Saddik, Shijian Lu, and Eric Xing. 3d semantic segmentation in the wild: Learning generalized models for adverse-condition point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9382-9392, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[108] Zeqi Xiao, Wenwei Zhang, Tai Wang, Chen Change Loy, Dahua Lin, and Jiangmiao Pang. Position-guided point cloud panoptic segmentation transformer. arXiv preprint arXiv:2303.13509, 2023." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[109] Shaoyuan Xie, Lingdong Kong, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Robobev: Towards robust bird's eye view perception under corruptions. arXiv preprint arXiv:2304.06719, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[110] Shaoyuan Xie, Lingdong Kong, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Benchmarking and analyzing bird's eye view perception robustness to corruptions. Preprint, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[111] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In European Conference on Computer Vision, pages 1-19, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 274 + ], + "type": "text", + "content": "[112] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In IEEE/CVF International Conference on Computer Vision, pages 16024-16033, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "text", + "content": "[113] Jingyi Xu, Weidong Yang, Lingdong Kong, Youquan Liu, Rui Zhang, Qingyuan Zhou, and Ben Fei. Visual foundation models boost cross-modal unsupervised domain adaptation for 3d semantic segmentation. arXiv preprint arXiv:2403.10001, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "type": "text", + "content": "[114] Xiang Xu, Lingdong Kong, Hui Shuai, and Qingshan Liu. Frnet: Frustum-range networks for scalable lidar segmentation. arXiv preprint arXiv:2312.04484, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "type": "text", + "content": "[115] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European Conference on Computer Vision, pages 677-695, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "type": "text", + "content": "[116] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d: Self-training for unsupervised domain adaptation on 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10368-10378, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 477, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 545, + 532 + ], + "type": "text", + "content": "[117] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: Towards a unified multi-task network for lidar perception. In AAAI Conference on Artificial Intelligence, pages 3231-3240, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "type": "text", + "content": "[118] Bo Zhang, Jiakang Yuan, Botian Shi, Tao Chen, Yikang Li, and Yu Qiao. Uni3d: A unified baseline for multi-dataset 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9253-9262, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "text", + "content": "[119] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianfeng Gao, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. arXiv preprint arXiv:2303.08131, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "text", + "content": "[120] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. In Advances in Neural Information Processing Systems, pages 10326-10338, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[121] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In IEEE/CVF Conference on Com" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14660" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 520 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 72, + 73, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 286, + 94 + ], + "type": "text", + "content": "puter Vision and Pattern Recognition, pages 9601-9610, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "type": "text", + "content": "[122] Zihui Zhang, Bo Yang, Bing Wang, and Bo Li. Growsp: Unsupervised semantic segmentation of 3d point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17619-17629, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "type": "text", + "content": "[123] Yiming Zhao, Lin Bai, and Xinming Huang. Fidnet: Lidar point cloud semantic segmentation with fully interpolation decoding. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4453-4458, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 229 + ], + "type": "text", + "content": "[124] Qiang Zhou, Yuang Liu, Chaohui Yu, Jingliang Li, Zhibin Wang, and Fan Wang. Lmseg: Language-guided multi-dataset segmentation. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "type": "text", + "content": "[125] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 276, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 286, + 319 + ], + "type": "text", + "content": "[126] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7571-7580, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 320, + 286, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 286, + 364 + ], + "type": "text", + "content": "[127] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-polararnet: Proposal-free lidar point cloud panoptic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13194-13203, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 365, + 286, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 286, + 419 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 286, + 419 + ], + "type": "text", + "content": "[128] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 421, + 286, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 421, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 49, + 421, + 286, + 475 + ], + "type": "text", + "content": "[129] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multisensor fusion for 3d lidar semantic segmentation. In IEEE/CVF International Conference on Computer Vision, pages 16280-16290, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 476, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 476, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 49, + 476, + 286, + 520 + ], + "type": "text", + "content": "[130] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. In Advances in Neural Information Processing Systems, 2023." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14661" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/4b27d7d8-e816-4974-8903-c5d6fd2b2750_content_list.json b/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/4b27d7d8-e816-4974-8903-c5d6fd2b2750_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..328447a04d911568531caabe35dbabf132dc0772 --- /dev/null +++ b/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/4b27d7d8-e816-4974-8903-c5d6fd2b2750_content_list.json @@ -0,0 +1,1423 @@ +[ + { + "type": "text", + "text": "Multi-Task Dense Prediction via Mixture of Low-Rank Experts", + "text_level": 1, + "bbox": [ + 166, + 130, + 803, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuqi Yang $^{1,2*}$ Peng-Tao Jiang $^{2*}$ Qibin Hou $^{1,3\\dagger}$ Hao Zhang $^{2}$ Jinwei Chen $^{2}$ Bo Li $^{2}$ $^{1}$ VCIP, CS, Nankai University $^{2}$ vivo Mobile Communication Co., Ltd $^{3}$ NKIARI, Shenzhen Futian yangyq2000@mail.nankai.edu.cn, pt.jiang@vivo.com, andrewhoux@gmail.com", + "bbox": [ + 86, + 179, + 898, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous multi-task dense prediction methods based on the Mixture of Experts (MoE) have received great performance but they neglect the importance of explicitly modeling the global relations among all tasks. In this paper, we present a novel decoder-focused method for multitask dense prediction, called Mixture-of-Low-Rank-Experts (MLoRE). To model the global task relationships, MLoRE adds a generic convolution path to the original MoE structure, where each task feature can go through this path for explicit parameter sharing. Furthermore, to control the parameters and computational cost brought by the increase in the number of experts, we take inspiration from LoRA and propose to leverage the low-rank format of a vanilla convolution in the expert network. Since the low-rank experts have fewer parameters and can be dynamically parameterized into the generic convolution, the parameters and computational cost do not change much with the increase of experts. Benefiting from this design, we increase the number of experts and its reception field to enlarge the representation capacity, facilitating multiple dense tasks learning in a unified network. Extensive experiments on the PASCAL-Context and NYUD-v2 benchmarks show that our MLoRE achieves superior performance compared to previous state-of-the-art methods on all metrics. Our code is available at https://github.com/YuqiYang213/MLoRE.", + "bbox": [ + 76, + 300, + 473, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 707, + 209, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Computer vision tasks, such as semantic segmentation [5, 29, 32, 51] and depth estimation [2, 37], have been significantly facilitated by deep learning techniques. Each vision task has its own elaborated deep models that usually follow a similar pipeline, i.e., feature extraction and prediction. Besides, some tasks also share relations. These facts motivate researchers to study Multi-Task Learning (MTL) that is able to unify different task models into a single one.", + "bbox": [ + 75, + 732, + 468, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/37a04ce7301efed64ff44816185c6f102aa24a509590b8abc6ab5e449839c131.jpg", + "image_caption": [ + "Figure 1. Performance comparison with state-of-the-art methods. Our MLoRE based on the proposed mixture of low-rank experts achieves superior performance on all tasks. $\\uparrow$ denotes higher is better. $\\downarrow$ denotes lower is better." + ], + "image_footnote": [], + "bbox": [ + 500, + 263, + 895, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The significant advantage of multi-task learning lies in that it can improve training and inference efficiency while keeping commensurate performance to each task model. Due to such an advantage, MTL models have been applied to several directions, including autonomous driving [23, 28, 52] and scene understanding [46, 48], etc.", + "bbox": [ + 496, + 595, + 893, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we focus on multi-task learning for dense scene understanding, in which each task predicts pixel-wise results. A research line of early works [4, 16, 33, 35, 45, 46, 48] focuses on designing delicate network architectures, including encoder-focused and decoder-focused methods. The encoder-focused methods [16, 33, 35] design handcrafted modules to share across task-specific encoders to construct task-generic features, while the decoder-focused methods [4, 45, 46, 48] propose tailored decoders to learn discriminative task-specific representations and build cross-task relationships.", + "bbox": [ + 496, + 688, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike the above methods focusing on designing static network architectures, some methods [6, 10, 14] introduce the mixture-of-experts (MoE) technique that sheds light on", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The first two authors contributed equally to this paper. Work was done when Yuqi Yang was an intern at vivo.", + "bbox": [ + 76, + 862, + 467, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Qibin Hou is the corresponding author.", + "bbox": [ + 96, + 887, + 308, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27927", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/ef3d696c3d423f981dccfac7b1520fe90c708406ab56da4a0e327a5a37ededde.jpg", + "table_caption": [ + "Table 1. Parameters and FLOPs of different settings in the standard MoE and the proposed MLoRE. The left setting presents the number of experts and the kernel size of the convolutions in the expert network." + ], + "table_footnote": [], + "table_body": "
SettingsParams (M)FLOPs (G)
MoEMLoREMoEMLoRE
5 experts, [1×1, 1×1]3.11.23.001.49
10 experts, [1×1, 1×1]4.71.64.491.58
15 experts, [1×1, 1×1]6.31.95.991.66
5 experts, [3×3, 1×1]14.93.414.247.12
10 experts, [3×3, 1×1]22.44.721.377.21
15 experts, [3×3, 1×1]29.96.028.497.29
", + "bbox": [ + 78, + 148, + 470, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "learning dynamic and automatic manner of task specialization and cooperations [10]. They utilize MoE to design the encoder blocks and dynamically select network paths for different tasks and different samples. However, compared to these encoder-focused methods, the utilization of MoE in the decoder is less studied, where only Ye et al. [50] first applied MoE to the decoder recently. It decodes task-specific features by dynamically assembling task-generic features from different experts and achieves superior performance than previous decoder-focused methods. This motivates us to delve deep into the MoE-based MTL decoder.", + "bbox": [ + 75, + 311, + 468, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Benefiting from the dynamic routing process, these MoE-based methods can significantly improve the diversity in both parameters and features, resulting in more discriminative task-specific features. However, such a paradigm still has some limitations. First, although MoE-based methods can build connections in a subset of tasks by sharing the same expert in the dynamic routing process, the chances of expert sharing among all the tasks are very low, potentially hindering the router from building global relationships among all tasks. Nevertheless, the global relationship modeling among all tasks has been proven useful for dense MTL in [48, 49]. Based on this fact, we argue it is essential to explicitly model global relationship among all tasks in MoE. Furthermore, the capacity of the task-generic feature spaces is highly related to the number of experts. As empirically demonstrated in [11, 15], increasing the number of experts has the potential to facilitate multi-task learning in a unified network. However, for existing dense multi-task learning, adding more expert networks will introduce more parameters and computational costs, which is a heavy burden for the whole model.", + "bbox": [ + 75, + 478, + 468, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issues, we propose a novel decoder-focused approach, which we call it Mixture of Low-Rank Experts (MLoRE). The core idea of the MLoRE framework is to explicitly model global relationships among all tasks in the MoE and free the MoE from heavily dense computation burdens when increasing the number of experts to enlarge the task-generic feature spaces and context. To address the", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "first issue, MLoRE builds upon the basic MoE structure and introduces a task-sharing generic convolution path that parallels with the MoE. Specifically, the backbone features are first projected to different task features and then all of them are fed into the generic convolution path and the original MoE's expert networks. By simply sharing the same generic path among all tasks, different tasks can be globally correlated with each other. Furthermore, to enhance the discrimination of task-specific features, we exclude some experts from the dynamic routing process and enable them to serve specific tasks.", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To increase the number of experts while not bringing too many parameters and FLOPs, we take inspiration from LoRA that the basic models adapted to different tasks only needs low-rank weight updates. To be specific, we transform the expert networks of the MoE to the different low-rank formats of a vanilla convolution, which saves more than $60\\%$ of the parameters compared to the standard MoE module, as shown in Tab. 1. In addition, to control the computational cost brought by the increasing number of expert networks, we do not use any non-linear activation functions in all the expert networks and the generic convolution path, which enables re-parameterization during inference. Through re-parameterization, the knowledge of experts can be injected into the generic convolution path, reducing the computational cost for dense tasks. To our knowledge, we are the first to use linear experts in MoE for multi-task dense prediction. To verify the effectiveness of our method, we conduct comprehensive experiments on the PASCAL-Context and NYUD-v2 datasets. Our method achieves new performance records on all tasks as shown in Fig. 1.", + "bbox": [ + 496, + 258, + 890, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are three-fold:", + "bbox": [ + 517, + 561, + 818, + 575 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We analyze the issues of MoE when applied in multi-task learning and propose a novel decoder-focused framework, MLoRE, which can explicitly model global relationships among all tasks and enlarge the capacity of feature representations without increasing the model size too much.", + "- We introduce a simple task-sharing generic path to the MoE structure and propose linear and low-rank expert networks inspired by LoRA. The generic convolution path and low-rank expert paths can be linearly combined, enabling re-parameterization at inference.", + "- Experiments on PASCAL-Context and NYUDv2 show that the proposed method clearly outperforms previous state-of-the-art MTL methods on all tasks." + ], + "bbox": [ + 500, + 577, + 890, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 790, + 640, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Dense Multi-Task Learning", + "text_level": 1, + "bbox": [ + 500, + 815, + 746, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In computer vision, multi-task learning (MTL) for dense prediction tasks has been widely studied. The previous methods can be divided into two categories, including optimization-based methods and architecture-based meth", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27928", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ods [44]. Optimization methods [8, 9, 17, 27, 55] facilitate MTL by utilizing different strategies to balance the influence of each task. Architecture-based methods aim to design a unified deep network for MTL. They can be further classified into two categories, encoder-focused methods, and decoder-focused methods. (i) Encoder-focused methods [3, 18, 33, 38] design multi-task backbones to extract features adapted to different tasks. Typical methods include cross-stitch networks [35], neural discriminative dimensionality reduction networks [16], multi-task attention networks [30], branched networks [43], and mixture-of-expert networks [10, 14]. (ii) Decoder-focused methods [4, 45, 46, 48-50, 53, 54, 56] share the same backbone and design delicate heads to extract task-specific features for each task and cross-task relationships. The advantage of decoder-focused methods lies in that they can benefit from powerful off-the-shelf vision backbones, such as DINOv2 [36]. Our method also falls into the decoder-focused category and studies how to produce task-specific features with the MoE technique.", + "bbox": [ + 75, + 90, + 472, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Mixture-of-Experts", + "text_level": 1, + "bbox": [ + 76, + 404, + 264, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mixture-of-Experts (MoE) [24, 25] learns multiple expert networks and a router network that controls the probability of each expert contributing to the final output. This technique is also used in multi-task learning, which can better adapt to the data diversity. Different expert networks learn different discriminative features. The router network learns hard/soft task-specific coefficients to dynamically assemble discriminative features for each task. Prior MoE-based multi-task methods [6, 10, 14] are mainly encoder-focused methods. They introduce MoE into the backbone blocks to sparsely activate different paths for different tasks in the inference stage. Recently, Ye et al. [50] first introduced the MoE technique to the decoder. They utilize the spatial context-aware gates to combine each pixel of the features from different expert networks. The above methods decompose the backbone feature into multiple generic feature spaces and assemble discriminative task-specific features from them. Unlike the above MoE-based MTL method, our method first builds global relationships among all tasks explicitly in the MoE structure rather than leaving this work implicitly done by the task-specific routers. Moreover, our proposed low-rank experts give MoE better efficiency compared to the naive MoE and this gap gradually gets larger as the number of experts increases.", + "bbox": [ + 75, + 428, + 472, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Low-Rank Structure", + "text_level": 1, + "bbox": [ + 76, + 801, + 272, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The low-rank structure is often used in deep learning for its efficiency [22, 40, 42, 47]. Recently, many methods in parameter-efficient adaptation [21, 26, 31, 41], such as LoRA [21], utilize the low-rank structure and have shown impressive results. LoRA takes inspiration from Aghajanyan et", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "al. [1] that the difference in weights between the pre-trained model and the adapted model resides on low intrinsic rank. It learns an extra low-rank matrix instead of tuning the whole layer for adaptation. More related to our work, earlier MTL methods [40, 47] utilize low-rank structure to model task-generic features and generate task-specific features by linear combination. Unlike them, our method utilizes the low-rank structure to control the computation budget when increasing the number of experts in MoE.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 238, + 591, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Framework", + "text_level": 1, + "bbox": [ + 500, + 263, + 689, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework follows the multi-scale architecture as in previous works [49, 50]. We utilize an off-the-shelf vision transformer (ViT) as the encoder and collect multiscale features from different layers. Formally, given an input image $\\mathbf{I}$ and a vision transformer $\\mathcal{F}$ , we can attain multi-scale feature sets $\\{\\mathbf{X}^l = \\mathcal{F}^l (\\mathbf{I})\\}$ from different layers, where $\\mathbf{X}^l\\in \\mathbb{R}^{L\\times C}$ . Here, $l$ , $L = H\\times W$ , and $C$ denote the layer index, the number of patches, and the feature dimension, respectively. $\\mathcal{F}^l (\\mathbf{I})$ is the output features of the $l$ -th transformer layer. The multi-scale features are fed into the decoder which includes two stacked MLoRE modules at each scale. For each task, the output features of MLoRE from different scales are concatenated together to generate the final task features for dense predictions.", + "bbox": [ + 496, + 286, + 893, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Preliminaries: Mixture of Experts", + "text_level": 1, + "bbox": [ + 500, + 506, + 799, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Before describing the details of the proposed Mixture-of-Low-Rank-Experts (MLoRE) module, we first introduce the basic form of MoE $f_{moe}(\\cdot)$ . Formally, suppose MoE contains $N$ experts and $T$ router networks, denoted as $\\mathbb{E} = \\{E_1,E_2,\\dots,E_N\\}$ and $\\mathbb{G} = \\{G_1,G_2,\\dots,G_T\\}$ , respectively. $N$ and $T$ are the number of experts and the number of tasks, respectively. The backbone feature $\\mathbf{X}^l$ from the $l$ -th layer is fed into the networks of $N$ experts and $T$ router networks, respectively. For convenience, the superscript $l$ is omitted in the following. For the $n$ -th expert, the discriminative output feature is generated by $\\mathbf{X}_n = E_n(\\mathbf{X})$ . In the meanwhile, MoE learns gates from the task-specific router networks for different tasks. For task $t$ , the gate value for each expert generated by the router network can be represented as $\\mathbf{g}^t = G_t(\\mathbf{X})$ , where $\\mathbf{g}^t\\in \\mathbb{R}^N$ . Finally, MoE generates task-specific feature $\\mathbf{S}^t$ for task $t$ by utilizing gates to combine the expert features, which can be formulated as", + "bbox": [ + 496, + 527, + 893, + 785 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} ^ {t} = f _ {\\text {m o e}} (\\mathbf {X}) = \\sum_ {n = 1} ^ {N} \\mathbf {g} _ {n} ^ {t} \\mathbf {X} _ {n}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 791, + 890, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For $\\mathrm{M}^3\\mathrm{ViT}$ [14] and Mod-Squad [10], they deactivate some experts according to the corresponding gate values for one inference and select the top- $k$ experts. The task-specific features are used to make predictions for each task.", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27929", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a24f7be17b8b52312e07257e3993109dfc5c625c26e2c4292c96f4630fa451a1.jpg", + "image_caption": [ + "Figure 2. Overall framework of the proposed method. The MLoRE modules are equipped at different layers, where the backbone features from different layers are fed into the MLoRE modules, respectively. At each selected layer, the backbone feature is first projected to different task features and then sent to the task-sharing convolution, task-sharing low-rank expert networks followed by the task-specific router network and task-specific low-rank expert networks. The outputs of these branches are accumulated to generate task-specific features. At each selected layer, we stack two MLoRE modules." + ], + "image_footnote": [], + "bbox": [ + 84, + 84, + 895, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The advantages of MoE lie in that it enables the dynamic encoding of features for each sample and each task and increases feature encoding diversity via multiple experts. However, when applying the MoE technique to build an MTL decoder, we find that it struggles with building global task relationships. Moreover, when increasing the number of experts to enlarge the capacity of feature representations and the context of the expert network, the parameters and computational cost increase accordingly. To address the above issues, we propose the Mixture-of-Low-Rank-Expert (MLoRE) module.", + "bbox": [ + 75, + 398, + 472, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Mixture of Low-Rank Experts", + "text_level": 1, + "bbox": [ + 76, + 574, + 346, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall pipeline of the proposed Mixture-of-Low-Rank-Expert (MLoRE) module is illustrated in Fig. 2. Considering building cross-task relations across the task-specific features, we first project the backbone features into task features using several lightweight convolutional layers. Then, the feature of each task is delivered to the task-sharing generic path and multiple task-sharing expert networks with different low-rank convolutions. The low-rank experts are selected according to the predictions of the task-specific router network for each task. Furthermore, except for building task-specific features based on task-specific routers, we introduce additional task-specific low-rank expert networks to assist in building more discriminative task-specific features. The features from the task-sharing generic path, task-sharing low-rank expert networks selected by task-specific routers and the task-specific expert networks are summed up to generate discriminative task-specific features. We introduce linearity into the MLoRE module and do not utilize any activation function in all paths, enabling reparameterization to reduce computational cost.", + "bbox": [ + 75, + 598, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To be specific, the backbone feature $\\mathbf{X}$ at the $l$ -th layer is first projected to multiple task splits corresponding to each task using $1\\times 1$ convolutions, which can be denoted as $\\{\\mathbf{X}^t = f_{t,1\\times 1}(\\mathbf{X}),t\\in [1,\\dots ,T]\\}$ , where $\\mathbf{X}^t\\in \\mathbb{R}^{C\\times H\\times W}$ . Then, the task feature is sent to three paths, i.e., task-sharing generic path $f_{g}(\\cdot)$ , task-sharing low-rank expert path $f_{lre}(\\cdot)$ with task-specific router network $f_{sr}^{t}(\\cdot)$ , and task-specific low-rank expert path $f_{se}(\\cdot)$ . The output task-specific feature $\\mathbf{S}^t$ is obtained by", + "bbox": [ + 496, + 398, + 893, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} ^ {t} = f _ {g} (\\mathbf {X} ^ {t}) + f _ {s r} ^ {t} \\left(f _ {l r e} (\\mathbf {X} ^ {t})\\right) + f _ {s e} ^ {t} (\\mathbf {X} ^ {t}). \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 544, + 890, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following, we introduce the network details of all these paths in the MLoRE module.", + "bbox": [ + 496, + 574, + 890, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task-sharing generic path contains a simple $3 \\times 3$ convolutional layer with a weight matrix $\\mathbf{W}_g \\in \\mathbb{R}^{3 \\times 3 \\times C \\times C}$ and a bias matrix $\\mathbf{b}_g \\in \\mathbb{R}^C$ . As all task features will go through this generic convolution, it will be optimized by the gradients of different tasks simultaneously, which can help extract common features among all tasks. During the training process, we stop the gradients of this path for further back-propagation. The gradient is back-propagated through the other two paths. We found such a simple operation can better ease the optimization process and well solve the gradient conflicts. Experimental results in Sec. 4.2 show that the simple task-sharing generic path can bring performance improvements on all tasks, demonstrating the effectiveness of the idea of explicitly building the relationship across all tasks from a global perspective.", + "bbox": [ + 496, + 609, + 893, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task-sharing low-rank expert path. We take inspiration from LoRA [21] and adopt low-rank convolution which is a low-rank format of a vanilla convolution. Each task-sharing expert network shares a similar structure consisting of a", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27930", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$3 \\times 3$ convolution and a $1 \\times 1$ convolution. The weights and bias of all task-sharing expert networks can be formulated as $\\{\\mathbf{W}_{l\\text{re}b}^{n}, \\mathbf{b}_{l\\text{re}b}^{n}, \\mathbf{W}_{l\\text{rea}}^{n}, \\mathbf{b}_{l\\text{rea}}^{n} | n \\in [1, \\dots, N]\\}$ , where $\\mathbf{W}_{l\\text{re}b}^{n} \\in \\mathbb{R}^{3 \\times 3 \\times C \\times r_{n}}$ , $\\mathbf{b}_{l\\text{re}b}^{n} \\in \\mathbb{R}^{r_{n}}$ , $\\mathbf{W}_{l\\text{rea}} \\in \\mathbb{R}^{1 \\times 1 \\times r_{n} \\times C}$ , and $\\mathbf{b}_{l\\text{rea}}^{n} \\in \\mathbb{R}^{C}$ ( $r_{n} \\ll C$ ). $r_{n}$ denotes the rank for the $n$ -th expert network. In our method, $r_{n}$ for different expert networks is different, aiming to improve the diversity of parameters and features. For each task, the task-specific router network $f_{sr}^{t}(\\cdot)$ learns gate values for these experts and activate the top- $k$ experts according to the gate values. The output features of all the activated experts are summed up and then sent to a BatchNorm layer to generate task-specific features. The BatchNorm layer contains four parameters, including the accumulated channel-wise mean $\\mu \\in \\mathbb{R}^{C}$ , the accumulated channel-wise standard deviation $\\sigma \\in \\mathbb{R}^{C}$ , the scaling factor $\\gamma \\in \\mathbb{R}^{C}$ and the $\\beta \\in \\mathbb{R}^{C}$ , respectively.", + "bbox": [ + 76, + 90, + 472, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task-specific low-rank expert path includes $T$ specific expert networks and each processes one task feature. For each specific expert network, we utilize a similar structure with task-sharing expert path, which contains a $3 \\times 3$ convolution with a weight matrix $\\mathbf{W}_{seb}^{t} \\in \\mathbb{R}^{3 \\times 3 \\times C \\times R}$ and a bias matrix $\\mathbf{b}_{seb}^{t} \\in \\mathbb{R}^{R}$ , followed by a $1 \\times 1$ convolution with a weight matrix $\\mathbf{W}_{sea}^{t} \\in \\mathbb{R}^{1 \\times 1 \\times R \\times C}$ and a bias matrix $\\mathbf{b}_{sea}^{t} \\in \\mathbb{R}^{C}$ . $R$ denotes the rank number ( $R \\ll C$ ). The task-specific expert path can enhance the distinctiveness of the task-specific features, which will be verified in the experiments.", + "bbox": [ + 76, + 339, + 472, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Router network. As shown in Fig. 2, to generate task-specific features from the task-sharing low-rank expert path, we learn task-specific router networks to generate gate values for each expert and utilize them as the weight of the linear combination of feature output from different experts. The router network for each task is usually the simple linear layers followed by an average pooling layer and a prediction layer. Specifically, the router network $f_{sr}^{t}(\\cdot)$ for the $t$ -th task is designed as follows. Our router network takes the task-specific features $\\mathbf{X}^t \\in \\mathbb{R}^{C \\times H \\times W}$ as input and feed them into two consecutive $1 \\times 1$ convolutions, mapping the channel dimension from $C$ to $C/4$ , followed by a global pooling layer. The output is a global feature vector $\\mathbf{X}_f \\in \\mathbb{R}^{\\frac{C}{4}}$ .", + "bbox": [ + 76, + 497, + 472, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, inspired by previous works [19, 20] showing that positional information is also important for modeling long-range spatial context, we introduce another parallel position-aware branch. Similarly, it consists of two linear layers. The first linear layer shrinks the feature along the spatial dimension, mapping the shape from $\\mathbb{R}^{C\\times HW}$ to $\\mathbb{R}^{C\\times 1}$ , which will then be transformed to $\\mathbb{R}^{\\frac{C}{4}}$ via the second linear layer. The output feature vectors of these two branches are concatenated along the final dimension and then sent to the final prediction layer followed by a Softmax function to produce the gate values $g_{t}$ for each expert.", + "bbox": [ + 76, + 696, + 472, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Re-parameterization during inference. We introduce linearity into the MLoRE module by removing all activation", + "bbox": [ + 76, + 869, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "functions, enabling the parameters of all paths to reparameterize to a simple $3 \\times 3$ convolution for each task at inference. We first parameterize the task-sharing low-rank expert path and then parameterize the parameters of all paths. According to [12], the weight and bias matrices in the task-sharing expert path can be combined and formulated as:", + "bbox": [ + 496, + 90, + 893, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} _ {l r e} ^ {t} = \\mathfrak {B} \\left(\\frac {\\gamma}{\\sigma}\\right) \\sum_ {k \\in \\mathbb {K} _ {t}} \\mathbf {g} _ {k} ^ {t} \\mathbf {W} _ {l r e b} ^ {k} \\mathbf {W} _ {l r e a} ^ {k}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 191, + 893, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} _ {l r e} ^ {t} = \\frac {\\gamma}{\\sigma} (\\sum_ {k \\in \\mathbb {K} _ {t}} \\mathbf {g} _ {k} ^ {t} (\\mathbf {b} _ {l r e b} ^ {k} \\mathbf {W} _ {l r e a} ^ {k} + \\mathbf {b} _ {l r e a} ^ {k}) - \\mu) + \\beta , (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 229, + 893, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathfrak{B}$ denotes the broadcast operation and $\\mathbb{K}_t$ denotes the index set of the activated experts selected by the router network for task $t$ . $\\mathbf{g}_k^t$ is the $k$ -th gate value predicted by the router. At the inference time, the weight matrix and bias matrix of all these three paths can be re-parameterized as", + "bbox": [ + 496, + 273, + 893, + 351 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} _ {r} ^ {t} = \\mathbf {W} _ {g} + \\mathbf {W} _ {s r} ^ {t} + \\mathbf {W} _ {s e b} ^ {t} \\mathbf {W} _ {s e a} ^ {t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 361, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} _ {r} ^ {t} = \\mathbf {b} _ {g} + \\mathbf {b} _ {s r} ^ {t} + \\mathbf {b} _ {s e b} ^ {t} \\mathbf {W} _ {s e a} ^ {t} + \\mathbf {b} _ {s e a} ^ {t}. \\qquad (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 381, + 890, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{W}_r^t$ and $\\mathbf{b}_r^t$ are the weight and bias of the re-parameterized convolution. Thus, Eqn. (2) can be reformulated as", + "bbox": [ + 496, + 410, + 893, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} ^ {t} = \\mathbf {X} ^ {t} * \\mathbf {W} _ {r} ^ {t} + \\mathfrak {B} \\left(\\mathbf {b} _ {r} ^ {t}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 452, + 890, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\oplus$ denotes the convolution operation and $\\mathbf{b}_r^t$ is with the same shape as $\\mathbf{X}^t$ via broadcast.", + "bbox": [ + 496, + 481, + 890, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 526, + 633, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 498, + 550, + 707, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. To demonstrate the effectiveness of our method, we evaluate the performance of our method on two popular multi-task datasets, including PASCAL-Context [7] and NYUD-v2 [39]. PASCAL-Context [7] contains high-quality annotations of several tasks, including semantic segmentation, human parsing, saliency detection, surface normals, and object boundary detection. There are 4,998 training images and 5,105 test images in this dataset. NYUDv2 [39] also provides high-quality multi-task annotations, including semantic segmentation, monocular depth estimation, surface normals, and object boundary detection. This dataset contains 795 training images and 654 test images.", + "bbox": [ + 496, + 579, + 893, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We introduce the evaluation metrics for the tasks mentioned above. Following previous multi-task works [48, 49], the mean intersection-over-union (mIoU) is used to evaluate semantic segmentation and human parsing. The root mean square error (RMSE) is used to evaluate the accuracy of monocular depth estimation. Saliency detection uses the maximum F-measure (maxF). Surface normal and object boundary detection adopt the mean error (mErr) and the optimal-dataset-scale F-measure", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27931", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/eaf5467696709639d2777ccf3ca68f62373d02423517a7ac6bcaa32b7d08eadd.jpg", + "table_caption": [ + "Table 2. Ablation study on different components in MLoRE on the PASCAL-Context dataset. Each row adds an extra setting to the above row. MoE: the standard mixture-of-experts structure; LoRE: task-sharing low-rank expert path; GC: task-sharing generic convolution path; SPE: task-specific expert path. $\\uparrow$ denotes higher is better. $\\downarrow$ denotes lower is better." + ], + "table_footnote": [], + "table_body": "
SettingsSemsegParsingSal.NormalBound.MTLFLOPs#Param
mIoU↑ mIoU↑ maxF↑ mErr ↓odsF↑ Δm ↑(G)(M)
Baseline77.3865.1585.0813.7969.87-3.41391115
+ MoE78.5666.7885.1813.5773.91-1.201834676
Baseline
+ LoRE78.3866.2185.1513.7173.53-1.71568213
+ GC79.2567.4385.2013.7074.38-0.88568243
+ SPE79.2667.8285.3113.6574.69-0.58568259
", + "bbox": [ + 76, + 174, + 470, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(odsF) as the evaluation metrics, respectively. In total, we evaluate the MTL gain $\\Delta_{m}$ across all tasks, following [34].", + "bbox": [ + 75, + 324, + 468, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training settings. We utilize ViT-large [13] as the backbone. The channel number of the decoder is set to 384. For the ablation study, the ViT-base network is set as the backbone. Following the previous work [49], the proposed MTL network is trained for 40,000 iterations with a batch size of 4 on both the two datasets. The optimizer and the loss functions for different tasks follow the previous work [49].", + "bbox": [ + 75, + 359, + 468, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 478, + 230, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we conduct extensive experiments to demonstrate the effectiveness of different components and find the best settings of different hyper-parameters. All the ablation experiments are conducted based on the ViT-based backbone if not specified otherwise. The baseline is built upon the ViT-base backbone with 12 layers, where the backbone features from the 3-rd, 6-th, 9-th, and 12-th layers are utilized as the multi-scale features, each of which is followed by a linear layer to project the channel dimension to the output channels for each task.", + "bbox": [ + 75, + 502, + 468, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of different components. We first conduct experiments to verify the effectiveness of different components of the MLoRE module. The quantitative results are shown in Tab. 2. We first examine the performance of the baseline with the standard MoE and the parameter size and FLOPs of their model. The expert networks in the standard MoE (15 experts) are similar to ours, each of which consists of a $3 \\times 3$ convolution and a $1 \\times 1$ convolution with ReLU among them. When adding the MoE to the baseline, we observe the MTL gain can be improved, but the parameters and FLOPs also expand about 5 times and 4 times, which is a heavy burden for the whole network. When adding the low-rank expert networks (LoRE) to the baseline, the performance is also improved, but the parameters and FLOPs are only $1/3$ and $1/3$ of the MoE-based model. When the low-rank property is applied to the expert net", + "bbox": [ + 75, + 657, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9c18f9cf1cc382e03a5ee457bc090eab8542b7344872bab0b6902d7ef1a3b319.jpg", + "image_caption": [ + "Figure 3. Ablation study on the number of experts $N$ and the number of activated experts $K$ . In the right figure, we also present the parameter change of the MLoRE module with the increase in the number of experts." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 692, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/389e55bcfd2f481bb3ef5ded0807d98532e12d53dd41320dd0696fd19b292a26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 89, + 890, + 196 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4577b6fbb9c3e67bdcf58c379d930855db1cbd5ed7099539b5359784b8a60d2f.jpg", + "table_caption": [ + "Table 3. Ablation on the rank setting in the task-sharing low-rank expert path of MLoRE on PASCAL-Context dataset." + ], + "table_footnote": [], + "table_body": "
Min/Max RankSemseg mIoU↑Parsing mIoU↑Saliency maxF↑Normal mErr↓Boundary odsF↑MTL Δm↑
16/1678.8468.0185.3213.6974.39-0.77
16/12879.2667.8285.3013.6574.69-0.58
128/12878.7966.9885.3513.6774.29-1.07
", + "bbox": [ + 514, + 303, + 879, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "works, the parameter size is reduced several times. When further introducing linearity in the expert networks by removing all activation functions, the computations are saved by re-parameterizing all experts to a single convolution.", + "bbox": [ + 496, + 410, + 890, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we also emphasize the importance of explicitly building global task correlations in the MoE and introduce the task-sharing generic path to achieve this goal. It can be seen that adding the task-sharing generic path to LoRE can further improve the performance and outperform the baseline with MoE on most metrics, which proves the effectiveness of modeling the global relationships among all tasks. In addition, adding a task-specific low-rank expert for each task also improves the performance, which demonstrates that the specialized expert can enhance the discrimination of task-specific features. We empirically set the rank of the task-specific low-rank expert to 64 in this paper.", + "bbox": [ + 496, + 472, + 892, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Number of task-sharing low-rank experts and top- $k$ selection. We ablate the number of low-rank experts in the MLoRE module and top- $k$ selection of the experts by the task-specific router networks. We first fix one and ablate another parameter to study their impact on multi-task performance. As shown in Fig. 3, when increasing the number of experts, the MTL gain of the model is significantly improved and allows us to achieve the best performance when the number of experts is 15. Further increasing the number of experts, we do not observe obvious performance gain. Besides, when fixing the number of experts to 15, we ablate the ratio of the activated experts. We observe that activating $60\\%$ experts for each task is the best choice in our experiments. When selecting all expert networks, the performance decreases by a large margin, which reflects the importance of sparsity for feature discrimination.", + "bbox": [ + 496, + 657, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27932", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/03e7c7917513671af0c74332c9ce3b99dc569c385da03763bf39f9fd37dbd1e8.jpg", + "image_caption": [ + "Figure 4. (a) The relations between tasks and low-rank experts. (b) The ratio of an expert activated by different numbers of tasks in the MLoRE module without the task-sharing generic path. We can see that without the task-sharing generic path, there is only a few experts can be activated by all five tasks. Horizontal coordinates represent the ranks of different experts." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 509, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a7079c070ff50d03056e70eaf56c1db42beec13c3ee2d4b92db3714f4dd0da2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 90, + 875, + 208 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rank number setting. The expert networks utilize the low-rank format of a vanilla $3 \\times 3$ convolution with 640 output channels. The rank $r$ of different expert networks also plays an important role. We study different settings, including 1) all experts with the same rank number 16, 2) all experts with the rank number 128, 3) all experts with the rank numbers from 16 to 128 with an increased step of 8. As shown in Tab. 3, it can be seen that selecting different rank numbers for expert networks achieves the best MTL gain. We analyze the expert networks with different rank numbers that can bring more feature diversity than the same rank numbers, which is more useful for assembling task-specific features, and utilize this as our settings.", + "bbox": [ + 75, + 270, + 472, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The relation visualizations between tasks and the low-rank experts are shown in Fig. 4(a). We count the relations from the second MLoRE module at the last stage and calculate the activation ratio of every expert selected by different tasks on the entire dataset. It can be seen that experts with different ranks tend to learn different subsets of tasks. Specifically, for this module, experts with lower ranks tend to learn shared knowledge for 3-4 related tasks, while experts with higher ranks tend to specialize in 1-2 tasks. Moreover, we also show the ratio of different experts to be activated by different numbers of tasks when the task-sharing generic path is not added to the MLoRE module in Fig. 4(b). It can be seen that in a fully dynamic manner, all these experts are seldom or even never activated by all the tasks in one sample. This proves the fact that almost no expert can learn global relationships across all tasks when utilizing the MoE in the decoder directly. This phenomenon strongly supports the necessity of the task-sharing generic path for explicitly modeling global task relationships.", + "bbox": [ + 75, + 470, + 472, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task-specific router network. The router network is important for generating task-specific gates, which decide how to activate the experts and assemble their features. We ablate several settings of the router network and the results are shown in Tab. 4. Adding the position-aware branch to the basic router network can improve the MTL gain by $+0.17$ . The position-aware branch can obtain more context information, which benefits for the router network. Moreover, when replacing the router's input from the learnable pa", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f6313c28d3a4f91497140fb562ad532b8f932f7c9c5755ee68cbb6dd4c735202.jpg", + "table_caption": [ + "Table 4. Ablation on the setting of the router network. basic: the basic router network. pos.: the positional-aware router network. w/o sample-dep: the input is the learnable parameters." + ], + "table_footnote": [], + "table_body": "
Router VariantSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑MTL Δm ↑
basic79.1567.4085.2113.5874.34-0.75
basic+pos.79.2667.8285.3113.6574.69-0.58
only pos.79.1067.7685.1113.7174.51-0.82
basic79.1567.4085.2113.5874.34-0.75
w/o sample-dep78.8667.3885.4113.6574.25-0.91
", + "bbox": [ + 501, + 314, + 893, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rameters to the sample features, the MTL gain increases by $+0.16$ , which demonstrates that dynamic information from samples is vital for gates.", + "bbox": [ + 498, + 454, + 893, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparisons with Other Methods", + "text_level": 1, + "bbox": [ + 498, + 511, + 799, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantitative comparisons with previous state-of-the-art (SOTA) methods are shown in Tab. 5 and Tab. 6. It can be seen that our method clearly outperforms previous methods in terms of all metrics on both the PASCAL-Context dataset and the NYUDv2 dataset. In particular, on the PASCAL-Context dataset, for semantic segmentation, human parsing, and boundary detection, the performance of our method outperforms the previous best method by $+0.52$ mIoU, $+1.10$ mIoU and $+1.92$ odsF, respectively.", + "bbox": [ + 496, + 536, + 893, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Previous methods, $\\mathrm{M}^3\\mathrm{ViT}$ [14], Mod-Squad [10], and TaskExpert [50] all utilize the MoE technique in their network. However, our method shows performance superior to theirs, which demonstrates the effectiveness of the MLoRE module. Compared with the decoder-focused method, TaskExpert, the performance of semantic segmentation, human parsing, and object boundary is significantly improved by $+0.77$ mIoU, $+1.10$ mIoU and $+2.12$ odsF, but using fewer parameters and FLOPs.", + "bbox": [ + 496, + 672, + 895, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we also present an intuitive visualization comparison among different methods in Fig. 5. Our method can generate better visualization results than the previous SoTA methods, especially for semantic segmentation, human parsing, and object boundary detection. More visualization comparisons are in the supplemental material.", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27933", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/16226faa1e84e8817cdb2665d821db4caf439c99085db9303790b5ee870899b1.jpg", + "table_caption": [ + "Table 5. Quantitative comparison of different methods on PASCAL-Context dataset. * denotes the reproduced performance of methods based on the ViT-large backbone in [50]." + ], + "table_footnote": [], + "table_body": "
MethodPublicationBackboneSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑FLOPs (G)#Param (M)
PAD-Net [46]CVPR'18HRNet1853.6059.6065.8015.3072.5012481
MTI-Net [45]ECCV'20HRNet1861.7060.1884.7814.2370.80161128
ATRC [4]ICCV'21HRNet1867.6762.9382.2914.2472.4221696
PAD-Net* [46]CVPR'18ViT-large78.0167.1279.2114.3772.60773330
MTI-Net* [45]ECCV'20ViT-large78.3167.4084.7514.6773.00774851
ATRC* [4]ICCV'21ViT-large77.1166.8481.2014.2372.10871340
InvPT [48]ECCV'22ViT-large79.0367.6184.8114.1573.00669423
TaskPrompter [49]ICLR'23ViT-large80.8968.8984.8313.7273.50497401
TaskExpert [50]ICCV'23ViT-large80.6469.4284.8713.5673.30622420
Ours-ViT-large81.4170.5284.9013.5175.42571407
", + "bbox": [ + 109, + 121, + 859, + 310 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/325b54e721e51c0e3ea13350955c9556a1ec802963ba129b1428e7364dcc65b6.jpg", + "table_caption": [ + "Table 6. Quantitative comparison of different methods on the NYUD-v2 dataset. Our method performs the best on all four tasks." + ], + "table_footnote": [], + "table_body": "
MethodBackboneSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑
PAD-Net [46]HRNet1836.610.624620.8876.38
MTI-Net [45]HRNet4845.970.536520.2777.86
ATRC [4]HRNet4846.330.536320.1877.94
InvPT [48]ViT-large53.560.518319.0478.10
TaskPrompter [49]ViT-large55.300.515218.4778.20
TaskExpert [50]ViT-large55.350.515718.5478.40
OursViT-large55.960.507618.3378.43
", + "bbox": [ + 78, + 366, + 468, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Efficient MTL Models", + "text_level": 1, + "bbox": [ + 76, + 520, + 284, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also apply our MLoRE module to the ViT-small backbone to check the performance of the efficient models. Specifically, the channel number of the decoder decreases from 384 to 192. As shown in Tab. 7, using about $35\\%$ GFLOPs of TaskExpert, our method can achieve a highly competitive result. In particular, the performance of the semantic segmentation and object boundary is improved by $0.6\\%$ mIoU and $1.01\\%$ odsF while the metrics of other tasks are close to TaskExpert. Furthermore, the parameters are fewer than TaskExpert by 11M.", + "bbox": [ + 75, + 545, + 468, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 709, + 207, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel decoder-focused multi-task learning method, called MLoRE. We delve deep into the standard mixture-of-experts (MoE) technique and improve it for dense multi-task learning from two aspects. First, to address the neglected global relationship modeling in the MoE, we add a simple generic convolution path to MoE, enabling different task features to be able to share this path. Furthermore, we apply the low-rank format of a vanilla convolution to different expert networks to free the MoE from high computational cost and a large number of parameters when increasing the number of experts. Experiments", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ca4a757f080bfad844bb4f5a5c6c306ef67a4cd23488774da1511b1fa0a6992f.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison among different methods, including InvPT [48], TaskPrompter [49], and ours. Best viewed with zoom-in. It can be seen that our method achieves better visual results than other methods on all five tasks thanks to the proposed MLoRE module." + ], + "image_footnote": [], + "bbox": [ + 501, + 330, + 890, + 498 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c5bdef0358164ab92c26fecf7b61bbfd5cc04e5321ef947e1758741d923ac5c6.jpg", + "table_caption": [ + "Table 7. Quantitative comparison of the MoE-based efficient models on PASCAL-Context dataset." + ], + "table_footnote": [], + "table_body": "
MethodSemseg mIoU ↑Parsing mIoU ↑Sal. maxF ↑Nor. mErr ↓Bound. odsF ↑FLOPs (G)Param (M)
M3ViT72.8062.1066.3014.5071.7042042
Mod-Squad74.1062.7066.9013.7072.0042052
TaskExpert75.0462.6884.6814.2268.8020455
ours75.6462.6584.7014.4369.817244
", + "bbox": [ + 501, + 638, + 890, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "demonstrate that the proposed method clearly outperforms previous state-of-the-art methods on all metrics.", + "bbox": [ + 498, + 763, + 890, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This research was supported by NSFC (NO. 62276145), the Fundamental Research Funds for the Central Universities (Nankai University, 070-63223049), CAST through Young Elite Scientist Sponsorship Program (No. YESS20210377). Computations were supported by the Supercomputing Center of Nankai University (NKSC).", + "bbox": [ + 498, + 809, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27934", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Armen Aghajanyan, Luke Zettlemoyer, and Sonal Gupta. Intrinsic dimensionality explains the effectiveness of language model fine-tuning. arXiv preprint arXiv:2012.13255, 2020. 3", + "[2] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4009-4018, 2021. 1", + "[3] David Bruggemann, Menelaos Kanakis, Stamatios Georgoulis, and Luc Van Gool. Automated search for resource-efficient branched multi-task networks. arXiv preprint arXiv:2008.10292, 2020. 3", + "[4] David Brüggemann, Menelaos Kanakis, Anton Obukhov, Stamatios Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In Int. Conf. Comput. Vis., pages 15869-15878, 2021. 1, 3, 8", + "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):834-848, 2017. 1", + "[6] Tianlong Chen, Xuxi Chen, Xianzhi Du, Abdullah Rashwan, Fan Yang, Huizhong Chen, Zhangyang Wang, and Yeqing Li. Adamv-moe: Adaptive multi-task vision mixture-of-experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17346-17357, 2023. 1, 3", + "[7] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1971-1978, 2014. 5", + "[8] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 3", + "[9] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. Advances in Neural Information Processing Systems, 33:2039-2050, 2020. 3", + "[10] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. Mod-squad: Designing mixtures of experts as modular multi-task learners. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11828-11837, 2023. 1, 2, 3, 7", + "[11] Aidan Clark, Diego De Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, et al. Unified scaling laws for routed language models. In Int. Conf. Mach. Learn., pages 4057-4086. PMLR, 2022. 2", + "[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Diverse branch block: Building a convolution as an inception-like unit. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10886-10895, 2021. 5" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] 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. In Int. Conf. Learn. Represent., 2020. 6", + "[14] Zhiwen Fan, Rishov Sarkar, Ziyu Jiang, Tianlong Chen, Kai Zou, Yu Cheng, Cong Hao, Zhangyang Wang, et al. $\\mathbf{M}^3$ vit: Mixture-of-experts vision transformer for efficient multi-task learning with model-accelerator co-design. In Adv. Neural Inform. Process. Syst., volume 35, pages 28441-28457, 2022. 1, 3, 7", + "[15] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022. 2", + "[16] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3205-3214, 2019. 1, 3", + "[17] Michelle Guo, Albert Haque, De-An Huang, Serena Yeung, and Li Fei-Fei. Dynamic task prioritization for multitask learning. In Proceedings of the European conference on computer vision (ECCV), pages 270–287, 2018. 3", + "[18] Pengsheng Guo, Chen-Yu Lee, and Daniel Ulbricht. Learning to branch for multi-task learning. In Int. Conf. Mach. Learn., pages 3854-3863. PMLR, 2020. 3", + "[19] Qibin Hou, Li Zhang, Ming-Ming Cheng, and Jiashi Feng. Strip pooling: Rethinking spatial pooling for scene parsing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4003-4012, 2020. 5", + "[20] Qibin Hou, Daquan Zhou, and Jiashi Feng. Coordinate attention for efficient mobile network design. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13713-13722, 2021. 5", + "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3, 4", + "[22] Yerlan Idelbayev and Miguel A Carreira-Perpinan. Low-rank compression of neural nets: Learning the rank of each layer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8049–8059, 2020. 3", + "[23] Keishi Ishihara, Anssi Kanervisto, Jun Miura, and Ville Hautamaki. Multi-task learning with attention for end-to-end autonomous driving. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2902-2911, 2021. 1", + "[24] Robert A Jacobs and Michael I Jordan. Learning piecewise control strategies in a modular neural network architecture. IEEE Transactions on Systems, Man, and Cybernetics, 23(2):337-345, 1993. 3", + "[25] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. *Neural Comput.*, 3(1):79-87, 1991. 3", + "[26] Shibo Jie and Zhi-Hong Deng. Fact: Factor-tuning for lightweight adaptation on vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 1060-1068, 2023. 3" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27935", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 3", + "[28] Zhihao Li, Toshiyuki Motoyoshi, Kazuma Sasaki, Tetsuya Ogata, and Shigeki Sugano. Rethinking self-driving: Multi-task knowledge for better generalization and accident explanation ability. arXiv preprint arXiv:1809.11100, 2018. 1", + "[29] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1925-1934, 2017. 1", + "[30] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1871-1880, 2019. 3", + "[31] Yen-Cheng Liu, Chih-Yao Ma, Junjiao Tian, Zijian He, and Zsolt Kira. Polyhistor: Parameter-efficient multi-task adaptation for dense vision tasks. Advances in Neural Information Processing Systems, 35:36889-36901, 2022. 3", + "[32] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3431-3440, 2015. 1", + "[33] Yongxi Lu, Abhishek Kumar, Shuangfei Zhai, Yu Cheng, Tara Javidi, and Rogerio Feris. Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5334-5343, 2017. 1, 3", + "[34] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1851-1860, 2019. 6", + "[35] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3994-4003, 2016. 1, 3", + "[36] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3", + "[37] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Int. Conf. Comput. Vis., pages 12179-12188, 2021. 1", + "[38] Sebastian Ruder, Joachim Bingel, Isabelle Augenstein, and Anders Søgaard. Latent multi-task architecture learning. In AAAI Conf. Artif. Intell., volume 33, pages 4822-4829, 2019. 3", + "[39] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In *Eur. Conf. Comput. Vis.*, pages 746-760. Springer, 2012. 5", + "[40] Chi Su, Fan Yang, Shiliang Zhang, Qi Tian, Larry S Davis, and Wen Gao. Multi-task learning with low rank attribute embedding for person re-identification. In Proceedings of the IEEE international conference on computer vision, pages 3739–3747, 2015. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5227-5237, 2022. 3", + "[42] Madeleine Udell, Corinne Horn, Reza Zadeh, Stephen Boyd, et al. Generalized low rank models. Foundations and Trends® in Machine Learning, 9(1):1-118, 2016. 3", + "[43] Simon Vandenhende, Stamatios Georgoulis, Bert De Brabandere, and Luc Van Gool. Branched multi-task networks: deciding what layers to share. In Brit. Mach. Vis. Conf., 2019. 3", + "[44] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 44(7):3614-3633, 2021. 3", + "[45] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In Eur. Conf. Comput. Vis., pages 527-543. Springer, 2020. 1, 3, 8", + "[46] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 675-684, 2018. 1, 3, 8", + "[47] Yongxin Yang and Timothy M Hospedales. Trace norm regularised deep multi-task learning. arXiv preprint arXiv:1606.04038, 2016. 3", + "[48] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. In Eur. Conf. Comput. Vis., pages 514-530. Springer, 2022. 1, 2, 3, 5, 8", + "[49] Hanrong Ye and Dan Xu. Taskprompter: Spatial-channel multi-task prompting for dense scene understanding. In Int. Conf. Learn. Represent., 2022. 2, 3, 5, 6, 8", + "[50] Hanrong Ye and Dan Xu. Taskexpert: Dynamically assembling multi-task representations with memorial mixture-of-experts. In Int. Conf. Comput. Vis., 2023. 2, 3, 7, 8", + "[51] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Eur. Conf. Comput. Vis., pages 325-341, 2018. 1", + "[52] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2636-2645, 2020. 1", + "[53] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Zequn Jie, Xiang Li, and Jian Yang. Joint task-recursive learning for semantic segmentation and depth estimation. In Eur. Conf. Comput. Vis., pages 235–251, 2018. 3", + "[54] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4106-4115, 2019. 3", + "[55] Xiangyun Zhao, Haoxiang Li, Xiaohui Shen, Xiaodan Liang, and Ying Wu. A modulation module for multi-task learning" + ], + "bbox": [ + 501, + 90, + 892, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27936", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with applications in image retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 401-416, 2018. 3", + "[56] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4514–4523, 2020. 3" + ], + "bbox": [ + 78, + 90, + 468, + 191 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27937", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/4b27d7d8-e816-4974-8903-c5d6fd2b2750_model.json b/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/4b27d7d8-e816-4974-8903-c5d6fd2b2750_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2a22d6b21bab437fe078f603106ef9d752c71d1b --- /dev/null +++ b/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/4b27d7d8-e816-4974-8903-c5d6fd2b2750_model.json @@ -0,0 +1,2092 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.804, + 0.154 + ], + "angle": 0, + "content": "Multi-Task Dense Prediction via Mixture of Low-Rank Experts" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.18, + 0.899, + 0.234 + ], + "angle": 0, + "content": "Yuqi Yang\\(^{1,2*}\\) Peng-Tao Jiang\\(^{2*}\\) Qibin Hou\\(^{1,3\\dagger}\\) Hao Zhang\\(^{2}\\) Jinwei Chen\\(^{2}\\) Bo Li\\(^{2}\\) \n\\(^{1}\\)VCIP, CS, Nankai University \\(^{2}\\)vivo Mobile Communication Co., Ltd \\(^{3}\\)NKIARI, Shenzhen Futian yangyq2000@mail.nankai.edu.cn, pt.jiang@vivo.com, andrewhoux@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.301, + 0.474, + 0.681 + ], + "angle": 0, + "content": "Previous multi-task dense prediction methods based on the Mixture of Experts (MoE) have received great performance but they neglect the importance of explicitly modeling the global relations among all tasks. In this paper, we present a novel decoder-focused method for multitask dense prediction, called Mixture-of-Low-Rank-Experts (MLoRE). To model the global task relationships, MLoRE adds a generic convolution path to the original MoE structure, where each task feature can go through this path for explicit parameter sharing. Furthermore, to control the parameters and computational cost brought by the increase in the number of experts, we take inspiration from LoRA and propose to leverage the low-rank format of a vanilla convolution in the expert network. Since the low-rank experts have fewer parameters and can be dynamically parameterized into the generic convolution, the parameters and computational cost do not change much with the increase of experts. Benefiting from this design, we increase the number of experts and its reception field to enlarge the representation capacity, facilitating multiple dense tasks learning in a unified network. Extensive experiments on the PASCAL-Context and NYUD-v2 benchmarks show that our MLoRE achieves superior performance compared to previous state-of-the-art methods on all metrics. Our code is available at https://github.com/YuqiYang213/MLoRE." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.708, + 0.21, + 0.723 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Computer vision tasks, such as semantic segmentation [5, 29, 32, 51] and depth estimation [2, 37], have been significantly facilitated by deep learning techniques. Each vision task has its own elaborated deep models that usually follow a similar pipeline, i.e., feature extraction and prediction. Besides, some tasks also share relations. These facts motivate researchers to study Multi-Task Learning (MTL) that is able to unify different task models into a single one." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.265, + 0.897, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.513, + 0.895, + 0.57 + ], + "angle": 0, + "content": "Figure 1. Performance comparison with state-of-the-art methods. Our MLoRE based on the proposed mixture of low-rank experts achieves superior performance on all tasks. \\(\\uparrow\\) denotes higher is better. \\(\\downarrow\\) denotes lower is better." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.895, + 0.688 + ], + "angle": 0, + "content": "The significant advantage of multi-task learning lies in that it can improve training and inference efficiency while keeping commensurate performance to each task model. Due to such an advantage, MTL models have been applied to several directions, including autonomous driving [23, 28, 52] and scene understanding [46, 48], etc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.854 + ], + "angle": 0, + "content": "In this paper, we focus on multi-task learning for dense scene understanding, in which each task predicts pixel-wise results. A research line of early works [4, 16, 33, 35, 45, 46, 48] focuses on designing delicate network architectures, including encoder-focused and decoder-focused methods. The encoder-focused methods [16, 33, 35] design handcrafted modules to share across task-specific encoders to construct task-generic features, while the decoder-focused methods [4, 45, 46, 48] propose tailored decoders to learn discriminative task-specific representations and build cross-task relationships." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Unlike the above methods focusing on designing static network architectures, some methods [6, 10, 14] introduce the mixture-of-experts (MoE) technique that sheds light on" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.863, + 0.468, + 0.887 + ], + "angle": 0, + "content": "*The first two authors contributed equally to this paper. Work was done when Yuqi Yang was an intern at vivo." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.309, + 0.9 + ], + "angle": 0, + "content": "†Qibin Hou is the corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.863, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27927" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.146 + ], + "angle": 0, + "content": "Table 1. Parameters and FLOPs of different settings in the standard MoE and the proposed MLoRE. The left setting presents the number of experts and the kernel size of the convolutions in the expert network." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.149, + 0.472, + 0.287 + ], + "angle": 0, + "content": "
SettingsParams (M)FLOPs (G)
MoEMLoREMoEMLoRE
5 experts, [1×1, 1×1]3.11.23.001.49
10 experts, [1×1, 1×1]4.71.64.491.58
15 experts, [1×1, 1×1]6.31.95.991.66
5 experts, [3×3, 1×1]14.93.414.247.12
10 experts, [3×3, 1×1]22.44.721.377.21
15 experts, [3×3, 1×1]29.96.028.497.29
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.47, + 0.478 + ], + "angle": 0, + "content": "learning dynamic and automatic manner of task specialization and cooperations [10]. They utilize MoE to design the encoder blocks and dynamically select network paths for different tasks and different samples. However, compared to these encoder-focused methods, the utilization of MoE in the decoder is less studied, where only Ye et al. [50] first applied MoE to the decoder recently. It decodes task-specific features by dynamically assembling task-generic features from different experts and achieves superior performance than previous decoder-focused methods. This motivates us to delve deep into the MoE-based MTL decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Benefiting from the dynamic routing process, these MoE-based methods can significantly improve the diversity in both parameters and features, resulting in more discriminative task-specific features. However, such a paradigm still has some limitations. First, although MoE-based methods can build connections in a subset of tasks by sharing the same expert in the dynamic routing process, the chances of expert sharing among all the tasks are very low, potentially hindering the router from building global relationships among all tasks. Nevertheless, the global relationship modeling among all tasks has been proven useful for dense MTL in [48, 49]. Based on this fact, we argue it is essential to explicitly model global relationship among all tasks in MoE. Furthermore, the capacity of the task-generic feature spaces is highly related to the number of experts. As empirically demonstrated in [11, 15], increasing the number of experts has the potential to facilitate multi-task learning in a unified network. However, for existing dense multi-task learning, adding more expert networks will introduce more parameters and computational costs, which is a heavy burden for the whole model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To address the above issues, we propose a novel decoder-focused approach, which we call it Mixture of Low-Rank Experts (MLoRE). The core idea of the MLoRE framework is to explicitly model global relationships among all tasks in the MoE and free the MoE from heavily dense computation burdens when increasing the number of experts to enlarge the task-generic feature spaces and context. To address the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "first issue, MLoRE builds upon the basic MoE structure and introduces a task-sharing generic convolution path that parallels with the MoE. Specifically, the backbone features are first projected to different task features and then all of them are fed into the generic convolution path and the original MoE's expert networks. By simply sharing the same generic path among all tasks, different tasks can be globally correlated with each other. Furthermore, to enhance the discrimination of task-specific features, we exclude some experts from the dynamic routing process and enable them to serve specific tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.56 + ], + "angle": 0, + "content": "To increase the number of experts while not bringing too many parameters and FLOPs, we take inspiration from LoRA that the basic models adapted to different tasks only needs low-rank weight updates. To be specific, we transform the expert networks of the MoE to the different low-rank formats of a vanilla convolution, which saves more than \\(60\\%\\) of the parameters compared to the standard MoE module, as shown in Tab. 1. In addition, to control the computational cost brought by the increasing number of expert networks, we do not use any non-linear activation functions in all the expert networks and the generic convolution path, which enables re-parameterization during inference. Through re-parameterization, the knowledge of experts can be injected into the generic convolution path, reducing the computational cost for dense tasks. To our knowledge, we are the first to use linear experts in MoE for multi-task dense prediction. To verify the effectiveness of our method, we conduct comprehensive experiments on the PASCAL-Context and NYUD-v2 datasets. Our method achieves new performance records on all tasks as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.562, + 0.819, + 0.576 + ], + "angle": 0, + "content": "In summary, our contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.578, + 0.891, + 0.653 + ], + "angle": 0, + "content": "- We analyze the issues of MoE when applied in multi-task learning and propose a novel decoder-focused framework, MLoRE, which can explicitly model global relationships among all tasks and enlarge the capacity of feature representations without increasing the model size too much." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.654, + 0.891, + 0.728 + ], + "angle": 0, + "content": "- We introduce a simple task-sharing generic path to the MoE structure and propose linear and low-rank expert networks inspired by LoRA. The generic convolution path and low-rank expert paths can be linearly combined, enabling re-parameterization at inference." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.729, + 0.891, + 0.772 + ], + "angle": 0, + "content": "- Experiments on PASCAL-Context and NYUDv2 show that the proposed method clearly outperforms previous state-of-the-art MTL methods on all tasks." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.578, + 0.891, + 0.772 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.791, + 0.642, + 0.807 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.748, + 0.833 + ], + "angle": 0, + "content": "2.1. Dense Multi-Task Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In computer vision, multi-task learning (MTL) for dense prediction tasks has been widely studied. The previous methods can be divided into two categories, including optimization-based methods and architecture-based meth" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27928" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.395 + ], + "angle": 0, + "content": "ods [44]. Optimization methods [8, 9, 17, 27, 55] facilitate MTL by utilizing different strategies to balance the influence of each task. Architecture-based methods aim to design a unified deep network for MTL. They can be further classified into two categories, encoder-focused methods, and decoder-focused methods. (i) Encoder-focused methods [3, 18, 33, 38] design multi-task backbones to extract features adapted to different tasks. Typical methods include cross-stitch networks [35], neural discriminative dimensionality reduction networks [16], multi-task attention networks [30], branched networks [43], and mixture-of-expert networks [10, 14]. (ii) Decoder-focused methods [4, 45, 46, 48-50, 53, 54, 56] share the same backbone and design delicate heads to extract task-specific features for each task and cross-task relationships. The advantage of decoder-focused methods lies in that they can benefit from powerful off-the-shelf vision backbones, such as DINOv2 [36]. Our method also falls into the decoder-focused category and studies how to produce task-specific features with the MoE technique." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.405, + 0.265, + 0.421 + ], + "angle": 0, + "content": "2.2. Mixture-of-Experts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.473, + 0.791 + ], + "angle": 0, + "content": "Mixture-of-Experts (MoE) [24, 25] learns multiple expert networks and a router network that controls the probability of each expert contributing to the final output. This technique is also used in multi-task learning, which can better adapt to the data diversity. Different expert networks learn different discriminative features. The router network learns hard/soft task-specific coefficients to dynamically assemble discriminative features for each task. Prior MoE-based multi-task methods [6, 10, 14] are mainly encoder-focused methods. They introduce MoE into the backbone blocks to sparsely activate different paths for different tasks in the inference stage. Recently, Ye et al. [50] first introduced the MoE technique to the decoder. They utilize the spatial context-aware gates to combine each pixel of the features from different expert networks. The above methods decompose the backbone feature into multiple generic feature spaces and assemble discriminative task-specific features from them. Unlike the above MoE-based MTL method, our method first builds global relationships among all tasks explicitly in the MoE structure rather than leaving this work implicitly done by the task-specific routers. Moreover, our proposed low-rank experts give MoE better efficiency compared to the naive MoE and this gap gradually gets larger as the number of experts increases." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.274, + 0.817 + ], + "angle": 0, + "content": "2.3. Low-Rank Structure" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.903 + ], + "angle": 0, + "content": "The low-rank structure is often used in deep learning for its efficiency [22, 40, 42, 47]. Recently, many methods in parameter-efficient adaptation [21, 26, 31, 41], such as LoRA [21], utilize the low-rank structure and have shown impressive results. LoRA takes inspiration from Aghajanyan et" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "al. [1] that the difference in weights between the pre-trained model and the adapted model resides on low intrinsic rank. It learns an extra low-rank matrix instead of tuning the whole layer for adaptation. More related to our work, earlier MTL methods [40, 47] utilize low-rank structure to model task-generic features and generate task-specific features by linear combination. Unlike them, our method utilizes the low-rank structure to control the computation budget when increasing the number of experts in MoE." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.239, + 0.593, + 0.255 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.264, + 0.691, + 0.279 + ], + "angle": 0, + "content": "3.1. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.895, + 0.501 + ], + "angle": 0, + "content": "The overall framework follows the multi-scale architecture as in previous works [49, 50]. We utilize an off-the-shelf vision transformer (ViT) as the encoder and collect multiscale features from different layers. Formally, given an input image \\(\\mathbf{I}\\) and a vision transformer \\(\\mathcal{F}\\), we can attain multi-scale feature sets \\(\\{\\mathbf{X}^l = \\mathcal{F}^l (\\mathbf{I})\\}\\) from different layers, where \\(\\mathbf{X}^l\\in \\mathbb{R}^{L\\times C}\\). Here, \\(l\\), \\(L = H\\times W\\), and \\(C\\) denote the layer index, the number of patches, and the feature dimension, respectively. \\(\\mathcal{F}^l (\\mathbf{I})\\) is the output features of the \\(l\\)-th transformer layer. The multi-scale features are fed into the decoder which includes two stacked MLoRE modules at each scale. For each task, the output features of MLoRE from different scales are concatenated together to generate the final task features for dense predictions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.507, + 0.8, + 0.523 + ], + "angle": 0, + "content": "3.2. Preliminaries: Mixture of Experts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.895, + 0.786 + ], + "angle": 0, + "content": "Before describing the details of the proposed Mixture-of-Low-Rank-Experts (MLoRE) module, we first introduce the basic form of MoE \\( f_{moe}(\\cdot) \\). Formally, suppose MoE contains \\( N \\) experts and \\( T \\) router networks, denoted as \\( \\mathbb{E} = \\{E_1,E_2,\\dots,E_N\\} \\) and \\( \\mathbb{G} = \\{G_1,G_2,\\dots,G_T\\} \\), respectively. \\( N \\) and \\( T \\) are the number of experts and the number of tasks, respectively. The backbone feature \\( \\mathbf{X}^l \\) from the \\( l \\)-th layer is fed into the networks of \\( N \\) experts and \\( T \\) router networks, respectively. For convenience, the superscript \\( l \\) is omitted in the following. For the \\( n \\)-th expert, the discriminative output feature is generated by \\( \\mathbf{X}_n = E_n(\\mathbf{X}) \\). In the meanwhile, MoE learns gates from the task-specific router networks for different tasks. For task \\( t \\), the gate value for each expert generated by the router network can be represented as \\( \\mathbf{g}^t = G_t(\\mathbf{X}) \\), where \\( \\mathbf{g}^t\\in \\mathbb{R}^N \\). Finally, MoE generates task-specific feature \\( \\mathbf{S}^t \\) for task \\( t \\) by utilizing gates to combine the expert features, which can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.792, + 0.892, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} ^ {t} = f _ {\\text {m o e}} (\\mathbf {X}) = \\sum_ {n = 1} ^ {N} \\mathbf {g} _ {n} ^ {t} \\mathbf {X} _ {n}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.902 + ], + "angle": 0, + "content": "For \\(\\mathrm{M}^3\\mathrm{ViT}\\) [14] and Mod-Squad [10], they deactivate some experts according to the corresponding gate values for one inference and select the top-\\(k\\) experts. The task-specific features are used to make predictions for each task." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27929" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.085, + 0.896, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.895, + 0.372 + ], + "angle": 0, + "content": "Figure 2. Overall framework of the proposed method. The MLoRE modules are equipped at different layers, where the backbone features from different layers are fed into the MLoRE modules, respectively. At each selected layer, the backbone feature is first projected to different task features and then sent to the task-sharing convolution, task-sharing low-rank expert networks followed by the task-specific router network and task-specific low-rank expert networks. The outputs of these branches are accumulated to generate task-specific features. At each selected layer, we stack two MLoRE modules." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.473, + 0.565 + ], + "angle": 0, + "content": "The advantages of MoE lie in that it enables the dynamic encoding of features for each sample and each task and increases feature encoding diversity via multiple experts. However, when applying the MoE technique to build an MTL decoder, we find that it struggles with building global task relationships. Moreover, when increasing the number of experts to enlarge the capacity of feature representations and the context of the expert network, the parameters and computational cost increase accordingly. To address the above issues, we propose the Mixture-of-Low-Rank-Expert (MLoRE) module." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.575, + 0.347, + 0.592 + ], + "angle": 0, + "content": "3.3. Mixture of Low-Rank Experts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The overall pipeline of the proposed Mixture-of-Low-Rank-Expert (MLoRE) module is illustrated in Fig. 2. Considering building cross-task relations across the task-specific features, we first project the backbone features into task features using several lightweight convolutional layers. Then, the feature of each task is delivered to the task-sharing generic path and multiple task-sharing expert networks with different low-rank convolutions. The low-rank experts are selected according to the predictions of the task-specific router network for each task. Furthermore, except for building task-specific features based on task-specific routers, we introduce additional task-specific low-rank expert networks to assist in building more discriminative task-specific features. The features from the task-sharing generic path, task-sharing low-rank expert networks selected by task-specific routers and the task-specific expert networks are summed up to generate discriminative task-specific features. We introduce linearity into the MLoRE module and do not utilize any activation function in all paths, enabling reparameterization to reduce computational cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.895, + 0.537 + ], + "angle": 0, + "content": "To be specific, the backbone feature \\(\\mathbf{X}\\) at the \\(l\\)-th layer is first projected to multiple task splits corresponding to each task using \\(1\\times 1\\) convolutions, which can be denoted as \\(\\{\\mathbf{X}^t = f_{t,1\\times 1}(\\mathbf{X}),t\\in [1,\\dots ,T]\\}\\), where \\(\\mathbf{X}^t\\in \\mathbb{R}^{C\\times H\\times W}\\). Then, the task feature is sent to three paths, i.e., task-sharing generic path \\(f_{g}(\\cdot)\\), task-sharing low-rank expert path \\(f_{lre}(\\cdot)\\) with task-specific router network \\(f_{sr}^{t}(\\cdot)\\), and task-specific low-rank expert path \\(f_{se}(\\cdot)\\). The output task-specific feature \\(\\mathbf{S}^t\\) is obtained by" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.545, + 0.892, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} ^ {t} = f _ {g} (\\mathbf {X} ^ {t}) + f _ {s r} ^ {t} \\left(f _ {l r e} (\\mathbf {X} ^ {t})\\right) + f _ {s e} ^ {t} (\\mathbf {X} ^ {t}). \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.604 + ], + "angle": 0, + "content": "In the following, we introduce the network details of all these paths in the MLoRE module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.895, + 0.837 + ], + "angle": 0, + "content": "Task-sharing generic path contains a simple \\(3 \\times 3\\) convolutional layer with a weight matrix \\(\\mathbf{W}_g \\in \\mathbb{R}^{3 \\times 3 \\times C \\times C}\\) and a bias matrix \\(\\mathbf{b}_g \\in \\mathbb{R}^C\\). As all task features will go through this generic convolution, it will be optimized by the gradients of different tasks simultaneously, which can help extract common features among all tasks. During the training process, we stop the gradients of this path for further back-propagation. The gradient is back-propagated through the other two paths. We found such a simple operation can better ease the optimization process and well solve the gradient conflicts. Experimental results in Sec. 4.2 show that the simple task-sharing generic path can bring performance improvements on all tasks, demonstrating the effectiveness of the idea of explicitly building the relationship across all tasks from a global perspective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Task-sharing low-rank expert path. We take inspiration from LoRA [21] and adopt low-rank convolution which is a low-rank format of a vanilla convolution. Each task-sharing expert network shares a similar structure consisting of a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27930" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.335 + ], + "angle": 0, + "content": "\\(3 \\times 3\\) convolution and a \\(1 \\times 1\\) convolution. The weights and bias of all task-sharing expert networks can be formulated as \\(\\{\\mathbf{W}_{l\\text{re}b}^{n}, \\mathbf{b}_{l\\text{re}b}^{n}, \\mathbf{W}_{l\\text{rea}}^{n}, \\mathbf{b}_{l\\text{rea}}^{n} | n \\in [1, \\dots, N]\\}\\), where \\(\\mathbf{W}_{l\\text{re}b}^{n} \\in \\mathbb{R}^{3 \\times 3 \\times C \\times r_{n}}\\), \\(\\mathbf{b}_{l\\text{re}b}^{n} \\in \\mathbb{R}^{r_{n}}\\), \\(\\mathbf{W}_{l\\text{rea}} \\in \\mathbb{R}^{1 \\times 1 \\times r_{n} \\times C}\\), and \\(\\mathbf{b}_{l\\text{rea}}^{n} \\in \\mathbb{R}^{C}\\) (\\(r_{n} \\ll C\\)). \\(r_{n}\\) denotes the rank for the \\(n\\)-th expert network. In our method, \\(r_{n}\\) for different expert networks is different, aiming to improve the diversity of parameters and features. For each task, the task-specific router network \\(f_{sr}^{t}(\\cdot)\\) learns gate values for these experts and activate the top-\\(k\\) experts according to the gate values. The output features of all the activated experts are summed up and then sent to a BatchNorm layer to generate task-specific features. The BatchNorm layer contains four parameters, including the accumulated channel-wise mean \\(\\mu \\in \\mathbb{R}^{C}\\), the accumulated channel-wise standard deviation \\(\\sigma \\in \\mathbb{R}^{C}\\), the scaling factor \\(\\gamma \\in \\mathbb{R}^{C}\\) and the \\(\\beta \\in \\mathbb{R}^{C}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.34, + 0.473, + 0.492 + ], + "angle": 0, + "content": "Task-specific low-rank expert path includes \\(T\\) specific expert networks and each processes one task feature. For each specific expert network, we utilize a similar structure with task-sharing expert path, which contains a \\(3 \\times 3\\) convolution with a weight matrix \\(\\mathbf{W}_{seb}^{t} \\in \\mathbb{R}^{3 \\times 3 \\times C \\times R}\\) and a bias matrix \\(\\mathbf{b}_{seb}^{t} \\in \\mathbb{R}^{R}\\), followed by a \\(1 \\times 1\\) convolution with a weight matrix \\(\\mathbf{W}_{sea}^{t} \\in \\mathbb{R}^{1 \\times 1 \\times R \\times C}\\) and a bias matrix \\(\\mathbf{b}_{sea}^{t} \\in \\mathbb{R}^{C}\\). \\(R\\) denotes the rank number (\\(R \\ll C\\)). The task-specific expert path can enhance the distinctiveness of the task-specific features, which will be verified in the experiments." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.498, + 0.473, + 0.698 + ], + "angle": 0, + "content": "Router network. As shown in Fig. 2, to generate task-specific features from the task-sharing low-rank expert path, we learn task-specific router networks to generate gate values for each expert and utilize them as the weight of the linear combination of feature output from different experts. The router network for each task is usually the simple linear layers followed by an average pooling layer and a prediction layer. Specifically, the router network \\( f_{sr}^{t}(\\cdot) \\) for the \\( t \\)-th task is designed as follows. Our router network takes the task-specific features \\( \\mathbf{X}^t \\in \\mathbb{R}^{C \\times H \\times W} \\) as input and feed them into two consecutive \\( 1 \\times 1 \\) convolutions, mapping the channel dimension from \\( C \\) to \\( C/4 \\), followed by a global pooling layer. The output is a global feature vector \\( \\mathbf{X}_f \\in \\mathbb{R}^{\\frac{C}{4}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.698, + 0.473, + 0.866 + ], + "angle": 0, + "content": "In addition, inspired by previous works [19, 20] showing that positional information is also important for modeling long-range spatial context, we introduce another parallel position-aware branch. Similarly, it consists of two linear layers. The first linear layer shrinks the feature along the spatial dimension, mapping the shape from \\(\\mathbb{R}^{C\\times HW}\\) to \\(\\mathbb{R}^{C\\times 1}\\), which will then be transformed to \\(\\mathbb{R}^{\\frac{C}{4}}\\) via the second linear layer. The output feature vectors of these two branches are concatenated along the final dimension and then sent to the final prediction layer followed by a Softmax function to produce the gate values \\(g_{t}\\) for each expert." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Re-parameterization during inference. We introduce linearity into the MLoRE module by removing all activation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.184 + ], + "angle": 0, + "content": "functions, enabling the parameters of all paths to reparameterize to a simple \\(3 \\times 3\\) convolution for each task at inference. We first parameterize the task-sharing low-rank expert path and then parameterize the parameters of all paths. According to [12], the weight and bias matrices in the task-sharing expert path can be combined and formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.193, + 0.895, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} _ {l r e} ^ {t} = \\mathfrak {B} \\left(\\frac {\\gamma}{\\sigma}\\right) \\sum_ {k \\in \\mathbb {K} _ {t}} \\mathbf {g} _ {k} ^ {t} \\mathbf {W} _ {l r e b} ^ {k} \\mathbf {W} _ {l r e a} ^ {k}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.23, + 0.895, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} _ {l r e} ^ {t} = \\frac {\\gamma}{\\sigma} (\\sum_ {k \\in \\mathbb {K} _ {t}} \\mathbf {g} _ {k} ^ {t} (\\mathbf {b} _ {l r e b} ^ {k} \\mathbf {W} _ {l r e a} ^ {k} + \\mathbf {b} _ {l r e a} ^ {k}) - \\mu) + \\beta , (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.895, + 0.352 + ], + "angle": 0, + "content": "where \\(\\mathfrak{B}\\) denotes the broadcast operation and \\(\\mathbb{K}_t\\) denotes the index set of the activated experts selected by the router network for task \\(t\\). \\(\\mathbf{g}_k^t\\) is the \\(k\\)-th gate value predicted by the router. At the inference time, the weight matrix and bias matrix of all these three paths can be re-parameterized as" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.362, + 0.892, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} _ {r} ^ {t} = \\mathbf {W} _ {g} + \\mathbf {W} _ {s r} ^ {t} + \\mathbf {W} _ {s e b} ^ {t} \\mathbf {W} _ {s e a} ^ {t}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.382, + 0.892, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} _ {r} ^ {t} = \\mathbf {b} _ {g} + \\mathbf {b} _ {s r} ^ {t} + \\mathbf {b} _ {s e b} ^ {t} \\mathbf {W} _ {s e a} ^ {t} + \\mathbf {b} _ {s e a} ^ {t}. \\qquad (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.895, + 0.443 + ], + "angle": 0, + "content": "\\(\\mathbf{W}_r^t\\) and \\(\\mathbf{b}_r^t\\) are the weight and bias of the re-parameterized convolution. Thus, Eqn. (2) can be reformulated as" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.453, + 0.892, + 0.471 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} ^ {t} = \\mathbf {X} ^ {t} * \\mathbf {W} _ {r} ^ {t} + \\mathfrak {B} \\left(\\mathbf {b} _ {r} ^ {t}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.892, + 0.513 + ], + "angle": 0, + "content": "where \\(\\oplus\\) denotes the convolution operation and \\(\\mathbf{b}_r^t\\) is with the same shape as \\(\\mathbf{X}^t\\) via broadcast." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.527, + 0.634, + 0.544 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.551, + 0.709, + 0.568 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.895, + 0.762 + ], + "angle": 0, + "content": "Datasets. To demonstrate the effectiveness of our method, we evaluate the performance of our method on two popular multi-task datasets, including PASCAL-Context [7] and NYUD-v2 [39]. PASCAL-Context [7] contains high-quality annotations of several tasks, including semantic segmentation, human parsing, saliency detection, surface normals, and object boundary detection. There are 4,998 training images and 5,105 test images in this dataset. NYUDv2 [39] also provides high-quality multi-task annotations, including semantic segmentation, monocular depth estimation, surface normals, and object boundary detection. This dataset contains 795 training images and 654 test images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Evaluation metrics. We introduce the evaluation metrics for the tasks mentioned above. Following previous multi-task works [48, 49], the mean intersection-over-union (mIoU) is used to evaluate semantic segmentation and human parsing. The root mean square error (RMSE) is used to evaluate the accuracy of monocular depth estimation. Saliency detection uses the maximum F-measure (maxF). Surface normal and object boundary detection adopt the mean error (mErr) and the optimal-dataset-scale F-measure" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "27931" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.173 + ], + "angle": 0, + "content": "Table 2. Ablation study on different components in MLoRE on the PASCAL-Context dataset. Each row adds an extra setting to the above row. MoE: the standard mixture-of-experts structure; LoRE: task-sharing low-rank expert path; GC: task-sharing generic convolution path; SPE: task-specific expert path. \\(\\uparrow\\) denotes higher is better. \\(\\downarrow\\) denotes lower is better." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.175, + 0.472, + 0.299 + ], + "angle": 0, + "content": "
SettingsSemsegParsingSal.NormalBound.MTLFLOPs#Param
mIoU↑ mIoU↑ maxF↑ mErr ↓odsF↑ Δm ↑(G)(M)
Baseline77.3865.1585.0813.7969.87-3.41391115
+ MoE78.5666.7885.1813.5773.91-1.201834676
Baseline
+ LoRE78.3866.2185.1513.7173.53-1.71568213
+ GC79.2567.4385.2013.7074.38-0.88568243
+ SPE79.2667.8285.3113.6574.69-0.58568259
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.47, + 0.357 + ], + "angle": 0, + "content": "(odsF) as the evaluation metrics, respectively. In total, we evaluate the MTL gain \\(\\Delta_{m}\\) across all tasks, following [34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.47, + 0.468 + ], + "angle": 0, + "content": "Training settings. We utilize ViT-large [13] as the backbone. The channel number of the decoder is set to 384. For the ablation study, the ViT-base network is set as the backbone. Following the previous work [49], the proposed MTL network is trained for 40,000 iterations with a batch size of 4 on both the two datasets. The optimizer and the loss functions for different tasks follow the previous work [49]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.479, + 0.231, + 0.496 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.655 + ], + "angle": 0, + "content": "In this subsection, we conduct extensive experiments to demonstrate the effectiveness of different components and find the best settings of different hyper-parameters. All the ablation experiments are conducted based on the ViT-based backbone if not specified otherwise. The baseline is built upon the ViT-base backbone with 12 layers, where the backbone features from the 3-rd, 6-th, 9-th, and 12-th layers are utilized as the multi-scale features, each of which is followed by a linear layer to project the channel dimension to the output channels for each task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of different components. We first conduct experiments to verify the effectiveness of different components of the MLoRE module. The quantitative results are shown in Tab. 2. We first examine the performance of the baseline with the standard MoE and the parameter size and FLOPs of their model. The expert networks in the standard MoE (15 experts) are similar to ours, each of which consists of a \\(3 \\times 3\\) convolution and a \\(1 \\times 1\\) convolution with ReLU among them. When adding the MoE to the baseline, we observe the MTL gain can be improved, but the parameters and FLOPs also expand about 5 times and 4 times, which is a heavy burden for the whole network. When adding the low-rank expert networks (LoRE) to the baseline, the performance is also improved, but the parameters and FLOPs are only \\(1/3\\) and \\(1/3\\) of the MoE-based model. When the low-rank property is applied to the expert net" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.693, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.09, + 0.892, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.202, + 0.894, + 0.259 + ], + "angle": 0, + "content": "Figure 3. Ablation study on the number of experts \\(N\\) and the number of activated experts \\(K\\). In the right figure, we also present the parameter change of the MLoRE module with the increase in the number of experts." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.275, + 0.892, + 0.303 + ], + "angle": 0, + "content": "Table 3. Ablation on the rank setting in the task-sharing low-rank expert path of MLoRE on PASCAL-Context dataset." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.304, + 0.88, + 0.384 + ], + "angle": 0, + "content": "
Min/Max RankSemseg mIoU↑Parsing mIoU↑Saliency maxF↑Normal mErr↓Boundary odsF↑MTL Δm↑
16/1678.8468.0185.3213.6974.39-0.77
16/12879.2667.8285.3013.6574.69-0.58
128/12878.7966.9885.3513.6774.29-1.07
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.472 + ], + "angle": 0, + "content": "works, the parameter size is reduced several times. When further introducing linearity in the expert networks by removing all activation functions, the computations are saved by re-parameterizing all experts to a single convolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.893, + 0.655 + ], + "angle": 0, + "content": "Furthermore, we also emphasize the importance of explicitly building global task correlations in the MoE and introduce the task-sharing generic path to achieve this goal. It can be seen that adding the task-sharing generic path to LoRE can further improve the performance and outperform the baseline with MoE on most metrics, which proves the effectiveness of modeling the global relationships among all tasks. In addition, adding a task-specific low-rank expert for each task also improves the performance, which demonstrates that the specialized expert can enhance the discrimination of task-specific features. We empirically set the rank of the task-specific low-rank expert to 64 in this paper." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Number of task-sharing low-rank experts and top-\\(k\\) selection. We ablate the number of low-rank experts in the MLoRE module and top-\\(k\\) selection of the experts by the task-specific router networks. We first fix one and ablate another parameter to study their impact on multi-task performance. As shown in Fig. 3, when increasing the number of experts, the MTL gain of the model is significantly improved and allows us to achieve the best performance when the number of experts is 15. Further increasing the number of experts, we do not observe obvious performance gain. Besides, when fixing the number of experts to 15, we ablate the ratio of the activated experts. We observe that activating \\(60\\%\\) experts for each task is the best choice in our experiments. When selecting all expert networks, the performance decreases by a large margin, which reflects the importance of sparsity for feature discrimination." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27932" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.51, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.091, + 0.876, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.215, + 0.895, + 0.259 + ], + "angle": 0, + "content": "Figure 4. (a) The relations between tasks and low-rank experts. (b) The ratio of an expert activated by different numbers of tasks in the MLoRE module without the task-sharing generic path. We can see that without the task-sharing generic path, there is only a few experts can be activated by all five tasks. Horizontal coordinates represent the ranks of different experts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.271, + 0.473, + 0.468 + ], + "angle": 0, + "content": "Rank number setting. The expert networks utilize the low-rank format of a vanilla \\(3 \\times 3\\) convolution with 640 output channels. The rank \\(r\\) of different expert networks also plays an important role. We study different settings, including 1) all experts with the same rank number 16, 2) all experts with the rank number 128, 3) all experts with the rank numbers from 16 to 128 with an increased step of 8. As shown in Tab. 3, it can be seen that selecting different rank numbers for expert networks achieves the best MTL gain. We analyze the expert networks with different rank numbers that can bring more feature diversity than the same rank numbers, which is more useful for assembling task-specific features, and utilize this as our settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.473, + 0.761 + ], + "angle": 0, + "content": "The relation visualizations between tasks and the low-rank experts are shown in Fig. 4(a). We count the relations from the second MLoRE module at the last stage and calculate the activation ratio of every expert selected by different tasks on the entire dataset. It can be seen that experts with different ranks tend to learn different subsets of tasks. Specifically, for this module, experts with lower ranks tend to learn shared knowledge for 3-4 related tasks, while experts with higher ranks tend to specialize in 1-2 tasks. Moreover, we also show the ratio of different experts to be activated by different numbers of tasks when the task-sharing generic path is not added to the MLoRE module in Fig. 4(b). It can be seen that in a fully dynamic manner, all these experts are seldom or even never activated by all the tasks in one sample. This proves the fact that almost no expert can learn global relationships across all tasks when utilizing the MoE in the decoder directly. This phenomenon strongly supports the necessity of the task-sharing generic path for explicitly modeling global task relationships." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Task-specific router network. The router network is important for generating task-specific gates, which decide how to activate the experts and assemble their features. We ablate several settings of the router network and the results are shown in Tab. 4. Adding the position-aware branch to the basic router network can improve the MTL gain by \\(+0.17\\). The position-aware branch can obtain more context information, which benefits for the router network. Moreover, when replacing the router's input from the learnable pa" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.269, + 0.895, + 0.312 + ], + "angle": 0, + "content": "Table 4. Ablation on the setting of the router network. basic: the basic router network. pos.: the positional-aware router network. w/o sample-dep: the input is the learnable parameters." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.315, + 0.895, + 0.428 + ], + "angle": 0, + "content": "
Router VariantSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑MTL Δm ↑
basic79.1567.4085.2113.5874.34-0.75
basic+pos.79.2667.8285.3113.6574.69-0.58
only pos.79.1067.7685.1113.7174.51-0.82
basic79.1567.4085.2113.5874.34-0.75
w/o sample-dep78.8667.3885.4113.6574.25-0.91
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.455, + 0.895, + 0.501 + ], + "angle": 0, + "content": "rameters to the sample features, the MTL gain increases by \\(+0.16\\), which demonstrates that dynamic information from samples is vital for gates." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.512, + 0.8, + 0.529 + ], + "angle": 0, + "content": "4.3. Comparisons with Other Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.895, + 0.673 + ], + "angle": 0, + "content": "The quantitative comparisons with previous state-of-the-art (SOTA) methods are shown in Tab. 5 and Tab. 6. It can be seen that our method clearly outperforms previous methods in terms of all metrics on both the PASCAL-Context dataset and the NYUDv2 dataset. In particular, on the PASCAL-Context dataset, for semantic segmentation, human parsing, and boundary detection, the performance of our method outperforms the previous best method by \\(+0.52\\) mIoU, \\(+1.10\\) mIoU and \\(+1.92\\) odsF, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.896, + 0.809 + ], + "angle": 0, + "content": "Previous methods, \\(\\mathrm{M}^3\\mathrm{ViT}\\) [14], Mod-Squad [10], and TaskExpert [50] all utilize the MoE technique in their network. However, our method shows performance superior to theirs, which demonstrates the effectiveness of the MLoRE module. Compared with the decoder-focused method, TaskExpert, the performance of semantic segmentation, human parsing, and object boundary is significantly improved by \\(+0.77\\) mIoU, \\(+1.10\\) mIoU and \\(+2.12\\) odsF, but using fewer parameters and FLOPs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Furthermore, we also present an intuitive visualization comparison among different methods in Fig. 5. Our method can generate better visualization results than the previous SoTA methods, especially for semantic segmentation, human parsing, and object boundary detection. More visualization comparisons are in the supplemental material." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27933" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison of different methods on PASCAL-Context dataset. * denotes the reproduced performance of methods based on the ViT-large backbone in [50]." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.122, + 0.861, + 0.311 + ], + "angle": 0, + "content": "
MethodPublicationBackboneSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑FLOPs (G)#Param (M)
PAD-Net [46]CVPR'18HRNet1853.6059.6065.8015.3072.5012481
MTI-Net [45]ECCV'20HRNet1861.7060.1884.7814.2370.80161128
ATRC [4]ICCV'21HRNet1867.6762.9382.2914.2472.4221696
PAD-Net* [46]CVPR'18ViT-large78.0167.1279.2114.3772.60773330
MTI-Net* [45]ECCV'20ViT-large78.3167.4084.7514.6773.00774851
ATRC* [4]ICCV'21ViT-large77.1166.8481.2014.2372.10871340
InvPT [48]ECCV'22ViT-large79.0367.6184.8114.1573.00669423
TaskPrompter [49]ICLR'23ViT-large80.8968.8984.8313.7273.50497401
TaskExpert [50]ICCV'23ViT-large80.6469.4284.8713.5673.30622420
Ours-ViT-large81.4170.5284.9013.5175.42571407
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.333, + 0.47, + 0.361 + ], + "angle": 0, + "content": "Table 6. Quantitative comparison of different methods on the NYUD-v2 dataset. Our method performs the best on all four tasks." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.367, + 0.47, + 0.495 + ], + "angle": 0, + "content": "
MethodBackboneSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑
PAD-Net [46]HRNet1836.610.624620.8876.38
MTI-Net [45]HRNet4845.970.536520.2777.86
ATRC [4]HRNet4846.330.536320.1877.94
InvPT [48]ViT-large53.560.518319.0478.10
TaskPrompter [49]ViT-large55.300.515218.4778.20
TaskExpert [50]ViT-large55.350.515718.5478.40
OursViT-large55.960.507618.3378.43
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.521, + 0.285, + 0.536 + ], + "angle": 0, + "content": "4.4. Efficient MTL Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.698 + ], + "angle": 0, + "content": "We also apply our MLoRE module to the ViT-small backbone to check the performance of the efficient models. Specifically, the channel number of the decoder decreases from 384 to 192. As shown in Tab. 7, using about \\(35\\%\\) GFLOPs of TaskExpert, our method can achieve a highly competitive result. In particular, the performance of the semantic segmentation and object boundary is improved by \\(0.6\\%\\) mIoU and \\(1.01\\%\\) odsF while the metrics of other tasks are close to TaskExpert. Furthermore, the parameters are fewer than TaskExpert by 11M." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.208, + 0.725 + ], + "angle": 0, + "content": "4.5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We present a novel decoder-focused multi-task learning method, called MLoRE. We delve deep into the standard mixture-of-experts (MoE) technique and improve it for dense multi-task learning from two aspects. First, to address the neglected global relationship modeling in the MoE, we add a simple generic convolution path to MoE, enabling different task features to be able to share this path. Furthermore, we apply the low-rank format of a vanilla convolution to different expert networks to free the MoE from high computational cost and a large number of parameters when increasing the number of experts. Experiments" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.332, + 0.892, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.504, + 0.892, + 0.573 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison among different methods, including InvPT [48], TaskPrompter [49], and ours. Best viewed with zoom-in. It can be seen that our method achieves better visual results than other methods on all five tasks thanks to the proposed MLoRE module." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.603, + 0.891, + 0.631 + ], + "angle": 0, + "content": "Table 7. Quantitative comparison of the MoE-based efficient models on PASCAL-Context dataset." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.639, + 0.892, + 0.737 + ], + "angle": 0, + "content": "
MethodSemseg mIoU ↑Parsing mIoU ↑Sal. maxF ↑Nor. mErr ↓Bound. odsF ↑FLOPs (G)Param (M)
M3ViT72.8062.1066.3014.5071.7042042
Mod-Squad74.1062.7066.9013.7072.0042052
TaskExpert75.0462.6884.6814.2268.8020455
ours75.6462.6584.7014.4369.817244
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.891, + 0.796 + ], + "angle": 0, + "content": "demonstrate that the proposed method clearly outperforms previous state-of-the-art methods on all metrics." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgments. This research was supported by NSFC (NO. 62276145), the Fundamental Research Funds for the Central Universities (Nankai University, 070-63223049), CAST through Young Elite Scientist Sponsorship Program (No. YESS20210377). Computations were supported by the Supercomputing Center of Nankai University (NKSC)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27934" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Armen Aghajanyan, Luke Zettlemoyer, and Sonal Gupta. Intrinsic dimensionality explains the effectiveness of language model fine-tuning. arXiv preprint arXiv:2012.13255, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4009-4018, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.284 + ], + "angle": 0, + "content": "[3] David Bruggemann, Menelaos Kanakis, Stamatios Georgoulis, and Luc Van Gool. Automated search for resource-efficient branched multi-task networks. arXiv preprint arXiv:2008.10292, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.471, + 0.34 + ], + "angle": 0, + "content": "[4] David Brüggemann, Menelaos Kanakis, Anton Obukhov, Stamatios Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In Int. Conf. Comput. Vis., pages 15869-15878, 2021. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.47, + 0.409 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):834-848, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[6] Tianlong Chen, Xuxi Chen, Xianzhi Du, Abdullah Rashwan, Fan Yang, Huizhong Chen, Zhangyang Wang, and Yeqing Li. Adamv-moe: Adaptive multi-task vision mixture-of-experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17346-17357, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[7] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1971-1978, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.633 + ], + "angle": 0, + "content": "[8] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[9] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. Advances in Neural Information Processing Systems, 33:2039-2050, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.705, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[10] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. Mod-squad: Designing mixtures of experts as modular multi-task learners. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11828-11837, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.776, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[11] Aidan Clark, Diego De Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, et al. Unified scaling laws for routed language models. In Int. Conf. Mach. Learn., pages 4057-4086. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.846, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Diverse branch block: Building a convolution as an inception-like unit. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10886-10895, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[13] 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. In Int. Conf. Learn. Represent., 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.894, + 0.259 + ], + "angle": 0, + "content": "[14] Zhiwen Fan, Rishov Sarkar, Ziyu Jiang, Tianlong Chen, Kai Zou, Yu Cheng, Cong Hao, Zhangyang Wang, et al. \\(\\mathbf{M}^3\\) vit: Mixture-of-experts vision transformer for efficient multi-task learning with model-accelerator co-design. In Adv. Neural Inform. Process. Syst., volume 35, pages 28441-28457, 2022. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.26, + 0.894, + 0.315 + ], + "angle": 0, + "content": "[15] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.894, + 0.384 + ], + "angle": 0, + "content": "[16] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3205-3214, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[17] Michelle Guo, Albert Haque, De-An Huang, Serena Yeung, and Li Fei-Fei. Dynamic task prioritization for multitask learning. In Proceedings of the European conference on computer vision (ECCV), pages 270–287, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[18] Pengsheng Guo, Chen-Yu Lee, and Daniel Ulbricht. Learning to branch for multi-task learning. In Int. Conf. Mach. Learn., pages 3854-3863. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[19] Qibin Hou, Li Zhang, Ming-Ming Cheng, and Jiashi Feng. Strip pooling: Rethinking spatial pooling for scene parsing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4003-4012, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.538, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[20] Qibin Hou, Daquan Zhou, and Jiashi Feng. Coordinate attention for efficient mobile network design. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13713-13722, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[22] Yerlan Idelbayev and Miguel A Carreira-Perpinan. Low-rank compression of neural nets: Learning the rank of each layer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8049–8059, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[23] Keishi Ishihara, Anssi Kanervisto, Jun Miura, and Ville Hautamaki. Multi-task learning with attention for end-to-end autonomous driving. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2902-2911, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[24] Robert A Jacobs and Michael I Jordan. Learning piecewise control strategies in a modular neural network architecture. IEEE Transactions on Systems, Man, and Cybernetics, 23(2):337-345, 1993. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[25] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. *Neural Comput.*, 3(1):79-87, 1991. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[26] Shibo Jie and Zhi-Hong Deng. Fact: Factor-tuning for lightweight adaptation on vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 1060-1068, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27935" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[27] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.161, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[28] Zhihao Li, Toshiyuki Motoyoshi, Kazuma Sasaki, Tetsuya Ogata, and Shigeki Sugano. Rethinking self-driving: Multi-task knowledge for better generalization and accident explanation ability. arXiv preprint arXiv:1809.11100, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.273 + ], + "angle": 0, + "content": "[29] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1925-1934, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.273, + 0.47, + 0.316 + ], + "angle": 0, + "content": "[30] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1871-1880, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.47, + 0.371 + ], + "angle": 0, + "content": "[31] Yen-Cheng Liu, Chih-Yao Ma, Junjiao Tian, Zijian He, and Zsolt Kira. Polyhistor: Parameter-efficient multi-task adaptation for dense vision tasks. Advances in Neural Information Processing Systems, 35:36889-36901, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.47, + 0.426 + ], + "angle": 0, + "content": "[32] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3431-3440, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.47, + 0.497 + ], + "angle": 0, + "content": "[33] Yongxi Lu, Abhishek Kumar, Shuangfei Zhai, Yu Cheng, Tara Javidi, and Rogerio Feris. Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5334-5343, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.47, + 0.552 + ], + "angle": 0, + "content": "[34] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1851-1860, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[35] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3994-4003, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.47, + 0.678 + ], + "angle": 0, + "content": "[36] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.47, + 0.719 + ], + "angle": 0, + "content": "[37] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Int. Conf. Comput. Vis., pages 12179-12188, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[38] Sebastian Ruder, Joachim Bingel, Isabelle Augenstein, and Anders Søgaard. Latent multi-task architecture learning. In AAAI Conf. Artif. Intell., volume 33, pages 4822-4829, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[39] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In *Eur. Conf. Comput. Vis.*, pages 746-760. Springer, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[40] Chi Su, Fan Yang, Shiliang Zhang, Qi Tian, Larry S Davis, and Wen Gao. Multi-task learning with low rank attribute embedding for person re-identification. In Proceedings of the IEEE international conference on computer vision, pages 3739–3747, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[41] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5227-5237, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.161, + 0.893, + 0.205 + ], + "angle": 0, + "content": "[42] Madeleine Udell, Corinne Horn, Reza Zadeh, Stephen Boyd, et al. Generalized low rank models. Foundations and Trends® in Machine Learning, 9(1):1-118, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.893, + 0.261 + ], + "angle": 0, + "content": "[43] Simon Vandenhende, Stamatios Georgoulis, Bert De Brabandere, and Luc Van Gool. Branched multi-task networks: deciding what layers to share. In Brit. Mach. Vis. Conf., 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.893, + 0.332 + ], + "angle": 0, + "content": "[44] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 44(7):3614-3633, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.893, + 0.389 + ], + "angle": 0, + "content": "[45] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In Eur. Conf. Comput. Vis., pages 527-543. Springer, 2020. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.893, + 0.459 + ], + "angle": 0, + "content": "[46] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 675-684, 2018. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.893, + 0.502 + ], + "angle": 0, + "content": "[47] Yongxin Yang and Timothy M Hospedales. Trace norm regularised deep multi-task learning. arXiv preprint arXiv:1606.04038, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.502, + 0.893, + 0.546 + ], + "angle": 0, + "content": "[48] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. In Eur. Conf. Comput. Vis., pages 514-530. Springer, 2022. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.893, + 0.589 + ], + "angle": 0, + "content": "[49] Hanrong Ye and Dan Xu. Taskprompter: Spatial-channel multi-task prompting for dense scene understanding. In Int. Conf. Learn. Represent., 2022. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.893, + 0.631 + ], + "angle": 0, + "content": "[50] Hanrong Ye and Dan Xu. Taskexpert: Dynamically assembling multi-task representations with memorial mixture-of-experts. In Int. Conf. Comput. Vis., 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.631, + 0.893, + 0.688 + ], + "angle": 0, + "content": "[51] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Eur. Conf. Comput. Vis., pages 325-341, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.893, + 0.759 + ], + "angle": 0, + "content": "[52] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2636-2645, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.893, + 0.815 + ], + "angle": 0, + "content": "[53] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Zequn Jie, Xiang Li, and Jian Yang. Joint task-recursive learning for semantic segmentation and depth estimation. In Eur. Conf. Comput. Vis., pages 235–251, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.815, + 0.893, + 0.872 + ], + "angle": 0, + "content": "[54] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4106-4115, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.893, + 0.903 + ], + "angle": 0, + "content": "[55] Xiangyun Zhao, Haoxiang Li, Xiaohui Shen, Xiaodan Liang, and Ying Wu. A modulation module for multi-task learning" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27936" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "with applications in image retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 401-416, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.192 + ], + "angle": 0, + "content": "[56] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4514–4523, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "27937" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/4b27d7d8-e816-4974-8903-c5d6fd2b2750_origin.pdf b/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/4b27d7d8-e816-4974-8903-c5d6fd2b2750_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4695a5d106c0e94a2e264738275b9c51ec95d6a5 --- /dev/null +++ b/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/4b27d7d8-e816-4974-8903-c5d6fd2b2750_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef64217a0492f8bb2621eea07d906b31c551ebfb1d8f06f85b5860a4fa9e438f +size 1966513 diff --git a/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/full.md b/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e02d13e16420a77a8f91f3b12a03a19e1bb951ed --- /dev/null +++ b/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/full.md @@ -0,0 +1,286 @@ +# Multi-Task Dense Prediction via Mixture of Low-Rank Experts + +Yuqi Yang $^{1,2*}$ Peng-Tao Jiang $^{2*}$ Qibin Hou $^{1,3\dagger}$ Hao Zhang $^{2}$ Jinwei Chen $^{2}$ Bo Li $^{2}$ $^{1}$ VCIP, CS, Nankai University $^{2}$ vivo Mobile Communication Co., Ltd $^{3}$ NKIARI, Shenzhen Futian yangyq2000@mail.nankai.edu.cn, pt.jiang@vivo.com, andrewhoux@gmail.com + +# Abstract + +Previous multi-task dense prediction methods based on the Mixture of Experts (MoE) have received great performance but they neglect the importance of explicitly modeling the global relations among all tasks. In this paper, we present a novel decoder-focused method for multitask dense prediction, called Mixture-of-Low-Rank-Experts (MLoRE). To model the global task relationships, MLoRE adds a generic convolution path to the original MoE structure, where each task feature can go through this path for explicit parameter sharing. Furthermore, to control the parameters and computational cost brought by the increase in the number of experts, we take inspiration from LoRA and propose to leverage the low-rank format of a vanilla convolution in the expert network. Since the low-rank experts have fewer parameters and can be dynamically parameterized into the generic convolution, the parameters and computational cost do not change much with the increase of experts. Benefiting from this design, we increase the number of experts and its reception field to enlarge the representation capacity, facilitating multiple dense tasks learning in a unified network. Extensive experiments on the PASCAL-Context and NYUD-v2 benchmarks show that our MLoRE achieves superior performance compared to previous state-of-the-art methods on all metrics. Our code is available at https://github.com/YuqiYang213/MLoRE. + +# 1. Introduction + +Computer vision tasks, such as semantic segmentation [5, 29, 32, 51] and depth estimation [2, 37], have been significantly facilitated by deep learning techniques. Each vision task has its own elaborated deep models that usually follow a similar pipeline, i.e., feature extraction and prediction. Besides, some tasks also share relations. These facts motivate researchers to study Multi-Task Learning (MTL) that is able to unify different task models into a single one. + +![](images/37a04ce7301efed64ff44816185c6f102aa24a509590b8abc6ab5e449839c131.jpg) +Figure 1. Performance comparison with state-of-the-art methods. Our MLoRE based on the proposed mixture of low-rank experts achieves superior performance on all tasks. $\uparrow$ denotes higher is better. $\downarrow$ denotes lower is better. + +The significant advantage of multi-task learning lies in that it can improve training and inference efficiency while keeping commensurate performance to each task model. Due to such an advantage, MTL models have been applied to several directions, including autonomous driving [23, 28, 52] and scene understanding [46, 48], etc. + +In this paper, we focus on multi-task learning for dense scene understanding, in which each task predicts pixel-wise results. A research line of early works [4, 16, 33, 35, 45, 46, 48] focuses on designing delicate network architectures, including encoder-focused and decoder-focused methods. The encoder-focused methods [16, 33, 35] design handcrafted modules to share across task-specific encoders to construct task-generic features, while the decoder-focused methods [4, 45, 46, 48] propose tailored decoders to learn discriminative task-specific representations and build cross-task relationships. + +Unlike the above methods focusing on designing static network architectures, some methods [6, 10, 14] introduce the mixture-of-experts (MoE) technique that sheds light on + +Table 1. Parameters and FLOPs of different settings in the standard MoE and the proposed MLoRE. The left setting presents the number of experts and the kernel size of the convolutions in the expert network. + +
SettingsParams (M)FLOPs (G)
MoEMLoREMoEMLoRE
5 experts, [1×1, 1×1]3.11.23.001.49
10 experts, [1×1, 1×1]4.71.64.491.58
15 experts, [1×1, 1×1]6.31.95.991.66
5 experts, [3×3, 1×1]14.93.414.247.12
10 experts, [3×3, 1×1]22.44.721.377.21
15 experts, [3×3, 1×1]29.96.028.497.29
+ +learning dynamic and automatic manner of task specialization and cooperations [10]. They utilize MoE to design the encoder blocks and dynamically select network paths for different tasks and different samples. However, compared to these encoder-focused methods, the utilization of MoE in the decoder is less studied, where only Ye et al. [50] first applied MoE to the decoder recently. It decodes task-specific features by dynamically assembling task-generic features from different experts and achieves superior performance than previous decoder-focused methods. This motivates us to delve deep into the MoE-based MTL decoder. + +Benefiting from the dynamic routing process, these MoE-based methods can significantly improve the diversity in both parameters and features, resulting in more discriminative task-specific features. However, such a paradigm still has some limitations. First, although MoE-based methods can build connections in a subset of tasks by sharing the same expert in the dynamic routing process, the chances of expert sharing among all the tasks are very low, potentially hindering the router from building global relationships among all tasks. Nevertheless, the global relationship modeling among all tasks has been proven useful for dense MTL in [48, 49]. Based on this fact, we argue it is essential to explicitly model global relationship among all tasks in MoE. Furthermore, the capacity of the task-generic feature spaces is highly related to the number of experts. As empirically demonstrated in [11, 15], increasing the number of experts has the potential to facilitate multi-task learning in a unified network. However, for existing dense multi-task learning, adding more expert networks will introduce more parameters and computational costs, which is a heavy burden for the whole model. + +To address the above issues, we propose a novel decoder-focused approach, which we call it Mixture of Low-Rank Experts (MLoRE). The core idea of the MLoRE framework is to explicitly model global relationships among all tasks in the MoE and free the MoE from heavily dense computation burdens when increasing the number of experts to enlarge the task-generic feature spaces and context. To address the + +first issue, MLoRE builds upon the basic MoE structure and introduces a task-sharing generic convolution path that parallels with the MoE. Specifically, the backbone features are first projected to different task features and then all of them are fed into the generic convolution path and the original MoE's expert networks. By simply sharing the same generic path among all tasks, different tasks can be globally correlated with each other. Furthermore, to enhance the discrimination of task-specific features, we exclude some experts from the dynamic routing process and enable them to serve specific tasks. + +To increase the number of experts while not bringing too many parameters and FLOPs, we take inspiration from LoRA that the basic models adapted to different tasks only needs low-rank weight updates. To be specific, we transform the expert networks of the MoE to the different low-rank formats of a vanilla convolution, which saves more than $60\%$ of the parameters compared to the standard MoE module, as shown in Tab. 1. In addition, to control the computational cost brought by the increasing number of expert networks, we do not use any non-linear activation functions in all the expert networks and the generic convolution path, which enables re-parameterization during inference. Through re-parameterization, the knowledge of experts can be injected into the generic convolution path, reducing the computational cost for dense tasks. To our knowledge, we are the first to use linear experts in MoE for multi-task dense prediction. To verify the effectiveness of our method, we conduct comprehensive experiments on the PASCAL-Context and NYUD-v2 datasets. Our method achieves new performance records on all tasks as shown in Fig. 1. + +In summary, our contributions are three-fold: + +- We analyze the issues of MoE when applied in multi-task learning and propose a novel decoder-focused framework, MLoRE, which can explicitly model global relationships among all tasks and enlarge the capacity of feature representations without increasing the model size too much. +- We introduce a simple task-sharing generic path to the MoE structure and propose linear and low-rank expert networks inspired by LoRA. The generic convolution path and low-rank expert paths can be linearly combined, enabling re-parameterization at inference. +- Experiments on PASCAL-Context and NYUDv2 show that the proposed method clearly outperforms previous state-of-the-art MTL methods on all tasks. + +# 2. Related Work + +# 2.1. Dense Multi-Task Learning + +In computer vision, multi-task learning (MTL) for dense prediction tasks has been widely studied. The previous methods can be divided into two categories, including optimization-based methods and architecture-based meth + +ods [44]. Optimization methods [8, 9, 17, 27, 55] facilitate MTL by utilizing different strategies to balance the influence of each task. Architecture-based methods aim to design a unified deep network for MTL. They can be further classified into two categories, encoder-focused methods, and decoder-focused methods. (i) Encoder-focused methods [3, 18, 33, 38] design multi-task backbones to extract features adapted to different tasks. Typical methods include cross-stitch networks [35], neural discriminative dimensionality reduction networks [16], multi-task attention networks [30], branched networks [43], and mixture-of-expert networks [10, 14]. (ii) Decoder-focused methods [4, 45, 46, 48-50, 53, 54, 56] share the same backbone and design delicate heads to extract task-specific features for each task and cross-task relationships. The advantage of decoder-focused methods lies in that they can benefit from powerful off-the-shelf vision backbones, such as DINOv2 [36]. Our method also falls into the decoder-focused category and studies how to produce task-specific features with the MoE technique. + +# 2.2. Mixture-of-Experts + +Mixture-of-Experts (MoE) [24, 25] learns multiple expert networks and a router network that controls the probability of each expert contributing to the final output. This technique is also used in multi-task learning, which can better adapt to the data diversity. Different expert networks learn different discriminative features. The router network learns hard/soft task-specific coefficients to dynamically assemble discriminative features for each task. Prior MoE-based multi-task methods [6, 10, 14] are mainly encoder-focused methods. They introduce MoE into the backbone blocks to sparsely activate different paths for different tasks in the inference stage. Recently, Ye et al. [50] first introduced the MoE technique to the decoder. They utilize the spatial context-aware gates to combine each pixel of the features from different expert networks. The above methods decompose the backbone feature into multiple generic feature spaces and assemble discriminative task-specific features from them. Unlike the above MoE-based MTL method, our method first builds global relationships among all tasks explicitly in the MoE structure rather than leaving this work implicitly done by the task-specific routers. Moreover, our proposed low-rank experts give MoE better efficiency compared to the naive MoE and this gap gradually gets larger as the number of experts increases. + +# 2.3. Low-Rank Structure + +The low-rank structure is often used in deep learning for its efficiency [22, 40, 42, 47]. Recently, many methods in parameter-efficient adaptation [21, 26, 31, 41], such as LoRA [21], utilize the low-rank structure and have shown impressive results. LoRA takes inspiration from Aghajanyan et + +al. [1] that the difference in weights between the pre-trained model and the adapted model resides on low intrinsic rank. It learns an extra low-rank matrix instead of tuning the whole layer for adaptation. More related to our work, earlier MTL methods [40, 47] utilize low-rank structure to model task-generic features and generate task-specific features by linear combination. Unlike them, our method utilizes the low-rank structure to control the computation budget when increasing the number of experts in MoE. + +# 3. Method + +# 3.1. Overall Framework + +The overall framework follows the multi-scale architecture as in previous works [49, 50]. We utilize an off-the-shelf vision transformer (ViT) as the encoder and collect multiscale features from different layers. Formally, given an input image $\mathbf{I}$ and a vision transformer $\mathcal{F}$ , we can attain multi-scale feature sets $\{\mathbf{X}^l = \mathcal{F}^l (\mathbf{I})\}$ from different layers, where $\mathbf{X}^l\in \mathbb{R}^{L\times C}$ . Here, $l$ , $L = H\times W$ , and $C$ denote the layer index, the number of patches, and the feature dimension, respectively. $\mathcal{F}^l (\mathbf{I})$ is the output features of the $l$ -th transformer layer. The multi-scale features are fed into the decoder which includes two stacked MLoRE modules at each scale. For each task, the output features of MLoRE from different scales are concatenated together to generate the final task features for dense predictions. + +# 3.2. Preliminaries: Mixture of Experts + +Before describing the details of the proposed Mixture-of-Low-Rank-Experts (MLoRE) module, we first introduce the basic form of MoE $f_{moe}(\cdot)$ . Formally, suppose MoE contains $N$ experts and $T$ router networks, denoted as $\mathbb{E} = \{E_1,E_2,\dots,E_N\}$ and $\mathbb{G} = \{G_1,G_2,\dots,G_T\}$ , respectively. $N$ and $T$ are the number of experts and the number of tasks, respectively. The backbone feature $\mathbf{X}^l$ from the $l$ -th layer is fed into the networks of $N$ experts and $T$ router networks, respectively. For convenience, the superscript $l$ is omitted in the following. For the $n$ -th expert, the discriminative output feature is generated by $\mathbf{X}_n = E_n(\mathbf{X})$ . In the meanwhile, MoE learns gates from the task-specific router networks for different tasks. For task $t$ , the gate value for each expert generated by the router network can be represented as $\mathbf{g}^t = G_t(\mathbf{X})$ , where $\mathbf{g}^t\in \mathbb{R}^N$ . Finally, MoE generates task-specific feature $\mathbf{S}^t$ for task $t$ by utilizing gates to combine the expert features, which can be formulated as + +$$ +\mathbf {S} ^ {t} = f _ {\text {m o e}} (\mathbf {X}) = \sum_ {n = 1} ^ {N} \mathbf {g} _ {n} ^ {t} \mathbf {X} _ {n}. \tag {1} +$$ + +For $\mathrm{M}^3\mathrm{ViT}$ [14] and Mod-Squad [10], they deactivate some experts according to the corresponding gate values for one inference and select the top- $k$ experts. The task-specific features are used to make predictions for each task. + +![](images/a24f7be17b8b52312e07257e3993109dfc5c625c26e2c4292c96f4630fa451a1.jpg) +Figure 2. Overall framework of the proposed method. The MLoRE modules are equipped at different layers, where the backbone features from different layers are fed into the MLoRE modules, respectively. At each selected layer, the backbone feature is first projected to different task features and then sent to the task-sharing convolution, task-sharing low-rank expert networks followed by the task-specific router network and task-specific low-rank expert networks. The outputs of these branches are accumulated to generate task-specific features. At each selected layer, we stack two MLoRE modules. + +The advantages of MoE lie in that it enables the dynamic encoding of features for each sample and each task and increases feature encoding diversity via multiple experts. However, when applying the MoE technique to build an MTL decoder, we find that it struggles with building global task relationships. Moreover, when increasing the number of experts to enlarge the capacity of feature representations and the context of the expert network, the parameters and computational cost increase accordingly. To address the above issues, we propose the Mixture-of-Low-Rank-Expert (MLoRE) module. + +# 3.3. Mixture of Low-Rank Experts + +The overall pipeline of the proposed Mixture-of-Low-Rank-Expert (MLoRE) module is illustrated in Fig. 2. Considering building cross-task relations across the task-specific features, we first project the backbone features into task features using several lightweight convolutional layers. Then, the feature of each task is delivered to the task-sharing generic path and multiple task-sharing expert networks with different low-rank convolutions. The low-rank experts are selected according to the predictions of the task-specific router network for each task. Furthermore, except for building task-specific features based on task-specific routers, we introduce additional task-specific low-rank expert networks to assist in building more discriminative task-specific features. The features from the task-sharing generic path, task-sharing low-rank expert networks selected by task-specific routers and the task-specific expert networks are summed up to generate discriminative task-specific features. We introduce linearity into the MLoRE module and do not utilize any activation function in all paths, enabling reparameterization to reduce computational cost. + +To be specific, the backbone feature $\mathbf{X}$ at the $l$ -th layer is first projected to multiple task splits corresponding to each task using $1\times 1$ convolutions, which can be denoted as $\{\mathbf{X}^t = f_{t,1\times 1}(\mathbf{X}),t\in [1,\dots ,T]\}$ , where $\mathbf{X}^t\in \mathbb{R}^{C\times H\times W}$ . Then, the task feature is sent to three paths, i.e., task-sharing generic path $f_{g}(\cdot)$ , task-sharing low-rank expert path $f_{lre}(\cdot)$ with task-specific router network $f_{sr}^{t}(\cdot)$ , and task-specific low-rank expert path $f_{se}(\cdot)$ . The output task-specific feature $\mathbf{S}^t$ is obtained by + +$$ +\mathbf {S} ^ {t} = f _ {g} (\mathbf {X} ^ {t}) + f _ {s r} ^ {t} \left(f _ {l r e} (\mathbf {X} ^ {t})\right) + f _ {s e} ^ {t} (\mathbf {X} ^ {t}). \qquad (2) +$$ + +In the following, we introduce the network details of all these paths in the MLoRE module. + +Task-sharing generic path contains a simple $3 \times 3$ convolutional layer with a weight matrix $\mathbf{W}_g \in \mathbb{R}^{3 \times 3 \times C \times C}$ and a bias matrix $\mathbf{b}_g \in \mathbb{R}^C$ . As all task features will go through this generic convolution, it will be optimized by the gradients of different tasks simultaneously, which can help extract common features among all tasks. During the training process, we stop the gradients of this path for further back-propagation. The gradient is back-propagated through the other two paths. We found such a simple operation can better ease the optimization process and well solve the gradient conflicts. Experimental results in Sec. 4.2 show that the simple task-sharing generic path can bring performance improvements on all tasks, demonstrating the effectiveness of the idea of explicitly building the relationship across all tasks from a global perspective. + +Task-sharing low-rank expert path. We take inspiration from LoRA [21] and adopt low-rank convolution which is a low-rank format of a vanilla convolution. Each task-sharing expert network shares a similar structure consisting of a + +$3 \times 3$ convolution and a $1 \times 1$ convolution. The weights and bias of all task-sharing expert networks can be formulated as $\{\mathbf{W}_{l\text{re}b}^{n}, \mathbf{b}_{l\text{re}b}^{n}, \mathbf{W}_{l\text{rea}}^{n}, \mathbf{b}_{l\text{rea}}^{n} | n \in [1, \dots, N]\}$ , where $\mathbf{W}_{l\text{re}b}^{n} \in \mathbb{R}^{3 \times 3 \times C \times r_{n}}$ , $\mathbf{b}_{l\text{re}b}^{n} \in \mathbb{R}^{r_{n}}$ , $\mathbf{W}_{l\text{rea}} \in \mathbb{R}^{1 \times 1 \times r_{n} \times C}$ , and $\mathbf{b}_{l\text{rea}}^{n} \in \mathbb{R}^{C}$ ( $r_{n} \ll C$ ). $r_{n}$ denotes the rank for the $n$ -th expert network. In our method, $r_{n}$ for different expert networks is different, aiming to improve the diversity of parameters and features. For each task, the task-specific router network $f_{sr}^{t}(\cdot)$ learns gate values for these experts and activate the top- $k$ experts according to the gate values. The output features of all the activated experts are summed up and then sent to a BatchNorm layer to generate task-specific features. The BatchNorm layer contains four parameters, including the accumulated channel-wise mean $\mu \in \mathbb{R}^{C}$ , the accumulated channel-wise standard deviation $\sigma \in \mathbb{R}^{C}$ , the scaling factor $\gamma \in \mathbb{R}^{C}$ and the $\beta \in \mathbb{R}^{C}$ , respectively. + +Task-specific low-rank expert path includes $T$ specific expert networks and each processes one task feature. For each specific expert network, we utilize a similar structure with task-sharing expert path, which contains a $3 \times 3$ convolution with a weight matrix $\mathbf{W}_{seb}^{t} \in \mathbb{R}^{3 \times 3 \times C \times R}$ and a bias matrix $\mathbf{b}_{seb}^{t} \in \mathbb{R}^{R}$ , followed by a $1 \times 1$ convolution with a weight matrix $\mathbf{W}_{sea}^{t} \in \mathbb{R}^{1 \times 1 \times R \times C}$ and a bias matrix $\mathbf{b}_{sea}^{t} \in \mathbb{R}^{C}$ . $R$ denotes the rank number ( $R \ll C$ ). The task-specific expert path can enhance the distinctiveness of the task-specific features, which will be verified in the experiments. + +Router network. As shown in Fig. 2, to generate task-specific features from the task-sharing low-rank expert path, we learn task-specific router networks to generate gate values for each expert and utilize them as the weight of the linear combination of feature output from different experts. The router network for each task is usually the simple linear layers followed by an average pooling layer and a prediction layer. Specifically, the router network $f_{sr}^{t}(\cdot)$ for the $t$ -th task is designed as follows. Our router network takes the task-specific features $\mathbf{X}^t \in \mathbb{R}^{C \times H \times W}$ as input and feed them into two consecutive $1 \times 1$ convolutions, mapping the channel dimension from $C$ to $C/4$ , followed by a global pooling layer. The output is a global feature vector $\mathbf{X}_f \in \mathbb{R}^{\frac{C}{4}}$ . + +In addition, inspired by previous works [19, 20] showing that positional information is also important for modeling long-range spatial context, we introduce another parallel position-aware branch. Similarly, it consists of two linear layers. The first linear layer shrinks the feature along the spatial dimension, mapping the shape from $\mathbb{R}^{C\times HW}$ to $\mathbb{R}^{C\times 1}$ , which will then be transformed to $\mathbb{R}^{\frac{C}{4}}$ via the second linear layer. The output feature vectors of these two branches are concatenated along the final dimension and then sent to the final prediction layer followed by a Softmax function to produce the gate values $g_{t}$ for each expert. + +Re-parameterization during inference. We introduce linearity into the MLoRE module by removing all activation + +functions, enabling the parameters of all paths to reparameterize to a simple $3 \times 3$ convolution for each task at inference. We first parameterize the task-sharing low-rank expert path and then parameterize the parameters of all paths. According to [12], the weight and bias matrices in the task-sharing expert path can be combined and formulated as: + +$$ +\mathbf {W} _ {l r e} ^ {t} = \mathfrak {B} \left(\frac {\gamma}{\sigma}\right) \sum_ {k \in \mathbb {K} _ {t}} \mathbf {g} _ {k} ^ {t} \mathbf {W} _ {l r e b} ^ {k} \mathbf {W} _ {l r e a} ^ {k}, \tag {3} +$$ + +$$ +\mathbf {b} _ {l r e} ^ {t} = \frac {\gamma}{\sigma} (\sum_ {k \in \mathbb {K} _ {t}} \mathbf {g} _ {k} ^ {t} (\mathbf {b} _ {l r e b} ^ {k} \mathbf {W} _ {l r e a} ^ {k} + \mathbf {b} _ {l r e a} ^ {k}) - \mu) + \beta , (4) +$$ + +where $\mathfrak{B}$ denotes the broadcast operation and $\mathbb{K}_t$ denotes the index set of the activated experts selected by the router network for task $t$ . $\mathbf{g}_k^t$ is the $k$ -th gate value predicted by the router. At the inference time, the weight matrix and bias matrix of all these three paths can be re-parameterized as + +$$ +\mathbf {W} _ {r} ^ {t} = \mathbf {W} _ {g} + \mathbf {W} _ {s r} ^ {t} + \mathbf {W} _ {s e b} ^ {t} \mathbf {W} _ {s e a} ^ {t}, \tag {5} +$$ + +$$ +\mathbf {b} _ {r} ^ {t} = \mathbf {b} _ {g} + \mathbf {b} _ {s r} ^ {t} + \mathbf {b} _ {s e b} ^ {t} \mathbf {W} _ {s e a} ^ {t} + \mathbf {b} _ {s e a} ^ {t}. \qquad (6) +$$ + +$\mathbf{W}_r^t$ and $\mathbf{b}_r^t$ are the weight and bias of the re-parameterized convolution. Thus, Eqn. (2) can be reformulated as + +$$ +\mathbf {S} ^ {t} = \mathbf {X} ^ {t} * \mathbf {W} _ {r} ^ {t} + \mathfrak {B} \left(\mathbf {b} _ {r} ^ {t}\right), \tag {7} +$$ + +where $\oplus$ denotes the convolution operation and $\mathbf{b}_r^t$ is with the same shape as $\mathbf{X}^t$ via broadcast. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. To demonstrate the effectiveness of our method, we evaluate the performance of our method on two popular multi-task datasets, including PASCAL-Context [7] and NYUD-v2 [39]. PASCAL-Context [7] contains high-quality annotations of several tasks, including semantic segmentation, human parsing, saliency detection, surface normals, and object boundary detection. There are 4,998 training images and 5,105 test images in this dataset. NYUDv2 [39] also provides high-quality multi-task annotations, including semantic segmentation, monocular depth estimation, surface normals, and object boundary detection. This dataset contains 795 training images and 654 test images. + +Evaluation metrics. We introduce the evaluation metrics for the tasks mentioned above. Following previous multi-task works [48, 49], the mean intersection-over-union (mIoU) is used to evaluate semantic segmentation and human parsing. The root mean square error (RMSE) is used to evaluate the accuracy of monocular depth estimation. Saliency detection uses the maximum F-measure (maxF). Surface normal and object boundary detection adopt the mean error (mErr) and the optimal-dataset-scale F-measure + +Table 2. Ablation study on different components in MLoRE on the PASCAL-Context dataset. Each row adds an extra setting to the above row. MoE: the standard mixture-of-experts structure; LoRE: task-sharing low-rank expert path; GC: task-sharing generic convolution path; SPE: task-specific expert path. $\uparrow$ denotes higher is better. $\downarrow$ denotes lower is better. + +
SettingsSemsegParsingSal.NormalBound.MTLFLOPs#Param
mIoU↑ mIoU↑ maxF↑ mErr ↓odsF↑ Δm ↑(G)(M)
Baseline77.3865.1585.0813.7969.87-3.41391115
+ MoE78.5666.7885.1813.5773.91-1.201834676
Baseline
+ LoRE78.3866.2185.1513.7173.53-1.71568213
+ GC79.2567.4385.2013.7074.38-0.88568243
+ SPE79.2667.8285.3113.6574.69-0.58568259
+ +(odsF) as the evaluation metrics, respectively. In total, we evaluate the MTL gain $\Delta_{m}$ across all tasks, following [34]. + +Training settings. We utilize ViT-large [13] as the backbone. The channel number of the decoder is set to 384. For the ablation study, the ViT-base network is set as the backbone. Following the previous work [49], the proposed MTL network is trained for 40,000 iterations with a batch size of 4 on both the two datasets. The optimizer and the loss functions for different tasks follow the previous work [49]. + +# 4.2. Ablation Study + +In this subsection, we conduct extensive experiments to demonstrate the effectiveness of different components and find the best settings of different hyper-parameters. All the ablation experiments are conducted based on the ViT-based backbone if not specified otherwise. The baseline is built upon the ViT-base backbone with 12 layers, where the backbone features from the 3-rd, 6-th, 9-th, and 12-th layers are utilized as the multi-scale features, each of which is followed by a linear layer to project the channel dimension to the output channels for each task. + +Effectiveness of different components. We first conduct experiments to verify the effectiveness of different components of the MLoRE module. The quantitative results are shown in Tab. 2. We first examine the performance of the baseline with the standard MoE and the parameter size and FLOPs of their model. The expert networks in the standard MoE (15 experts) are similar to ours, each of which consists of a $3 \times 3$ convolution and a $1 \times 1$ convolution with ReLU among them. When adding the MoE to the baseline, we observe the MTL gain can be improved, but the parameters and FLOPs also expand about 5 times and 4 times, which is a heavy burden for the whole network. When adding the low-rank expert networks (LoRE) to the baseline, the performance is also improved, but the parameters and FLOPs are only $1/3$ and $1/3$ of the MoE-based model. When the low-rank property is applied to the expert net + +![](images/9c18f9cf1cc382e03a5ee457bc090eab8542b7344872bab0b6902d7ef1a3b319.jpg) +Figure 3. Ablation study on the number of experts $N$ and the number of activated experts $K$ . In the right figure, we also present the parameter change of the MLoRE module with the increase in the number of experts. + +![](images/389e55bcfd2f481bb3ef5ded0807d98532e12d53dd41320dd0696fd19b292a26.jpg) + +Table 3. Ablation on the rank setting in the task-sharing low-rank expert path of MLoRE on PASCAL-Context dataset. + +
Min/Max RankSemseg mIoU↑Parsing mIoU↑Saliency maxF↑Normal mErr↓Boundary odsF↑MTL Δm↑
16/1678.8468.0185.3213.6974.39-0.77
16/12879.2667.8285.3013.6574.69-0.58
128/12878.7966.9885.3513.6774.29-1.07
+ +works, the parameter size is reduced several times. When further introducing linearity in the expert networks by removing all activation functions, the computations are saved by re-parameterizing all experts to a single convolution. + +Furthermore, we also emphasize the importance of explicitly building global task correlations in the MoE and introduce the task-sharing generic path to achieve this goal. It can be seen that adding the task-sharing generic path to LoRE can further improve the performance and outperform the baseline with MoE on most metrics, which proves the effectiveness of modeling the global relationships among all tasks. In addition, adding a task-specific low-rank expert for each task also improves the performance, which demonstrates that the specialized expert can enhance the discrimination of task-specific features. We empirically set the rank of the task-specific low-rank expert to 64 in this paper. + +Number of task-sharing low-rank experts and top- $k$ selection. We ablate the number of low-rank experts in the MLoRE module and top- $k$ selection of the experts by the task-specific router networks. We first fix one and ablate another parameter to study their impact on multi-task performance. As shown in Fig. 3, when increasing the number of experts, the MTL gain of the model is significantly improved and allows us to achieve the best performance when the number of experts is 15. Further increasing the number of experts, we do not observe obvious performance gain. Besides, when fixing the number of experts to 15, we ablate the ratio of the activated experts. We observe that activating $60\%$ experts for each task is the best choice in our experiments. When selecting all expert networks, the performance decreases by a large margin, which reflects the importance of sparsity for feature discrimination. + +![](images/03e7c7917513671af0c74332c9ce3b99dc569c385da03763bf39f9fd37dbd1e8.jpg) +Figure 4. (a) The relations between tasks and low-rank experts. (b) The ratio of an expert activated by different numbers of tasks in the MLoRE module without the task-sharing generic path. We can see that without the task-sharing generic path, there is only a few experts can be activated by all five tasks. Horizontal coordinates represent the ranks of different experts. + +![](images/1a7079c070ff50d03056e70eaf56c1db42beec13c3ee2d4b92db3714f4dd0da2.jpg) + +Rank number setting. The expert networks utilize the low-rank format of a vanilla $3 \times 3$ convolution with 640 output channels. The rank $r$ of different expert networks also plays an important role. We study different settings, including 1) all experts with the same rank number 16, 2) all experts with the rank number 128, 3) all experts with the rank numbers from 16 to 128 with an increased step of 8. As shown in Tab. 3, it can be seen that selecting different rank numbers for expert networks achieves the best MTL gain. We analyze the expert networks with different rank numbers that can bring more feature diversity than the same rank numbers, which is more useful for assembling task-specific features, and utilize this as our settings. + +The relation visualizations between tasks and the low-rank experts are shown in Fig. 4(a). We count the relations from the second MLoRE module at the last stage and calculate the activation ratio of every expert selected by different tasks on the entire dataset. It can be seen that experts with different ranks tend to learn different subsets of tasks. Specifically, for this module, experts with lower ranks tend to learn shared knowledge for 3-4 related tasks, while experts with higher ranks tend to specialize in 1-2 tasks. Moreover, we also show the ratio of different experts to be activated by different numbers of tasks when the task-sharing generic path is not added to the MLoRE module in Fig. 4(b). It can be seen that in a fully dynamic manner, all these experts are seldom or even never activated by all the tasks in one sample. This proves the fact that almost no expert can learn global relationships across all tasks when utilizing the MoE in the decoder directly. This phenomenon strongly supports the necessity of the task-sharing generic path for explicitly modeling global task relationships. + +Task-specific router network. The router network is important for generating task-specific gates, which decide how to activate the experts and assemble their features. We ablate several settings of the router network and the results are shown in Tab. 4. Adding the position-aware branch to the basic router network can improve the MTL gain by $+0.17$ . The position-aware branch can obtain more context information, which benefits for the router network. Moreover, when replacing the router's input from the learnable pa + +Table 4. Ablation on the setting of the router network. basic: the basic router network. pos.: the positional-aware router network. w/o sample-dep: the input is the learnable parameters. + +
Router VariantSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑MTL Δm ↑
basic79.1567.4085.2113.5874.34-0.75
basic+pos.79.2667.8285.3113.6574.69-0.58
only pos.79.1067.7685.1113.7174.51-0.82
basic79.1567.4085.2113.5874.34-0.75
w/o sample-dep78.8667.3885.4113.6574.25-0.91
+ +rameters to the sample features, the MTL gain increases by $+0.16$ , which demonstrates that dynamic information from samples is vital for gates. + +# 4.3. Comparisons with Other Methods + +The quantitative comparisons with previous state-of-the-art (SOTA) methods are shown in Tab. 5 and Tab. 6. It can be seen that our method clearly outperforms previous methods in terms of all metrics on both the PASCAL-Context dataset and the NYUDv2 dataset. In particular, on the PASCAL-Context dataset, for semantic segmentation, human parsing, and boundary detection, the performance of our method outperforms the previous best method by $+0.52$ mIoU, $+1.10$ mIoU and $+1.92$ odsF, respectively. + +Previous methods, $\mathrm{M}^3\mathrm{ViT}$ [14], Mod-Squad [10], and TaskExpert [50] all utilize the MoE technique in their network. However, our method shows performance superior to theirs, which demonstrates the effectiveness of the MLoRE module. Compared with the decoder-focused method, TaskExpert, the performance of semantic segmentation, human parsing, and object boundary is significantly improved by $+0.77$ mIoU, $+1.10$ mIoU and $+2.12$ odsF, but using fewer parameters and FLOPs. + +Furthermore, we also present an intuitive visualization comparison among different methods in Fig. 5. Our method can generate better visualization results than the previous SoTA methods, especially for semantic segmentation, human parsing, and object boundary detection. More visualization comparisons are in the supplemental material. + +Table 5. Quantitative comparison of different methods on PASCAL-Context dataset. * denotes the reproduced performance of methods based on the ViT-large backbone in [50]. + +
MethodPublicationBackboneSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑FLOPs (G)#Param (M)
PAD-Net [46]CVPR'18HRNet1853.6059.6065.8015.3072.5012481
MTI-Net [45]ECCV'20HRNet1861.7060.1884.7814.2370.80161128
ATRC [4]ICCV'21HRNet1867.6762.9382.2914.2472.4221696
PAD-Net* [46]CVPR'18ViT-large78.0167.1279.2114.3772.60773330
MTI-Net* [45]ECCV'20ViT-large78.3167.4084.7514.6773.00774851
ATRC* [4]ICCV'21ViT-large77.1166.8481.2014.2372.10871340
InvPT [48]ECCV'22ViT-large79.0367.6184.8114.1573.00669423
TaskPrompter [49]ICLR'23ViT-large80.8968.8984.8313.7273.50497401
TaskExpert [50]ICCV'23ViT-large80.6469.4284.8713.5673.30622420
Ours-ViT-large81.4170.5284.9013.5175.42571407
+ +Table 6. Quantitative comparison of different methods on the NYUD-v2 dataset. Our method performs the best on all four tasks. + +
MethodBackboneSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑
PAD-Net [46]HRNet1836.610.624620.8876.38
MTI-Net [45]HRNet4845.970.536520.2777.86
ATRC [4]HRNet4846.330.536320.1877.94
InvPT [48]ViT-large53.560.518319.0478.10
TaskPrompter [49]ViT-large55.300.515218.4778.20
TaskExpert [50]ViT-large55.350.515718.5478.40
OursViT-large55.960.507618.3378.43
+ +# 4.4. Efficient MTL Models + +We also apply our MLoRE module to the ViT-small backbone to check the performance of the efficient models. Specifically, the channel number of the decoder decreases from 384 to 192. As shown in Tab. 7, using about $35\%$ GFLOPs of TaskExpert, our method can achieve a highly competitive result. In particular, the performance of the semantic segmentation and object boundary is improved by $0.6\%$ mIoU and $1.01\%$ odsF while the metrics of other tasks are close to TaskExpert. Furthermore, the parameters are fewer than TaskExpert by 11M. + +# 4.5. Conclusions + +We present a novel decoder-focused multi-task learning method, called MLoRE. We delve deep into the standard mixture-of-experts (MoE) technique and improve it for dense multi-task learning from two aspects. First, to address the neglected global relationship modeling in the MoE, we add a simple generic convolution path to MoE, enabling different task features to be able to share this path. Furthermore, we apply the low-rank format of a vanilla convolution to different expert networks to free the MoE from high computational cost and a large number of parameters when increasing the number of experts. Experiments + +![](images/ca4a757f080bfad844bb4f5a5c6c306ef67a4cd23488774da1511b1fa0a6992f.jpg) +Figure 5. Qualitative comparison among different methods, including InvPT [48], TaskPrompter [49], and ours. Best viewed with zoom-in. It can be seen that our method achieves better visual results than other methods on all five tasks thanks to the proposed MLoRE module. + +Table 7. Quantitative comparison of the MoE-based efficient models on PASCAL-Context dataset. + +
MethodSemseg mIoU ↑Parsing mIoU ↑Sal. maxF ↑Nor. mErr ↓Bound. odsF ↑FLOPs (G)Param (M)
M3ViT72.8062.1066.3014.5071.7042042
Mod-Squad74.1062.7066.9013.7072.0042052
TaskExpert75.0462.6884.6814.2268.8020455
ours75.6462.6584.7014.4369.817244
+ +demonstrate that the proposed method clearly outperforms previous state-of-the-art methods on all metrics. + +Acknowledgments. This research was supported by NSFC (NO. 62276145), the Fundamental Research Funds for the Central Universities (Nankai University, 070-63223049), CAST through Young Elite Scientist Sponsorship Program (No. YESS20210377). Computations were supported by the Supercomputing Center of Nankai University (NKSC). + +# References + +[1] Armen Aghajanyan, Luke Zettlemoyer, and Sonal Gupta. Intrinsic dimensionality explains the effectiveness of language model fine-tuning. arXiv preprint arXiv:2012.13255, 2020. 3 +[2] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4009-4018, 2021. 1 +[3] David Bruggemann, Menelaos Kanakis, Stamatios Georgoulis, and Luc Van Gool. Automated search for resource-efficient branched multi-task networks. arXiv preprint arXiv:2008.10292, 2020. 3 +[4] David Brüggemann, Menelaos Kanakis, Anton Obukhov, Stamatios Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In Int. Conf. Comput. Vis., pages 15869-15878, 2021. 1, 3, 8 +[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):834-848, 2017. 1 +[6] Tianlong Chen, Xuxi Chen, Xianzhi Du, Abdullah Rashwan, Fan Yang, Huizhong Chen, Zhangyang Wang, and Yeqing Li. Adamv-moe: Adaptive multi-task vision mixture-of-experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17346-17357, 2023. 1, 3 +[7] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1971-1978, 2014. 5 +[8] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 3 +[9] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. Advances in Neural Information Processing Systems, 33:2039-2050, 2020. 3 +[10] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. Mod-squad: Designing mixtures of experts as modular multi-task learners. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11828-11837, 2023. 1, 2, 3, 7 +[11] Aidan Clark, Diego De Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, et al. Unified scaling laws for routed language models. In Int. Conf. Mach. Learn., pages 4057-4086. PMLR, 2022. 2 +[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Diverse branch block: Building a convolution as an inception-like unit. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10886-10895, 2021. 5 + +[13] 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. In Int. Conf. Learn. Represent., 2020. 6 +[14] Zhiwen Fan, Rishov Sarkar, Ziyu Jiang, Tianlong Chen, Kai Zou, Yu Cheng, Cong Hao, Zhangyang Wang, et al. $\mathbf{M}^3$ vit: Mixture-of-experts vision transformer for efficient multi-task learning with model-accelerator co-design. In Adv. Neural Inform. Process. Syst., volume 35, pages 28441-28457, 2022. 1, 3, 7 +[15] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022. 2 +[16] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3205-3214, 2019. 1, 3 +[17] Michelle Guo, Albert Haque, De-An Huang, Serena Yeung, and Li Fei-Fei. Dynamic task prioritization for multitask learning. In Proceedings of the European conference on computer vision (ECCV), pages 270–287, 2018. 3 +[18] Pengsheng Guo, Chen-Yu Lee, and Daniel Ulbricht. Learning to branch for multi-task learning. In Int. Conf. Mach. Learn., pages 3854-3863. PMLR, 2020. 3 +[19] Qibin Hou, Li Zhang, Ming-Ming Cheng, and Jiashi Feng. Strip pooling: Rethinking spatial pooling for scene parsing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4003-4012, 2020. 5 +[20] Qibin Hou, Daquan Zhou, and Jiashi Feng. Coordinate attention for efficient mobile network design. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13713-13722, 2021. 5 +[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3, 4 +[22] Yerlan Idelbayev and Miguel A Carreira-Perpinan. Low-rank compression of neural nets: Learning the rank of each layer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8049–8059, 2020. 3 +[23] Keishi Ishihara, Anssi Kanervisto, Jun Miura, and Ville Hautamaki. Multi-task learning with attention for end-to-end autonomous driving. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2902-2911, 2021. 1 +[24] Robert A Jacobs and Michael I Jordan. Learning piecewise control strategies in a modular neural network architecture. IEEE Transactions on Systems, Man, and Cybernetics, 23(2):337-345, 1993. 3 +[25] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. *Neural Comput.*, 3(1):79-87, 1991. 3 +[26] Shibo Jie and Zhi-Hong Deng. Fact: Factor-tuning for lightweight adaptation on vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 1060-1068, 2023. 3 + +[27] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 3 +[28] Zhihao Li, Toshiyuki Motoyoshi, Kazuma Sasaki, Tetsuya Ogata, and Shigeki Sugano. Rethinking self-driving: Multi-task knowledge for better generalization and accident explanation ability. arXiv preprint arXiv:1809.11100, 2018. 1 +[29] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1925-1934, 2017. 1 +[30] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1871-1880, 2019. 3 +[31] Yen-Cheng Liu, Chih-Yao Ma, Junjiao Tian, Zijian He, and Zsolt Kira. Polyhistor: Parameter-efficient multi-task adaptation for dense vision tasks. Advances in Neural Information Processing Systems, 35:36889-36901, 2022. 3 +[32] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3431-3440, 2015. 1 +[33] Yongxi Lu, Abhishek Kumar, Shuangfei Zhai, Yu Cheng, Tara Javidi, and Rogerio Feris. Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5334-5343, 2017. 1, 3 +[34] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1851-1860, 2019. 6 +[35] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3994-4003, 2016. 1, 3 +[36] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3 +[37] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Int. Conf. Comput. Vis., pages 12179-12188, 2021. 1 +[38] Sebastian Ruder, Joachim Bingel, Isabelle Augenstein, and Anders Søgaard. Latent multi-task architecture learning. In AAAI Conf. Artif. Intell., volume 33, pages 4822-4829, 2019. 3 +[39] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In *Eur. Conf. Comput. Vis.*, pages 746-760. Springer, 2012. 5 +[40] Chi Su, Fan Yang, Shiliang Zhang, Qi Tian, Larry S Davis, and Wen Gao. Multi-task learning with low rank attribute embedding for person re-identification. In Proceedings of the IEEE international conference on computer vision, pages 3739–3747, 2015. 3 + +[41] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5227-5237, 2022. 3 +[42] Madeleine Udell, Corinne Horn, Reza Zadeh, Stephen Boyd, et al. Generalized low rank models. Foundations and Trends® in Machine Learning, 9(1):1-118, 2016. 3 +[43] Simon Vandenhende, Stamatios Georgoulis, Bert De Brabandere, and Luc Van Gool. Branched multi-task networks: deciding what layers to share. In Brit. Mach. Vis. Conf., 2019. 3 +[44] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 44(7):3614-3633, 2021. 3 +[45] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In Eur. Conf. Comput. Vis., pages 527-543. Springer, 2020. 1, 3, 8 +[46] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 675-684, 2018. 1, 3, 8 +[47] Yongxin Yang and Timothy M Hospedales. Trace norm regularised deep multi-task learning. arXiv preprint arXiv:1606.04038, 2016. 3 +[48] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. In Eur. Conf. Comput. Vis., pages 514-530. Springer, 2022. 1, 2, 3, 5, 8 +[49] Hanrong Ye and Dan Xu. Taskprompter: Spatial-channel multi-task prompting for dense scene understanding. In Int. Conf. Learn. Represent., 2022. 2, 3, 5, 6, 8 +[50] Hanrong Ye and Dan Xu. Taskexpert: Dynamically assembling multi-task representations with memorial mixture-of-experts. In Int. Conf. Comput. Vis., 2023. 2, 3, 7, 8 +[51] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Eur. Conf. Comput. Vis., pages 325-341, 2018. 1 +[52] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2636-2645, 2020. 1 +[53] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Zequn Jie, Xiang Li, and Jian Yang. Joint task-recursive learning for semantic segmentation and depth estimation. In Eur. Conf. Comput. Vis., pages 235–251, 2018. 3 +[54] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4106-4115, 2019. 3 +[55] Xiangyun Zhao, Haoxiang Li, Xiaohui Shen, Xiaodan Liang, and Ying Wu. A modulation module for multi-task learning + +with applications in image retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 401-416, 2018. 3 +[56] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4514–4523, 2020. 3 \ No newline at end of file diff --git a/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/images.zip b/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2d411e4766cbdee6ac1f088a863a00ac498ac296 --- /dev/null +++ b/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46664655c14453a2aecdc1cb55319121419de18be4afb3b6501e55bf1d5c8d66 +size 592022 diff --git a/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/layout.json b/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..43997a8460b9ba71a853a0bc93cf8773540bfbc1 --- /dev/null +++ b/2024/Multi-Task Dense Prediction via Mixture of Low-Rank Experts/layout.json @@ -0,0 +1,8936 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "type": "text", + "content": "Multi-Task Dense Prediction via Mixture of Low-Rank Experts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "text", + "content": "Yuqi Yang" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "text", + "content": " Peng-Tao Jiang" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "text", + "content": " Qibin Hou" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "text", + "content": " Hao Zhang" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "text", + "content": " Jinwei Chen" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "text", + "content": " Bo Li" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "text", + "content": "VCIP, CS, Nankai University " + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "text", + "content": "vivo Mobile Communication Co., Ltd " + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 550, + 185 + ], + "type": "text", + "content": "NKIARI, Shenzhen Futian yangyq2000@mail.nankai.edu.cn, pt.jiang@vivo.com, andrewhoux@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 238, + 290, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 290, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 290, + 539 + ], + "type": "text", + "content": "Previous multi-task dense prediction methods based on the Mixture of Experts (MoE) have received great performance but they neglect the importance of explicitly modeling the global relations among all tasks. In this paper, we present a novel decoder-focused method for multitask dense prediction, called Mixture-of-Low-Rank-Experts (MLoRE). To model the global task relationships, MLoRE adds a generic convolution path to the original MoE structure, where each task feature can go through this path for explicit parameter sharing. Furthermore, to control the parameters and computational cost brought by the increase in the number of experts, we take inspiration from LoRA and propose to leverage the low-rank format of a vanilla convolution in the expert network. Since the low-rank experts have fewer parameters and can be dynamically parameterized into the generic convolution, the parameters and computational cost do not change much with the increase of experts. Benefiting from this design, we increase the number of experts and its reception field to enlarge the representation capacity, facilitating multiple dense tasks learning in a unified network. Extensive experiments on the PASCAL-Context and NYUD-v2 benchmarks show that our MLoRE achieves superior performance compared to previous state-of-the-art methods on all metrics. Our code is available at https://github.com/YuqiYang213/MLoRE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 560, + 128, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 128, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 128, + 572 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 580, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 676 + ], + "type": "text", + "content": "Computer vision tasks, such as semantic segmentation [5, 29, 32, 51] and depth estimation [2, 37], have been significantly facilitated by deep learning techniques. Each vision task has its own elaborated deep models that usually follow a similar pipeline, i.e., feature extraction and prediction. Besides, some tasks also share relations. These facts motivate researchers to study Multi-Task Learning (MTL) that is able to unify different task models into a single one." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 209, + 548, + 404 + ], + "blocks": [ + { + "bbox": [ + 306, + 209, + 548, + 404 + ], + "lines": [ + { + "bbox": [ + 306, + 209, + 548, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 548, + 404 + ], + "type": "image", + "image_path": "37a04ce7301efed64ff44816185c6f102aa24a509590b8abc6ab5e449839c131.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 406, + 547, + 451 + ], + "lines": [ + { + "bbox": [ + 304, + 406, + 547, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 547, + 451 + ], + "type": "text", + "content": "Figure 1. Performance comparison with state-of-the-art methods. Our MLoRE based on the proposed mixture of low-rank experts achieves superior performance on all tasks. " + }, + { + "bbox": [ + 304, + 406, + 547, + 451 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 304, + 406, + 547, + 451 + ], + "type": "text", + "content": " denotes higher is better. " + }, + { + "bbox": [ + 304, + 406, + 547, + 451 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 304, + 406, + 547, + 451 + ], + "type": "text", + "content": " denotes lower is better." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 472, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 544 + ], + "type": "text", + "content": "The significant advantage of multi-task learning lies in that it can improve training and inference efficiency while keeping commensurate performance to each task model. Due to such an advantage, MTL models have been applied to several directions, including autonomous driving [23, 28, 52] and scene understanding [46, 48], etc." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 676 + ], + "type": "text", + "content": "In this paper, we focus on multi-task learning for dense scene understanding, in which each task predicts pixel-wise results. A research line of early works [4, 16, 33, 35, 45, 46, 48] focuses on designing delicate network architectures, including encoder-focused and decoder-focused methods. The encoder-focused methods [16, 33, 35] design handcrafted modules to share across task-specific encoders to construct task-generic features, while the decoder-focused methods [4, 45, 46, 48] propose tailored decoders to learn discriminative task-specific representations and build cross-task relationships." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Unlike the above methods focusing on designing static network architectures, some methods [6, 10, 14] introduce the mixture-of-experts (MoE) technique that sheds light on" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 683, + 286, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 286, + 702 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 286, + 702 + ], + "type": "text", + "content": "*The first two authors contributed equally to this paper. Work was done when Yuqi Yang was an intern at vivo." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 703, + 189, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 189, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 189, + 712 + ], + "type": "text", + "content": "†Qibin Hou is the corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27927" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 118, + 288, + 227 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": "Table 1. Parameters and FLOPs of different settings in the standard MoE and the proposed MLoRE. The left setting presents the number of experts and the kernel size of the convolutions in the expert network." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 118, + 288, + 227 + ], + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 227 + ], + "type": "table", + "html": "
SettingsParams (M)FLOPs (G)
MoEMLoREMoEMLoRE
5 experts, [1×1, 1×1]3.11.23.001.49
10 experts, [1×1, 1×1]4.71.64.491.58
15 experts, [1×1, 1×1]6.31.95.991.66
5 experts, [3×3, 1×1]14.93.414.247.12
10 experts, [3×3, 1×1]22.44.721.377.21
15 experts, [3×3, 1×1]29.96.028.497.29
", + "image_path": "ef3d696c3d423f981dccfac7b1520fe90c708406ab56da4a0e327a5a37ededde.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 247, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 287, + 378 + ], + "type": "text", + "content": "learning dynamic and automatic manner of task specialization and cooperations [10]. They utilize MoE to design the encoder blocks and dynamically select network paths for different tasks and different samples. However, compared to these encoder-focused methods, the utilization of MoE in the decoder is less studied, where only Ye et al. [50] first applied MoE to the decoder recently. It decodes task-specific features by dynamically assembling task-generic features from different experts and achieves superior performance than previous decoder-focused methods. This motivates us to delve deep into the MoE-based MTL decoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 379, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 628 + ], + "type": "text", + "content": "Benefiting from the dynamic routing process, these MoE-based methods can significantly improve the diversity in both parameters and features, resulting in more discriminative task-specific features. However, such a paradigm still has some limitations. First, although MoE-based methods can build connections in a subset of tasks by sharing the same expert in the dynamic routing process, the chances of expert sharing among all the tasks are very low, potentially hindering the router from building global relationships among all tasks. Nevertheless, the global relationship modeling among all tasks has been proven useful for dense MTL in [48, 49]. Based on this fact, we argue it is essential to explicitly model global relationship among all tasks in MoE. Furthermore, the capacity of the task-generic feature spaces is highly related to the number of experts. As empirically demonstrated in [11, 15], increasing the number of experts has the potential to facilitate multi-task learning in a unified network. However, for existing dense multi-task learning, adding more expert networks will introduce more parameters and computational costs, which is a heavy burden for the whole model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "To address the above issues, we propose a novel decoder-focused approach, which we call it Mixture of Low-Rank Experts (MLoRE). The core idea of the MLoRE framework is to explicitly model global relationships among all tasks in the MoE and free the MoE from heavily dense computation burdens when increasing the number of experts to enlarge the task-generic feature spaces and context. To address the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "first issue, MLoRE builds upon the basic MoE structure and introduces a task-sharing generic convolution path that parallels with the MoE. Specifically, the backbone features are first projected to different task features and then all of them are fed into the generic convolution path and the original MoE's expert networks. By simply sharing the same generic path among all tasks, different tasks can be globally correlated with each other. Furthermore, to enhance the discrimination of task-specific features, we exclude some experts from the dynamic routing process and enable them to serve specific tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 205, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 443 + ], + "type": "text", + "content": "To increase the number of experts while not bringing too many parameters and FLOPs, we take inspiration from LoRA that the basic models adapted to different tasks only needs low-rank weight updates. To be specific, we transform the expert networks of the MoE to the different low-rank formats of a vanilla convolution, which saves more than " + }, + { + "bbox": [ + 304, + 205, + 545, + 443 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 205, + 545, + 443 + ], + "type": "text", + "content": " of the parameters compared to the standard MoE module, as shown in Tab. 1. In addition, to control the computational cost brought by the increasing number of expert networks, we do not use any non-linear activation functions in all the expert networks and the generic convolution path, which enables re-parameterization during inference. Through re-parameterization, the knowledge of experts can be injected into the generic convolution path, reducing the computational cost for dense tasks. To our knowledge, we are the first to use linear experts in MoE for multi-task dense prediction. To verify the effectiveness of our method, we conduct comprehensive experiments on the PASCAL-Context and NYUD-v2 datasets. Our method achieves new performance records on all tasks as shown in Fig. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 445, + 501, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 501, + 456 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 501, + 456 + ], + "type": "text", + "content": "In summary, our contributions are three-fold:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 457, + 545, + 611 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 457, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 457, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 457, + 545, + 517 + ], + "type": "text", + "content": "- We analyze the issues of MoE when applied in multi-task learning and propose a novel decoder-focused framework, MLoRE, which can explicitly model global relationships among all tasks and enlarge the capacity of feature representations without increasing the model size too much." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 517, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 545, + 576 + ], + "type": "text", + "content": "- We introduce a simple task-sharing generic path to the MoE structure and propose linear and low-rank expert networks inspired by LoRA. The generic convolution path and low-rank expert paths can be linearly combined, enabling re-parameterization at inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 577, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 577, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 577, + 545, + 611 + ], + "type": "text", + "content": "- Experiments on PASCAL-Context and NYUDv2 show that the proposed method clearly outperforms previous state-of-the-art MTL methods on all tasks." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 626, + 392, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 626, + 392, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 626, + 392, + 639 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 646, + 457, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 457, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 457, + 659 + ], + "type": "text", + "content": "2.1. Dense Multi-Task Learning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "In computer vision, multi-task learning (MTL) for dense prediction tasks has been widely studied. The previous methods can be divided into two categories, including optimization-based methods and architecture-based meth" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27928" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": "ods [44]. Optimization methods [8, 9, 17, 27, 55] facilitate MTL by utilizing different strategies to balance the influence of each task. Architecture-based methods aim to design a unified deep network for MTL. They can be further classified into two categories, encoder-focused methods, and decoder-focused methods. (i) Encoder-focused methods [3, 18, 33, 38] design multi-task backbones to extract features adapted to different tasks. Typical methods include cross-stitch networks [35], neural discriminative dimensionality reduction networks [16], multi-task attention networks [30], branched networks [43], and mixture-of-expert networks [10, 14]. (ii) Decoder-focused methods [4, 45, 46, 48-50, 53, 54, 56] share the same backbone and design delicate heads to extract task-specific features for each task and cross-task relationships. The advantage of decoder-focused methods lies in that they can benefit from powerful off-the-shelf vision backbones, such as DINOv2 [36]. Our method also falls into the decoder-focused category and studies how to produce task-specific features with the MoE technique." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 320, + 162, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 320, + 162, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 320, + 162, + 333 + ], + "type": "text", + "content": "2.2. Mixture-of-Experts" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 339, + 289, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 289, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 289, + 626 + ], + "type": "text", + "content": "Mixture-of-Experts (MoE) [24, 25] learns multiple expert networks and a router network that controls the probability of each expert contributing to the final output. This technique is also used in multi-task learning, which can better adapt to the data diversity. Different expert networks learn different discriminative features. The router network learns hard/soft task-specific coefficients to dynamically assemble discriminative features for each task. Prior MoE-based multi-task methods [6, 10, 14] are mainly encoder-focused methods. They introduce MoE into the backbone blocks to sparsely activate different paths for different tasks in the inference stage. Recently, Ye et al. [50] first introduced the MoE technique to the decoder. They utilize the spatial context-aware gates to combine each pixel of the features from different expert networks. The above methods decompose the backbone feature into multiple generic feature spaces and assemble discriminative task-specific features from them. Unlike the above MoE-based MTL method, our method first builds global relationships among all tasks explicitly in the MoE structure rather than leaving this work implicitly done by the task-specific routers. Moreover, our proposed low-rank experts give MoE better efficiency compared to the naive MoE and this gap gradually gets larger as the number of experts increases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 635, + 167, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 167, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 167, + 647 + ], + "type": "text", + "content": "2.3. Low-Rank Structure" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "The low-rank structure is often used in deep learning for its efficiency [22, 40, 42, 47]. Recently, many methods in parameter-efficient adaptation [21, 26, 31, 41], such as LoRA [21], utilize the low-rank structure and have shown impressive results. LoRA takes inspiration from Aghajanyan et" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "al. [1] that the difference in weights between the pre-trained model and the adapted model resides on low intrinsic rank. It learns an extra low-rank matrix instead of tuning the whole layer for adaptation. More related to our work, earlier MTL methods [40, 47] utilize low-rank structure to model task-generic features and generate task-specific features by linear combination. Unlike them, our method utilizes the low-rank structure to control the computation budget when increasing the number of experts in MoE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 189, + 362, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 189, + 362, + 201 + ], + "spans": [ + { + "bbox": [ + 306, + 189, + 362, + 201 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 209, + 422, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 209, + 422, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 422, + 220 + ], + "type": "text", + "content": "3.1. Overall Framework" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "text", + "content": "The overall framework follows the multi-scale architecture as in previous works [49, 50]. We utilize an off-the-shelf vision transformer (ViT) as the encoder and collect multiscale features from different layers. Formally, given an input image " + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "text", + "content": " and a vision transformer " + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "text", + "content": ", we can attain multi-scale feature sets " + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}^l = \\mathcal{F}^l (\\mathbf{I})\\}" + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "text", + "content": " from different layers, where " + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l\\in \\mathbb{R}^{L\\times C}" + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "inline_equation", + "content": "L = H\\times W" + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "text", + "content": " denote the layer index, the number of patches, and the feature dimension, respectively. " + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^l (\\mathbf{I})" + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "text", + "content": " is the output features of the " + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 227, + 547, + 396 + ], + "type": "text", + "content": "-th transformer layer. The multi-scale features are fed into the decoder which includes two stacked MLoRE modules at each scale. For each task, the output features of MLoRE from different scales are concatenated together to generate the final task features for dense predictions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 401, + 489, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 401, + 489, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 489, + 414 + ], + "type": "text", + "content": "3.2. Preliminaries: Mixture of Experts" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": "Before describing the details of the proposed Mixture-of-Low-Rank-Experts (MLoRE) module, we first introduce the basic form of MoE " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "f_{moe}(\\cdot)" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": ". Formally, suppose MoE contains " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": " experts and " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": " router networks, denoted as " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "\\mathbb{E} = \\{E_1,E_2,\\dots,E_N\\}" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "\\mathbb{G} = \\{G_1,G_2,\\dots,G_T\\}" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": " are the number of experts and the number of tasks, respectively. The backbone feature " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": "-th layer is fed into the networks of " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": " experts and " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": " router networks, respectively. For convenience, the superscript " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": " is omitted in the following. For the " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": "-th expert, the discriminative output feature is generated by " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_n = E_n(\\mathbf{X})" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": ". In the meanwhile, MoE learns gates from the task-specific router networks for different tasks. For task " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": ", the gate value for each expert generated by the router network can be represented as " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^t = G_t(\\mathbf{X})" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^t\\in \\mathbb{R}^N" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": ". Finally, MoE generates task-specific feature " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^t" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": " for task " + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": " by utilizing gates to combine the expert features, which can be formulated as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 364, + 627, + 545, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 627, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 364, + 627, + 545, + 659 + ], + "type": "interline_equation", + "content": "\\mathbf {S} ^ {t} = f _ {\\text {m o e}} (\\mathbf {X}) = \\sum_ {n = 1} ^ {N} \\mathbf {g} _ {n} ^ {t} \\mathbf {X} _ {n}. \\tag {1}", + "image_path": "5058c40a7273c3f352af11f262f659f221ca5e834baf24cbad123861701fa4db.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{M}^3\\mathrm{ViT}" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " [14] and Mod-Squad [10], they deactivate some experts according to the corresponding gate values for one inference and select the top-" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " experts. The task-specific features are used to make predictions for each task." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27929" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 67, + 548, + 231 + ], + "blocks": [ + { + "bbox": [ + 52, + 67, + 548, + 231 + ], + "lines": [ + { + "bbox": [ + 52, + 67, + 548, + 231 + ], + "spans": [ + { + "bbox": [ + 52, + 67, + 548, + 231 + ], + "type": "image", + "image_path": "a24f7be17b8b52312e07257e3993109dfc5c625c26e2c4292c96f4630fa451a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 547, + 294 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 294 + ], + "type": "text", + "content": "Figure 2. Overall framework of the proposed method. The MLoRE modules are equipped at different layers, where the backbone features from different layers are fed into the MLoRE modules, respectively. At each selected layer, the backbone feature is first projected to different task features and then sent to the task-sharing convolution, task-sharing low-rank expert networks followed by the task-specific router network and task-specific low-rank expert networks. The outputs of these branches are accumulated to generate task-specific features. At each selected layer, we stack two MLoRE modules." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 289, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 289, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 289, + 447 + ], + "type": "text", + "content": "The advantages of MoE lie in that it enables the dynamic encoding of features for each sample and each task and increases feature encoding diversity via multiple experts. However, when applying the MoE technique to build an MTL decoder, we find that it struggles with building global task relationships. Moreover, when increasing the number of experts to enlarge the capacity of feature representations and the context of the expert network, the parameters and computational cost increase accordingly. To address the above issues, we propose the Mixture-of-Low-Rank-Expert (MLoRE) module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 455, + 212, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 212, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 212, + 468 + ], + "type": "text", + "content": "3.3. Mixture of Low-Rank Experts" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": "The overall pipeline of the proposed Mixture-of-Low-Rank-Expert (MLoRE) module is illustrated in Fig. 2. Considering building cross-task relations across the task-specific features, we first project the backbone features into task features using several lightweight convolutional layers. Then, the feature of each task is delivered to the task-sharing generic path and multiple task-sharing expert networks with different low-rank convolutions. The low-rank experts are selected according to the predictions of the task-specific router network for each task. Furthermore, except for building task-specific features based on task-specific routers, we introduce additional task-specific low-rank expert networks to assist in building more discriminative task-specific features. The features from the task-sharing generic path, task-sharing low-rank expert networks selected by task-specific routers and the task-specific expert networks are summed up to generate discriminative task-specific features. We introduce linearity into the MLoRE module and do not utilize any activation function in all paths, enabling reparameterization to reduce computational cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "content": "To be specific, the backbone feature " + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "content": "-th layer is first projected to multiple task splits corresponding to each task using " + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "content": " convolutions, which can be denoted as " + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}^t = f_{t,1\\times 1}(\\mathbf{X}),t\\in [1,\\dots ,T]\\}" + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "content": ". Then, the task feature is sent to three paths, i.e., task-sharing generic path " + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "inline_equation", + "content": "f_{g}(\\cdot)" + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "content": ", task-sharing low-rank expert path " + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "inline_equation", + "content": "f_{lre}(\\cdot)" + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "content": " with task-specific router network " + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "inline_equation", + "content": "f_{sr}^{t}(\\cdot)" + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "content": ", and task-specific low-rank expert path " + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "inline_equation", + "content": "f_{se}(\\cdot)" + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "content": ". The output task-specific feature " + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^t" + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "content": " is obtained by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 337, + 431, + 545, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 431, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 337, + 431, + 545, + 447 + ], + "type": "interline_equation", + "content": "\\mathbf {S} ^ {t} = f _ {g} (\\mathbf {X} ^ {t}) + f _ {s r} ^ {t} \\left(f _ {l r e} (\\mathbf {X} ^ {t})\\right) + f _ {s e} ^ {t} (\\mathbf {X} ^ {t}). \\qquad (2)", + "image_path": "f85e5bd356ba77e6fc10de3af07119207cddeb266907d2a66382560421950808.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 455, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 478 + ], + "type": "text", + "content": "In the following, we introduce the network details of all these paths in the MLoRE module." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 483, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 662 + ], + "type": "text", + "content": "Task-sharing generic path contains a simple " + }, + { + "bbox": [ + 304, + 483, + 547, + 662 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 483, + 547, + 662 + ], + "type": "text", + "content": " convolutional layer with a weight matrix " + }, + { + "bbox": [ + 304, + 483, + 547, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_g \\in \\mathbb{R}^{3 \\times 3 \\times C \\times C}" + }, + { + "bbox": [ + 304, + 483, + 547, + 662 + ], + "type": "text", + "content": " and a bias matrix " + }, + { + "bbox": [ + 304, + 483, + 547, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_g \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 483, + 547, + 662 + ], + "type": "text", + "content": ". As all task features will go through this generic convolution, it will be optimized by the gradients of different tasks simultaneously, which can help extract common features among all tasks. During the training process, we stop the gradients of this path for further back-propagation. The gradient is back-propagated through the other two paths. We found such a simple operation can better ease the optimization process and well solve the gradient conflicts. Experimental results in Sec. 4.2 show that the simple task-sharing generic path can bring performance improvements on all tasks, demonstrating the effectiveness of the idea of explicitly building the relationship across all tasks from a global perspective." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Task-sharing low-rank expert path. We take inspiration from LoRA [21] and adopt low-rank convolution which is a low-rank format of a vanilla convolution. Each task-sharing expert network shares a similar structure consisting of a" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27930" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": " convolution and a " + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": " convolution. The weights and bias of all task-sharing expert networks can be formulated as " + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{W}_{l\\text{re}b}^{n}, \\mathbf{b}_{l\\text{re}b}^{n}, \\mathbf{W}_{l\\text{rea}}^{n}, \\mathbf{b}_{l\\text{rea}}^{n} | n \\in [1, \\dots, N]\\}" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{l\\text{re}b}^{n} \\in \\mathbb{R}^{3 \\times 3 \\times C \\times r_{n}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{l\\text{re}b}^{n} \\in \\mathbb{R}^{r_{n}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{l\\text{rea}} \\in \\mathbb{R}^{1 \\times 1 \\times r_{n} \\times C}" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{l\\text{rea}}^{n} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "r_{n} \\ll C" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "r_{n}" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": " denotes the rank for the " + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": "-th expert network. In our method, " + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "r_{n}" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": " for different expert networks is different, aiming to improve the diversity of parameters and features. For each task, the task-specific router network " + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "f_{sr}^{t}(\\cdot)" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": " learns gate values for these experts and activate the top-" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": " experts according to the gate values. The output features of all the activated experts are summed up and then sent to a BatchNorm layer to generate task-specific features. The BatchNorm layer contains four parameters, including the accumulated channel-wise mean " + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": ", the accumulated channel-wise standard deviation " + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": ", the scaling factor " + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 47, + 72, + 289, + 265 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "text", + "content": "Task-specific low-rank expert path includes " + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "text", + "content": " specific expert networks and each processes one task feature. For each specific expert network, we utilize a similar structure with task-sharing expert path, which contains a " + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "text", + "content": " convolution with a weight matrix " + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{seb}^{t} \\in \\mathbb{R}^{3 \\times 3 \\times C \\times R}" + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "text", + "content": " and a bias matrix " + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{seb}^{t} \\in \\mathbb{R}^{R}" + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "text", + "content": ", followed by a " + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "text", + "content": " convolution with a weight matrix " + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{sea}^{t} \\in \\mathbb{R}^{1 \\times 1 \\times R \\times C}" + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "text", + "content": " and a bias matrix " + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{sea}^{t} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "text", + "content": " denotes the rank number (" + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "inline_equation", + "content": "R \\ll C" + }, + { + "bbox": [ + 47, + 269, + 289, + 389 + ], + "type": "text", + "content": "). The task-specific expert path can enhance the distinctiveness of the task-specific features, which will be verified in the experiments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "text", + "content": "Router network. As shown in Fig. 2, to generate task-specific features from the task-sharing low-rank expert path, we learn task-specific router networks to generate gate values for each expert and utilize them as the weight of the linear combination of feature output from different experts. The router network for each task is usually the simple linear layers followed by an average pooling layer and a prediction layer. Specifically, the router network " + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "inline_equation", + "content": "f_{sr}^{t}(\\cdot)" + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "text", + "content": "-th task is designed as follows. Our router network takes the task-specific features " + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "text", + "content": " as input and feed them into two consecutive " + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "text", + "content": " convolutions, mapping the channel dimension from " + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "inline_equation", + "content": "C/4" + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "text", + "content": ", followed by a global pooling layer. The output is a global feature vector " + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_f \\in \\mathbb{R}^{\\frac{C}{4}}" + }, + { + "bbox": [ + 47, + 394, + 289, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 552, + 289, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 289, + 685 + ], + "type": "text", + "content": "In addition, inspired by previous works [19, 20] showing that positional information is also important for modeling long-range spatial context, we introduce another parallel position-aware branch. Similarly, it consists of two linear layers. The first linear layer shrinks the feature along the spatial dimension, mapping the shape from " + }, + { + "bbox": [ + 47, + 552, + 289, + 685 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times HW}" + }, + { + "bbox": [ + 47, + 552, + 289, + 685 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 552, + 289, + 685 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times 1}" + }, + { + "bbox": [ + 47, + 552, + 289, + 685 + ], + "type": "text", + "content": ", which will then be transformed to " + }, + { + "bbox": [ + 47, + 552, + 289, + 685 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\frac{C}{4}}" + }, + { + "bbox": [ + 47, + 552, + 289, + 685 + ], + "type": "text", + "content": " via the second linear layer. The output feature vectors of these two branches are concatenated along the final dimension and then sent to the final prediction layer followed by a Softmax function to produce the gate values " + }, + { + "bbox": [ + 47, + 552, + 289, + 685 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 47, + 552, + 289, + 685 + ], + "type": "text", + "content": " for each expert." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "content": "Re-parameterization during inference. We introduce linearity into the MLoRE module by removing all activation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "content": "functions, enabling the parameters of all paths to reparameterize to a simple " + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "content": " convolution for each task at inference. We first parameterize the task-sharing low-rank expert path and then parameterize the parameters of all paths. According to [12], the weight and bias matrices in the task-sharing expert path can be combined and formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 152, + 547, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 547, + 180 + ], + "type": "interline_equation", + "content": "\\mathbf {W} _ {l r e} ^ {t} = \\mathfrak {B} \\left(\\frac {\\gamma}{\\sigma}\\right) \\sum_ {k \\in \\mathbb {K} _ {t}} \\mathbf {g} _ {k} ^ {t} \\mathbf {W} _ {l r e b} ^ {k} \\mathbf {W} _ {l r e a} ^ {k}, \\tag {3}", + "image_path": "ba00398b5cef25416541db862691aee97f27c7c4c68fdc66fcd55c4233882dec.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 182, + 547, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 182, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 318, + 182, + 547, + 209 + ], + "type": "interline_equation", + "content": "\\mathbf {b} _ {l r e} ^ {t} = \\frac {\\gamma}{\\sigma} (\\sum_ {k \\in \\mathbb {K} _ {t}} \\mathbf {g} _ {k} ^ {t} (\\mathbf {b} _ {l r e b} ^ {k} \\mathbf {W} _ {l r e a} ^ {k} + \\mathbf {b} _ {l r e a} ^ {k}) - \\mu) + \\beta , (4)", + "image_path": "47e76ea3125e79abcf3b4798ffab5c6fef8011d3b8a866908ceda8c84abbda8c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 217, + 547, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 547, + 278 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 217, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\mathfrak{B}" + }, + { + "bbox": [ + 304, + 217, + 547, + 278 + ], + "type": "text", + "content": " denotes the broadcast operation and " + }, + { + "bbox": [ + 304, + 217, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\mathbb{K}_t" + }, + { + "bbox": [ + 304, + 217, + 547, + 278 + ], + "type": "text", + "content": " denotes the index set of the activated experts selected by the router network for task " + }, + { + "bbox": [ + 304, + 217, + 547, + 278 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 217, + 547, + 278 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 217, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_k^t" + }, + { + "bbox": [ + 304, + 217, + 547, + 278 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 217, + 547, + 278 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 217, + 547, + 278 + ], + "type": "text", + "content": "-th gate value predicted by the router. At the inference time, the weight matrix and bias matrix of all these three paths can be re-parameterized as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 346, + 286, + 545, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 286, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 346, + 286, + 545, + 300 + ], + "type": "interline_equation", + "content": "\\mathbf {W} _ {r} ^ {t} = \\mathbf {W} _ {g} + \\mathbf {W} _ {s r} ^ {t} + \\mathbf {W} _ {s e b} ^ {t} \\mathbf {W} _ {s e a} ^ {t}, \\tag {5}", + "image_path": "fd3ee3669990332131f7a6f38cbbef2476df4dba758a24c6502a10a17f108ebb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 352, + 302, + 545, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 302, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 352, + 302, + 545, + 317 + ], + "type": "interline_equation", + "content": "\\mathbf {b} _ {r} ^ {t} = \\mathbf {b} _ {g} + \\mathbf {b} _ {s r} ^ {t} + \\mathbf {b} _ {s e b} ^ {t} \\mathbf {W} _ {s e a} ^ {t} + \\mathbf {b} _ {s e a} ^ {t}. \\qquad (6)", + "image_path": "2450a5ad2efaf609c48d360e2d485f4265ca7cded5e144dddd0a70b73563bd49.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 325, + 547, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 547, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_r^t" + }, + { + "bbox": [ + 304, + 325, + 547, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 325, + 547, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_r^t" + }, + { + "bbox": [ + 304, + 325, + 547, + 350 + ], + "type": "text", + "content": " are the weight and bias of the re-parameterized convolution. Thus, Eqn. (2) can be reformulated as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 370, + 358, + 545, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 358, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 370, + 358, + 545, + 373 + ], + "type": "interline_equation", + "content": "\\mathbf {S} ^ {t} = \\mathbf {X} ^ {t} * \\mathbf {W} _ {r} ^ {t} + \\mathfrak {B} \\left(\\mathbf {b} _ {r} ^ {t}\\right), \\tag {7}", + "image_path": "4b7a3fd080e49307da85aeabdfbcb2ad82944b303d010f865c2653adb5bcf5f8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 381, + 545, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 545, + 406 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 381, + 545, + 406 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 381, + 545, + 406 + ], + "type": "text", + "content": " denotes the convolution operation and " + }, + { + "bbox": [ + 304, + 381, + 545, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_r^t" + }, + { + "bbox": [ + 304, + 381, + 545, + 406 + ], + "type": "text", + "content": " is with the same shape as " + }, + { + "bbox": [ + 304, + 381, + 545, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t" + }, + { + "bbox": [ + 304, + 381, + 545, + 406 + ], + "type": "text", + "content": " via broadcast." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 417, + 388, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 417, + 388, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 417, + 388, + 430 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 436, + 433, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 436, + 433, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 436, + 433, + 449 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 459, + 547, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 547, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 547, + 603 + ], + "type": "text", + "content": "Datasets. To demonstrate the effectiveness of our method, we evaluate the performance of our method on two popular multi-task datasets, including PASCAL-Context [7] and NYUD-v2 [39]. PASCAL-Context [7] contains high-quality annotations of several tasks, including semantic segmentation, human parsing, saliency detection, surface normals, and object boundary detection. There are 4,998 training images and 5,105 test images in this dataset. NYUDv2 [39] also provides high-quality multi-task annotations, including semantic segmentation, monocular depth estimation, surface normals, and object boundary detection. This dataset contains 795 training images and 654 test images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Evaluation metrics. We introduce the evaluation metrics for the tasks mentioned above. Following previous multi-task works [48, 49], the mean intersection-over-union (mIoU) is used to evaluate semantic segmentation and human parsing. The root mean square error (RMSE) is used to evaluate the accuracy of monocular depth estimation. Saliency detection uses the maximum F-measure (maxF). Surface normal and object boundary detection adopt the mean error (mErr) and the optimal-dataset-scale F-measure" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "27931" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 138, + 288, + 236 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 137 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 137 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 137 + ], + "type": "text", + "content": "Table 2. Ablation study on different components in MLoRE on the PASCAL-Context dataset. Each row adds an extra setting to the above row. MoE: the standard mixture-of-experts structure; LoRE: task-sharing low-rank expert path; GC: task-sharing generic convolution path; SPE: task-specific expert path. " + }, + { + "bbox": [ + 47, + 70, + 288, + 137 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 47, + 70, + 288, + 137 + ], + "type": "text", + "content": " denotes higher is better. " + }, + { + "bbox": [ + 47, + 70, + 288, + 137 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 47, + 70, + 288, + 137 + ], + "type": "text", + "content": " denotes lower is better." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 138, + 288, + 236 + ], + "lines": [ + { + "bbox": [ + 47, + 138, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 138, + 288, + 236 + ], + "type": "table", + "html": "
SettingsSemsegParsingSal.NormalBound.MTLFLOPs#Param
mIoU↑ mIoU↑ maxF↑ mErr ↓odsF↑ Δm ↑(G)(M)
Baseline77.3865.1585.0813.7969.87-3.41391115
+ MoE78.5666.7885.1813.5773.91-1.201834676
Baseline
+ LoRE78.3866.2185.1513.7173.53-1.71568213
+ GC79.2567.4385.2013.7074.38-0.88568243
+ SPE79.2667.8285.3113.6574.69-0.58568259
", + "image_path": "eaf5467696709639d2777ccf3ca68f62373d02423517a7ac6bcaa32b7d08eadd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 257, + 287, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 282 + ], + "type": "text", + "content": "(odsF) as the evaluation metrics, respectively. In total, we evaluate the MTL gain " + }, + { + "bbox": [ + 46, + 257, + 287, + 282 + ], + "type": "inline_equation", + "content": "\\Delta_{m}" + }, + { + "bbox": [ + 46, + 257, + 287, + 282 + ], + "type": "text", + "content": " across all tasks, following [34]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 285, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 370 + ], + "type": "text", + "content": "Training settings. We utilize ViT-large [13] as the backbone. The channel number of the decoder is set to 384. For the ablation study, the ViT-base network is set as the backbone. Following the previous work [49], the proposed MTL network is trained for 40,000 iterations with a batch size of 4 on both the two datasets. The optimizer and the loss functions for different tasks follow the previous work [49]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 379, + 141, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 379, + 141, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 379, + 141, + 392 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 398, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 518 + ], + "type": "text", + "content": "In this subsection, we conduct extensive experiments to demonstrate the effectiveness of different components and find the best settings of different hyper-parameters. All the ablation experiments are conducted based on the ViT-based backbone if not specified otherwise. The baseline is built upon the ViT-base backbone with 12 layers, where the backbone features from the 3-rd, 6-th, 9-th, and 12-th layers are utilized as the multi-scale features, each of which is followed by a linear layer to project the channel dimension to the output channels for each task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": "Effectiveness of different components. We first conduct experiments to verify the effectiveness of different components of the MLoRE module. The quantitative results are shown in Tab. 2. We first examine the performance of the baseline with the standard MoE and the parameter size and FLOPs of their model. The expert networks in the standard MoE (15 experts) are similar to ours, each of which consists of a " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " convolution and a " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " convolution with ReLU among them. When adding the MoE to the baseline, we observe the MTL gain can be improved, but the parameters and FLOPs also expand about 5 times and 4 times, which is a heavy burden for the whole network. When adding the low-rank expert networks (LoRE) to the baseline, the performance is also improved, but the parameters and FLOPs are only " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " of the MoE-based model. When the low-rank property is applied to the expert net" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 424, + 156 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 424, + 156 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 424, + 156 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 424, + 156 + ], + "type": "image", + "image_path": "9c18f9cf1cc382e03a5ee457bc090eab8542b7344872bab0b6902d7ef1a3b319.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 159, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 304, + 159, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 547, + 205 + ], + "type": "text", + "content": "Figure 3. Ablation study on the number of experts " + }, + { + "bbox": [ + 304, + 159, + 547, + 205 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 159, + 547, + 205 + ], + "type": "text", + "content": " and the number of activated experts " + }, + { + "bbox": [ + 304, + 159, + 547, + 205 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 159, + 547, + 205 + ], + "type": "text", + "content": ". In the right figure, we also present the parameter change of the MLoRE module with the increase in the number of experts." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 424, + 71, + 545, + 156 + ], + "blocks": [ + { + "bbox": [ + 424, + 71, + 545, + 156 + ], + "lines": [ + { + "bbox": [ + 424, + 71, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 424, + 71, + 545, + 156 + ], + "type": "image", + "image_path": "389e55bcfd2f481bb3ef5ded0807d98532e12d53dd41320dd0696fd19b292a26.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 240, + 538, + 304 + ], + "blocks": [ + { + "bbox": [ + 305, + 217, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 305, + 217, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 545, + 239 + ], + "type": "text", + "content": "Table 3. Ablation on the rank setting in the task-sharing low-rank expert path of MLoRE on PASCAL-Context dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 240, + 538, + 304 + ], + "lines": [ + { + "bbox": [ + 315, + 240, + 538, + 304 + ], + "spans": [ + { + "bbox": [ + 315, + 240, + 538, + 304 + ], + "type": "table", + "html": "
Min/Max RankSemseg mIoU↑Parsing mIoU↑Saliency maxF↑Normal mErr↓Boundary odsF↑MTL Δm↑
16/1678.8468.0185.3213.6974.39-0.77
16/12879.2667.8285.3013.6574.69-0.58
128/12878.7966.9885.3513.6774.29-1.07
", + "image_path": "4577b6fbb9c3e67bdcf58c379d930855db1cbd5ed7099539b5359784b8a60d2f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 325, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 373 + ], + "type": "text", + "content": "works, the parameter size is reduced several times. When further introducing linearity in the expert networks by removing all activation functions, the computations are saved by re-parameterizing all experts to a single convolution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 374, + 546, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 546, + 518 + ], + "type": "text", + "content": "Furthermore, we also emphasize the importance of explicitly building global task correlations in the MoE and introduce the task-sharing generic path to achieve this goal. It can be seen that adding the task-sharing generic path to LoRE can further improve the performance and outperform the baseline with MoE on most metrics, which proves the effectiveness of modeling the global relationships among all tasks. In addition, adding a task-specific low-rank expert for each task also improves the performance, which demonstrates that the specialized expert can enhance the discrimination of task-specific features. We empirically set the rank of the task-specific low-rank expert to 64 in this paper." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": "Number of task-sharing low-rank experts and top-" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " selection. We ablate the number of low-rank experts in the MLoRE module and top-" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " selection of the experts by the task-specific router networks. We first fix one and ablate another parameter to study their impact on multi-task performance. As shown in Fig. 3, when increasing the number of experts, the MTL gain of the model is significantly improved and allows us to achieve the best performance when the number of experts is 15. Further increasing the number of experts, we do not observe obvious performance gain. Besides, when fixing the number of experts to 15, we ablate the ratio of the activated experts. We observe that activating " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " experts for each task is the best choice in our experiments. When selecting all expert networks, the performance decreases by a large margin, which reflects the importance of sparsity for feature discrimination." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27932" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 312, + 165 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 312, + 165 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 312, + 165 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 312, + 165 + ], + "type": "image", + "image_path": "03e7c7917513671af0c74332c9ce3b99dc569c385da03763bf39f9fd37dbd1e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 170, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 170, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 547, + 205 + ], + "type": "text", + "content": "Figure 4. (a) The relations between tasks and low-rank experts. (b) The ratio of an expert activated by different numbers of tasks in the MLoRE module without the task-sharing generic path. We can see that without the task-sharing generic path, there is only a few experts can be activated by all five tasks. Horizontal coordinates represent the ranks of different experts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 536, + 165 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 536, + 165 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 536, + 165 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 536, + 165 + ], + "type": "image", + "image_path": "1a7079c070ff50d03056e70eaf56c1db42beec13c3ee2d4b92db3714f4dd0da2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 214, + 289, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 214, + 289, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 289, + 370 + ], + "type": "text", + "content": "Rank number setting. The expert networks utilize the low-rank format of a vanilla " + }, + { + "bbox": [ + 46, + 214, + 289, + 370 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 214, + 289, + 370 + ], + "type": "text", + "content": " convolution with 640 output channels. The rank " + }, + { + "bbox": [ + 46, + 214, + 289, + 370 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 214, + 289, + 370 + ], + "type": "text", + "content": " of different expert networks also plays an important role. We study different settings, including 1) all experts with the same rank number 16, 2) all experts with the rank number 128, 3) all experts with the rank numbers from 16 to 128 with an increased step of 8. As shown in Tab. 3, it can be seen that selecting different rank numbers for expert networks achieves the best MTL gain. We analyze the expert networks with different rank numbers that can bring more feature diversity than the same rank numbers, which is more useful for assembling task-specific features, and utilize this as our settings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 373, + 289, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 289, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 289, + 602 + ], + "type": "text", + "content": "The relation visualizations between tasks and the low-rank experts are shown in Fig. 4(a). We count the relations from the second MLoRE module at the last stage and calculate the activation ratio of every expert selected by different tasks on the entire dataset. It can be seen that experts with different ranks tend to learn different subsets of tasks. Specifically, for this module, experts with lower ranks tend to learn shared knowledge for 3-4 related tasks, while experts with higher ranks tend to specialize in 1-2 tasks. Moreover, we also show the ratio of different experts to be activated by different numbers of tasks when the task-sharing generic path is not added to the MLoRE module in Fig. 4(b). It can be seen that in a fully dynamic manner, all these experts are seldom or even never activated by all the tasks in one sample. This proves the fact that almost no expert can learn global relationships across all tasks when utilizing the MoE in the decoder directly. This phenomenon strongly supports the necessity of the task-sharing generic path for explicitly modeling global task relationships." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "Task-specific router network. The router network is important for generating task-specific gates, which decide how to activate the experts and assemble their features. We ablate several settings of the router network and the results are shown in Tab. 4. Adding the position-aware branch to the basic router network can improve the MTL gain by " + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "inline_equation", + "content": "+0.17" + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": ". The position-aware branch can obtain more context information, which benefits for the router network. Moreover, when replacing the router's input from the learnable pa" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 249, + 547, + 338 + ], + "blocks": [ + { + "bbox": [ + 305, + 213, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 305, + 213, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 547, + 247 + ], + "type": "text", + "content": "Table 4. Ablation on the setting of the router network. basic: the basic router network. pos.: the positional-aware router network. w/o sample-dep: the input is the learnable parameters." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 249, + 547, + 338 + ], + "lines": [ + { + "bbox": [ + 307, + 249, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 249, + 547, + 338 + ], + "type": "table", + "html": "
Router VariantSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑MTL Δm ↑
basic79.1567.4085.2113.5874.34-0.75
basic+pos.79.2667.8285.3113.6574.69-0.58
only pos.79.1067.7685.1113.7174.51-0.82
basic79.1567.4085.2113.5874.34-0.75
w/o sample-dep78.8667.3885.4113.6574.25-0.91
", + "image_path": "f6313c28d3a4f91497140fb562ad532b8f932f7c9c5755ee68cbb6dd4c735202.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 360, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 360, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 360, + 547, + 396 + ], + "type": "text", + "content": "rameters to the sample features, the MTL gain increases by " + }, + { + "bbox": [ + 305, + 360, + 547, + 396 + ], + "type": "inline_equation", + "content": "+0.16" + }, + { + "bbox": [ + 305, + 360, + 547, + 396 + ], + "type": "text", + "content": ", which demonstrates that dynamic information from samples is vital for gates." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 405, + 489, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 405, + 489, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 405, + 489, + 418 + ], + "type": "text", + "content": "4.3. Comparisons with Other Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 425, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 547, + 533 + ], + "type": "text", + "content": "The quantitative comparisons with previous state-of-the-art (SOTA) methods are shown in Tab. 5 and Tab. 6. It can be seen that our method clearly outperforms previous methods in terms of all metrics on both the PASCAL-Context dataset and the NYUDv2 dataset. In particular, on the PASCAL-Context dataset, for semantic segmentation, human parsing, and boundary detection, the performance of our method outperforms the previous best method by " + }, + { + "bbox": [ + 304, + 425, + 547, + 533 + ], + "type": "inline_equation", + "content": "+0.52" + }, + { + "bbox": [ + 304, + 425, + 547, + 533 + ], + "type": "text", + "content": " mIoU, " + }, + { + "bbox": [ + 304, + 425, + 547, + 533 + ], + "type": "inline_equation", + "content": "+1.10" + }, + { + "bbox": [ + 304, + 425, + 547, + 533 + ], + "type": "text", + "content": " mIoU and " + }, + { + "bbox": [ + 304, + 425, + 547, + 533 + ], + "type": "inline_equation", + "content": "+1.92" + }, + { + "bbox": [ + 304, + 425, + 547, + 533 + ], + "type": "text", + "content": " odsF, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 533, + 548, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 548, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 548, + 640 + ], + "type": "text", + "content": "Previous methods, " + }, + { + "bbox": [ + 304, + 533, + 548, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{M}^3\\mathrm{ViT}" + }, + { + "bbox": [ + 304, + 533, + 548, + 640 + ], + "type": "text", + "content": " [14], Mod-Squad [10], and TaskExpert [50] all utilize the MoE technique in their network. However, our method shows performance superior to theirs, which demonstrates the effectiveness of the MLoRE module. Compared with the decoder-focused method, TaskExpert, the performance of semantic segmentation, human parsing, and object boundary is significantly improved by " + }, + { + "bbox": [ + 304, + 533, + 548, + 640 + ], + "type": "inline_equation", + "content": "+0.77" + }, + { + "bbox": [ + 304, + 533, + 548, + 640 + ], + "type": "text", + "content": " mIoU, " + }, + { + "bbox": [ + 304, + 533, + 548, + 640 + ], + "type": "inline_equation", + "content": "+1.10" + }, + { + "bbox": [ + 304, + 533, + 548, + 640 + ], + "type": "text", + "content": " mIoU and " + }, + { + "bbox": [ + 304, + 533, + 548, + 640 + ], + "type": "inline_equation", + "content": "+2.12" + }, + { + "bbox": [ + 304, + 533, + 548, + 640 + ], + "type": "text", + "content": " odsF, but using fewer parameters and FLOPs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Furthermore, we also present an intuitive visualization comparison among different methods in Fig. 5. Our method can generate better visualization results than the previous SoTA methods, especially for semantic segmentation, human parsing, and object boundary detection. More visualization comparisons are in the supplemental material." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27933" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 96, + 526, + 246 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 5. Quantitative comparison of different methods on PASCAL-Context dataset. * denotes the reproduced performance of methods based on the ViT-large backbone in [50]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 96, + 526, + 246 + ], + "lines": [ + { + "bbox": [ + 67, + 96, + 526, + 246 + ], + "spans": [ + { + "bbox": [ + 67, + 96, + 526, + 246 + ], + "type": "table", + "html": "
MethodPublicationBackboneSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑FLOPs (G)#Param (M)
PAD-Net [46]CVPR'18HRNet1853.6059.6065.8015.3072.5012481
MTI-Net [45]ECCV'20HRNet1861.7060.1884.7814.2370.80161128
ATRC [4]ICCV'21HRNet1867.6762.9382.2914.2472.4221696
PAD-Net* [46]CVPR'18ViT-large78.0167.1279.2114.3772.60773330
MTI-Net* [45]ECCV'20ViT-large78.3167.4084.7514.6773.00774851
ATRC* [4]ICCV'21ViT-large77.1166.8481.2014.2372.10871340
InvPT [48]ECCV'22ViT-large79.0367.6184.8114.1573.00669423
TaskPrompter [49]ICLR'23ViT-large80.8968.8984.8313.7273.50497401
TaskExpert [50]ICCV'23ViT-large80.6469.4284.8713.5673.30622420
Ours-ViT-large81.4170.5284.9013.5175.42571407
", + "image_path": "16226faa1e84e8817cdb2665d821db4caf439c99085db9303790b5ee870899b1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 290, + 287, + 392 + ], + "blocks": [ + { + "bbox": [ + 47, + 263, + 287, + 285 + ], + "lines": [ + { + "bbox": [ + 47, + 263, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 287, + 285 + ], + "type": "text", + "content": "Table 6. Quantitative comparison of different methods on the NYUD-v2 dataset. Our method performs the best on all four tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 290, + 287, + 392 + ], + "lines": [ + { + "bbox": [ + 48, + 290, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 290, + 287, + 392 + ], + "type": "table", + "html": "
MethodBackboneSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑
PAD-Net [46]HRNet1836.610.624620.8876.38
MTI-Net [45]HRNet4845.970.536520.2777.86
ATRC [4]HRNet4846.330.536320.1877.94
InvPT [48]ViT-large53.560.518319.0478.10
TaskPrompter [49]ViT-large55.300.515218.4778.20
TaskExpert [50]ViT-large55.350.515718.5478.40
OursViT-large55.960.507618.3378.43
", + "image_path": "325b54e721e51c0e3ea13350955c9556a1ec802963ba129b1428e7364dcc65b6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 412, + 174, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 174, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 174, + 424 + ], + "type": "text", + "content": "4.4. Efficient MTL Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 432, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 552 + ], + "type": "text", + "content": "We also apply our MLoRE module to the ViT-small backbone to check the performance of the efficient models. Specifically, the channel number of the decoder decreases from 384 to 192. As shown in Tab. 7, using about " + }, + { + "bbox": [ + 46, + 432, + 287, + 552 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 46, + 432, + 287, + 552 + ], + "type": "text", + "content": " GFLOPs of TaskExpert, our method can achieve a highly competitive result. In particular, the performance of the semantic segmentation and object boundary is improved by " + }, + { + "bbox": [ + 46, + 432, + 287, + 552 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 46, + 432, + 287, + 552 + ], + "type": "text", + "content": " mIoU and " + }, + { + "bbox": [ + 46, + 432, + 287, + 552 + ], + "type": "inline_equation", + "content": "1.01\\%" + }, + { + "bbox": [ + 46, + 432, + 287, + 552 + ], + "type": "text", + "content": " odsF while the metrics of other tasks are close to TaskExpert. Furthermore, the parameters are fewer than TaskExpert by 11M." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 562, + 127, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 127, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 127, + 574 + ], + "type": "text", + "content": "4.5. Conclusions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "We present a novel decoder-focused multi-task learning method, called MLoRE. We delve deep into the standard mixture-of-experts (MoE) technique and improve it for dense multi-task learning from two aspects. First, to address the neglected global relationship modeling in the MoE, we add a simple generic convolution path to MoE, enabling different task features to be able to share this path. Furthermore, we apply the low-rank format of a vanilla convolution to different expert networks to free the MoE from high computational cost and a large number of parameters when increasing the number of experts. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 262, + 545, + 395 + ], + "blocks": [ + { + "bbox": [ + 307, + 262, + 545, + 395 + ], + "lines": [ + { + "bbox": [ + 307, + 262, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 545, + 395 + ], + "type": "image", + "image_path": "ca4a757f080bfad844bb4f5a5c6c306ef67a4cd23488774da1511b1fa0a6992f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 399, + 545, + 453 + ], + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 453 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison among different methods, including InvPT [48], TaskPrompter [49], and ours. Best viewed with zoom-in. It can be seen that our method achieves better visual results than other methods on all five tasks thanks to the proposed MLoRE module." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 506, + 545, + 583 + ], + "blocks": [ + { + "bbox": [ + 305, + 477, + 545, + 499 + ], + "lines": [ + { + "bbox": [ + 305, + 477, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 545, + 499 + ], + "type": "text", + "content": "Table 7. Quantitative comparison of the MoE-based efficient models on PASCAL-Context dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 506, + 545, + 583 + ], + "lines": [ + { + "bbox": [ + 307, + 506, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 307, + 506, + 545, + 583 + ], + "type": "table", + "html": "
MethodSemseg mIoU ↑Parsing mIoU ↑Sal. maxF ↑Nor. mErr ↓Bound. odsF ↑FLOPs (G)Param (M)
M3ViT72.8062.1066.3014.5071.7042042
Mod-Squad74.1062.7066.9013.7072.0042052
TaskExpert75.0462.6884.6814.2268.8020455
ours75.6462.6584.7014.4369.817244
", + "image_path": "c5bdef0358164ab92c26fecf7b61bbfd5cc04e5321ef947e1758741d923ac5c6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 605, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 630 + ], + "type": "text", + "content": "demonstrate that the proposed method clearly outperforms previous state-of-the-art methods on all metrics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgments. This research was supported by NSFC (NO. 62276145), the Fundamental Research Funds for the Central Universities (Nankai University, 070-63223049), CAST through Young Elite Scientist Sponsorship Program (No. YESS20210377). Computations were supported by the Supercomputing Center of Nankai University (NKSC)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27934" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Armen Aghajanyan, Luke Zettlemoyer, and Sonal Gupta. Intrinsic dimensionality explains the effectiveness of language model fine-tuning. arXiv preprint arXiv:2012.13255, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "text", + "content": "[2] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4009-4018, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "type": "text", + "content": "[3] David Bruggemann, Menelaos Kanakis, Stamatios Georgoulis, and Luc Van Gool. Automated search for resource-efficient branched multi-task networks. arXiv preprint arXiv:2008.10292, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 225, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 288, + 269 + ], + "type": "text", + "content": "[4] David Brüggemann, Menelaos Kanakis, Anton Obukhov, Stamatios Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In Int. Conf. Comput. Vis., pages 15869-15878, 2021. 1, 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):834-848, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 390 + ], + "type": "text", + "content": "[6] Tianlong Chen, Xuxi Chen, Xianzhi Du, Abdullah Rashwan, Fan Yang, Huizhong Chen, Zhangyang Wang, and Yeqing Li. Adamv-moe: Adaptive multi-task vision mixture-of-experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17346-17357, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 392, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 446 + ], + "type": "text", + "content": "[7] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1971-1978, 2014. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 288, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 501 + ], + "type": "text", + "content": "[8] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 557 + ], + "type": "text", + "content": "[9] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. Advances in Neural Information Processing Systems, 33:2039-2050, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 613 + ], + "type": "text", + "content": "[10] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. Mod-squad: Designing mixtures of experts as modular multi-task learners. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11828-11837, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 669 + ], + "type": "text", + "content": "[11] Aidan Clark, Diego De Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, et al. Unified scaling laws for routed language models. In Int. Conf. Mach. Learn., pages 4057-4086. PMLR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "text", + "content": "[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Diverse branch block: Building a convolution as an inception-like unit. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10886-10895, 2021. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "text", + "content": "[13] 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. In Int. Conf. Learn. Represent., 2020. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 205 + ], + "type": "text", + "content": "[14] Zhiwen Fan, Rishov Sarkar, Ziyu Jiang, Tianlong Chen, Kai Zou, Yu Cheng, Cong Hao, Zhangyang Wang, et al. " + }, + { + "bbox": [ + 307, + 140, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3" + }, + { + "bbox": [ + 307, + 140, + 547, + 205 + ], + "type": "text", + "content": " vit: Mixture-of-experts vision transformer for efficient multi-task learning with model-accelerator co-design. In Adv. Neural Inform. Process. Syst., volume 35, pages 28441-28457, 2022. 1, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 205, + 547, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 205, + 547, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 547, + 249 + ], + "type": "text", + "content": "[15] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 250, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 547, + 304 + ], + "type": "text", + "content": "[16] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3205-3214, 2019. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 304, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 349 + ], + "type": "text", + "content": "[17] Michelle Guo, Albert Haque, De-An Huang, Serena Yeung, and Li Fei-Fei. Dynamic task prioritization for multitask learning. In Proceedings of the European conference on computer vision (ECCV), pages 270–287, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 349, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 545, + 381 + ], + "type": "text", + "content": "[18] Pengsheng Guo, Chen-Yu Lee, and Daniel Ulbricht. Learning to branch for multi-task learning. In Int. Conf. Mach. Learn., pages 3854-3863. PMLR, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "type": "text", + "content": "[19] Qibin Hou, Li Zhang, Ming-Ming Cheng, and Jiashi Feng. Strip pooling: Rethinking spatial pooling for scene parsing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4003-4012, 2020. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "type": "text", + "content": "[20] Qibin Hou, Daquan Zhou, and Jiashi Feng. Coordinate attention for efficient mobile network design. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13713-13722, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 459, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 545, + 503 + ], + "type": "text", + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 503, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 545, + 548 + ], + "type": "text", + "content": "[22] Yerlan Idelbayev and Miguel A Carreira-Perpinan. Low-rank compression of neural nets: Learning the rank of each layer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8049–8059, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "type": "text", + "content": "[23] Keishi Ishihara, Anssi Kanervisto, Jun Miura, and Ville Hautamaki. Multi-task learning with attention for end-to-end autonomous driving. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2902-2911, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "type": "text", + "content": "[24] Robert A Jacobs and Michael I Jordan. Learning piecewise control strategies in a modular neural network architecture. IEEE Transactions on Systems, Man, and Cybernetics, 23(2):337-345, 1993. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 669 + ], + "type": "text", + "content": "[25] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. *Neural Comput.*, 3(1):79-87, 1991. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 670, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 714 + ], + "type": "text", + "content": "[26] Shibo Jie and Zhi-Hong Deng. Fact: Factor-tuning for lightweight adaptation on vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 1060-1068, 2023. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27935" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[27] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 127, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 127, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 127, + 287, + 172 + ], + "type": "text", + "content": "[28] Zhihao Li, Toshiyuki Motoyoshi, Kazuma Sasaki, Tetsuya Ogata, and Shigeki Sugano. Rethinking self-driving: Multi-task knowledge for better generalization and accident explanation ability. arXiv preprint arXiv:1809.11100, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "text", + "content": "[29] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1925-1934, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 216, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 216, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 216, + 287, + 250 + ], + "type": "text", + "content": "[30] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1871-1880, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "type": "text", + "content": "[31] Yen-Cheng Liu, Chih-Yao Ma, Junjiao Tian, Zijian He, and Zsolt Kira. Polyhistor: Parameter-efficient multi-task adaptation for dense vision tasks. Advances in Neural Information Processing Systems, 35:36889-36901, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 293, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 337 + ], + "type": "text", + "content": "[32] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3431-3440, 2015. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 337, + 287, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 287, + 393 + ], + "type": "text", + "content": "[33] Yongxi Lu, Abhishek Kumar, Shuangfei Zhai, Yu Cheng, Tara Javidi, and Rogerio Feris. Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5334-5343, 2017. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "text", + "content": "[34] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1851-1860, 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "text", + "content": "[35] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3994-4003, 2016. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 536 + ], + "type": "text", + "content": "[36] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "type": "text", + "content": "[37] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Int. Conf. Comput. Vis., pages 12179-12188, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "text", + "content": "[38] Sebastian Ruder, Joachim Bingel, Isabelle Augenstein, and Anders Søgaard. Latent multi-task architecture learning. In AAAI Conf. Artif. Intell., volume 33, pages 4822-4829, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 658 + ], + "type": "text", + "content": "[39] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In *Eur. Conf. Comput. Vis.*, pages 746-760. Springer, 2012. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[40] Chi Su, Fan Yang, Shiliang Zhang, Qi Tian, Larry S Davis, and Wen Gao. Multi-task learning with low rank attribute embedding for person re-identification. In Proceedings of the IEEE international conference on computer vision, pages 3739–3747, 2015. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 715 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 127 + ], + "type": "text", + "content": "[41] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5227-5237, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 127, + 546, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 127, + 546, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 127, + 546, + 162 + ], + "type": "text", + "content": "[42] Madeleine Udell, Corinne Horn, Reza Zadeh, Stephen Boyd, et al. Generalized low rank models. Foundations and Trends® in Machine Learning, 9(1):1-118, 2016. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 162, + 546, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 546, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 546, + 206 + ], + "type": "text", + "content": "[43] Simon Vandenhende, Stamatios Georgoulis, Bert De Brabandere, and Luc Van Gool. Branched multi-task networks: deciding what layers to share. In Brit. Mach. Vis. Conf., 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 206, + 546, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 546, + 262 + ], + "type": "text", + "content": "[44] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 44(7):3614-3633, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 262, + 546, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 546, + 308 + ], + "type": "text", + "content": "[45] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In Eur. Conf. Comput. Vis., pages 527-543. Springer, 2020. 1, 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 546, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 546, + 363 + ], + "type": "text", + "content": "[46] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 675-684, 2018. 1, 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 363, + 546, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 363, + 546, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 546, + 397 + ], + "type": "text", + "content": "[47] Yongxin Yang and Timothy M Hospedales. Trace norm regularised deep multi-task learning. arXiv preprint arXiv:1606.04038, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 397, + 546, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 397, + 546, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 397, + 546, + 432 + ], + "type": "text", + "content": "[48] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. In Eur. Conf. Comput. Vis., pages 514-530. Springer, 2022. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 432, + 546, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 546, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 546, + 466 + ], + "type": "text", + "content": "[49] Hanrong Ye and Dan Xu. Taskprompter: Spatial-channel multi-task prompting for dense scene understanding. In Int. Conf. Learn. Represent., 2022. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 466, + 546, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 546, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 546, + 499 + ], + "type": "text", + "content": "[50] Hanrong Ye and Dan Xu. Taskexpert: Dynamically assembling multi-task representations with memorial mixture-of-experts. In Int. Conf. Comput. Vis., 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 499, + 546, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 499, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 499, + 546, + 544 + ], + "type": "text", + "content": "[51] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Eur. Conf. Comput. Vis., pages 325-341, 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 544, + 546, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 546, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 546, + 601 + ], + "type": "text", + "content": "[52] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2636-2645, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 601, + 546, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 546, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 546, + 645 + ], + "type": "text", + "content": "[53] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Zequn Jie, Xiang Li, and Jian Yang. Joint task-recursive learning for semantic segmentation and depth estimation. In Eur. Conf. Comput. Vis., pages 235–251, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 645, + 546, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 645, + 546, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 645, + 546, + 690 + ], + "type": "text", + "content": "[54] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4106-4115, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 690, + 546, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 546, + 715 + ], + "type": "text", + "content": "[55] Xiangyun Zhao, Haoxiang Li, Xiaohui Shen, Xiaodan Liang, and Ying Wu. A modulation module for multi-task learning" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "27936" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 152 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "with applications in image retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 401-416, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "text", + "content": "[56] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4514–4523, 2020. 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27937" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/cb273d05-390f-4fb9-aa74-74b5a46402e1_content_list.json b/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/cb273d05-390f-4fb9-aa74-74b5a46402e1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6ebc8254be169d8142baefe4f3c764f432a3a6bd --- /dev/null +++ b/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/cb273d05-390f-4fb9-aa74-74b5a46402e1_content_list.json @@ -0,0 +1,1781 @@ +[ + { + "type": "text", + "text": "Multi-View Attentive Contextualization for Multi-View 3D Object Detection", + "text_level": 1, + "bbox": [ + 101, + 130, + 867, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xianpeng Liu $^{1}$ , Ce Zheng $^{2}$ , Ming Qian $^{3}$ , Nan Xue $^{3}$ , Chen Chen $^{2}$ , Zhebin Zhang $^{4}$ , Chen Li $^{4}$ , Tianfu Wu $^{1}$ $^{1}$ North Carolina State University $^{2}$ University of Central Florida \n $^{3}$ Ant Group $^{4}$ OPPO U.S. Research Center \nhttps://xianpeng919.github.io/mvacon", + "bbox": [ + 86, + 179, + 906, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 312, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Multi-View Attentive Contextualization (MvACon), a simple yet effective method for improving 2D-to-3D feature lifting in query-based multi-view 3D (MV3D) object detection. Despite remarkable progress witnessed in the field of query-based MV3D object detection, prior art often suffers from either the lack of exploiting high-resolution 2D features in dense attention-based lifting, due to high computational costs, or from insufficiently dense grounding of 3D queries to multi-scale 2D features in sparse attention-based lifting. Our proposed MvACon hits the two birds with one stone using a representationally dense yet computationally sparse attentive feature contextualization scheme that is agnostic to specific 2D-to-3D feature lifting approaches. In experiments, the proposed MvA-Con is thoroughly tested on the nuScenes benchmark, using both the BEVFormer and its recent 3D deformable attention (DFA3D) variant, as well as the PETR, showing consistent detection performance improvement, especially in enhancing performance in location, orientation, and velocity prediction. It is also tested on the Waymo-mini benchmark using BEVFormer with similar improvement. We qualitatively and quantitatively show that global cluster-based contexts effectively encode dense scene-level contexts for MV3D object detection. The promising results of our proposed MvA-Con reinforces the adage in computer vision – “(contextualized) feature matters”.", + "bbox": [ + 76, + 316, + 473, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 739, + 209, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camera-based 3D object detection is a pivotal area of research in computer vision, particularly owing to its appealing applications in cost-effective autonomous systems, such as autonomous driving and robot autonomy. Recently, significant progress has been witnessed in the field of multiview 3D (MV3D) object detection, especially with the advent of end-to-end MV3D detection approaches [38]. One crucial component in the end-to-end MV3D detection system is the 2D-to-3D feature lifting module, which converts", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4b3c92ee1a1a867ad581bc546d93dce8b7fdd55aefd35d6ca4915a4d2798e344.jpg", + "image_caption": [ + "(a) Projected BEV anchor" + ], + "image_footnote": [], + "bbox": [ + 504, + 286, + 694, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/72f9097ae567920622785e2aca5192658191fac136935847c787a75bb3068327.jpg", + "image_caption": [ + "(b) BEVFormer [25]" + ], + "image_footnote": [], + "bbox": [ + 699, + 286, + 888, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f81bff84d620c7860b1cf7e2d1b47f57ac88b21ca09e94f489362b0d58ca13c1.jpg", + "image_caption": [ + "(c) Learned clusters" + ], + "image_footnote": [], + "bbox": [ + 503, + 387, + 694, + 473 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc7d655eb37f6fe469d90b6b5c093a6ce9f567b774320dfb3c5904332d09cf92.jpg", + "image_caption": [ + "(d) Our MvACon", + "Figure 1. The effects of our proposed MvACon in the 2D-to-3D feature lifting. Consider a point (red) on the car in (a), which is projected from a 3D BEV anchor point. In lifting 2D features to ground the 3D BEV anchor point, vanilla BEVFormer [25] utilizes a predefined number of deformable points, with offsets learned through a 6-layer cross-attention module relative to the projection point. (b) shows the deformed points after the final cross-attention layer, most of which have low attention weights, indicating the model's uncertainty or inability to consolidate contributions effectively for good lifting. Our MvACon tackles this issue with clustering-based attention, as visualized in (c). (d) shows the deformed points, where we observe not only high-confidence points on the car but also on the building. We further observe that the points on the building remain stable across encoding layers (see Fig. 4) and consecutive frames (see suppl.). With those high-confident deformed points in a spatiotemporally stable configuration, our MvACon may induce a local object-context aware coordinate system that helps the overall performance, especially the estimation of velocity and orientation, as we quantitatively observed in experiments. See text for details." + ], + "image_footnote": [], + "bbox": [ + 699, + 387, + 888, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "perspective 2D multi-view image feature maps to 3D feature representation. It aims to counter the complete loss of depth information in individual 2D images by exploiting multi-view clues. However, a significant challenge in practical applications like autonomous driving is the often insufficient field-of-view overlap across views, making it difficult to effectively address the loss of depth information.", + "bbox": [ + 496, + 779, + 893, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To perform 2D-to-3D feature lifting, recent methods", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16688", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[15, 25, 29, 57] aim to learn a unified 3D space representation using 3D anchors that are either sparsely or uniformly sampled. These methods generally fall into two categories subject to the interaction of 3D anchors with 2D features and the feature aggregation strategy. (1) The Lift-Splat-Shoot (LSS) method [14, 15, 23, 24, 45] first lifts 2D features into 3D (pseudo-LiDAR) space using the outer product with the estimated depth, then assigns them to the nearest 3D anchors. (2) In contrast, the query-based design [25, 29, 57], pioneered by the DETR method [4] for end-to-end 2D object detection, adopts 3D anchors as queries and uses 2D image features as keys and values. They interact and aggregate via spatial cross-attention in the expressive Transformer architecture [50]. These two paradigms have been widely used in downstream tasks like map segmentation [30] and occupancy prediction [25]. This paper primarily focuses on the query-based detection paradigm. One reason is that LSS-based methods often encounter excessive computational complexity and issues with error propagation and depth estimation magnification post-lifting, potentially capping their performance. However, the query-based design also grapples with heavy computation costs or limited 3D information awareness, depending on their Transformer design.", + "bbox": [ + 76, + 90, + 472, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we focus on addressing limitations of two main paradigms of query based MV3D object detectors in a unified way (elaborated in Sec. 3). In particular, we introduce multi-view attentive contextualization (MvA-Con) to address limitations of decoder-only dense attention methods like PETR [29], which lack high-resolution features due to computational constraints, and to simultaneously address the issue of sparsely grounded 3D anchors in encoder-decoder 2D/3D deformable attention methods such as BEVFormer [25] and DFA3D [19]. Our proposed MvA-Con aims to be representationally dense while computationally sparse. To achieve this, we expand the conventional three-component paradigm of MV3D object detection to a four-component setup (Fig. 2): (1) 2D image representation learning through a feature backbone shared across views, (2) MvACon for attentive contextualization of the 2D features, (3) 2D-to-3D feature lifting, and (4) a 3D object detection head or decoder that utilizes these lifted features. This modular design allows our MvACon to remain agnostic to specific 2D-to-3D feature lifting strategies and aligns with the classic adage in representation learning and computer vision: -('contextualized') feature matters'.", + "bbox": [ + 76, + 458, + 472, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More specifically, our approach contextualizes original feature maps extracted from the backbone network using a cluster-attention operation. This builds upon the recently proposed patch-to-cluster attention (PaCa) [12]. For perspective-based decoder-only detectors like PETR, we apply cluster contextualization before the feature maps are fed into the decoder. For encoder-decoder based detec", + "bbox": [ + 75, + 795, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tors, such as BEVFormer and DFA3D, we incorporate cluster contextualization within the spatial cross-attention operation. Through extensive experiments, we demonstrate that our proposed MvACon effectively and consistently enhances query-based MV3D object detectors by encoding more useful contexts, thereby facilitating better 2D-to-3D feature lifting. Rigorously controlled experiments reveal that, for perspective-based decoder-only detectors, the cluster attention contextualization significantly improves localization and velocity prediction. In the case of encoder-decoder based detectors, it effectively reduces errors in location, orientation, and velocity.", + "bbox": [ + 496, + 90, + 893, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are:", + "bbox": [ + 517, + 273, + 787, + 286 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We analyze and address the limitation of 2D-to-3D feature lifting in the prior art, that is the lack of sufficient 3D representational power due to their local 3D awareness.", + "- We propose MvACon (Multi-view Attentive Contextualization) to induce the global 3D awareness in an easy-to-integrate way to enhance the 2D-to-3D feature lifting in both decoder-only based MV3D object detectors and encoder-decoder based ones.", + "- We show consistent performance improvement of our MvACon on the challenging NuScenes [3] dataset using three baseline query-based MV3D object detectors, as well as on the Waymo-mini [11] benchmark." + ], + "bbox": [ + 500, + 287, + 890, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 482, + 640, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camera-based 3D Object Detection. Camera-based 3D object detection can be primarily categorized into two settings: single-view and multi-view. In the realm of monocular 3D object detection research, addressing the challenge of inaccurate object localization [37] is critical. Researchers have exerted considerable effort to utilize monocular depth cues. This includes transforming inputs into pseudo-lidar point clouds [35, 36, 56] and explicitly incorporating depth into models [8, 9, 16]. Another significant research direction involves the explicit use of geometric priors, encompassing approaches like key-point constraints [6, 20, 33], shape projection relationships [22, 34, 40, 64], and temporal depth estimation [54]. Innovations in monocular 3D object detection also include novel loss modules [7, 47, 65], 3D-aware backbones [2, 17, 18], and second-stage detection paradigms [28, 44]. In multi-view settings, the configuration often closely resembles that of monocular setups due to the limited field-of-view overlap between the different camera views. Therefore, MV3D object detection focuses on addressing the challenge of learning universal representation for multi-view sensors. It has benefited from advancements in various techniques such as view lifting [5, 19, 25, 29, 45, 57], depth encoding [19, 23, 24, 46], and temporal modeling [14, 25, 26, 30, 43, 52]. In our multi-view approach, we aim at addressing the challenge of multi-view representation learning with focus on en", + "bbox": [ + 496, + 507, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16689", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "hancing the view lifting module within query-based detection methods by empowering original features with cluster-based contextual features.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Representation Learning in Camera-based 3D Object Detection. Camera-based 3D object detection is inherently a data-intensive task due to its ill-posed nature, the expansive search space in 3D, and the scarcity of labeled data in scenes. Consequently, developing robust representations for this task is both critical and challenging. Early research in monocular 3D object detection has demonstrated the utility of depth contexts [41, 42] and projection contexts [27] in enhancing detection capabilities. Recent advances also highlight the effectiveness of scene-level representations, such as density fields [39], in improving 3D representation learning [59]. In the domain of multi-view research, most leading methods utilize backbones pre-trained with projection contexts (e.g., FCOS3D [53] weights) or depth contexts (e.g., DD3D [41] weights). However, these pre-trained weights may not fully leverage the capabilities of newer backbone network designs. Recent studies have begun to explore alternatives to this pre-trained paradigm, including the integration of an auxiliary projection context branch in end-to-end training [60]. Our work aims to enhance network representation by explicitly incorporating scene-level cluster context as supplementary information during the view lifting stage in query-based MV3D object detectors.", + "bbox": [ + 75, + 140, + 472, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision Transformers. Since the pioneer work of ViT [10], extensive research [31] has been dedicated to enhancing the representational abilities of neural networks for visual tasks. It has been established that CNNs and Transformers can mutually augment each other's capabilities, as evidenced in designs like Transformer-enhanced CNNs [48, 58] and CNN-enhanced Transformers [21, 49, 61]. Additionally, a significant branch of visual Transformer research focuses on developing new attention mechanisms tailored to the locality bias in vision tasks. Notable examples include HaloNet [51], SWin [32], Deformable Attention [67], and VOLO [62], all introducing innovative local attention mechanisms to mitigate the quadratic computational cost associated with visual inputs. Concurrently, models like TNT [13], ViL [63], PVTv2-linear [55], POTTER [66], and PaCa [12] explore the integration of local and global contexts. Inspired by the progress of Vision Transformers, our work addresses the limitations of two prevalent paradigms in query-based MV3D detectors caused by their attention mechanisms.", + "bbox": [ + 75, + 491, + 472, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 813, + 184, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of images $I_{i} \\in \\mathbb{R}^{3 \\times H \\times W}$ from $N$ cameras with known extrinsics $T_{i} \\in \\mathrm{SE}(3)$ and intrinsics $K_{i} \\in \\mathbb{R}^{3 \\times 3}$ , MV3D object detection aims to infer the label (e.g., Car, Pedestrian, Barrier) and the 3D bounding box for each ob", + "bbox": [ + 75, + 838, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ject instance in the scene. In this section, we first delve into the pipeline of query-based MV3D object detection in Sec. 3.1. We then analyze the pros and cons of the core 2D-to-3D feature lifting component in two state-of-the-art MV3D object detection methods in Sec. 3.2. Finally, we present our proposed MvACon in Sec. 3.3.", + "bbox": [ + 496, + 90, + 890, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Query-based MV3D Object Detection", + "text_level": 1, + "bbox": [ + 498, + 193, + 825, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For better understanding, we explain the query-based MV3D object detection pipeline, shown in Fig. 2, in a reverse manner. The 3D detection head typically builds upon DETR3D [57], which is based on the original DETR [4]. Initially, it defines a sufficiently large number, $O$ , of latent $C$ -dimensional 3D object queries, $Q_{O,C}$ . These object queries update using Keys and Values derived from multi-view inputs, followed by a classification head predicting object labels and a bounding box regression head determining the 3D bounding boxes. Obviously, the key challenge lies in how to transform / lift multi-view 2D inputs into 3D-aware Keys and Values.", + "bbox": [ + 496, + 218, + 890, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, a feature backbone is trained to extract deep 2D features from the multi-view input images. The complexity arises from different design choices for feature lifting. It mainly involves two aspects: representation and computation. From the representational perspective, multiscale feature pyramids, crucial in 2D object detection, become even more essential in 3D object detection. Computationally, handling multi-view inputs is already demanding. Adding multi-scale feature pyramids without careful optimization can significantly increase the computational load. There are two strategies in the state-of-the-art development of query-based MV3D object detection.", + "bbox": [ + 496, + 400, + 892, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Decoder-Only Architectures: Single-Scale Multi-View 2D Features with Dense Attention. These designs are among the most straightforward. They involves using multi-view 2D feature maps from the last layer of the feature backbone, which are then concatenated and flattened along the spatial dimensions to form Keys and Values. The latent object queries, $Q_{O,C}$ , are updated using a vanilla Transformer (i.e., each object query attends to every 2D location in the multi-view inputs, termed dense attention). However, this basic approach often fails as it does not encode any 3D-aware information. To address this, The PETR [29] introduces a physically-meaningful 3D position transformation as positional encoding, added to the multi-view 2D feature maps before concatenation and flattening. It has proven effective for query-based MV3D object detection, as illustrated in Option 1 in Fig. 2).", + "bbox": [ + 496, + 582, + 892, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoder-Decoder Architectures: Multi-Scale Multi-View 2D Features to Latent BEV Queries with Sparse Attention. The BEV (Bird's Eye View) representation acts as a unified, grid-based and ego-centric scene representation with predefined grid sizes (e.g., $200 \\times 200$ ) on the XZ", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16690", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f59ccbc492e5079cd54861e6c4fb9222a64fbfa244efb5d7d034a31258211189.jpg", + "image_caption": [ + "Figure 2. Overview of a query-based MV3D object detection pipeline with our proposed MvACon. Our proposed MvACon is a plug-and-play module for two state-of-the-art query-based MV3D object detection paradigms (e.g., PETR [29] and BEVFormer [25] respectively), which computes attentively contextualized features to facilitate better 2D-to-3D feature lifting in the two paradigms. See text for details." + ], + "image_footnote": [], + "bbox": [ + 86, + 85, + 880, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "plane. Geometrically, the BEV grid can be treated as a pillar-based point-cloud representation, collapsed along the Y-axis. A predefined number of points along the pillar (Y-axis) direction in the BEV grid will be uniformly sampled. These points form a uniform geometry prior for the underlying 3D scene and serve as BEV anchors to elevate 2D features into the BEV space. With known camera poses, these sampled points can be projected to each view at multiple scales. However, due to the uniform geometry prior in the projection, the projected points require deformability to better align with data observations. The BEVFormer [25] addresses this by introducing a sparse deformable attention mechanism (see Option 2 in Fig. 2). To counter the uniform geometry prior further, it learns a small predefined number of offsets, rather than directly deforming the projected points on each view, to lift 2D features from those deformed points with attentive weights. Latent BEV queries are introduced in learning these offsets and attentive weights. The BEV encoder's role is to refine the BEV queries, enabling them to provide meaningful offsets and attentive weights for lifting 2D features to 3D BEV anchors. Finally, embedded BEV queries as Keys/Values in the 3D detection head (i.e., decoder) update the latent object queries, $Q_{O,C}$ , e.g., through sparse deformable attention as in the BEVFormer, before predicting the 3D object detection results.", + "bbox": [ + 76, + 330, + 472, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. The Limitation of 2D-to-3D Feature Lifting in the Prior Art", + "text_level": 1, + "bbox": [ + 76, + 722, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although they both have shown remarkable progress for MV3D object detection, the PETR pipeline and the BEV-Former pipeline have a common limitation in their 2D-to-3D feature lifting, that is the local or shallow 3D awareness, rather than the desirable counterpart, the global and semantic meaningful 3D awareness.", + "bbox": [ + 75, + 762, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the PETR pipeline, consider the 2D feature map of the $n$ -th view, $F_{h\\times w\\times c}^{n}$ , where $(h,w)$ are the spatial sizes, height and width respectively, and $c$ the feature dimen", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sion of the backbone. The 3D position transform converts the (shared) camera frustum discretized as a mesh grid of sizes $(h,w,D)$ to the 3D space based on the known camera poses, where $D$ is the discretized depth levels. After the conversion, each 3D point is represented by a normalized 3D coordinate in the homogeneous form, i.e., $(x,y,z,1)$ . So the 3D position transformation results in the positional encoding $P_{h\\times w\\times 4\\cdot D}^{n}$ . Both $F_{h\\times w\\times c}^{n}$ and $P_{h\\times w\\times 4\\cdot D}^{n}$ are projected into the space of the same dimensionality, $d$ using a linear layer and a Multi-layer Perceptron (MLP) with the ReLU nonlinearity respectively, we have $F_{h\\times w\\times d}^{n}$ and $P_{h\\times w\\times d}^{n}$ which are summed in an element-wise way. Consider the latent representation of the 3D position in $P_{h\\times w\\times d}^{n}$ with the depth grid fused by the MLP, it is grounded to one feature point in $F_{h\\times w\\times d}^{n}$ , leading to the local 3D awareness.", + "bbox": [ + 496, + 330, + 890, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the BEVFormer pipeline, as we show in Fig. 1, the projected BEV anchor is often grounded to some low-confidence and scattered deformed points. Although the grounding may not be spatially local, they are often semantically shallow.", + "bbox": [ + 496, + 556, + 890, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Our Proposed MvACon Method", + "text_level": 1, + "bbox": [ + 500, + 645, + 782, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to address the local or shallow 3D awareness stated above by introducing an easy-to-integrate (mostly plug-and-play) module to learn the global and semantically meaningful 3D-awareness, as illustrated in Fig. 2. The basic idea of our MvACon is to attentively contextualize the 2D features in the 2D-to-3D lifting.", + "bbox": [ + 496, + 669, + 890, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the PETR pipeline, our idea is to contextualize the individual 2D feature map, $F_{h \\times w \\times d}^{n}$ ,", + "bbox": [ + 496, + 761, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {h \\times w \\times d} ^ {n} = \\operatorname {M v A C o n} \\left(F _ {h \\times w \\times d} ^ {n}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 792, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where after the contextualization every feature point in $\\mathbf{F}_{h\\times w\\times d}^n$ can connect to the entire map $F_{h\\times w\\times d}^n$ , inducing the global 3D awareness for grounding the positional encoding $P_{h\\times w\\times d}^n$ .", + "bbox": [ + 496, + 809, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the BEVFormer pipeline, our idea is to contextualize the multi-scale feature maps, e.g., the $l$ -th layer of the fea", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16691", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ture pyramid of the $n$ -th view, $F_{h\\times w\\times c}^{n,l}$", + "bbox": [ + 76, + 89, + 339, + 108 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {h \\times w \\times c} ^ {n, l} = \\operatorname {M v A C o n} \\left(\\left\\{F _ {h \\times w \\times c} ^ {n, l} \\right\\} _ {l = 1} ^ {L}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 108, + 467, + 127 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where after the contextualization every feature point in $\\mathbf{F}_{h\\times w\\times c}^{n,l}$ can connect to the entire $L$ -layer feature pyramid $\\{F_{h\\times w\\times c}^{n,l}\\}_{l = 1}^{L}$ , inducing the global 3D awareness for grounding projected BEV anchors on the $n$ -the view.", + "bbox": [ + 76, + 128, + 467, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve the global contextualization effect, we adapt the recently proposed Patch-to-Cluster attention (PaCa) [12] method. The core idea of PaCa is to leverage a learnable clustering module to cluster a feature map into a predefined number $M$ of clusters. For notional simplicity, consider a feature map $F_{h\\times w\\times c}$ as $N = h\\times w$ tokens $F_{N\\times c}$ , the clustering assignment is computed by,", + "bbox": [ + 76, + 191, + 467, + 296 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {N, M} = \\operatorname {S o f t m a x} (\\operatorname {C l u s t e r i n g} (F _ {N, c})), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 297, + 467, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where Clustering() can be implemented in different ways (see our ablation studies in Tab. 6), and the Softmax is along the token dimension. Then, we compute $M$ clusters by,", + "bbox": [ + 76, + 315, + 467, + 359 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {M, c} = \\mathrm {L N} \\left(\\mathcal {C} _ {N, M} ^ {\\top} \\cdot F _ {N, c}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 361, + 467, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{LN}()$ is the layer normalization [1].", + "bbox": [ + 76, + 377, + 356, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the PaCa-based MvAcon is defined by,", + "bbox": [ + 96, + 393, + 395, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {N, c} ^ {\\prime} = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {N , c} \\cdot K _ {M , c} ^ {\\top}}{\\sqrt {c}}\\right) \\cdot V _ {M, c} + F _ {N, c}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 407, + 467, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Q_{N,c}$ is the linear projection of $F_{N,c}, K_{M,c}$ and $V_{M,c}$ are from the clusters $z_{M,c}$ . The second term is the shortcut. The multi-head PaCa can be straightforwardly defined. For Eqn. 2, we concatenate the clusters from all the pyramid layers before computing the Key and the Value. Here, the PaCa module is of linear complexity.", + "bbox": [ + 76, + 443, + 467, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 547, + 207, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 573, + 264, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset and Metrics We evaluate our MvAcon on the challenging large-scale NuScenes dataset [3] and Waymo dataset [11]. The NuScenes dataset includes 1,000 scene sequences, which are divided into training, validation, and testing subsets in a 700/150/150 split. Each sequence in the NuScenes dataset is a 20-second video clip, annotated at a rate of 2 frames per second (FPS). The NuScenes dataset employs a comprehensive suite of evaluation metrics for assessing detection performance. These metrics comprise mean Average Precision (mAP), mean Average Translation Error (mATE), mean Average Scale Error (mASE), mean Average Orientation Error (mAOE), mean Average Velocity Error (mAVE), mean Average Attribution Error (mAAE), and the NuScenes Detection Score (NDS). The Waymo dataset contains 798 training sequences and 202 validation sequences. We use a subset of the training set (Waymomini) by sampling every third frame from the training sequences following [25].", + "bbox": [ + 76, + 597, + 467, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details We leverage open-source code bases (PETR [29], BEVFormer [25], and DFA3D [19]) in", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "our experiments. To ensure a fair and stringent comparison, we maintain all original configurations of these methods, making only one modification: the addition of an attentive contextualization module. We conduct qualitative analysis and ablation study on the BEVFormer-base model. We train all models for 24 epochs using 8 NVIDIA Tesla A100 GPUs, following the configurations and settings outlined in previous works [19, 25, 29].", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. The Effectiveness of MvACon across Different Methods", + "text_level": 1, + "bbox": [ + 498, + 220, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our proposed MvA-Con method, we first apply it to two typical query-based MV3D object detection paradigms: the perspective-based decoder-only detector (PETR [29]) and the encoder-decoder based detector (BEVFormer [25]). We choose these as our baselines because state-of-the-art (SOTA) query-based MV3D detectors [19, 30, 52, 60] primarily follow these two paradigms. We also test our method on DFA3D [19] to demonstrate its generalizability to SOTA methods.", + "bbox": [ + 496, + 258, + 890, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the NuScenes dataset, Table 1 shows that our proposed MvACon consistently improves performance across different detectors. Specifically, for the perspective-based decoder-only detector PETR, it improves the baseline by 0.8 NDS. For the encoder-decoder based detector BEVFormer, our method achieves an improvement of 1.3 in NDS on average. On a more advanced, depth-context enhanced BEVFormer (DFA3D), our method further improves performance by up to 0.5 NDS. Notably, our MvACon achieves significant improvement in location (mAP, mATE), orientation (mAOE), and velocity prediction (mAVE) for encoder-decoder based detectors. It also markedly enhances performance in location (mAP, mATE) and velocity (mAVE) prediction for the perspective-based decoder-only detector.", + "bbox": [ + 496, + 396, + 890, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the Waymo dataset, since there are few released codes for MV3D detectors on Waymo except for the BEVFormer, we only test BEVFormer on Waymo-mini following its settings with results shown in Table 2. Our MvACon shows consistent improvement on Waymo metrics.", + "bbox": [ + 496, + 607, + 890, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. How Does Our MvACon Work?", + "text_level": 1, + "bbox": [ + 498, + 691, + 779, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We elaborate on the effects of our MvACon by providing detailed analyses during the 2D-to-3D feature lifting process. We first demonstrate what the learned cluster contexts encode, then show how these contexts affect the behavior of deformable points during feature lifting. Lastly, we illustrate how our MvACon improves detection results by presenting a qualitative comparison within a scene. We select BEVFormer-base as our analysis target due to its incorporation of six layers of deformable attention modules in the encoder. More qualitative analysis is provided in the supplementary materials.", + "bbox": [ + 496, + 715, + 890, + 881 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "What do the Learned Cluster Contexts Encode? We vi-", + "bbox": [ + 500, + 885, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16692", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d9b0f9ec2c20605534ddef99ffe26296ef9d22350491fc62eb85643b4e7e12a0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
PETR-VovNet-99 [29]37.874.627.248.890.621.242.6
PETR-VovNet-99-MvACon38.2 (+0.5)73.927.050.584.121.343.4 (+0.8)
BEVFormer-t [25]25.290.029.465.565.721.635.4
BEVFormer-t-MvACon25.9 (+0.7)88.428.864.660.522.536.5 (+1.1)
BEVFormer-s [25]37.072.128.040.743.622.047.9
BEVFormer-s-MvACon39.3 (+2.3)71.327.740.142.019.749.6 (+1.7)
BEVFormer-b [25]41.667.327.437.239.419.851.7
BEVFormer-b-MvACon42.6 (+1.0)66.427.635.036.220.052.8 (+1.1)
DFA3D-s [19]40.172.127.941.139.119.650.1
DFA3D-s-MvACon40.171.027.438.337.220.850.6 (+0.5)
DFA3D-b [19]43.065.427.137.434.120.553.1
DFA3D-b-MvACon43.2 (+0.2)66.427.534.432.320.753.5 (+0.4)
", + "bbox": [ + 181, + 82, + 787, + 275 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c0aa0f973601d562baf867b1786bce2e4d9992d5b22c24c6d19d5b0565ee4474.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 328, + 313, + 404 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3e9ee985e76f829fcd96abfa656ad52caef07299387e4e36ea72cfec2d6cdf49.jpg", + "image_caption": [ + "Figure 3. Visualization results of learned cluster contexts in our MvACon on the NuScenes validation set. We sum all the learned clusters along the channel and upsample it to the original image resolution through bilinear interpolation. We observed that the learned cluster context encodes abundant context information in the scene. We provide details with raw images in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 143, + 405, + 313, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f24cf33f389162a391e4b02ad7503f079bc6ec580145ec613ecaec4be3c7096c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 315, + 328, + 480, + 404 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7d78c0edf5d69db9010c2f652450b12b3432ef79dbdd716dfdc75b32c3f18498.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 315, + 405, + 480, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9b0b78e6189103ed8c556c121f5244ecc9bbc0364c5c9eeefb5ea707b212bd27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 328, + 650, + 404 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0d8d9fc9e6f66c02159cd0eb1ac1fb0333032de245e690a71bb8cdcd17832197.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 405, + 650, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a56d1c6240209c0b7350f78a1ed48bf598f448dc478a2108e4a56ae991f00ccf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 328, + 818, + 404 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3c589eba87906b82b521e299d80e8cdf1990b8745ee5df20ed56809d052b1352.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 405, + 818, + 481 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3144b4b6226437684ebd86a546cfc05b6c13aacb33bd6a40331035193a56b2d1.jpg", + "table_caption": [ + "Table 1. Comparisons of our method with baslines on the NuScenes validation set. BEVFormer-t/s/b refers to BEVFormer-tiny/small/base in the BEVFormer's open-source codes. 'DFA3D' refers to the adaptation of 2D deformable attention into a (depth-weighted) 3D deformable attention within the BEVFormer model, as adopted in [19]." + ], + "table_footnote": [], + "table_body": "
MethodLET-mAPL↑LET-mAPH↑
BEVFormer-ResNet101 [25]34.946.3
BEVFormer-ResNet101-MvACon35.7 (+0.8)47.5 (+1.2)
", + "bbox": [ + 99, + 536, + 446, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparisons on the Waymo-mini.", + "bbox": [ + 143, + 590, + 401, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sualize the learned cluster context in a heatmap, by summing all clusters along the channel and then upsampling to the original image resolution using bilinear interpolation. The results, shown in Fig. 3, reveal that despite the low resolution of the original summed heatmap, it can accurately locate foreground objects in the scene after upsampling. This suggests that the cluster contexts encode the foreground layouts in a scene. Furthermore, we note that the upsampled heatmap shows a high response to various foreground objects, even when they are spatially close in 2D. This is attributed to the effectiveness of our MvACon in dense scenes, as demonstrated and analyzed later in our experiments.", + "bbox": [ + 75, + 622, + 468, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "How do Cluster Contexts Affect the Behavior of Deformable Points? We demonstrate the dynamics of deformable points in Fig. 4. These points originate from one 2D reference point, which is projected from a 3D BEV anchor point in the BEVFormer encoder. As the encoding", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layers deepen in the vanilla BEVFormer, we observe that most deformable points have low attention weights, suggesting the model's uncertainty about the relevance of the selected context. Consequently, these contexts contribute minimally during the encoding of BEV query features in the feature lifting process. In contrast, deformable points predicted by our method maintain high confidence weights on foreground objects (e.g., cars), as well as on surrounding buildings. We also note that points on buildings remain stable across encoding layers and consecutive frames (see the supplementary). These high-confidence deformed points in our MvACon may foster a local object-context aware coordinate system, enhancing overall performance, including the estimation of velocity and orientation. This observation aligns with our quantitative findings in Table 1.", + "bbox": [ + 496, + 539, + 890, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons with the Baseline Method. We provide qualitative comparison with BEVFormer in Fig. 5. Our method exhibits superior performance in dense scenarios where objects are crowded and challenging to localize. We attribute this enhanced performance to the rich foreground layout context in the scene, facilitated by our attentive contextualization module. Additional qualitative comparisons are available in the supplementary.", + "bbox": [ + 496, + 779, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16693", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ea8e463d6a162de83c5c59461ee9eff5ca7f6088b46876653e7c8d36107398dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 95, + 382, + 190 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5ab9c2bd43a57241e3d3282aa56f8c992107df2f89f37056b8180bd12ba3ad4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 95, + 598, + 190 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3f0a40011a876d1f04a9438d4072a5a65c86ad481ced787ba7be0c1124015fb7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 95, + 810, + 190 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82a96e50fa205cc7b1d84d5e613e3453a601b9a90c04d56cbbd711cda24fcb9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 191, + 382, + 285 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/70ad81e688f2c4c764f5e312ef38fba4f1403e3d956f0da28772c7c3abc462ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 191, + 598, + 285 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec8d0c83009dcc5348cf026119890974632859e4c47f40946d9dddfac979f401.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 191, + 810, + 285 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0803f15021e679c4855e4156aae7e86d2f7a06441cad50fd4ff5a9a290c56d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 287, + 382, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1aa1fb988b350f72349578cfed2dbb8cbcdb03432482af17434e64ff4e79afd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 290, + 598, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d307767d3e5d58525008b9b76b67cd6d38f68cb22260f6ee1723f5281a699a24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 291, + 810, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/246726bcd7652b194da8259cba91b0c8080c7eed480c107abda1c9d0929374c5.jpg", + "image_caption": [ + "Figure 4. Visualization results of the deformable points originating from a 2D reference point, which is projected from a 3D BEV anchor point in the BEVFormer encoder, on NuScenes validation set. We utilize the same BEV anchor point as demonstrated in Fig. 1. From left to right and up to bottom, we display the deformable points output from each layer (#1-#6) in the encoder, respectively." + ], + "image_footnote": [], + "bbox": [ + 143, + 387, + 382, + 481 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b2bfd2901d36a70ac5abb4bc8e03529ea044f310ba5d0d04cb95e4652df426af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 387, + 598, + 481 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/30201d8134e19dc120cc42b9215218fffe89654e81f79cdae7290c978908fc9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 387, + 810, + 481 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc9568dc4ce23aa7eaf48d326f7f561c6f32b107e432a17ea178486f864f0cc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 535, + 339, + 698 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c03d03ac0013399ac7ba2bcc9ed0dd23491a7bc8945cfbcaba171bea6f82447.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 535, + 509, + 614 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18f5e56b228531861d0e143ce986f9ee3918cc0a9d7bb550387cc57bcb114820.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 537, + 674, + 613 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/210aa166ae1c98c05cd71a15d95ed115479a3fb470a8b3b9ba929b58e8561c02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 537, + 841, + 613 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d5bccb46a4b166489d57ecfda6f703addcb729d02a0c0f09ee192ada2acc644.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 616, + 509, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f62eb02409fedcab9e8b5d61e6a5fea1b6f58b59722a222453c9914a11e04e41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 616, + 674, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd64fd307686675c626386891f762e3913f88f24ea5d84a2d1a7c1ad797abc42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 616, + 841, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0f13c5f3c184899e1ed6ed39fcba720db8dc763f03494f4ec1a80c4e89429ade.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons between BEVFormer and our MvACon method on NuScenes validation set." + ], + "image_footnote": [], + "bbox": [ + 109, + 699, + 338, + 861 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0fd7e1520e178ad8c353679dbaf3e11c555eb934c84a38344a6bdb2626d7b9f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 699, + 509, + 782 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1470954d2e27c7807a52607f89565c582b53f982c228ca7519f20ad619f11c14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 699, + 674, + 782 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6b303c60dbe6837e90ea6ab9c65cca735e7f05470de43c2148e6f8577df7ea2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 699, + 841, + 782 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c360e5184c406e4b73793fee739843b2ca95c17dfd4869ec08aa43ec27fb51e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 784, + 509, + 861 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/380357fd83a567b48d3def6b532c979c48039e1ae1aed50dce8cb86af7e32276.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 784, + 674, + 861 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94faaa01fa51e26324087c0d894d46685166bde06bf62b964f0b44bd222a57d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 784, + 841, + 861 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16694", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 90, + 240, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Different Contextualization Methods. Table 3 demonstrates the impact of various contextual methods on detection performance. We selected three representative contexts: local shift window-based context (SWin [32]), global pooling-based context (PVTV2-linear [55]), and global cluster-based context (PaCa [12]). The results show that the local shift window-based context offers minimal improvements in detection performance, which can be attributed to similar local contexts already provided by convolutional backbone networks. Conversely, enhancing the original feature with global contexts, as observed in PVTV2-linear and PaCa experiments, leads to better performance. Notably, orientation and velocity show significant improvements with these global contexts. The inclusion of cluster contexts further enhances improvements in location, orientation, and velocity prediction, as evidenced in the global cluster experiment.", + "bbox": [ + 76, + 112, + 467, + 369 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c592b04b2b45ceb3e12ae95a32f75d48e1ad851e9e19018751ca156914d2002a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
BEVFormer-b [25]41.667.327.437.239.419.851.7
Shift Window (SWin [32])41.5 (-0.1)67.127.837.839.620.151.5 (-0.2)
Global Pooling (PVTV2-linear [55])41.6 (+0.0)66.527.536.837.619.652.0 (+0.3)
Global Cluster (PaCa [12])42.6 (+1.0)66.427.635.036.220.052.8 (+1.1)
", + "bbox": [ + 83, + 376, + 460, + 422 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/093ff578e6c5cf801e58b2ab3513c30112b7b8f394f3d8f96e5a6a5236cc9d62.jpg", + "table_caption": [ + "Table 3. Ablation study on different contextualization methods." + ], + "table_footnote": [], + "table_body": "
ContextmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
Local (BEVFormer-b [25])41.667.327.437.239.419.851.7
Global Cluster41.666.427.438.335.119.552.1
Local + Global Cluster42.6 (+1.0)66.427.635.036.220.052.8 (+1.1)
", + "bbox": [ + 83, + 457, + 460, + 500 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5121107333bb6260a7d653d6a7b4f5096ef17106f5f17686a93d44c716ddfd0e.jpg", + "table_caption": [ + "Table 4. Ablation study on the relationship between local contexts and global cluster-based contexts." + ], + "table_footnote": [], + "table_body": "
Method#layers#clustersCross-level\\( \\mathbf{{mAP}} \\dagger \\)\\( \\mathbf{{mATE}} \\downarrow \\)\\( \\mathbf{{mASE}} \\downarrow \\)\\( \\mathbf{{mAOE}} \\downarrow \\)\\( \\mathbf{{mAVE}} \\downarrow \\)\\( \\mathbf{{mAAE}} \\downarrow \\)\\( \\mathbf{{NDS}} \\uparrow \\)
BEVFormer-b [25]---41.667.327.437.239.419.851.7
Baseline 13100✓41.165.627.137.736.418.952.0
Baseline 2650✓41.366.327.336.134.118.852.4
Baseline 36100-42.466.827.637.136.319.152.5
MvACon6100✓42.666.427.635.036.220.052.8
", + "bbox": [ + 83, + 547, + 460, + 597 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/23d1e5e6b6d9d9fa5f275e915d4f42bddc992cdde2d9e172014d46005a2fe010.jpg", + "table_caption": [ + "Table 5. Ablation study on the structure of our attentive contextualization module." + ], + "table_footnote": [], + "table_body": "
Clustering MethodmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
BEVFormer-b [25]41.667.327.437.239.419.851.7
Linear41.965.627.238.234.819.352.4
MLP42.466.427.338.235.019.252.6
Conv42.666.427.635.036.220.052.8
", + "bbox": [ + 83, + 645, + 460, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study on clustering operations in the attentive contextualization module.", + "bbox": [ + 76, + 708, + 467, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Relationship between Local Contexts and Global Cluster Contexts. Table 4 reveals the relationship between local and global cluster contexts in enhancing feature learning for view lifting. The exclusive use of global cluster context results in improved velocity prediction, while local attention contributes to better orientation prediction results. Combining these two contexts enhances predictions in location, orientation, and velocity. This highlights the complementary role that global cluster contexts play in feature encoding for view lifting.", + "bbox": [ + 76, + 750, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Structure of the Attentive Contextualization Module. Table 5 shows the impact of the structure of our attentive contextualization method. Baseline 1 demonstrates that attentive contextualization can efficiently encode feature information. With only three layers, MvACon achieves an improvement of 0.3 NDS compared to the vanilla BEVFormer. Baseline 2 indicates that attentive contextualization requires a sufficient number of clusters to extract abundant cluster contexts in the scene. Baseline 3 suggests that attending to clusters across the feature map aids in improving orientation, velocity, and mAP prediction. Table 6 illustrates that the convolution operation for clustering yields better results in terms of orientation, mAP, and NDS prediction. Conversely, point-based operations (such as a linear layer or multi-layer perceptron) demonstrate superior performance in location and velocity prediction.", + "bbox": [ + 496, + 90, + 890, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 345, + 617, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents Multi-View Attentive Contextualization (MvACon) for improving query-based multi-view 3D (MV3D) object detection. It addresses the limitations of two main paradigms of query based MV3D object detector in a unified way: decoder-only dense attention methods like PETR, which lack high-resolution features due to computational constraints, and encoder-decoder sparse 2D/3D deformable attention methods such as BEVFormer and DFA3D. Our MvACon contextualizes the original feature maps extracted from the backbone network using a cluster-attention operation built on the recently proposed patch-to-cluster attention (PaCa). In experiments, we show that our MvACon effectively and consistently enhances query-based MV3D object detectors by encoding more useful contexts, thereby facilitating better 2D-to-3D feature lifting. Rigorously controlled experiments reveal that, for decoder-only detectors, the cluster attention contextualization significantly improves localization and velocity prediction. For encoder-decoder based detectors, it effectively reduces errors in location, orientation, and velocity.", + "bbox": [ + 496, + 371, + 890, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 686, + 658, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "X. Liu and T. Wu were partly supported by NSF IIS-1909644, ARO Grant W911NF1810295, ARO Grant W911NF2210010, NSF CMMI-2024688, NSF IUSE2013451, and a research gift fund from the Innopeak Technology, Inc. (an affiliate of OPPO). The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ARO, NSF, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon.", + "bbox": [ + 496, + 712, + 890, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16695", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5", + "[2] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In ICCV, pages 9287-9296, 2019. 2", + "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 5", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2, 3", + "[5] Dian Chen, Jie Li, Vitor Guizilini, Rares Andrei Ambrus, and Adrien Gaidon. Viewpoint equivariance for multiview 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9213-9222, 2023. 2", + "[6] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2781-2790, 2022. 2", + "[7] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. Monopair: Monocular 3d object detection using pairwise spatial relationships. In CVPR, pages 12093-12102, 2020. 2", + "[8] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3d object detection in autonomous driving. In CVPR, 2022. 2", + "[9] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3d object detection. In CVPR Workshops, pages 1000-1001, 2020. 2", + "[10] 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. 3", + "[11] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 2, 5", + "[12] Ryan Grainger, Thomas Paniagua, Xi Song, Naresh Cuntoor, Mun Wai Lee, and Tianfu Wu. Paca-vit: Learning patch-to-cluster attention in vision transformers. In Proceedings of" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18568-18578, 2023. 2, 3, 5, 8", + "[13] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems, 34:15908-15919, 2021. 3", + "[14] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 2", + "[15] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 2", + "[16] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In CVPR, 2022. 2", + "[17] Abhinav Kumar, Garrick Brazil, and Xiaoming Liu. Groomed-nms: Grouped mathematically differentiable nms for monocular 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8973-8983, 2021. 2", + "[18] Abhinav Kumar, Garrick Brazil, Enrique Corona, Armin Parchami, and Xiaoming Liu. Deviant: Depth equivariant network for monocular 3d object detection. In ECCV, 2022. 2", + "[19] Hongyang Li, Hao Zhang, Zhaoyang Zeng, Shilong Liu, Feng Li, Tianhe Ren, and Lei Zhang. Dfa3d: 3d deformable attention for 2d-to-3d feature lifting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6684–6693, 2023. 2, 5, 6", + "[20] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. Rtm3d: Real-time monocular 3d detection from object keypoints for autonomous driving. In European Conference on Computer Vision, pages 644-660. Springer, 2020. 2", + "[21] Yawei Li, Kai Zhang, Jiezhang Cao, Radu Timofte, and Luc Van Gool. Localvit: Bringing locality to vision transformers. arXiv preprint arXiv:2104.05707, 2021. 3", + "[22] Yingyan Li, Yuntao Chen, Jiawei He, and Zhaoxiang Zhang. Densely constrained depth estimator for monocular 3d object detection. ECCV, 2022. 2", + "[23] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 2", + "[24] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2", + "[25] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 4, 5, 6, 8" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16696", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 2", + "[27] Xianpeng Liu, Nan Xue, and Tianfu Wu. Learning auxiliary monocular contexts helps monocular 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1810-1818, 2022. 3", + "[28] Xianpeng Liu, Ce Zheng, Kelvin B Cheng, Nan Xue, GuoJun Qi, and Tianfu Wu. Monocular 3d object detection with bounding box denoising in 3d by perceiver. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6436-6446, 2023. 2", + "[29] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022. 2, 3, 4, 5, 6, 1", + "[30] Yingfei Liu, Junjie Yan, Fan Jia, Shuailin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 2, 5", + "[31] Yang Liu, Yao Zhang, Yixin Wang, Feng Hou, Jin Yuan, Jiang Tian, Yang Zhang, Zhongchao Shi, Jianping Fan, and Zhiqiang He. A survey of visual transformers. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3", + "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 8", + "[33] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15641-15650, 2021. 2", + "[34] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. arXiv:2107.13774, 2021. 2", + "[35] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In ICCV, pages 6851-6860, 2019. 2", + "[36] Xinzhu Ma, Shinan Liu, Zhiyi Xia, Hongwen Zhang, Xingyu Zeng, and Wanli Ouyang. Rethinking pseudo-lidar representation. In ECCV, pages 311-327. Springer, 2020. 2", + "[37] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3d object detection. In CVPR, pages 4721-4730, 2021. 2", + "[38] Yuexin Ma, Tai Wang, Xuyang Bai, Huitong Yang, Yuenan Hou, Yaming Wang, Yu Qiao, Ruigang Yang, Dinesh Manocha, and Xinge Zhu. Vision-centric bev perception: A survey. arXiv preprint arXiv:2208.02797, 2022. 1", + "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[40] Arsalan Mousavian, Dragomir Anguelov, John Flynn, and Jana Kosecka. 3d bounding box estimation using deep learning and geometry. In CVPR, pages 7074-7082, 2017. 2", + "[41] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021. 3", + "[42] Dennis Park, Jie Li, Dian Chen, Vitor Guizilini, and Adrien Gaidon. Depth is all you need for monocular 3d detection. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7024-7031, 2023. 3", + "[43] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 2", + "[44] Liang Peng, Xiaopei Wu, Zheng Yang, Haifeng Liu, and Deng Cai. Did-m3d: Decoupling instance depth for monocular 3d object detection. In ECCV, 2022. 2", + "[45] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2", + "[46] Changyong Shu, Fisher Yu, and Yifan Liu. 3d point positional encoding for multi-camera 3d object detection transformers. arXiv preprint arXiv:2211.14710, 2022. 2", + "[47] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In ICCV, pages 1991-1999, 2019. 2", + "[48] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16519-16529, 2021. 3", + "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 3", + "[50] 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. 2", + "[51] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12894–12904, 2021. 3", + "[52] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. arXiv preprint arXiv:2303.11926, 2023. 2, 5" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16697", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021. 3", + "[54] Tai Wang, Pang Jiangmiao, and Lin Dahua. Monocular 3d object detection with depth from motion. In ECCV, 2022. 2", + "[55] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 3, 8", + "[56] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In CVPR, pages 8445-8453, 2019. 2", + "[57] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2, 3", + "[58] Bichen Wu, Chenfeng Xu, Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Zhicheng Yan, Masayoshi Tomizuka, Joseph Gonzalez, Kurt Keutzer, and Peter Vajda. Visual transformers: Token-based image representation and processing for computer vision. arXiv preprint arXiv:2006.03677, 2020. 3", + "[59] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerflike representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 3", + "[60] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 3, 5", + "[61] Kun Yuan, Shaopeng Guo, Ziwei Liu, Aojun Zhou, Fengwei Yu, and Wei Wu. Incorporating convolution designs into visual transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 579-588, 2021. 3", + "[62] Li Yuan, Qibin Hou, Zihang Jiang, Jiashi Feng, and Shuicheng Yan. Volo: Vision outlooker for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 45(5):6575-6586, 2022. 3", + "[63] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2998-3008, 2021. 3", + "[64] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In CVPR, pages 3289-3298, 2021. 2", + "[65] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Dalong Du, Jie Zhou, and Jiwen Lu. Dimension embeddings for monocular 3d object detection. In CVPR, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Ce Zheng, Xianpeng Liu, Guo-Jun Qi, and Chen Chen. Potter: Pooling attention transformer for efficient human mesh recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1620, 2023. 3", + "[67] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3" + ], + "bbox": [ + 501, + 92, + 890, + 217 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16698", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/cb273d05-390f-4fb9-aa74-74b5a46402e1_model.json b/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/cb273d05-390f-4fb9-aa74-74b5a46402e1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d74952ae2e1772c1af32c802fc37037b14d87fad --- /dev/null +++ b/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/cb273d05-390f-4fb9-aa74-74b5a46402e1_model.json @@ -0,0 +1,2510 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.131, + 0.868, + 0.152 + ], + "angle": 0, + "content": "Multi-View Attentive Contextualization for Multi-View 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.18, + 0.907, + 0.253 + ], + "angle": 0, + "content": "Xianpeng Liu\\(^{1}\\), Ce Zheng\\(^{2}\\), Ming Qian\\(^{3}\\), Nan Xue\\(^{3}\\), Chen Chen\\(^{2}\\), Zhebin Zhang\\(^{4}\\), Chen Li\\(^{4}\\), Tianfu Wu\\(^{1}\\) \n\\(^{1}\\)North Carolina State University \\(^{2}\\)University of Central Florida \n\\(^{3}\\)Ant Group \\(^{4}\\)OPPO U.S. Research Center \nhttps://xianpeng919.github.io/mvacon" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.313, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.474, + 0.712 + ], + "angle": 0, + "content": "We present Multi-View Attentive Contextualization (MvACon), a simple yet effective method for improving 2D-to-3D feature lifting in query-based multi-view 3D (MV3D) object detection. Despite remarkable progress witnessed in the field of query-based MV3D object detection, prior art often suffers from either the lack of exploiting high-resolution 2D features in dense attention-based lifting, due to high computational costs, or from insufficiently dense grounding of 3D queries to multi-scale 2D features in sparse attention-based lifting. Our proposed MvACon hits the two birds with one stone using a representationally dense yet computationally sparse attentive feature contextualization scheme that is agnostic to specific 2D-to-3D feature lifting approaches. In experiments, the proposed MvA-Con is thoroughly tested on the nuScenes benchmark, using both the BEVFormer and its recent 3D deformable attention (DFA3D) variant, as well as the PETR, showing consistent detection performance improvement, especially in enhancing performance in location, orientation, and velocity prediction. It is also tested on the Waymo-mini benchmark using BEVFormer with similar improvement. We qualitatively and quantitatively show that global cluster-based contexts effectively encode dense scene-level contexts for MV3D object detection. The promising results of our proposed MvA-Con reinforces the adage in computer vision – “(contextualized) feature matters”." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.74, + 0.21, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Camera-based 3D object detection is a pivotal area of research in computer vision, particularly owing to its appealing applications in cost-effective autonomous systems, such as autonomous driving and robot autonomy. Recently, significant progress has been witnessed in the field of multiview 3D (MV3D) object detection, especially with the advent of end-to-end MV3D detection approaches [38]. One crucial component in the end-to-end MV3D detection system is the 2D-to-3D feature lifting module, which converts" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.287, + 0.695, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.374, + 0.669, + 0.386 + ], + "angle": 0, + "content": "(a) Projected BEV anchor" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.287, + 0.89, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.374, + 0.85, + 0.386 + ], + "angle": 0, + "content": "(b) BEVFormer [25]" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.388, + 0.695, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.476, + 0.652, + 0.487 + ], + "angle": 0, + "content": "(c) Learned clusters" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.388, + 0.89, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.748, + 0.475, + 0.842, + 0.487 + ], + "angle": 0, + "content": "(d) Our MvACon" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.49, + 0.895, + 0.768 + ], + "angle": 0, + "content": "Figure 1. The effects of our proposed MvACon in the 2D-to-3D feature lifting. Consider a point (red) on the car in (a), which is projected from a 3D BEV anchor point. In lifting 2D features to ground the 3D BEV anchor point, vanilla BEVFormer [25] utilizes a predefined number of deformable points, with offsets learned through a 6-layer cross-attention module relative to the projection point. (b) shows the deformed points after the final cross-attention layer, most of which have low attention weights, indicating the model's uncertainty or inability to consolidate contributions effectively for good lifting. Our MvACon tackles this issue with clustering-based attention, as visualized in (c). (d) shows the deformed points, where we observe not only high-confidence points on the car but also on the building. We further observe that the points on the building remain stable across encoding layers (see Fig. 4) and consecutive frames (see suppl.). With those high-confident deformed points in a spatiotemporally stable configuration, our MvACon may induce a local object-context aware coordinate system that helps the overall performance, especially the estimation of velocity and orientation, as we quantitatively observed in experiments. See text for details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.885 + ], + "angle": 0, + "content": "perspective 2D multi-view image feature maps to 3D feature representation. It aims to counter the complete loss of depth information in individual 2D images by exploiting multi-view clues. However, a significant challenge in practical applications like autonomous driving is the often insufficient field-of-view overlap across views, making it difficult to effectively address the loss of depth information." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To perform 2D-to-3D feature lifting, recent methods" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16688" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.455 + ], + "angle": 0, + "content": "[15, 25, 29, 57] aim to learn a unified 3D space representation using 3D anchors that are either sparsely or uniformly sampled. These methods generally fall into two categories subject to the interaction of 3D anchors with 2D features and the feature aggregation strategy. (1) The Lift-Splat-Shoot (LSS) method [14, 15, 23, 24, 45] first lifts 2D features into 3D (pseudo-LiDAR) space using the outer product with the estimated depth, then assigns them to the nearest 3D anchors. (2) In contrast, the query-based design [25, 29, 57], pioneered by the DETR method [4] for end-to-end 2D object detection, adopts 3D anchors as queries and uses 2D image features as keys and values. They interact and aggregate via spatial cross-attention in the expressive Transformer architecture [50]. These two paradigms have been widely used in downstream tasks like map segmentation [30] and occupancy prediction [25]. This paper primarily focuses on the query-based detection paradigm. One reason is that LSS-based methods often encounter excessive computational complexity and issues with error propagation and depth estimation magnification post-lifting, potentially capping their performance. However, the query-based design also grapples with heavy computation costs or limited 3D information awareness, depending on their Transformer design." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.459, + 0.473, + 0.791 + ], + "angle": 0, + "content": "In this paper, we focus on addressing limitations of two main paradigms of query based MV3D object detectors in a unified way (elaborated in Sec. 3). In particular, we introduce multi-view attentive contextualization (MvA-Con) to address limitations of decoder-only dense attention methods like PETR [29], which lack high-resolution features due to computational constraints, and to simultaneously address the issue of sparsely grounded 3D anchors in encoder-decoder 2D/3D deformable attention methods such as BEVFormer [25] and DFA3D [19]. Our proposed MvA-Con aims to be representationally dense while computationally sparse. To achieve this, we expand the conventional three-component paradigm of MV3D object detection to a four-component setup (Fig. 2): (1) 2D image representation learning through a feature backbone shared across views, (2) MvACon for attentive contextualization of the 2D features, (3) 2D-to-3D feature lifting, and (4) a 3D object detection head or decoder that utilizes these lifted features. This modular design allows our MvACon to remain agnostic to specific 2D-to-3D feature lifting strategies and aligns with the classic adage in representation learning and computer vision: -('contextualized') feature matters'." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.901 + ], + "angle": 0, + "content": "More specifically, our approach contextualizes original feature maps extracted from the backbone network using a cluster-attention operation. This builds upon the recently proposed patch-to-cluster attention (PaCa) [12]. For perspective-based decoder-only detectors like PETR, we apply cluster contextualization before the feature maps are fed into the decoder. For encoder-decoder based detec" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.273 + ], + "angle": 0, + "content": "tors, such as BEVFormer and DFA3D, we incorporate cluster contextualization within the spatial cross-attention operation. Through extensive experiments, we demonstrate that our proposed MvACon effectively and consistently enhances query-based MV3D object detectors by encoding more useful contexts, thereby facilitating better 2D-to-3D feature lifting. Rigorously controlled experiments reveal that, for perspective-based decoder-only detectors, the cluster attention contextualization significantly improves localization and velocity prediction. In the case of encoder-decoder based detectors, it effectively reduces errors in location, orientation, and velocity." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.274, + 0.788, + 0.287 + ], + "angle": 0, + "content": "In summary, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.289, + 0.891, + 0.332 + ], + "angle": 0, + "content": "- We analyze and address the limitation of 2D-to-3D feature lifting in the prior art, that is the lack of sufficient 3D representational power due to their local 3D awareness." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.334, + 0.892, + 0.407 + ], + "angle": 0, + "content": "- We propose MvACon (Multi-view Attentive Contextualization) to induce the global 3D awareness in an easy-to-integrate way to enhance the 2D-to-3D feature lifting in both decoder-only based MV3D object detectors and encoder-decoder based ones." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.409, + 0.892, + 0.469 + ], + "angle": 0, + "content": "- We show consistent performance improvement of our MvACon on the challenging NuScenes [3] dataset using three baseline query-based MV3D object detectors, as well as on the Waymo-mini [11] benchmark." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.289, + 0.892, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.483, + 0.642, + 0.499 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Camera-based 3D Object Detection. Camera-based 3D object detection can be primarily categorized into two settings: single-view and multi-view. In the realm of monocular 3D object detection research, addressing the challenge of inaccurate object localization [37] is critical. Researchers have exerted considerable effort to utilize monocular depth cues. This includes transforming inputs into pseudo-lidar point clouds [35, 36, 56] and explicitly incorporating depth into models [8, 9, 16]. Another significant research direction involves the explicit use of geometric priors, encompassing approaches like key-point constraints [6, 20, 33], shape projection relationships [22, 34, 40, 64], and temporal depth estimation [54]. Innovations in monocular 3D object detection also include novel loss modules [7, 47, 65], 3D-aware backbones [2, 17, 18], and second-stage detection paradigms [28, 44]. In multi-view settings, the configuration often closely resembles that of monocular setups due to the limited field-of-view overlap between the different camera views. Therefore, MV3D object detection focuses on addressing the challenge of learning universal representation for multi-view sensors. It has benefited from advancements in various techniques such as view lifting [5, 19, 25, 29, 45, 57], depth encoding [19, 23, 24, 46], and temporal modeling [14, 25, 26, 30, 43, 52]. In our multi-view approach, we aim at addressing the challenge of multi-view representation learning with focus on en" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16689" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "hancing the view lifting module within query-based detection methods by empowering original features with cluster-based contextual features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.141, + 0.473, + 0.489 + ], + "angle": 0, + "content": "Representation Learning in Camera-based 3D Object Detection. Camera-based 3D object detection is inherently a data-intensive task due to its ill-posed nature, the expansive search space in 3D, and the scarcity of labeled data in scenes. Consequently, developing robust representations for this task is both critical and challenging. Early research in monocular 3D object detection has demonstrated the utility of depth contexts [41, 42] and projection contexts [27] in enhancing detection capabilities. Recent advances also highlight the effectiveness of scene-level representations, such as density fields [39], in improving 3D representation learning [59]. In the domain of multi-view research, most leading methods utilize backbones pre-trained with projection contexts (e.g., FCOS3D [53] weights) or depth contexts (e.g., DD3D [41] weights). However, these pre-trained weights may not fully leverage the capabilities of newer backbone network designs. Recent studies have begun to explore alternatives to this pre-trained paradigm, including the integration of an auxiliary projection context branch in end-to-end training [60]. Our work aims to enhance network representation by explicitly incorporating scene-level cluster context as supplementary information during the view lifting stage in query-based MV3D object detectors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.473, + 0.794 + ], + "angle": 0, + "content": "Vision Transformers. Since the pioneer work of ViT [10], extensive research [31] has been dedicated to enhancing the representational abilities of neural networks for visual tasks. It has been established that CNNs and Transformers can mutually augment each other's capabilities, as evidenced in designs like Transformer-enhanced CNNs [48, 58] and CNN-enhanced Transformers [21, 49, 61]. Additionally, a significant branch of visual Transformer research focuses on developing new attention mechanisms tailored to the locality bias in vision tasks. Notable examples include HaloNet [51], SWin [32], Deformable Attention [67], and VOLO [62], all introducing innovative local attention mechanisms to mitigate the quadratic computational cost associated with visual inputs. Concurrently, models like TNT [13], ViL [63], PVTv2-linear [55], POTTER [66], and PaCa [12] explore the integration of local and global contexts. Inspired by the progress of Vision Transformers, our work addresses the limitations of two prevalent paradigms in query-based MV3D detectors caused by their attention mechanisms." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.814, + 0.186, + 0.831 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.839, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Given a set of images \\( I_{i} \\in \\mathbb{R}^{3 \\times H \\times W} \\) from \\( N \\) cameras with known extrinsics \\( T_{i} \\in \\mathrm{SE}(3) \\) and intrinsics \\( K_{i} \\in \\mathbb{R}^{3 \\times 3} \\), MV3D object detection aims to infer the label (e.g., Car, Pedestrian, Barrier) and the 3D bounding box for each ob" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "ject instance in the scene. In this section, we first delve into the pipeline of query-based MV3D object detection in Sec. 3.1. We then analyze the pros and cons of the core 2D-to-3D feature lifting component in two state-of-the-art MV3D object detection methods in Sec. 3.2. Finally, we present our proposed MvACon in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.194, + 0.826, + 0.211 + ], + "angle": 0, + "content": "3.1. Query-based MV3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.892, + 0.4 + ], + "angle": 0, + "content": "For better understanding, we explain the query-based MV3D object detection pipeline, shown in Fig. 2, in a reverse manner. The 3D detection head typically builds upon DETR3D [57], which is based on the original DETR [4]. Initially, it defines a sufficiently large number, \\( O \\), of latent \\( C \\)-dimensional 3D object queries, \\( Q_{O,C} \\). These object queries update using Keys and Values derived from multi-view inputs, followed by a classification head predicting object labels and a bounding box regression head determining the 3D bounding boxes. Obviously, the key challenge lies in how to transform / lift multi-view 2D inputs into 3D-aware Keys and Values." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.893, + 0.581 + ], + "angle": 0, + "content": "To this end, a feature backbone is trained to extract deep 2D features from the multi-view input images. The complexity arises from different design choices for feature lifting. It mainly involves two aspects: representation and computation. From the representational perspective, multiscale feature pyramids, crucial in 2D object detection, become even more essential in 3D object detection. Computationally, handling multi-view inputs is already demanding. Adding multi-scale feature pyramids without careful optimization can significantly increase the computational load. There are two strategies in the state-of-the-art development of query-based MV3D object detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.893, + 0.825 + ], + "angle": 0, + "content": "Decoder-Only Architectures: Single-Scale Multi-View 2D Features with Dense Attention. These designs are among the most straightforward. They involves using multi-view 2D feature maps from the last layer of the feature backbone, which are then concatenated and flattened along the spatial dimensions to form Keys and Values. The latent object queries, \\( Q_{O,C} \\), are updated using a vanilla Transformer (i.e., each object query attends to every 2D location in the multi-view inputs, termed dense attention). However, this basic approach often fails as it does not encode any 3D-aware information. To address this, The PETR [29] introduces a physically-meaningful 3D position transformation as positional encoding, added to the multi-view 2D feature maps before concatenation and flattening. It has proven effective for query-based MV3D object detection, as illustrated in Option 1 in Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Encoder-Decoder Architectures: Multi-Scale Multi-View 2D Features to Latent BEV Queries with Sparse Attention. The BEV (Bird's Eye View) representation acts as a unified, grid-based and ego-centric scene representation with predefined grid sizes (e.g., \\(200 \\times 200\\)) on the XZ" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16690" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.087, + 0.882, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.27, + 0.894, + 0.312 + ], + "angle": 0, + "content": "Figure 2. Overview of a query-based MV3D object detection pipeline with our proposed MvACon. Our proposed MvACon is a plug-and-play module for two state-of-the-art query-based MV3D object detection paradigms (e.g., PETR [29] and BEVFormer [25] respectively), which computes attentively contextualized features to facilitate better 2D-to-3D feature lifting in the two paradigms. See text for details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.331, + 0.473, + 0.709 + ], + "angle": 0, + "content": "plane. Geometrically, the BEV grid can be treated as a pillar-based point-cloud representation, collapsed along the Y-axis. A predefined number of points along the pillar (Y-axis) direction in the BEV grid will be uniformly sampled. These points form a uniform geometry prior for the underlying 3D scene and serve as BEV anchors to elevate 2D features into the BEV space. With known camera poses, these sampled points can be projected to each view at multiple scales. However, due to the uniform geometry prior in the projection, the projected points require deformability to better align with data observations. The BEVFormer [25] addresses this by introducing a sparse deformable attention mechanism (see Option 2 in Fig. 2). To counter the uniform geometry prior further, it learns a small predefined number of offsets, rather than directly deforming the projected points on each view, to lift 2D features from those deformed points with attentive weights. Latent BEV queries are introduced in learning these offsets and attentive weights. The BEV encoder's role is to refine the BEV queries, enabling them to provide meaningful offsets and attentive weights for lifting 2D features to 3D BEV anchors. Finally, embedded BEV queries as Keys/Values in the 3D detection head (i.e., decoder) update the latent object queries, \\( Q_{O,C} \\), e.g., through sparse deformable attention as in the BEVFormer, before predicting the 3D object detection results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.723, + 0.47, + 0.755 + ], + "angle": 0, + "content": "3.2. The Limitation of 2D-to-3D Feature Lifting in the Prior Art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Although they both have shown remarkable progress for MV3D object detection, the PETR pipeline and the BEV-Former pipeline have a common limitation in their 2D-to-3D feature lifting, that is the local or shallow 3D awareness, rather than the desirable counterpart, the global and semantic meaningful 3D awareness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In the PETR pipeline, consider the 2D feature map of the \\(n\\)-th view, \\(F_{h\\times w\\times c}^{n}\\), where \\((h,w)\\) are the spatial sizes, height and width respectively, and \\(c\\) the feature dimen" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.331, + 0.892, + 0.558 + ], + "angle": 0, + "content": "sion of the backbone. The 3D position transform converts the (shared) camera frustum discretized as a mesh grid of sizes \\((h,w,D)\\) to the 3D space based on the known camera poses, where \\(D\\) is the discretized depth levels. After the conversion, each 3D point is represented by a normalized 3D coordinate in the homogeneous form, i.e., \\((x,y,z,1)\\). So the 3D position transformation results in the positional encoding \\(P_{h\\times w\\times 4\\cdot D}^{n}\\). Both \\(F_{h\\times w\\times c}^{n}\\) and \\(P_{h\\times w\\times 4\\cdot D}^{n}\\) are projected into the space of the same dimensionality, \\(d\\) using a linear layer and a Multi-layer Perceptron (MLP) with the ReLU nonlinearity respectively, we have \\(F_{h\\times w\\times d}^{n}\\) and \\(P_{h\\times w\\times d}^{n}\\) which are summed in an element-wise way. Consider the latent representation of the 3D position in \\(P_{h\\times w\\times d}^{n}\\) with the depth grid fused by the MLP, it is grounded to one feature point in \\(F_{h\\times w\\times d}^{n}\\), leading to the local 3D awareness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.634 + ], + "angle": 0, + "content": "In the BEVFormer pipeline, as we show in Fig. 1, the projected BEV anchor is often grounded to some low-confidence and scattered deformed points. Although the grounding may not be spatially local, they are often semantically shallow." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.783, + 0.663 + ], + "angle": 0, + "content": "3.3. Our Proposed MvACon Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.761 + ], + "angle": 0, + "content": "Our goal is to address the local or shallow 3D awareness stated above by introducing an easy-to-integrate (mostly plug-and-play) module to learn the global and semantically meaningful 3D-awareness, as illustrated in Fig. 2. The basic idea of our MvACon is to attentively contextualize the 2D features in the 2D-to-3D lifting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.892, + 0.793 + ], + "angle": 0, + "content": "In the PETR pipeline, our idea is to contextualize the individual 2D feature map, \\( F_{h \\times w \\times d}^{n} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.794, + 0.891, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {h \\times w \\times d} ^ {n} = \\operatorname {M v A C o n} \\left(F _ {h \\times w \\times d} ^ {n}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.871 + ], + "angle": 0, + "content": "where after the contextualization every feature point in \\(\\mathbf{F}_{h\\times w\\times d}^n\\) can connect to the entire map \\(F_{h\\times w\\times d}^n\\), inducing the global 3D awareness for grounding the positional encoding \\(P_{h\\times w\\times d}^n\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the BEVFormer pipeline, our idea is to contextualize the multi-scale feature maps, e.g., the \\(l\\)-th layer of the fea" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16691" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.341, + 0.109 + ], + "angle": 0, + "content": "ture pyramid of the \\(n\\) -th view, \\(F_{h\\times w\\times c}^{n,l}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.109, + 0.468, + 0.128 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {h \\times w \\times c} ^ {n, l} = \\operatorname {M v A C o n} \\left(\\left\\{F _ {h \\times w \\times c} ^ {n, l} \\right\\} _ {l = 1} ^ {L}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.129, + 0.468, + 0.191 + ], + "angle": 0, + "content": "where after the contextualization every feature point in \\(\\mathbf{F}_{h\\times w\\times c}^{n,l}\\) can connect to the entire \\(L\\)-layer feature pyramid \\(\\{F_{h\\times w\\times c}^{n,l}\\}_{l = 1}^{L}\\), inducing the global 3D awareness for grounding projected BEV anchors on the \\(n\\)-the view." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.192, + 0.468, + 0.297 + ], + "angle": 0, + "content": "To achieve the global contextualization effect, we adapt the recently proposed Patch-to-Cluster attention (PaCa) [12] method. The core idea of PaCa is to leverage a learnable clustering module to cluster a feature map into a predefined number \\(M\\) of clusters. For notional simplicity, consider a feature map \\(F_{h\\times w\\times c}\\) as \\(N = h\\times w\\) tokens \\(F_{N\\times c}\\), the clustering assignment is computed by," + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.299, + 0.468, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {N, M} = \\operatorname {S o f t m a x} (\\operatorname {C l u s t e r i n g} (F _ {N, c})), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.316, + 0.468, + 0.36 + ], + "angle": 0, + "content": "where Clustering() can be implemented in different ways (see our ablation studies in Tab. 6), and the Softmax is along the token dimension. Then, we compute \\(M\\) clusters by," + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.362, + 0.468, + 0.378 + ], + "angle": 0, + "content": "\\[\nz _ {M, c} = \\mathrm {L N} \\left(\\mathcal {C} _ {N, M} ^ {\\top} \\cdot F _ {N, c}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.378, + 0.357, + 0.393 + ], + "angle": 0, + "content": "where \\(\\mathrm{LN}()\\) is the layer normalization [1]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.394, + 0.397, + 0.409 + ], + "angle": 0, + "content": "Then, the PaCa-based MvAcon is defined by," + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.409, + 0.468, + 0.443 + ], + "angle": 0, + "content": "\\[\nF _ {N, c} ^ {\\prime} = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {N , c} \\cdot K _ {M , c} ^ {\\top}}{\\sqrt {c}}\\right) \\cdot V _ {M, c} + F _ {N, c}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.444, + 0.468, + 0.533 + ], + "angle": 0, + "content": "where \\( Q_{N,c} \\) is the linear projection of \\( F_{N,c}, K_{M,c} \\) and \\( V_{M,c} \\) are from the clusters \\( z_{M,c} \\). The second term is the shortcut. The multi-head PaCa can be straightforwardly defined. For Eqn. 2, we concatenate the clusters from all the pyramid layers before computing the Key and the Value. Here, the PaCa module is of linear complexity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.549, + 0.208, + 0.566 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.574, + 0.266, + 0.59 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.598, + 0.468, + 0.869 + ], + "angle": 0, + "content": "Dataset and Metrics We evaluate our MvAcon on the challenging large-scale NuScenes dataset [3] and Waymo dataset [11]. The NuScenes dataset includes 1,000 scene sequences, which are divided into training, validation, and testing subsets in a 700/150/150 split. Each sequence in the NuScenes dataset is a 20-second video clip, annotated at a rate of 2 frames per second (FPS). The NuScenes dataset employs a comprehensive suite of evaluation metrics for assessing detection performance. These metrics comprise mean Average Precision (mAP), mean Average Translation Error (mATE), mean Average Scale Error (mASE), mean Average Orientation Error (mAOE), mean Average Velocity Error (mAVE), mean Average Attribution Error (mAAE), and the NuScenes Detection Score (NDS). The Waymo dataset contains 798 training sequences and 202 validation sequences. We use a subset of the training set (Waymomini) by sampling every third frame from the training sequences following [25]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Implementation Details We leverage open-source code bases (PETR [29], BEVFormer [25], and DFA3D [19]) in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "our experiments. To ensure a fair and stringent comparison, we maintain all original configurations of these methods, making only one modification: the addition of an attentive contextualization module. We conduct qualitative analysis and ablation study on the BEVFormer-base model. We train all models for 24 epochs using 8 NVIDIA Tesla A100 GPUs, following the configurations and settings outlined in previous works [19, 25, 29]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.222, + 0.892, + 0.251 + ], + "angle": 0, + "content": "4.2. The Effectiveness of MvACon across Different Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.892, + 0.396 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our proposed MvA-Con method, we first apply it to two typical query-based MV3D object detection paradigms: the perspective-based decoder-only detector (PETR [29]) and the encoder-decoder based detector (BEVFormer [25]). We choose these as our baselines because state-of-the-art (SOTA) query-based MV3D detectors [19, 30, 52, 60] primarily follow these two paradigms. We also test our method on DFA3D [19] to demonstrate its generalizability to SOTA methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.607 + ], + "angle": 0, + "content": "On the NuScenes dataset, Table 1 shows that our proposed MvACon consistently improves performance across different detectors. Specifically, for the perspective-based decoder-only detector PETR, it improves the baseline by 0.8 NDS. For the encoder-decoder based detector BEVFormer, our method achieves an improvement of 1.3 in NDS on average. On a more advanced, depth-context enhanced BEVFormer (DFA3D), our method further improves performance by up to 0.5 NDS. Notably, our MvACon achieves significant improvement in location (mAP, mATE), orientation (mAOE), and velocity prediction (mAVE) for encoder-decoder based detectors. It also markedly enhances performance in location (mAP, mATE) and velocity (mAVE) prediction for the perspective-based decoder-only detector." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.683 + ], + "angle": 0, + "content": "On the Waymo dataset, since there are few released codes for MV3D detectors on Waymo except for the BEVFormer, we only test BEVFormer on Waymo-mini following its settings with results shown in Table 2. Our MvACon shows consistent improvement on Waymo metrics." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.693, + 0.78, + 0.707 + ], + "angle": 0, + "content": "4.3. How Does Our MvACon Work?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.716, + 0.892, + 0.882 + ], + "angle": 0, + "content": "We elaborate on the effects of our MvACon by providing detailed analyses during the 2D-to-3D feature lifting process. We first demonstrate what the learned cluster contexts encode, then show how these contexts affect the behavior of deformable points during feature lifting. Lastly, we illustrate how our MvACon improves detection results by presenting a qualitative comparison within a scene. We select BEVFormer-base as our analysis target due to its incorporation of six layers of deformable attention modules in the encoder. More qualitative analysis is provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.9 + ], + "angle": 0, + "content": "What do the Learned Cluster Contexts Encode? We vi-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16692" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.182, + 0.083, + 0.788, + 0.276 + ], + "angle": 0, + "content": "
MethodmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
PETR-VovNet-99 [29]37.874.627.248.890.621.242.6
PETR-VovNet-99-MvACon38.2 (+0.5)73.927.050.584.121.343.4 (+0.8)
BEVFormer-t [25]25.290.029.465.565.721.635.4
BEVFormer-t-MvACon25.9 (+0.7)88.428.864.660.522.536.5 (+1.1)
BEVFormer-s [25]37.072.128.040.743.622.047.9
BEVFormer-s-MvACon39.3 (+2.3)71.327.740.142.019.749.6 (+1.7)
BEVFormer-b [25]41.667.327.437.239.419.851.7
BEVFormer-b-MvACon42.6 (+1.0)66.427.635.036.220.052.8 (+1.1)
DFA3D-s [19]40.172.127.941.139.119.650.1
DFA3D-s-MvACon40.171.027.438.337.220.850.6 (+0.5)
DFA3D-b [19]43.065.427.137.434.120.553.1
DFA3D-b-MvACon43.2 (+0.2)66.427.534.432.320.753.5 (+0.4)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.281, + 0.895, + 0.325 + ], + "angle": 0, + "content": "Table 1. Comparisons of our method with baslines on the NuScenes validation set. BEVFormer-t/s/b refers to BEVFormer-tiny/small/base in the BEVFormer's open-source codes. 'DFA3D' refers to the adaptation of 2D deformable attention into a (depth-weighted) 3D deformable attention within the BEVFormer model, as adopted in [19]." + }, + { + "type": "image", + "bbox": [ + 0.145, + 0.329, + 0.314, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.145, + 0.406, + 0.314, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.329, + 0.482, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.406, + 0.482, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.329, + 0.651, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.406, + 0.651, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.329, + 0.82, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.406, + 0.82, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.484, + 0.893, + 0.527 + ], + "angle": 0, + "content": "Figure 3. Visualization results of learned cluster contexts in our MvACon on the NuScenes validation set. We sum all the learned clusters along the channel and upsample it to the original image resolution through bilinear interpolation. We observed that the learned cluster context encodes abundant context information in the scene. We provide details with raw images in the supplementary." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.537, + 0.447, + 0.588 + ], + "angle": 0, + "content": "
MethodLET-mAPL↑LET-mAPH↑
BEVFormer-ResNet101 [25]34.946.3
BEVFormer-ResNet101-MvACon35.7 (+0.8)47.5 (+1.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.144, + 0.592, + 0.402, + 0.607 + ], + "angle": 0, + "content": "Table 2. Comparisons on the Waymo-mini." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.47, + 0.82 + ], + "angle": 0, + "content": "sualize the learned cluster context in a heatmap, by summing all clusters along the channel and then upsampling to the original image resolution using bilinear interpolation. The results, shown in Fig. 3, reveal that despite the low resolution of the original summed heatmap, it can accurately locate foreground objects in the scene after upsampling. This suggests that the cluster contexts encode the foreground layouts in a scene. Furthermore, we note that the upsampled heatmap shows a high response to various foreground objects, even when they are spatially close in 2D. This is attributed to the effectiveness of our MvACon in dense scenes, as demonstrated and analyzed later in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "How do Cluster Contexts Affect the Behavior of Deformable Points? We demonstrate the dynamics of deformable points in Fig. 4. These points originate from one 2D reference point, which is projected from a 3D BEV anchor point in the BEVFormer encoder. As the encoding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.766 + ], + "angle": 0, + "content": "layers deepen in the vanilla BEVFormer, we observe that most deformable points have low attention weights, suggesting the model's uncertainty about the relevance of the selected context. Consequently, these contexts contribute minimally during the encoding of BEV query features in the feature lifting process. In contrast, deformable points predicted by our method maintain high confidence weights on foreground objects (e.g., cars), as well as on surrounding buildings. We also note that points on buildings remain stable across encoding layers and consecutive frames (see the supplementary). These high-confidence deformed points in our MvACon may foster a local object-context aware coordinate system, enhancing overall performance, including the estimation of velocity and orientation. This observation aligns with our quantitative findings in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Qualitative Comparisons with the Baseline Method. We provide qualitative comparison with BEVFormer in Fig. 5. Our method exhibits superior performance in dense scenarios where objects are crowded and challenging to localize. We attribute this enhanced performance to the rich foreground layout context in the scene, facilitated by our attentive contextualization module. Additional qualitative comparisons are available in the supplementary." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16693" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.096, + 0.383, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.096, + 0.599, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.097, + 0.811, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.192, + 0.383, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.192, + 0.599, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.192, + 0.811, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.289, + 0.383, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.291, + 0.599, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.292, + 0.811, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.388, + 0.383, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.388, + 0.599, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.388, + 0.811, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.489, + 0.892, + 0.53 + ], + "angle": 0, + "content": "Figure 4. Visualization results of the deformable points originating from a 2D reference point, which is projected from a 3D BEV anchor point in the BEVFormer encoder, on NuScenes validation set. We utilize the same BEV anchor point as demonstrated in Fig. 1. From left to right and up to bottom, we display the deformable points output from each layer (#1-#6) in the encoder, respectively." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.536, + 0.34, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.536, + 0.511, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.538, + 0.676, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.538, + 0.842, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.617, + 0.511, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.617, + 0.676, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.617, + 0.842, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.7, + 0.339, + 0.862 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.7, + 0.511, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.7, + 0.676, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.7, + 0.842, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.785, + 0.511, + 0.862 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.785, + 0.676, + 0.862 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.785, + 0.842, + 0.862 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.868, + 0.812, + 0.882 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons between BEVFormer and our MvACon method on NuScenes validation set." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16694" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.241, + 0.106 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.113, + 0.468, + 0.371 + ], + "angle": 0, + "content": "Effectiveness of Different Contextualization Methods. Table 3 demonstrates the impact of various contextual methods on detection performance. We selected three representative contexts: local shift window-based context (SWin [32]), global pooling-based context (PVTV2-linear [55]), and global cluster-based context (PaCa [12]). The results show that the local shift window-based context offers minimal improvements in detection performance, which can be attributed to similar local contexts already provided by convolutional backbone networks. Conversely, enhancing the original feature with global contexts, as observed in PVTV2-linear and PaCa experiments, leads to better performance. Notably, orientation and velocity show significant improvements with these global contexts. The inclusion of cluster contexts further enhances improvements in location, orientation, and velocity prediction, as evidenced in the global cluster experiment." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.377, + 0.462, + 0.424 + ], + "angle": 0, + "content": "
MethodmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
BEVFormer-b [25]41.667.327.437.239.419.851.7
Shift Window (SWin [32])41.5 (-0.1)67.127.837.839.620.151.5 (-0.2)
Global Pooling (PVTV2-linear [55])41.6 (+0.0)66.527.536.837.619.652.0 (+0.3)
Global Cluster (PaCa [12])42.6 (+1.0)66.427.635.036.220.052.8 (+1.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.085, + 0.428, + 0.46, + 0.442 + ], + "angle": 0, + "content": "Table 3. Ablation study on different contextualization methods." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.458, + 0.462, + 0.5 + ], + "angle": 0, + "content": "
ContextmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
Local (BEVFormer-b [25])41.667.327.437.239.419.851.7
Global Cluster41.666.427.438.335.119.552.1
Local + Global Cluster42.6 (+1.0)66.427.635.036.220.052.8 (+1.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.505, + 0.468, + 0.533 + ], + "angle": 0, + "content": "Table 4. Ablation study on the relationship between local contexts and global cluster-based contexts." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.548, + 0.462, + 0.598 + ], + "angle": 0, + "content": "
Method#layers#clustersCross-level\\( \\mathbf{{mAP}} \\dagger \\)\\( \\mathbf{{mATE}} \\downarrow \\)\\( \\mathbf{{mASE}} \\downarrow \\)\\( \\mathbf{{mAOE}} \\downarrow \\)\\( \\mathbf{{mAVE}} \\downarrow \\)\\( \\mathbf{{mAAE}} \\downarrow \\)\\( \\mathbf{{NDS}} \\uparrow \\)
BEVFormer-b [25]---41.667.327.437.239.419.851.7
Baseline 13100✓41.165.627.137.736.418.952.0
Baseline 2650✓41.366.327.336.134.118.852.4
Baseline 36100-42.466.827.637.136.319.152.5
MvACon6100✓42.666.427.635.036.220.052.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.603, + 0.468, + 0.63 + ], + "angle": 0, + "content": "Table 5. Ablation study on the structure of our attentive contextualization module." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.646, + 0.462, + 0.704 + ], + "angle": 0, + "content": "
Clustering MethodmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
BEVFormer-b [25]41.667.327.437.239.419.851.7
Linear41.965.627.238.234.819.352.4
MLP42.466.427.338.235.019.252.6
Conv42.666.427.635.036.220.052.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.709, + 0.468, + 0.736 + ], + "angle": 0, + "content": "Table 6. Ablation study on clustering operations in the attentive contextualization module." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Relationship between Local Contexts and Global Cluster Contexts. Table 4 reveals the relationship between local and global cluster contexts in enhancing feature learning for view lifting. The exclusive use of global cluster context results in improved velocity prediction, while local attention contributes to better orientation prediction results. Combining these two contexts enhances predictions in location, orientation, and velocity. This highlights the complementary role that global cluster contexts play in feature encoding for view lifting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Structure of the Attentive Contextualization Module. Table 5 shows the impact of the structure of our attentive contextualization method. Baseline 1 demonstrates that attentive contextualization can efficiently encode feature information. With only three layers, MvACon achieves an improvement of 0.3 NDS compared to the vanilla BEVFormer. Baseline 2 indicates that attentive contextualization requires a sufficient number of clusters to extract abundant cluster contexts in the scene. Baseline 3 suggests that attending to clusters across the feature map aids in improving orientation, velocity, and mAP prediction. Table 6 illustrates that the convolution operation for clustering yields better results in terms of orientation, mAP, and NDS prediction. Conversely, point-based operations (such as a linear layer or multi-layer perceptron) demonstrate superior performance in location and velocity prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.347, + 0.618, + 0.362 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.675 + ], + "angle": 0, + "content": "This paper presents Multi-View Attentive Contextualization (MvACon) for improving query-based multi-view 3D (MV3D) object detection. It addresses the limitations of two main paradigms of query based MV3D object detector in a unified way: decoder-only dense attention methods like PETR, which lack high-resolution features due to computational constraints, and encoder-decoder sparse 2D/3D deformable attention methods such as BEVFormer and DFA3D. Our MvACon contextualizes the original feature maps extracted from the backbone network using a cluster-attention operation built on the recently proposed patch-to-cluster attention (PaCa). In experiments, we show that our MvACon effectively and consistently enhances query-based MV3D object detectors by encoding more useful contexts, thereby facilitating better 2D-to-3D feature lifting. Rigorously controlled experiments reveal that, for decoder-only detectors, the cluster attention contextualization significantly improves localization and velocity prediction. For encoder-decoder based detectors, it effectively reduces errors in location, orientation, and velocity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.688, + 0.659, + 0.705 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.893 + ], + "angle": 0, + "content": "X. Liu and T. Wu were partly supported by NSF IIS-1909644, ARO Grant W911NF1810295, ARO Grant W911NF2210010, NSF CMMI-2024688, NSF IUSE2013451, and a research gift fund from the Innopeak Technology, Inc. (an affiliate of OPPO). The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ARO, NSF, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16695" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.471, + 0.2 + ], + "angle": 0, + "content": "[2] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In ICCV, pages 9287-9296, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.472, + 0.285 + ], + "angle": 0, + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.47, + 0.355 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.47, + 0.429 + ], + "angle": 0, + "content": "[5] Dian Chen, Jie Li, Vitor Guizilini, Rares Andrei Ambrus, and Adrien Gaidon. Viewpoint equivariance for multiview 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9213-9222, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2781-2790, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.469, + 0.57 + ], + "angle": 0, + "content": "[7] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. Monopair: Monocular 3d object detection using pairwise spatial relationships. In CVPR, pages 12093-12102, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.469, + 0.614 + ], + "angle": 0, + "content": "[8] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3d object detection in autonomous driving. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.618, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[9] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3d object detection. In CVPR Workshops, pages 1000-1001, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[10] 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. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[11] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Ryan Grainger, Thomas Paniagua, Xi Song, Naresh Cuntoor, Mun Wai Lee, and Tianfu Wu. Paca-vit: Learning patch-to-cluster attention in vision transformers. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18568-18578, 2023. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.124, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[13] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems, 34:15908-15919, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.181, + 0.892, + 0.222 + ], + "angle": 0, + "content": "[14] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.225, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[15] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.283, + 0.892, + 0.325 + ], + "angle": 0, + "content": "[16] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.327, + 0.892, + 0.396 + ], + "angle": 0, + "content": "[17] Abhinav Kumar, Garrick Brazil, and Xiaoming Liu. Groomed-nms: Grouped mathematically differentiable nms for monocular 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8973-8983, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.399, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[18] Abhinav Kumar, Garrick Brazil, Enrique Corona, Armin Parchami, and Xiaoming Liu. Deviant: Depth equivariant network for monocular 3d object detection. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.457, + 0.892, + 0.526 + ], + "angle": 0, + "content": "[19] Hongyang Li, Hao Zhang, Zhaoyang Zeng, Shilong Liu, Feng Li, Tianhe Ren, and Lei Zhang. Dfa3d: 3d deformable attention for 2d-to-3d feature lifting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6684–6693, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.529, + 0.892, + 0.584 + ], + "angle": 0, + "content": "[20] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. Rtm3d: Real-time monocular 3d detection from object keypoints for autonomous driving. In European Conference on Computer Vision, pages 644-660. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.586, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[21] Yawei Li, Kai Zhang, Jiezhang Cao, Radu Timofte, and Luc Van Gool. Localvit: Bringing locality to vision transformers. arXiv preprint arXiv:2104.05707, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.631, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[22] Yingyan Li, Yuntao Chen, Jiawei He, and Zhaoxiang Zhang. Densely constrained depth estimator for monocular 3d object detection. ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[23] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[24] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 4, 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16696" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[26] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[27] Xianpeng Liu, Nan Xue, and Tianfu Wu. Learning auxiliary monocular contexts helps monocular 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1810-1818, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[28] Xianpeng Liu, Ce Zheng, Kelvin B Cheng, Nan Xue, GuoJun Qi, and Tianfu Wu. Monocular 3d object detection with bounding box denoising in 3d by perceiver. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6436-6446, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[29] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022. 2, 3, 4, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[30] Yingfei Liu, Junjie Yan, Fan Jia, Shuailin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[31] Yang Liu, Yao Zhang, Yixin Wang, Feng Hou, Jin Yuan, Jiang Tian, Yang Zhang, Zhongchao Shi, Jianping Fan, and Zhiqiang He. A survey of visual transformers. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.462, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.533, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[33] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15641-15650, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.603, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[34] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. arXiv:2107.13774, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.66, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[35] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In ICCV, pages 6851-6860, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.717, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[36] Xinzhu Ma, Shinan Liu, Zhiyi Xia, Hongwen Zhang, Xingyu Zeng, and Wanli Ouyang. Rethinking pseudo-lidar representation. In ECCV, pages 311-327. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.76, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[37] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3d object detection. In CVPR, pages 4721-4730, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[38] Yuexin Ma, Tai Wang, Xuyang Bai, Huitong Yang, Yuenan Hou, Yaming Wang, Yu Qiao, Ruigang Yang, Dinesh Manocha, and Xinge Zhu. Vision-centric bev perception: A survey. arXiv preprint arXiv:2208.02797, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[40] Arsalan Mousavian, Dragomir Anguelov, John Flynn, and Jana Kosecka. 3d bounding box estimation using deep learning and geometry. In CVPR, pages 7074-7082, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[41] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[42] Dennis Park, Jie Li, Dian Chen, Vitor Guizilini, and Adrien Gaidon. Depth is all you need for monocular 3d detection. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7024-7031, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[43] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[44] Liang Peng, Xiaopei Wu, Zheng Yang, Haifeng Liu, and Deng Cai. Did-m3d: Decoupling instance depth for monocular 3d object detection. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[45] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.471, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[46] Changyong Shu, Fisher Yu, and Yifan Liu. 3d point positional encoding for multi-camera 3d object detection transformers. arXiv preprint arXiv:2211.14710, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[47] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In ICCV, pages 1991-1999, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[48] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16519-16529, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[50] 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. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[51] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12894–12904, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[52] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. arXiv preprint arXiv:2303.11926, 2023. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16697" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[53] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.175 + ], + "angle": 0, + "content": "[54] Tai Wang, Pang Jiangmiao, and Lin Dahua. Monocular 3d object detection with depth from motion. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.178, + 0.468, + 0.232 + ], + "angle": 0, + "content": "[55] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.234, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[56] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In CVPR, pages 8445-8453, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.302, + 0.468, + 0.37 + ], + "angle": 0, + "content": "[57] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.373, + 0.468, + 0.453 + ], + "angle": 0, + "content": "[58] Bichen Wu, Chenfeng Xu, Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Zhicheng Yan, Masayoshi Tomizuka, Joseph Gonzalez, Kurt Keutzer, and Peter Vajda. Visual transformers: Token-based image representation and processing for computer vision. arXiv preprint arXiv:2006.03677, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.456, + 0.468, + 0.524 + ], + "angle": 0, + "content": "[59] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerflike representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.526, + 0.468, + 0.62 + ], + "angle": 0, + "content": "[60] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.624, + 0.468, + 0.689 + ], + "angle": 0, + "content": "[61] Kun Yuan, Shaopeng Guo, Ziwei Liu, Aojun Zhou, Fengwei Yu, and Wei Wu. Incorporating convolution designs into visual transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 579-588, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.693, + 0.468, + 0.747 + ], + "angle": 0, + "content": "[62] Li Yuan, Qibin Hou, Zihang Jiang, Jiashi Feng, and Shuicheng Yan. Volo: Vision outlooker for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 45(5):6575-6586, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.749, + 0.468, + 0.816 + ], + "angle": 0, + "content": "[63] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2998-3008, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.818, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[64] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In CVPR, pages 3289-3298, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[65] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Dalong Du, Jie Zhou, and Jiwen Lu. Dimension embeddings for monocular 3d object detection. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[66] Ce Zheng, Xianpeng Liu, Guo-Jun Qi, and Chen Chen. Potter: Pooling attention transformer for efficient human mesh recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1620, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[67] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16698" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/cb273d05-390f-4fb9-aa74-74b5a46402e1_origin.pdf b/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/cb273d05-390f-4fb9-aa74-74b5a46402e1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8335c8f50ffd88f2c1fc074e884e1ba09fd1d9c8 --- /dev/null +++ b/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/cb273d05-390f-4fb9-aa74-74b5a46402e1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04f10f91be924954339d61cba1bdd4085605297ce5618dd09a47a55fc2862086 +size 4251829 diff --git a/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/full.md b/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fb98b013c89482a721092dae45526fca3ce38223 --- /dev/null +++ b/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/full.md @@ -0,0 +1,349 @@ +# Multi-View Attentive Contextualization for Multi-View 3D Object Detection + +Xianpeng Liu $^{1}$ , Ce Zheng $^{2}$ , Ming Qian $^{3}$ , Nan Xue $^{3}$ , Chen Chen $^{2}$ , Zhebin Zhang $^{4}$ , Chen Li $^{4}$ , Tianfu Wu $^{1}$ $^{1}$ North Carolina State University $^{2}$ University of Central Florida + $^{3}$ Ant Group $^{4}$ OPPO U.S. Research Center +https://xianpeng919.github.io/mvacon + +# Abstract + +We present Multi-View Attentive Contextualization (MvACon), a simple yet effective method for improving 2D-to-3D feature lifting in query-based multi-view 3D (MV3D) object detection. Despite remarkable progress witnessed in the field of query-based MV3D object detection, prior art often suffers from either the lack of exploiting high-resolution 2D features in dense attention-based lifting, due to high computational costs, or from insufficiently dense grounding of 3D queries to multi-scale 2D features in sparse attention-based lifting. Our proposed MvACon hits the two birds with one stone using a representationally dense yet computationally sparse attentive feature contextualization scheme that is agnostic to specific 2D-to-3D feature lifting approaches. In experiments, the proposed MvA-Con is thoroughly tested on the nuScenes benchmark, using both the BEVFormer and its recent 3D deformable attention (DFA3D) variant, as well as the PETR, showing consistent detection performance improvement, especially in enhancing performance in location, orientation, and velocity prediction. It is also tested on the Waymo-mini benchmark using BEVFormer with similar improvement. We qualitatively and quantitatively show that global cluster-based contexts effectively encode dense scene-level contexts for MV3D object detection. The promising results of our proposed MvA-Con reinforces the adage in computer vision – “(contextualized) feature matters”. + +# 1. Introduction + +Camera-based 3D object detection is a pivotal area of research in computer vision, particularly owing to its appealing applications in cost-effective autonomous systems, such as autonomous driving and robot autonomy. Recently, significant progress has been witnessed in the field of multiview 3D (MV3D) object detection, especially with the advent of end-to-end MV3D detection approaches [38]. One crucial component in the end-to-end MV3D detection system is the 2D-to-3D feature lifting module, which converts + +![](images/4b3c92ee1a1a867ad581bc546d93dce8b7fdd55aefd35d6ca4915a4d2798e344.jpg) +(a) Projected BEV anchor + +![](images/72f9097ae567920622785e2aca5192658191fac136935847c787a75bb3068327.jpg) +(b) BEVFormer [25] + +![](images/f81bff84d620c7860b1cf7e2d1b47f57ac88b21ca09e94f489362b0d58ca13c1.jpg) +(c) Learned clusters + +![](images/cc7d655eb37f6fe469d90b6b5c093a6ce9f567b774320dfb3c5904332d09cf92.jpg) +(d) Our MvACon +Figure 1. The effects of our proposed MvACon in the 2D-to-3D feature lifting. Consider a point (red) on the car in (a), which is projected from a 3D BEV anchor point. In lifting 2D features to ground the 3D BEV anchor point, vanilla BEVFormer [25] utilizes a predefined number of deformable points, with offsets learned through a 6-layer cross-attention module relative to the projection point. (b) shows the deformed points after the final cross-attention layer, most of which have low attention weights, indicating the model's uncertainty or inability to consolidate contributions effectively for good lifting. Our MvACon tackles this issue with clustering-based attention, as visualized in (c). (d) shows the deformed points, where we observe not only high-confidence points on the car but also on the building. We further observe that the points on the building remain stable across encoding layers (see Fig. 4) and consecutive frames (see suppl.). With those high-confident deformed points in a spatiotemporally stable configuration, our MvACon may induce a local object-context aware coordinate system that helps the overall performance, especially the estimation of velocity and orientation, as we quantitatively observed in experiments. See text for details. + +perspective 2D multi-view image feature maps to 3D feature representation. It aims to counter the complete loss of depth information in individual 2D images by exploiting multi-view clues. However, a significant challenge in practical applications like autonomous driving is the often insufficient field-of-view overlap across views, making it difficult to effectively address the loss of depth information. + +To perform 2D-to-3D feature lifting, recent methods + +[15, 25, 29, 57] aim to learn a unified 3D space representation using 3D anchors that are either sparsely or uniformly sampled. These methods generally fall into two categories subject to the interaction of 3D anchors with 2D features and the feature aggregation strategy. (1) The Lift-Splat-Shoot (LSS) method [14, 15, 23, 24, 45] first lifts 2D features into 3D (pseudo-LiDAR) space using the outer product with the estimated depth, then assigns them to the nearest 3D anchors. (2) In contrast, the query-based design [25, 29, 57], pioneered by the DETR method [4] for end-to-end 2D object detection, adopts 3D anchors as queries and uses 2D image features as keys and values. They interact and aggregate via spatial cross-attention in the expressive Transformer architecture [50]. These two paradigms have been widely used in downstream tasks like map segmentation [30] and occupancy prediction [25]. This paper primarily focuses on the query-based detection paradigm. One reason is that LSS-based methods often encounter excessive computational complexity and issues with error propagation and depth estimation magnification post-lifting, potentially capping their performance. However, the query-based design also grapples with heavy computation costs or limited 3D information awareness, depending on their Transformer design. + +In this paper, we focus on addressing limitations of two main paradigms of query based MV3D object detectors in a unified way (elaborated in Sec. 3). In particular, we introduce multi-view attentive contextualization (MvA-Con) to address limitations of decoder-only dense attention methods like PETR [29], which lack high-resolution features due to computational constraints, and to simultaneously address the issue of sparsely grounded 3D anchors in encoder-decoder 2D/3D deformable attention methods such as BEVFormer [25] and DFA3D [19]. Our proposed MvA-Con aims to be representationally dense while computationally sparse. To achieve this, we expand the conventional three-component paradigm of MV3D object detection to a four-component setup (Fig. 2): (1) 2D image representation learning through a feature backbone shared across views, (2) MvACon for attentive contextualization of the 2D features, (3) 2D-to-3D feature lifting, and (4) a 3D object detection head or decoder that utilizes these lifted features. This modular design allows our MvACon to remain agnostic to specific 2D-to-3D feature lifting strategies and aligns with the classic adage in representation learning and computer vision: -('contextualized') feature matters'. + +More specifically, our approach contextualizes original feature maps extracted from the backbone network using a cluster-attention operation. This builds upon the recently proposed patch-to-cluster attention (PaCa) [12]. For perspective-based decoder-only detectors like PETR, we apply cluster contextualization before the feature maps are fed into the decoder. For encoder-decoder based detec + +tors, such as BEVFormer and DFA3D, we incorporate cluster contextualization within the spatial cross-attention operation. Through extensive experiments, we demonstrate that our proposed MvACon effectively and consistently enhances query-based MV3D object detectors by encoding more useful contexts, thereby facilitating better 2D-to-3D feature lifting. Rigorously controlled experiments reveal that, for perspective-based decoder-only detectors, the cluster attention contextualization significantly improves localization and velocity prediction. In the case of encoder-decoder based detectors, it effectively reduces errors in location, orientation, and velocity. + +In summary, our main contributions are: + +- We analyze and address the limitation of 2D-to-3D feature lifting in the prior art, that is the lack of sufficient 3D representational power due to their local 3D awareness. +- We propose MvACon (Multi-view Attentive Contextualization) to induce the global 3D awareness in an easy-to-integrate way to enhance the 2D-to-3D feature lifting in both decoder-only based MV3D object detectors and encoder-decoder based ones. +- We show consistent performance improvement of our MvACon on the challenging NuScenes [3] dataset using three baseline query-based MV3D object detectors, as well as on the Waymo-mini [11] benchmark. + +# 2. Related Work + +Camera-based 3D Object Detection. Camera-based 3D object detection can be primarily categorized into two settings: single-view and multi-view. In the realm of monocular 3D object detection research, addressing the challenge of inaccurate object localization [37] is critical. Researchers have exerted considerable effort to utilize monocular depth cues. This includes transforming inputs into pseudo-lidar point clouds [35, 36, 56] and explicitly incorporating depth into models [8, 9, 16]. Another significant research direction involves the explicit use of geometric priors, encompassing approaches like key-point constraints [6, 20, 33], shape projection relationships [22, 34, 40, 64], and temporal depth estimation [54]. Innovations in monocular 3D object detection also include novel loss modules [7, 47, 65], 3D-aware backbones [2, 17, 18], and second-stage detection paradigms [28, 44]. In multi-view settings, the configuration often closely resembles that of monocular setups due to the limited field-of-view overlap between the different camera views. Therefore, MV3D object detection focuses on addressing the challenge of learning universal representation for multi-view sensors. It has benefited from advancements in various techniques such as view lifting [5, 19, 25, 29, 45, 57], depth encoding [19, 23, 24, 46], and temporal modeling [14, 25, 26, 30, 43, 52]. In our multi-view approach, we aim at addressing the challenge of multi-view representation learning with focus on en + +hancing the view lifting module within query-based detection methods by empowering original features with cluster-based contextual features. + +Representation Learning in Camera-based 3D Object Detection. Camera-based 3D object detection is inherently a data-intensive task due to its ill-posed nature, the expansive search space in 3D, and the scarcity of labeled data in scenes. Consequently, developing robust representations for this task is both critical and challenging. Early research in monocular 3D object detection has demonstrated the utility of depth contexts [41, 42] and projection contexts [27] in enhancing detection capabilities. Recent advances also highlight the effectiveness of scene-level representations, such as density fields [39], in improving 3D representation learning [59]. In the domain of multi-view research, most leading methods utilize backbones pre-trained with projection contexts (e.g., FCOS3D [53] weights) or depth contexts (e.g., DD3D [41] weights). However, these pre-trained weights may not fully leverage the capabilities of newer backbone network designs. Recent studies have begun to explore alternatives to this pre-trained paradigm, including the integration of an auxiliary projection context branch in end-to-end training [60]. Our work aims to enhance network representation by explicitly incorporating scene-level cluster context as supplementary information during the view lifting stage in query-based MV3D object detectors. + +Vision Transformers. Since the pioneer work of ViT [10], extensive research [31] has been dedicated to enhancing the representational abilities of neural networks for visual tasks. It has been established that CNNs and Transformers can mutually augment each other's capabilities, as evidenced in designs like Transformer-enhanced CNNs [48, 58] and CNN-enhanced Transformers [21, 49, 61]. Additionally, a significant branch of visual Transformer research focuses on developing new attention mechanisms tailored to the locality bias in vision tasks. Notable examples include HaloNet [51], SWin [32], Deformable Attention [67], and VOLO [62], all introducing innovative local attention mechanisms to mitigate the quadratic computational cost associated with visual inputs. Concurrently, models like TNT [13], ViL [63], PVTv2-linear [55], POTTER [66], and PaCa [12] explore the integration of local and global contexts. Inspired by the progress of Vision Transformers, our work addresses the limitations of two prevalent paradigms in query-based MV3D detectors caused by their attention mechanisms. + +# 3. Approach + +Given a set of images $I_{i} \in \mathbb{R}^{3 \times H \times W}$ from $N$ cameras with known extrinsics $T_{i} \in \mathrm{SE}(3)$ and intrinsics $K_{i} \in \mathbb{R}^{3 \times 3}$ , MV3D object detection aims to infer the label (e.g., Car, Pedestrian, Barrier) and the 3D bounding box for each ob + +ject instance in the scene. In this section, we first delve into the pipeline of query-based MV3D object detection in Sec. 3.1. We then analyze the pros and cons of the core 2D-to-3D feature lifting component in two state-of-the-art MV3D object detection methods in Sec. 3.2. Finally, we present our proposed MvACon in Sec. 3.3. + +# 3.1. Query-based MV3D Object Detection + +For better understanding, we explain the query-based MV3D object detection pipeline, shown in Fig. 2, in a reverse manner. The 3D detection head typically builds upon DETR3D [57], which is based on the original DETR [4]. Initially, it defines a sufficiently large number, $O$ , of latent $C$ -dimensional 3D object queries, $Q_{O,C}$ . These object queries update using Keys and Values derived from multi-view inputs, followed by a classification head predicting object labels and a bounding box regression head determining the 3D bounding boxes. Obviously, the key challenge lies in how to transform / lift multi-view 2D inputs into 3D-aware Keys and Values. + +To this end, a feature backbone is trained to extract deep 2D features from the multi-view input images. The complexity arises from different design choices for feature lifting. It mainly involves two aspects: representation and computation. From the representational perspective, multiscale feature pyramids, crucial in 2D object detection, become even more essential in 3D object detection. Computationally, handling multi-view inputs is already demanding. Adding multi-scale feature pyramids without careful optimization can significantly increase the computational load. There are two strategies in the state-of-the-art development of query-based MV3D object detection. + +Decoder-Only Architectures: Single-Scale Multi-View 2D Features with Dense Attention. These designs are among the most straightforward. They involves using multi-view 2D feature maps from the last layer of the feature backbone, which are then concatenated and flattened along the spatial dimensions to form Keys and Values. The latent object queries, $Q_{O,C}$ , are updated using a vanilla Transformer (i.e., each object query attends to every 2D location in the multi-view inputs, termed dense attention). However, this basic approach often fails as it does not encode any 3D-aware information. To address this, The PETR [29] introduces a physically-meaningful 3D position transformation as positional encoding, added to the multi-view 2D feature maps before concatenation and flattening. It has proven effective for query-based MV3D object detection, as illustrated in Option 1 in Fig. 2). + +Encoder-Decoder Architectures: Multi-Scale Multi-View 2D Features to Latent BEV Queries with Sparse Attention. The BEV (Bird's Eye View) representation acts as a unified, grid-based and ego-centric scene representation with predefined grid sizes (e.g., $200 \times 200$ ) on the XZ + +![](images/f59ccbc492e5079cd54861e6c4fb9222a64fbfa244efb5d7d034a31258211189.jpg) +Figure 2. Overview of a query-based MV3D object detection pipeline with our proposed MvACon. Our proposed MvACon is a plug-and-play module for two state-of-the-art query-based MV3D object detection paradigms (e.g., PETR [29] and BEVFormer [25] respectively), which computes attentively contextualized features to facilitate better 2D-to-3D feature lifting in the two paradigms. See text for details. + +plane. Geometrically, the BEV grid can be treated as a pillar-based point-cloud representation, collapsed along the Y-axis. A predefined number of points along the pillar (Y-axis) direction in the BEV grid will be uniformly sampled. These points form a uniform geometry prior for the underlying 3D scene and serve as BEV anchors to elevate 2D features into the BEV space. With known camera poses, these sampled points can be projected to each view at multiple scales. However, due to the uniform geometry prior in the projection, the projected points require deformability to better align with data observations. The BEVFormer [25] addresses this by introducing a sparse deformable attention mechanism (see Option 2 in Fig. 2). To counter the uniform geometry prior further, it learns a small predefined number of offsets, rather than directly deforming the projected points on each view, to lift 2D features from those deformed points with attentive weights. Latent BEV queries are introduced in learning these offsets and attentive weights. The BEV encoder's role is to refine the BEV queries, enabling them to provide meaningful offsets and attentive weights for lifting 2D features to 3D BEV anchors. Finally, embedded BEV queries as Keys/Values in the 3D detection head (i.e., decoder) update the latent object queries, $Q_{O,C}$ , e.g., through sparse deformable attention as in the BEVFormer, before predicting the 3D object detection results. + +# 3.2. The Limitation of 2D-to-3D Feature Lifting in the Prior Art + +Although they both have shown remarkable progress for MV3D object detection, the PETR pipeline and the BEV-Former pipeline have a common limitation in their 2D-to-3D feature lifting, that is the local or shallow 3D awareness, rather than the desirable counterpart, the global and semantic meaningful 3D awareness. + +In the PETR pipeline, consider the 2D feature map of the $n$ -th view, $F_{h\times w\times c}^{n}$ , where $(h,w)$ are the spatial sizes, height and width respectively, and $c$ the feature dimen + +sion of the backbone. The 3D position transform converts the (shared) camera frustum discretized as a mesh grid of sizes $(h,w,D)$ to the 3D space based on the known camera poses, where $D$ is the discretized depth levels. After the conversion, each 3D point is represented by a normalized 3D coordinate in the homogeneous form, i.e., $(x,y,z,1)$ . So the 3D position transformation results in the positional encoding $P_{h\times w\times 4\cdot D}^{n}$ . Both $F_{h\times w\times c}^{n}$ and $P_{h\times w\times 4\cdot D}^{n}$ are projected into the space of the same dimensionality, $d$ using a linear layer and a Multi-layer Perceptron (MLP) with the ReLU nonlinearity respectively, we have $F_{h\times w\times d}^{n}$ and $P_{h\times w\times d}^{n}$ which are summed in an element-wise way. Consider the latent representation of the 3D position in $P_{h\times w\times d}^{n}$ with the depth grid fused by the MLP, it is grounded to one feature point in $F_{h\times w\times d}^{n}$ , leading to the local 3D awareness. + +In the BEVFormer pipeline, as we show in Fig. 1, the projected BEV anchor is often grounded to some low-confidence and scattered deformed points. Although the grounding may not be spatially local, they are often semantically shallow. + +# 3.3. Our Proposed MvACon Method + +Our goal is to address the local or shallow 3D awareness stated above by introducing an easy-to-integrate (mostly plug-and-play) module to learn the global and semantically meaningful 3D-awareness, as illustrated in Fig. 2. The basic idea of our MvACon is to attentively contextualize the 2D features in the 2D-to-3D lifting. + +In the PETR pipeline, our idea is to contextualize the individual 2D feature map, $F_{h \times w \times d}^{n}$ , + +$$ +\mathbf {F} _ {h \times w \times d} ^ {n} = \operatorname {M v A C o n} \left(F _ {h \times w \times d} ^ {n}\right), \tag {1} +$$ + +where after the contextualization every feature point in $\mathbf{F}_{h\times w\times d}^n$ can connect to the entire map $F_{h\times w\times d}^n$ , inducing the global 3D awareness for grounding the positional encoding $P_{h\times w\times d}^n$ . + +In the BEVFormer pipeline, our idea is to contextualize the multi-scale feature maps, e.g., the $l$ -th layer of the fea + +ture pyramid of the $n$ -th view, $F_{h\times w\times c}^{n,l}$ + +$$ +\mathbf {F} _ {h \times w \times c} ^ {n, l} = \operatorname {M v A C o n} \left(\left\{F _ {h \times w \times c} ^ {n, l} \right\} _ {l = 1} ^ {L}\right), \tag {2} +$$ + +where after the contextualization every feature point in $\mathbf{F}_{h\times w\times c}^{n,l}$ can connect to the entire $L$ -layer feature pyramid $\{F_{h\times w\times c}^{n,l}\}_{l = 1}^{L}$ , inducing the global 3D awareness for grounding projected BEV anchors on the $n$ -the view. + +To achieve the global contextualization effect, we adapt the recently proposed Patch-to-Cluster attention (PaCa) [12] method. The core idea of PaCa is to leverage a learnable clustering module to cluster a feature map into a predefined number $M$ of clusters. For notional simplicity, consider a feature map $F_{h\times w\times c}$ as $N = h\times w$ tokens $F_{N\times c}$ , the clustering assignment is computed by, + +$$ +\mathcal {C} _ {N, M} = \operatorname {S o f t m a x} (\operatorname {C l u s t e r i n g} (F _ {N, c})), \tag {3} +$$ + +where Clustering() can be implemented in different ways (see our ablation studies in Tab. 6), and the Softmax is along the token dimension. Then, we compute $M$ clusters by, + +$$ +z _ {M, c} = \mathrm {L N} \left(\mathcal {C} _ {N, M} ^ {\top} \cdot F _ {N, c}\right), \tag {4} +$$ + +where $\mathrm{LN}()$ is the layer normalization [1]. + +Then, the PaCa-based MvAcon is defined by, + +$$ +F _ {N, c} ^ {\prime} = \operatorname {S o f t m a x} \left(\frac {Q _ {N , c} \cdot K _ {M , c} ^ {\top}}{\sqrt {c}}\right) \cdot V _ {M, c} + F _ {N, c}, \tag {5} +$$ + +where $Q_{N,c}$ is the linear projection of $F_{N,c}, K_{M,c}$ and $V_{M,c}$ are from the clusters $z_{M,c}$ . The second term is the shortcut. The multi-head PaCa can be straightforwardly defined. For Eqn. 2, we concatenate the clusters from all the pyramid layers before computing the Key and the Value. Here, the PaCa module is of linear complexity. + +# 4. Experiments + +# 4.1. Experimental Setup + +Dataset and Metrics We evaluate our MvAcon on the challenging large-scale NuScenes dataset [3] and Waymo dataset [11]. The NuScenes dataset includes 1,000 scene sequences, which are divided into training, validation, and testing subsets in a 700/150/150 split. Each sequence in the NuScenes dataset is a 20-second video clip, annotated at a rate of 2 frames per second (FPS). The NuScenes dataset employs a comprehensive suite of evaluation metrics for assessing detection performance. These metrics comprise mean Average Precision (mAP), mean Average Translation Error (mATE), mean Average Scale Error (mASE), mean Average Orientation Error (mAOE), mean Average Velocity Error (mAVE), mean Average Attribution Error (mAAE), and the NuScenes Detection Score (NDS). The Waymo dataset contains 798 training sequences and 202 validation sequences. We use a subset of the training set (Waymomini) by sampling every third frame from the training sequences following [25]. + +Implementation Details We leverage open-source code bases (PETR [29], BEVFormer [25], and DFA3D [19]) in + +our experiments. To ensure a fair and stringent comparison, we maintain all original configurations of these methods, making only one modification: the addition of an attentive contextualization module. We conduct qualitative analysis and ablation study on the BEVFormer-base model. We train all models for 24 epochs using 8 NVIDIA Tesla A100 GPUs, following the configurations and settings outlined in previous works [19, 25, 29]. + +# 4.2. The Effectiveness of MvACon across Different Methods + +To demonstrate the effectiveness of our proposed MvA-Con method, we first apply it to two typical query-based MV3D object detection paradigms: the perspective-based decoder-only detector (PETR [29]) and the encoder-decoder based detector (BEVFormer [25]). We choose these as our baselines because state-of-the-art (SOTA) query-based MV3D detectors [19, 30, 52, 60] primarily follow these two paradigms. We also test our method on DFA3D [19] to demonstrate its generalizability to SOTA methods. + +On the NuScenes dataset, Table 1 shows that our proposed MvACon consistently improves performance across different detectors. Specifically, for the perspective-based decoder-only detector PETR, it improves the baseline by 0.8 NDS. For the encoder-decoder based detector BEVFormer, our method achieves an improvement of 1.3 in NDS on average. On a more advanced, depth-context enhanced BEVFormer (DFA3D), our method further improves performance by up to 0.5 NDS. Notably, our MvACon achieves significant improvement in location (mAP, mATE), orientation (mAOE), and velocity prediction (mAVE) for encoder-decoder based detectors. It also markedly enhances performance in location (mAP, mATE) and velocity (mAVE) prediction for the perspective-based decoder-only detector. + +On the Waymo dataset, since there are few released codes for MV3D detectors on Waymo except for the BEVFormer, we only test BEVFormer on Waymo-mini following its settings with results shown in Table 2. Our MvACon shows consistent improvement on Waymo metrics. + +# 4.3. How Does Our MvACon Work? + +We elaborate on the effects of our MvACon by providing detailed analyses during the 2D-to-3D feature lifting process. We first demonstrate what the learned cluster contexts encode, then show how these contexts affect the behavior of deformable points during feature lifting. Lastly, we illustrate how our MvACon improves detection results by presenting a qualitative comparison within a scene. We select BEVFormer-base as our analysis target due to its incorporation of six layers of deformable attention modules in the encoder. More qualitative analysis is provided in the supplementary materials. + +What do the Learned Cluster Contexts Encode? We vi- + +
MethodmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
PETR-VovNet-99 [29]37.874.627.248.890.621.242.6
PETR-VovNet-99-MvACon38.2 (+0.5)73.927.050.584.121.343.4 (+0.8)
BEVFormer-t [25]25.290.029.465.565.721.635.4
BEVFormer-t-MvACon25.9 (+0.7)88.428.864.660.522.536.5 (+1.1)
BEVFormer-s [25]37.072.128.040.743.622.047.9
BEVFormer-s-MvACon39.3 (+2.3)71.327.740.142.019.749.6 (+1.7)
BEVFormer-b [25]41.667.327.437.239.419.851.7
BEVFormer-b-MvACon42.6 (+1.0)66.427.635.036.220.052.8 (+1.1)
DFA3D-s [19]40.172.127.941.139.119.650.1
DFA3D-s-MvACon40.171.027.438.337.220.850.6 (+0.5)
DFA3D-b [19]43.065.427.137.434.120.553.1
DFA3D-b-MvACon43.2 (+0.2)66.427.534.432.320.753.5 (+0.4)
+ +![](images/c0aa0f973601d562baf867b1786bce2e4d9992d5b22c24c6d19d5b0565ee4474.jpg) + +![](images/3e9ee985e76f829fcd96abfa656ad52caef07299387e4e36ea72cfec2d6cdf49.jpg) +Figure 3. Visualization results of learned cluster contexts in our MvACon on the NuScenes validation set. We sum all the learned clusters along the channel and upsample it to the original image resolution through bilinear interpolation. We observed that the learned cluster context encodes abundant context information in the scene. We provide details with raw images in the supplementary. + +![](images/f24cf33f389162a391e4b02ad7503f079bc6ec580145ec613ecaec4be3c7096c.jpg) + +![](images/7d78c0edf5d69db9010c2f652450b12b3432ef79dbdd716dfdc75b32c3f18498.jpg) + +![](images/9b0b78e6189103ed8c556c121f5244ecc9bbc0364c5c9eeefb5ea707b212bd27.jpg) + +![](images/0d8d9fc9e6f66c02159cd0eb1ac1fb0333032de245e690a71bb8cdcd17832197.jpg) + +![](images/a56d1c6240209c0b7350f78a1ed48bf598f448dc478a2108e4a56ae991f00ccf.jpg) + +![](images/3c589eba87906b82b521e299d80e8cdf1990b8745ee5df20ed56809d052b1352.jpg) + +Table 1. Comparisons of our method with baslines on the NuScenes validation set. BEVFormer-t/s/b refers to BEVFormer-tiny/small/base in the BEVFormer's open-source codes. 'DFA3D' refers to the adaptation of 2D deformable attention into a (depth-weighted) 3D deformable attention within the BEVFormer model, as adopted in [19]. + +
MethodLET-mAPL↑LET-mAPH↑
BEVFormer-ResNet101 [25]34.946.3
BEVFormer-ResNet101-MvACon35.7 (+0.8)47.5 (+1.2)
+ +Table 2. Comparisons on the Waymo-mini. + +sualize the learned cluster context in a heatmap, by summing all clusters along the channel and then upsampling to the original image resolution using bilinear interpolation. The results, shown in Fig. 3, reveal that despite the low resolution of the original summed heatmap, it can accurately locate foreground objects in the scene after upsampling. This suggests that the cluster contexts encode the foreground layouts in a scene. Furthermore, we note that the upsampled heatmap shows a high response to various foreground objects, even when they are spatially close in 2D. This is attributed to the effectiveness of our MvACon in dense scenes, as demonstrated and analyzed later in our experiments. + +How do Cluster Contexts Affect the Behavior of Deformable Points? We demonstrate the dynamics of deformable points in Fig. 4. These points originate from one 2D reference point, which is projected from a 3D BEV anchor point in the BEVFormer encoder. As the encoding + +layers deepen in the vanilla BEVFormer, we observe that most deformable points have low attention weights, suggesting the model's uncertainty about the relevance of the selected context. Consequently, these contexts contribute minimally during the encoding of BEV query features in the feature lifting process. In contrast, deformable points predicted by our method maintain high confidence weights on foreground objects (e.g., cars), as well as on surrounding buildings. We also note that points on buildings remain stable across encoding layers and consecutive frames (see the supplementary). These high-confidence deformed points in our MvACon may foster a local object-context aware coordinate system, enhancing overall performance, including the estimation of velocity and orientation. This observation aligns with our quantitative findings in Table 1. + +Qualitative Comparisons with the Baseline Method. We provide qualitative comparison with BEVFormer in Fig. 5. Our method exhibits superior performance in dense scenarios where objects are crowded and challenging to localize. We attribute this enhanced performance to the rich foreground layout context in the scene, facilitated by our attentive contextualization module. Additional qualitative comparisons are available in the supplementary. + +![](images/ea8e463d6a162de83c5c59461ee9eff5ca7f6088b46876653e7c8d36107398dd.jpg) + +![](images/d5ab9c2bd43a57241e3d3282aa56f8c992107df2f89f37056b8180bd12ba3ad4.jpg) + +![](images/3f0a40011a876d1f04a9438d4072a5a65c86ad481ced787ba7be0c1124015fb7.jpg) + +![](images/82a96e50fa205cc7b1d84d5e613e3453a601b9a90c04d56cbbd711cda24fcb9b.jpg) + +![](images/70ad81e688f2c4c764f5e312ef38fba4f1403e3d956f0da28772c7c3abc462ab.jpg) + +![](images/ec8d0c83009dcc5348cf026119890974632859e4c47f40946d9dddfac979f401.jpg) + +![](images/a0803f15021e679c4855e4156aae7e86d2f7a06441cad50fd4ff5a9a290c56d3.jpg) + +![](images/1aa1fb988b350f72349578cfed2dbb8cbcdb03432482af17434e64ff4e79afd0.jpg) + +![](images/d307767d3e5d58525008b9b76b67cd6d38f68cb22260f6ee1723f5281a699a24.jpg) + +![](images/246726bcd7652b194da8259cba91b0c8080c7eed480c107abda1c9d0929374c5.jpg) +Figure 4. Visualization results of the deformable points originating from a 2D reference point, which is projected from a 3D BEV anchor point in the BEVFormer encoder, on NuScenes validation set. We utilize the same BEV anchor point as demonstrated in Fig. 1. From left to right and up to bottom, we display the deformable points output from each layer (#1-#6) in the encoder, respectively. + +![](images/b2bfd2901d36a70ac5abb4bc8e03529ea044f310ba5d0d04cb95e4652df426af.jpg) + +![](images/30201d8134e19dc120cc42b9215218fffe89654e81f79cdae7290c978908fc9d.jpg) + +![](images/bc9568dc4ce23aa7eaf48d326f7f561c6f32b107e432a17ea178486f864f0cc4.jpg) + +![](images/9c03d03ac0013399ac7ba2bcc9ed0dd23491a7bc8945cfbcaba171bea6f82447.jpg) + +![](images/18f5e56b228531861d0e143ce986f9ee3918cc0a9d7bb550387cc57bcb114820.jpg) + +![](images/210aa166ae1c98c05cd71a15d95ed115479a3fb470a8b3b9ba929b58e8561c02.jpg) + +![](images/3d5bccb46a4b166489d57ecfda6f703addcb729d02a0c0f09ee192ada2acc644.jpg) + +![](images/f62eb02409fedcab9e8b5d61e6a5fea1b6f58b59722a222453c9914a11e04e41.jpg) + +![](images/dd64fd307686675c626386891f762e3913f88f24ea5d84a2d1a7c1ad797abc42.jpg) + +![](images/0f13c5f3c184899e1ed6ed39fcba720db8dc763f03494f4ec1a80c4e89429ade.jpg) +Figure 5. Qualitative comparisons between BEVFormer and our MvACon method on NuScenes validation set. + +![](images/0fd7e1520e178ad8c353679dbaf3e11c555eb934c84a38344a6bdb2626d7b9f4.jpg) + +![](images/1470954d2e27c7807a52607f89565c582b53f982c228ca7519f20ad619f11c14.jpg) + +![](images/a6b303c60dbe6837e90ea6ab9c65cca735e7f05470de43c2148e6f8577df7ea2.jpg) + +![](images/7c360e5184c406e4b73793fee739843b2ca95c17dfd4869ec08aa43ec27fb51e.jpg) + +![](images/380357fd83a567b48d3def6b532c979c48039e1ae1aed50dce8cb86af7e32276.jpg) + +![](images/94faaa01fa51e26324087c0d894d46685166bde06bf62b964f0b44bd222a57d3.jpg) + +# 4.4. Ablation Studies + +Effectiveness of Different Contextualization Methods. Table 3 demonstrates the impact of various contextual methods on detection performance. We selected three representative contexts: local shift window-based context (SWin [32]), global pooling-based context (PVTV2-linear [55]), and global cluster-based context (PaCa [12]). The results show that the local shift window-based context offers minimal improvements in detection performance, which can be attributed to similar local contexts already provided by convolutional backbone networks. Conversely, enhancing the original feature with global contexts, as observed in PVTV2-linear and PaCa experiments, leads to better performance. Notably, orientation and velocity show significant improvements with these global contexts. The inclusion of cluster contexts further enhances improvements in location, orientation, and velocity prediction, as evidenced in the global cluster experiment. + +
MethodmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
BEVFormer-b [25]41.667.327.437.239.419.851.7
Shift Window (SWin [32])41.5 (-0.1)67.127.837.839.620.151.5 (-0.2)
Global Pooling (PVTV2-linear [55])41.6 (+0.0)66.527.536.837.619.652.0 (+0.3)
Global Cluster (PaCa [12])42.6 (+1.0)66.427.635.036.220.052.8 (+1.1)
+ +Table 3. Ablation study on different contextualization methods. + +
ContextmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
Local (BEVFormer-b [25])41.667.327.437.239.419.851.7
Global Cluster41.666.427.438.335.119.552.1
Local + Global Cluster42.6 (+1.0)66.427.635.036.220.052.8 (+1.1)
+ +Table 4. Ablation study on the relationship between local contexts and global cluster-based contexts. + +
Method#layers#clustersCross-level\( \mathbf{{mAP}} \dagger \)\( \mathbf{{mATE}} \downarrow \)\( \mathbf{{mASE}} \downarrow \)\( \mathbf{{mAOE}} \downarrow \)\( \mathbf{{mAVE}} \downarrow \)\( \mathbf{{mAAE}} \downarrow \)\( \mathbf{{NDS}} \uparrow \)
BEVFormer-b [25]---41.667.327.437.239.419.851.7
Baseline 13100✓41.165.627.137.736.418.952.0
Baseline 2650✓41.366.327.336.134.118.852.4
Baseline 36100-42.466.827.637.136.319.152.5
MvACon6100✓42.666.427.635.036.220.052.8
+ +Table 5. Ablation study on the structure of our attentive contextualization module. + +
Clustering MethodmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
BEVFormer-b [25]41.667.327.437.239.419.851.7
Linear41.965.627.238.234.819.352.4
MLP42.466.427.338.235.019.252.6
Conv42.666.427.635.036.220.052.8
+ +Table 6. Ablation study on clustering operations in the attentive contextualization module. + +Relationship between Local Contexts and Global Cluster Contexts. Table 4 reveals the relationship between local and global cluster contexts in enhancing feature learning for view lifting. The exclusive use of global cluster context results in improved velocity prediction, while local attention contributes to better orientation prediction results. Combining these two contexts enhances predictions in location, orientation, and velocity. This highlights the complementary role that global cluster contexts play in feature encoding for view lifting. + +Structure of the Attentive Contextualization Module. Table 5 shows the impact of the structure of our attentive contextualization method. Baseline 1 demonstrates that attentive contextualization can efficiently encode feature information. With only three layers, MvACon achieves an improvement of 0.3 NDS compared to the vanilla BEVFormer. Baseline 2 indicates that attentive contextualization requires a sufficient number of clusters to extract abundant cluster contexts in the scene. Baseline 3 suggests that attending to clusters across the feature map aids in improving orientation, velocity, and mAP prediction. Table 6 illustrates that the convolution operation for clustering yields better results in terms of orientation, mAP, and NDS prediction. Conversely, point-based operations (such as a linear layer or multi-layer perceptron) demonstrate superior performance in location and velocity prediction. + +# 5. Conclusion + +This paper presents Multi-View Attentive Contextualization (MvACon) for improving query-based multi-view 3D (MV3D) object detection. It addresses the limitations of two main paradigms of query based MV3D object detector in a unified way: decoder-only dense attention methods like PETR, which lack high-resolution features due to computational constraints, and encoder-decoder sparse 2D/3D deformable attention methods such as BEVFormer and DFA3D. Our MvACon contextualizes the original feature maps extracted from the backbone network using a cluster-attention operation built on the recently proposed patch-to-cluster attention (PaCa). In experiments, we show that our MvACon effectively and consistently enhances query-based MV3D object detectors by encoding more useful contexts, thereby facilitating better 2D-to-3D feature lifting. Rigorously controlled experiments reveal that, for decoder-only detectors, the cluster attention contextualization significantly improves localization and velocity prediction. For encoder-decoder based detectors, it effectively reduces errors in location, orientation, and velocity. + +# Acknowledgments + +X. Liu and T. Wu were partly supported by NSF IIS-1909644, ARO Grant W911NF1810295, ARO Grant W911NF2210010, NSF CMMI-2024688, NSF IUSE2013451, and a research gift fund from the Innopeak Technology, Inc. (an affiliate of OPPO). The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ARO, NSF, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. + +# References + +[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5 +[2] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In ICCV, pages 9287-9296, 2019. 2 +[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 5 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2, 3 +[5] Dian Chen, Jie Li, Vitor Guizilini, Rares Andrei Ambrus, and Adrien Gaidon. Viewpoint equivariance for multiview 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9213-9222, 2023. 2 +[6] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2781-2790, 2022. 2 +[7] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. Monopair: Monocular 3d object detection using pairwise spatial relationships. In CVPR, pages 12093-12102, 2020. 2 +[8] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3d object detection in autonomous driving. In CVPR, 2022. 2 +[9] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3d object detection. In CVPR Workshops, pages 1000-1001, 2020. 2 +[10] 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. 3 +[11] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 2, 5 +[12] Ryan Grainger, Thomas Paniagua, Xi Song, Naresh Cuntoor, Mun Wai Lee, and Tianfu Wu. Paca-vit: Learning patch-to-cluster attention in vision transformers. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18568-18578, 2023. 2, 3, 5, 8 +[13] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems, 34:15908-15919, 2021. 3 +[14] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 2 +[15] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 2 +[16] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In CVPR, 2022. 2 +[17] Abhinav Kumar, Garrick Brazil, and Xiaoming Liu. Groomed-nms: Grouped mathematically differentiable nms for monocular 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8973-8983, 2021. 2 +[18] Abhinav Kumar, Garrick Brazil, Enrique Corona, Armin Parchami, and Xiaoming Liu. Deviant: Depth equivariant network for monocular 3d object detection. In ECCV, 2022. 2 +[19] Hongyang Li, Hao Zhang, Zhaoyang Zeng, Shilong Liu, Feng Li, Tianhe Ren, and Lei Zhang. Dfa3d: 3d deformable attention for 2d-to-3d feature lifting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6684–6693, 2023. 2, 5, 6 +[20] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. Rtm3d: Real-time monocular 3d detection from object keypoints for autonomous driving. In European Conference on Computer Vision, pages 644-660. Springer, 2020. 2 +[21] Yawei Li, Kai Zhang, Jiezhang Cao, Radu Timofte, and Luc Van Gool. Localvit: Bringing locality to vision transformers. arXiv preprint arXiv:2104.05707, 2021. 3 +[22] Yingyan Li, Yuntao Chen, Jiawei He, and Zhaoxiang Zhang. Densely constrained depth estimator for monocular 3d object detection. ECCV, 2022. 2 +[23] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 2 +[24] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2 +[25] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 4, 5, 6, 8 + +[26] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 2 +[27] Xianpeng Liu, Nan Xue, and Tianfu Wu. Learning auxiliary monocular contexts helps monocular 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1810-1818, 2022. 3 +[28] Xianpeng Liu, Ce Zheng, Kelvin B Cheng, Nan Xue, GuoJun Qi, and Tianfu Wu. Monocular 3d object detection with bounding box denoising in 3d by perceiver. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6436-6446, 2023. 2 +[29] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022. 2, 3, 4, 5, 6, 1 +[30] Yingfei Liu, Junjie Yan, Fan Jia, Shuailin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 2, 5 +[31] Yang Liu, Yao Zhang, Yixin Wang, Feng Hou, Jin Yuan, Jiang Tian, Yang Zhang, Zhongchao Shi, Jianping Fan, and Zhiqiang He. A survey of visual transformers. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3 +[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 8 +[33] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15641-15650, 2021. 2 +[34] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. arXiv:2107.13774, 2021. 2 +[35] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In ICCV, pages 6851-6860, 2019. 2 +[36] Xinzhu Ma, Shinan Liu, Zhiyi Xia, Hongwen Zhang, Xingyu Zeng, and Wanli Ouyang. Rethinking pseudo-lidar representation. In ECCV, pages 311-327. Springer, 2020. 2 +[37] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3d object detection. In CVPR, pages 4721-4730, 2021. 2 +[38] Yuexin Ma, Tai Wang, Xuyang Bai, Huitong Yang, Yuenan Hou, Yaming Wang, Yu Qiao, Ruigang Yang, Dinesh Manocha, and Xinge Zhu. Vision-centric bev perception: A survey. arXiv preprint arXiv:2208.02797, 2022. 1 +[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: + +Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[40] Arsalan Mousavian, Dragomir Anguelov, John Flynn, and Jana Kosecka. 3d bounding box estimation using deep learning and geometry. In CVPR, pages 7074-7082, 2017. 2 +[41] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021. 3 +[42] Dennis Park, Jie Li, Dian Chen, Vitor Guizilini, and Adrien Gaidon. Depth is all you need for monocular 3d detection. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7024-7031, 2023. 3 +[43] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 2 +[44] Liang Peng, Xiaopei Wu, Zheng Yang, Haifeng Liu, and Deng Cai. Did-m3d: Decoupling instance depth for monocular 3d object detection. In ECCV, 2022. 2 +[45] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2 +[46] Changyong Shu, Fisher Yu, and Yifan Liu. 3d point positional encoding for multi-camera 3d object detection transformers. arXiv preprint arXiv:2211.14710, 2022. 2 +[47] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In ICCV, pages 1991-1999, 2019. 2 +[48] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16519-16529, 2021. 3 +[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 3 +[50] 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. 2 +[51] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12894–12904, 2021. 3 +[52] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. arXiv preprint arXiv:2303.11926, 2023. 2, 5 + +[53] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021. 3 +[54] Tai Wang, Pang Jiangmiao, and Lin Dahua. Monocular 3d object detection with depth from motion. In ECCV, 2022. 2 +[55] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 3, 8 +[56] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In CVPR, pages 8445-8453, 2019. 2 +[57] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2, 3 +[58] Bichen Wu, Chenfeng Xu, Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Zhicheng Yan, Masayoshi Tomizuka, Joseph Gonzalez, Kurt Keutzer, and Peter Vajda. Visual transformers: Token-based image representation and processing for computer vision. arXiv preprint arXiv:2006.03677, 2020. 3 +[59] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerflike representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 3 +[60] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 3, 5 +[61] Kun Yuan, Shaopeng Guo, Ziwei Liu, Aojun Zhou, Fengwei Yu, and Wei Wu. Incorporating convolution designs into visual transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 579-588, 2021. 3 +[62] Li Yuan, Qibin Hou, Zihang Jiang, Jiashi Feng, and Shuicheng Yan. Volo: Vision outlooker for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 45(5):6575-6586, 2022. 3 +[63] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2998-3008, 2021. 3 +[64] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In CVPR, pages 3289-3298, 2021. 2 +[65] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Dalong Du, Jie Zhou, and Jiwen Lu. Dimension embeddings for monocular 3d object detection. In CVPR, 2022. 2 + +[66] Ce Zheng, Xianpeng Liu, Guo-Jun Qi, and Chen Chen. Potter: Pooling attention transformer for efficient human mesh recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1620, 2023. 3 +[67] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3 \ No newline at end of file diff --git a/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/images.zip b/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..259eca1607783473616b0e4df9c39aebc6e561ac --- /dev/null +++ b/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f1e69f13ed273eb3616bc9045f55e1f06df323b907965c9def6158062a9731a +size 606292 diff --git a/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/layout.json b/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ddfa7840e0948b58e0f318716cb3b0ac73e9eb41 --- /dev/null +++ b/2024/Multi-View Attentive Contextualization for Multi-View 3D Object Detection/layout.json @@ -0,0 +1,9143 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 103, + 531, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 103, + 531, + 120 + ], + "spans": [ + { + "bbox": [ + 62, + 103, + 531, + 120 + ], + "type": "text", + "content": "Multi-View Attentive Contextualization for Multi-View 3D Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "text", + "content": "Xianpeng Liu" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "text", + "content": ", Ce Zheng" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "text", + "content": ", Ming Qian" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "text", + "content": ", Nan Xue" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "text", + "content": ", Chen Chen" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "text", + "content": ", Zhebin Zhang" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "text", + "content": ", Chen Li" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "text", + "content": ", Tianfu Wu" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "text", + "content": "North Carolina State University " + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "text", + "content": "University of Central Florida \n" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "text", + "content": "Ant Group " + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 142, + 555, + 200 + ], + "type": "text", + "content": "OPPO U.S. Research Center \nhttps://xianpeng919.github.io/mvacon" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "type": "text", + "content": "We present Multi-View Attentive Contextualization (MvACon), a simple yet effective method for improving 2D-to-3D feature lifting in query-based multi-view 3D (MV3D) object detection. Despite remarkable progress witnessed in the field of query-based MV3D object detection, prior art often suffers from either the lack of exploiting high-resolution 2D features in dense attention-based lifting, due to high computational costs, or from insufficiently dense grounding of 3D queries to multi-scale 2D features in sparse attention-based lifting. Our proposed MvACon hits the two birds with one stone using a representationally dense yet computationally sparse attentive feature contextualization scheme that is agnostic to specific 2D-to-3D feature lifting approaches. In experiments, the proposed MvA-Con is thoroughly tested on the nuScenes benchmark, using both the BEVFormer and its recent 3D deformable attention (DFA3D) variant, as well as the PETR, showing consistent detection performance improvement, especially in enhancing performance in location, orientation, and velocity prediction. It is also tested on the Waymo-mini benchmark using BEVFormer with similar improvement. We qualitatively and quantitatively show that global cluster-based contexts effectively encode dense scene-level contexts for MV3D object detection. The promising results of our proposed MvA-Con reinforces the adage in computer vision – “(contextualized) feature matters”." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Camera-based 3D object detection is a pivotal area of research in computer vision, particularly owing to its appealing applications in cost-effective autonomous systems, such as autonomous driving and robot autonomy. Recently, significant progress has been witnessed in the field of multiview 3D (MV3D) object detection, especially with the advent of end-to-end MV3D detection approaches [38]. One crucial component in the end-to-end MV3D detection system is the 2D-to-3D feature lifting module, which converts" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 227, + 425, + 293 + ], + "blocks": [ + { + "bbox": [ + 309, + 227, + 425, + 293 + ], + "lines": [ + { + "bbox": [ + 309, + 227, + 425, + 293 + ], + "spans": [ + { + "bbox": [ + 309, + 227, + 425, + 293 + ], + "type": "image", + "image_path": "4b3c92ee1a1a867ad581bc546d93dce8b7fdd55aefd35d6ca4915a4d2798e344.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 296, + 409, + 305 + ], + "lines": [ + { + "bbox": [ + 324, + 296, + 409, + 305 + ], + "spans": [ + { + "bbox": [ + 324, + 296, + 409, + 305 + ], + "type": "text", + "content": "(a) Projected BEV anchor" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 428, + 227, + 544, + 293 + ], + "blocks": [ + { + "bbox": [ + 428, + 227, + 544, + 293 + ], + "lines": [ + { + "bbox": [ + 428, + 227, + 544, + 293 + ], + "spans": [ + { + "bbox": [ + 428, + 227, + 544, + 293 + ], + "type": "image", + "image_path": "72f9097ae567920622785e2aca5192658191fac136935847c787a75bb3068327.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 296, + 520, + 305 + ], + "lines": [ + { + "bbox": [ + 452, + 296, + 520, + 305 + ], + "spans": [ + { + "bbox": [ + 452, + 296, + 520, + 305 + ], + "type": "text", + "content": "(b) BEVFormer [25]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 307, + 425, + 375 + ], + "blocks": [ + { + "bbox": [ + 308, + 307, + 425, + 375 + ], + "lines": [ + { + "bbox": [ + 308, + 307, + 425, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 307, + 425, + 375 + ], + "type": "image", + "image_path": "f81bff84d620c7860b1cf7e2d1b47f57ac88b21ca09e94f489362b0d58ca13c1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 376, + 399, + 385 + ], + "lines": [ + { + "bbox": [ + 333, + 376, + 399, + 385 + ], + "spans": [ + { + "bbox": [ + 333, + 376, + 399, + 385 + ], + "type": "text", + "content": "(c) Learned clusters" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 428, + 307, + 544, + 373 + ], + "blocks": [ + { + "bbox": [ + 428, + 307, + 544, + 373 + ], + "lines": [ + { + "bbox": [ + 428, + 307, + 544, + 373 + ], + "spans": [ + { + "bbox": [ + 428, + 307, + 544, + 373 + ], + "type": "image", + "image_path": "cc7d655eb37f6fe469d90b6b5c093a6ce9f567b774320dfb3c5904332d09cf92.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 376, + 515, + 385 + ], + "lines": [ + { + "bbox": [ + 457, + 376, + 515, + 385 + ], + "spans": [ + { + "bbox": [ + 457, + 376, + 515, + 385 + ], + "type": "text", + "content": "(d) Our MvACon" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 388, + 547, + 608 + ], + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 608 + ], + "type": "text", + "content": "Figure 1. The effects of our proposed MvACon in the 2D-to-3D feature lifting. Consider a point (red) on the car in (a), which is projected from a 3D BEV anchor point. In lifting 2D features to ground the 3D BEV anchor point, vanilla BEVFormer [25] utilizes a predefined number of deformable points, with offsets learned through a 6-layer cross-attention module relative to the projection point. (b) shows the deformed points after the final cross-attention layer, most of which have low attention weights, indicating the model's uncertainty or inability to consolidate contributions effectively for good lifting. Our MvACon tackles this issue with clustering-based attention, as visualized in (c). (d) shows the deformed points, where we observe not only high-confidence points on the car but also on the building. We further observe that the points on the building remain stable across encoding layers (see Fig. 4) and consecutive frames (see suppl.). With those high-confident deformed points in a spatiotemporally stable configuration, our MvACon may induce a local object-context aware coordinate system that helps the overall performance, especially the estimation of velocity and orientation, as we quantitatively observed in experiments. See text for details." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 700 + ], + "type": "text", + "content": "perspective 2D multi-view image feature maps to 3D feature representation. It aims to counter the complete loss of depth information in individual 2D images by exploiting multi-view clues. However, a significant challenge in practical applications like autonomous driving is the often insufficient field-of-view overlap across views, making it difficult to effectively address the loss of depth information." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "To perform 2D-to-3D feature lifting, recent methods" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16688" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": "[15, 25, 29, 57] aim to learn a unified 3D space representation using 3D anchors that are either sparsely or uniformly sampled. These methods generally fall into two categories subject to the interaction of 3D anchors with 2D features and the feature aggregation strategy. (1) The Lift-Splat-Shoot (LSS) method [14, 15, 23, 24, 45] first lifts 2D features into 3D (pseudo-LiDAR) space using the outer product with the estimated depth, then assigns them to the nearest 3D anchors. (2) In contrast, the query-based design [25, 29, 57], pioneered by the DETR method [4] for end-to-end 2D object detection, adopts 3D anchors as queries and uses 2D image features as keys and values. They interact and aggregate via spatial cross-attention in the expressive Transformer architecture [50]. These two paradigms have been widely used in downstream tasks like map segmentation [30] and occupancy prediction [25]. This paper primarily focuses on the query-based detection paradigm. One reason is that LSS-based methods often encounter excessive computational complexity and issues with error propagation and depth estimation magnification post-lifting, potentially capping their performance. However, the query-based design also grapples with heavy computation costs or limited 3D information awareness, depending on their Transformer design." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 363, + 289, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 289, + 626 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 289, + 626 + ], + "type": "text", + "content": "In this paper, we focus on addressing limitations of two main paradigms of query based MV3D object detectors in a unified way (elaborated in Sec. 3). In particular, we introduce multi-view attentive contextualization (MvA-Con) to address limitations of decoder-only dense attention methods like PETR [29], which lack high-resolution features due to computational constraints, and to simultaneously address the issue of sparsely grounded 3D anchors in encoder-decoder 2D/3D deformable attention methods such as BEVFormer [25] and DFA3D [19]. Our proposed MvA-Con aims to be representationally dense while computationally sparse. To achieve this, we expand the conventional three-component paradigm of MV3D object detection to a four-component setup (Fig. 2): (1) 2D image representation learning through a feature backbone shared across views, (2) MvACon for attentive contextualization of the 2D features, (3) 2D-to-3D feature lifting, and (4) a 3D object detection head or decoder that utilizes these lifted features. This modular design allows our MvACon to remain agnostic to specific 2D-to-3D feature lifting strategies and aligns with the classic adage in representation learning and computer vision: -('contextualized') feature matters'." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": "More specifically, our approach contextualizes original feature maps extracted from the backbone network using a cluster-attention operation. This builds upon the recently proposed patch-to-cluster attention (PaCa) [12]. For perspective-based decoder-only detectors like PETR, we apply cluster contextualization before the feature maps are fed into the decoder. For encoder-decoder based detec" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "content": "tors, such as BEVFormer and DFA3D, we incorporate cluster contextualization within the spatial cross-attention operation. Through extensive experiments, we demonstrate that our proposed MvACon effectively and consistently enhances query-based MV3D object detectors by encoding more useful contexts, thereby facilitating better 2D-to-3D feature lifting. Rigorously controlled experiments reveal that, for perspective-based decoder-only detectors, the cluster attention contextualization significantly improves localization and velocity prediction. In the case of encoder-decoder based detectors, it effectively reduces errors in location, orientation, and velocity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 217, + 482, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 482, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 482, + 227 + ], + "type": "text", + "content": "In summary, our main contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 228, + 545, + 371 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 228, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 262 + ], + "type": "text", + "content": "- We analyze and address the limitation of 2D-to-3D feature lifting in the prior art, that is the lack of sufficient 3D representational power due to their local 3D awareness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 264, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 264, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 545, + 322 + ], + "type": "text", + "content": "- We propose MvACon (Multi-view Attentive Contextualization) to induce the global 3D awareness in an easy-to-integrate way to enhance the 2D-to-3D feature lifting in both decoder-only based MV3D object detectors and encoder-decoder based ones." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 323, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 545, + 371 + ], + "type": "text", + "content": "- We show consistent performance improvement of our MvACon on the challenging NuScenes [3] dataset using three baseline query-based MV3D object detectors, as well as on the Waymo-mini [11] benchmark." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 382, + 392, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 382, + 392, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 392, + 395 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": "Camera-based 3D Object Detection. Camera-based 3D object detection can be primarily categorized into two settings: single-view and multi-view. In the realm of monocular 3D object detection research, addressing the challenge of inaccurate object localization [37] is critical. Researchers have exerted considerable effort to utilize monocular depth cues. This includes transforming inputs into pseudo-lidar point clouds [35, 36, 56] and explicitly incorporating depth into models [8, 9, 16]. Another significant research direction involves the explicit use of geometric priors, encompassing approaches like key-point constraints [6, 20, 33], shape projection relationships [22, 34, 40, 64], and temporal depth estimation [54]. Innovations in monocular 3D object detection also include novel loss modules [7, 47, 65], 3D-aware backbones [2, 17, 18], and second-stage detection paradigms [28, 44]. In multi-view settings, the configuration often closely resembles that of monocular setups due to the limited field-of-view overlap between the different camera views. Therefore, MV3D object detection focuses on addressing the challenge of learning universal representation for multi-view sensors. It has benefited from advancements in various techniques such as view lifting [5, 19, 25, 29, 45, 57], depth encoding [19, 23, 24, 46], and temporal modeling [14, 25, 26, 30, 43, 52]. In our multi-view approach, we aim at addressing the challenge of multi-view representation learning with focus on en" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16689" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "hancing the view lifting module within query-based detection methods by empowering original features with cluster-based contextual features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 111, + 289, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 111, + 289, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 111, + 289, + 387 + ], + "type": "text", + "content": "Representation Learning in Camera-based 3D Object Detection. Camera-based 3D object detection is inherently a data-intensive task due to its ill-posed nature, the expansive search space in 3D, and the scarcity of labeled data in scenes. Consequently, developing robust representations for this task is both critical and challenging. Early research in monocular 3D object detection has demonstrated the utility of depth contexts [41, 42] and projection contexts [27] in enhancing detection capabilities. Recent advances also highlight the effectiveness of scene-level representations, such as density fields [39], in improving 3D representation learning [59]. In the domain of multi-view research, most leading methods utilize backbones pre-trained with projection contexts (e.g., FCOS3D [53] weights) or depth contexts (e.g., DD3D [41] weights). However, these pre-trained weights may not fully leverage the capabilities of newer backbone network designs. Recent studies have begun to explore alternatives to this pre-trained paradigm, including the integration of an auxiliary projection context branch in end-to-end training [60]. Our work aims to enhance network representation by explicitly incorporating scene-level cluster context as supplementary information during the view lifting stage in query-based MV3D object detectors." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 389, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 289, + 628 + ], + "type": "text", + "content": "Vision Transformers. Since the pioneer work of ViT [10], extensive research [31] has been dedicated to enhancing the representational abilities of neural networks for visual tasks. It has been established that CNNs and Transformers can mutually augment each other's capabilities, as evidenced in designs like Transformer-enhanced CNNs [48, 58] and CNN-enhanced Transformers [21, 49, 61]. Additionally, a significant branch of visual Transformer research focuses on developing new attention mechanisms tailored to the locality bias in vision tasks. Notable examples include HaloNet [51], SWin [32], Deformable Attention [67], and VOLO [62], all introducing innovative local attention mechanisms to mitigate the quadratic computational cost associated with visual inputs. Concurrently, models like TNT [13], ViL [63], PVTv2-linear [55], POTTER [66], and PaCa [12] explore the integration of local and global contexts. Inspired by the progress of Vision Transformers, our work addresses the limitations of two prevalent paradigms in query-based MV3D detectors caused by their attention mechanisms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 644, + 113, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 644, + 113, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 644, + 113, + 658 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "type": "text", + "content": "Given a set of images " + }, + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "I_{i} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "type": "text", + "content": " cameras with known extrinsics " + }, + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "T_{i} \\in \\mathrm{SE}(3)" + }, + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "type": "text", + "content": " and intrinsics " + }, + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "K_{i} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "type": "text", + "content": ", MV3D object detection aims to infer the label (e.g., Car, Pedestrian, Barrier) and the 3D bounding box for each ob" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "ject instance in the scene. In this section, we first delve into the pipeline of query-based MV3D object detection in Sec. 3.1. We then analyze the pros and cons of the core 2D-to-3D feature lifting component in two state-of-the-art MV3D object detection methods in Sec. 3.2. Finally, we present our proposed MvACon in Sec. 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 153, + 505, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 153, + 505, + 167 + ], + "type": "text", + "content": "3.1. Query-based MV3D Object Detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 173, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 545, + 316 + ], + "type": "text", + "content": "For better understanding, we explain the query-based MV3D object detection pipeline, shown in Fig. 2, in a reverse manner. The 3D detection head typically builds upon DETR3D [57], which is based on the original DETR [4]. Initially, it defines a sufficiently large number, " + }, + { + "bbox": [ + 304, + 173, + 545, + 316 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 173, + 545, + 316 + ], + "type": "text", + "content": ", of latent " + }, + { + "bbox": [ + 304, + 173, + 545, + 316 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 173, + 545, + 316 + ], + "type": "text", + "content": "-dimensional 3D object queries, " + }, + { + "bbox": [ + 304, + 173, + 545, + 316 + ], + "type": "inline_equation", + "content": "Q_{O,C}" + }, + { + "bbox": [ + 304, + 173, + 545, + 316 + ], + "type": "text", + "content": ". These object queries update using Keys and Values derived from multi-view inputs, followed by a classification head predicting object labels and a bounding box regression head determining the 3D bounding boxes. Obviously, the key challenge lies in how to transform / lift multi-view 2D inputs into 3D-aware Keys and Values." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 317, + 546, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 546, + 460 + ], + "type": "text", + "content": "To this end, a feature backbone is trained to extract deep 2D features from the multi-view input images. The complexity arises from different design choices for feature lifting. It mainly involves two aspects: representation and computation. From the representational perspective, multiscale feature pyramids, crucial in 2D object detection, become even more essential in 3D object detection. Computationally, handling multi-view inputs is already demanding. Adding multi-scale feature pyramids without careful optimization can significantly increase the computational load. There are two strategies in the state-of-the-art development of query-based MV3D object detection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 461, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 546, + 653 + ], + "type": "text", + "content": "Decoder-Only Architectures: Single-Scale Multi-View 2D Features with Dense Attention. These designs are among the most straightforward. They involves using multi-view 2D feature maps from the last layer of the feature backbone, which are then concatenated and flattened along the spatial dimensions to form Keys and Values. The latent object queries, " + }, + { + "bbox": [ + 304, + 461, + 546, + 653 + ], + "type": "inline_equation", + "content": "Q_{O,C}" + }, + { + "bbox": [ + 304, + 461, + 546, + 653 + ], + "type": "text", + "content": ", are updated using a vanilla Transformer (i.e., each object query attends to every 2D location in the multi-view inputs, termed dense attention). However, this basic approach often fails as it does not encode any 3D-aware information. To address this, The PETR [29] introduces a physically-meaningful 3D position transformation as positional encoding, added to the multi-view 2D feature maps before concatenation and flattening. It has proven effective for query-based MV3D object detection, as illustrated in Option 1 in Fig. 2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Encoder-Decoder Architectures: Multi-Scale Multi-View 2D Features to Latent BEV Queries with Sparse Attention. The BEV (Bird's Eye View) representation acts as a unified, grid-based and ego-centric scene representation with predefined grid sizes (e.g., " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "200 \\times 200" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": ") on the XZ" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16690" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 68, + 539, + 212 + ], + "blocks": [ + { + "bbox": [ + 53, + 68, + 539, + 212 + ], + "lines": [ + { + "bbox": [ + 53, + 68, + 539, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 68, + 539, + 212 + ], + "type": "image", + "image_path": "f59ccbc492e5079cd54861e6c4fb9222a64fbfa244efb5d7d034a31258211189.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "text", + "content": "Figure 2. Overview of a query-based MV3D object detection pipeline with our proposed MvACon. Our proposed MvACon is a plug-and-play module for two state-of-the-art query-based MV3D object detection paradigms (e.g., PETR [29] and BEVFormer [25] respectively), which computes attentively contextualized features to facilitate better 2D-to-3D feature lifting in the two paradigms. See text for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 262, + 289, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 289, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 289, + 561 + ], + "type": "text", + "content": "plane. Geometrically, the BEV grid can be treated as a pillar-based point-cloud representation, collapsed along the Y-axis. A predefined number of points along the pillar (Y-axis) direction in the BEV grid will be uniformly sampled. These points form a uniform geometry prior for the underlying 3D scene and serve as BEV anchors to elevate 2D features into the BEV space. With known camera poses, these sampled points can be projected to each view at multiple scales. However, due to the uniform geometry prior in the projection, the projected points require deformability to better align with data observations. The BEVFormer [25] addresses this by introducing a sparse deformable attention mechanism (see Option 2 in Fig. 2). To counter the uniform geometry prior further, it learns a small predefined number of offsets, rather than directly deforming the projected points on each view, to lift 2D features from those deformed points with attentive weights. Latent BEV queries are introduced in learning these offsets and attentive weights. The BEV encoder's role is to refine the BEV queries, enabling them to provide meaningful offsets and attentive weights for lifting 2D features to 3D BEV anchors. Finally, embedded BEV queries as Keys/Values in the 3D detection head (i.e., decoder) update the latent object queries, " + }, + { + "bbox": [ + 47, + 262, + 289, + 561 + ], + "type": "inline_equation", + "content": "Q_{O,C}" + }, + { + "bbox": [ + 47, + 262, + 289, + 561 + ], + "type": "text", + "content": ", e.g., through sparse deformable attention as in the BEVFormer, before predicting the 3D object detection results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 572, + 287, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 287, + 597 + ], + "type": "text", + "content": "3.2. The Limitation of 2D-to-3D Feature Lifting in the Prior Art" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 604, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 287, + 676 + ], + "type": "text", + "content": "Although they both have shown remarkable progress for MV3D object detection, the PETR pipeline and the BEV-Former pipeline have a common limitation in their 2D-to-3D feature lifting, that is the local or shallow 3D awareness, rather than the desirable counterpart, the global and semantic meaningful 3D awareness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "In the PETR pipeline, consider the 2D feature map of the " + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "-th view, " + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "F_{h\\times w\\times c}^{n}" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": " are the spatial sizes, height and width respectively, and " + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": " the feature dimen" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "text", + "content": "sion of the backbone. The 3D position transform converts the (shared) camera frustum discretized as a mesh grid of sizes " + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "inline_equation", + "content": "(h,w,D)" + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "text", + "content": " to the 3D space based on the known camera poses, where " + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "text", + "content": " is the discretized depth levels. After the conversion, each 3D point is represented by a normalized 3D coordinate in the homogeneous form, i.e., " + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "inline_equation", + "content": "(x,y,z,1)" + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "text", + "content": ". So the 3D position transformation results in the positional encoding " + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "inline_equation", + "content": "P_{h\\times w\\times 4\\cdot D}^{n}" + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "text", + "content": ". Both " + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "inline_equation", + "content": "F_{h\\times w\\times c}^{n}" + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "inline_equation", + "content": "P_{h\\times w\\times 4\\cdot D}^{n}" + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "text", + "content": " are projected into the space of the same dimensionality, " + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "text", + "content": " using a linear layer and a Multi-layer Perceptron (MLP) with the ReLU nonlinearity respectively, we have " + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "inline_equation", + "content": "F_{h\\times w\\times d}^{n}" + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "inline_equation", + "content": "P_{h\\times w\\times d}^{n}" + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "text", + "content": " which are summed in an element-wise way. Consider the latent representation of the 3D position in " + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "inline_equation", + "content": "P_{h\\times w\\times d}^{n}" + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "text", + "content": " with the depth grid fused by the MLP, it is grounded to one feature point in " + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "inline_equation", + "content": "F_{h\\times w\\times d}^{n}" + }, + { + "bbox": [ + 304, + 262, + 545, + 441 + ], + "type": "text", + "content": ", leading to the local 3D awareness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 441, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 502 + ], + "type": "text", + "content": "In the BEVFormer pipeline, as we show in Fig. 1, the projected BEV anchor is often grounded to some low-confidence and scattered deformed points. Although the grounding may not be spatially local, they are often semantically shallow." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 511, + 479, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 479, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 479, + 525 + ], + "type": "text", + "content": "3.3. Our Proposed MvACon Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 530, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 602 + ], + "type": "text", + "content": "Our goal is to address the local or shallow 3D awareness stated above by introducing an easy-to-integrate (mostly plug-and-play) module to learn the global and semantically meaningful 3D-awareness, as illustrated in Fig. 2. The basic idea of our MvACon is to attentively contextualize the 2D features in the 2D-to-3D lifting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 603, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 545, + 628 + ], + "type": "text", + "content": "In the PETR pipeline, our idea is to contextualize the individual 2D feature map, " + }, + { + "bbox": [ + 304, + 603, + 545, + 628 + ], + "type": "inline_equation", + "content": "F_{h \\times w \\times d}^{n}" + }, + { + "bbox": [ + 304, + 603, + 545, + 628 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 359, + 628, + 545, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 628, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 359, + 628, + 545, + 641 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {h \\times w \\times d} ^ {n} = \\operatorname {M v A C o n} \\left(F _ {h \\times w \\times d} ^ {n}\\right), \\tag {1}", + "image_path": "4563554e965c3e6c39aa81c56a5079bedfc366762619456a1d7d7e198217ebe7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": "where after the contextualization every feature point in " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{h\\times w\\times d}^n" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": " can connect to the entire map " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "F_{h\\times w\\times d}^n" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": ", inducing the global 3D awareness for grounding the positional encoding " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "P_{h\\times w\\times d}^n" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "In the BEVFormer pipeline, our idea is to contextualize the multi-scale feature maps, e.g., the " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "-th layer of the fea" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16691" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 208, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 208, + 86 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 208, + 86 + ], + "type": "text", + "content": "ture pyramid of the " + }, + { + "bbox": [ + 47, + 71, + 208, + 86 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 71, + 208, + 86 + ], + "type": "text", + "content": " -th view, " + }, + { + "bbox": [ + 47, + 71, + 208, + 86 + ], + "type": "inline_equation", + "content": "F_{h\\times w\\times c}^{n,l}" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 86, + 286, + 101 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 86, + 286, + 101 + ], + "spans": [ + { + "bbox": [ + 89, + 86, + 286, + 101 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {h \\times w \\times c} ^ {n, l} = \\operatorname {M v A C o n} \\left(\\left\\{F _ {h \\times w \\times c} ^ {n, l} \\right\\} _ {l = 1} ^ {L}\\right), \\tag {2}", + "image_path": "45e924f170143c0932022c257d9b3511fa91047284ec8acc63d57c88c884eeb0.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 102, + 286, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 286, + 151 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 286, + 151 + ], + "type": "text", + "content": "where after the contextualization every feature point in " + }, + { + "bbox": [ + 47, + 102, + 286, + 151 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{h\\times w\\times c}^{n,l}" + }, + { + "bbox": [ + 47, + 102, + 286, + 151 + ], + "type": "text", + "content": " can connect to the entire " + }, + { + "bbox": [ + 47, + 102, + 286, + 151 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 102, + 286, + 151 + ], + "type": "text", + "content": "-layer feature pyramid " + }, + { + "bbox": [ + 47, + 102, + 286, + 151 + ], + "type": "inline_equation", + "content": "\\{F_{h\\times w\\times c}^{n,l}\\}_{l = 1}^{L}" + }, + { + "bbox": [ + 47, + 102, + 286, + 151 + ], + "type": "text", + "content": ", inducing the global 3D awareness for grounding projected BEV anchors on the " + }, + { + "bbox": [ + 47, + 102, + 286, + 151 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 102, + 286, + 151 + ], + "type": "text", + "content": "-the view." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 152, + 286, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 286, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 286, + 235 + ], + "type": "text", + "content": "To achieve the global contextualization effect, we adapt the recently proposed Patch-to-Cluster attention (PaCa) [12] method. The core idea of PaCa is to leverage a learnable clustering module to cluster a feature map into a predefined number " + }, + { + "bbox": [ + 47, + 152, + 286, + 235 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 152, + 286, + 235 + ], + "type": "text", + "content": " of clusters. For notional simplicity, consider a feature map " + }, + { + "bbox": [ + 47, + 152, + 286, + 235 + ], + "type": "inline_equation", + "content": "F_{h\\times w\\times c}" + }, + { + "bbox": [ + 47, + 152, + 286, + 235 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 152, + 286, + 235 + ], + "type": "inline_equation", + "content": "N = h\\times w" + }, + { + "bbox": [ + 47, + 152, + 286, + 235 + ], + "type": "text", + "content": " tokens " + }, + { + "bbox": [ + 47, + 152, + 286, + 235 + ], + "type": "inline_equation", + "content": "F_{N\\times c}" + }, + { + "bbox": [ + 47, + 152, + 286, + 235 + ], + "type": "text", + "content": ", the clustering assignment is computed by," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 90, + 236, + 286, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 236, + 286, + 249 + ], + "spans": [ + { + "bbox": [ + 90, + 236, + 286, + 249 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {N, M} = \\operatorname {S o f t m a x} (\\operatorname {C l u s t e r i n g} (F _ {N, c})), \\tag {3}", + "image_path": "fa756e13fe658a9c57e76898e4aaf1604942ff3a6e87f686ab37f5036073d190.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 250, + 286, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 250, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 250, + 286, + 285 + ], + "type": "text", + "content": "where Clustering() can be implemented in different ways (see our ablation studies in Tab. 6), and the Softmax is along the token dimension. Then, we compute " + }, + { + "bbox": [ + 47, + 250, + 286, + 285 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 250, + 286, + 285 + ], + "type": "text", + "content": " clusters by," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 286, + 286, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 286, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 112, + 286, + 286, + 299 + ], + "type": "interline_equation", + "content": "z _ {M, c} = \\mathrm {L N} \\left(\\mathcal {C} _ {N, M} ^ {\\top} \\cdot F _ {N, c}\\right), \\tag {4}", + "image_path": "37de5cc426f08cc4a611ba413e8b2beb70f64f17f98c6db966bcb12deb22e0ff.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 299, + 218, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 218, + 311 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 218, + 311 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 299, + 218, + 311 + ], + "type": "inline_equation", + "content": "\\mathrm{LN}()" + }, + { + "bbox": [ + 47, + 299, + 218, + 311 + ], + "type": "text", + "content": " is the layer normalization [1]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 312, + 242, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 312, + 242, + 323 + ], + "spans": [ + { + "bbox": [ + 59, + 312, + 242, + 323 + ], + "type": "text", + "content": "Then, the PaCa-based MvAcon is defined by," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 323, + 286, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 323, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 70, + 323, + 286, + 350 + ], + "type": "interline_equation", + "content": "F _ {N, c} ^ {\\prime} = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {N , c} \\cdot K _ {M , c} ^ {\\top}}{\\sqrt {c}}\\right) \\cdot V _ {M, c} + F _ {N, c}, \\tag {5}", + "image_path": "d843f71719948cfa0377cf1a4ece272d24a27bc675ae1e653a94cfa0b23b1b8a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 351, + 286, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 351, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 286, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 351, + 286, + 422 + ], + "type": "inline_equation", + "content": "Q_{N,c}" + }, + { + "bbox": [ + 47, + 351, + 286, + 422 + ], + "type": "text", + "content": " is the linear projection of " + }, + { + "bbox": [ + 47, + 351, + 286, + 422 + ], + "type": "inline_equation", + "content": "F_{N,c}, K_{M,c}" + }, + { + "bbox": [ + 47, + 351, + 286, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 351, + 286, + 422 + ], + "type": "inline_equation", + "content": "V_{M,c}" + }, + { + "bbox": [ + 47, + 351, + 286, + 422 + ], + "type": "text", + "content": " are from the clusters " + }, + { + "bbox": [ + 47, + 351, + 286, + 422 + ], + "type": "inline_equation", + "content": "z_{M,c}" + }, + { + "bbox": [ + 47, + 351, + 286, + 422 + ], + "type": "text", + "content": ". The second term is the shortcut. The multi-head PaCa can be straightforwardly defined. For Eqn. 2, we concatenate the clusters from all the pyramid layers before computing the Key and the Value. Here, the PaCa module is of linear complexity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 434, + 127, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 127, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 127, + 448 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 454, + 162, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 162, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 162, + 467 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 473, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 473, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 47, + 473, + 286, + 688 + ], + "type": "text", + "content": "Dataset and Metrics We evaluate our MvAcon on the challenging large-scale NuScenes dataset [3] and Waymo dataset [11]. The NuScenes dataset includes 1,000 scene sequences, which are divided into training, validation, and testing subsets in a 700/150/150 split. Each sequence in the NuScenes dataset is a 20-second video clip, annotated at a rate of 2 frames per second (FPS). The NuScenes dataset employs a comprehensive suite of evaluation metrics for assessing detection performance. These metrics comprise mean Average Precision (mAP), mean Average Translation Error (mATE), mean Average Scale Error (mASE), mean Average Orientation Error (mAOE), mean Average Velocity Error (mAVE), mean Average Attribution Error (mAAE), and the NuScenes Detection Score (NDS). The Waymo dataset contains 798 training sequences and 202 validation sequences. We use a subset of the training set (Waymomini) by sampling every third frame from the training sequences following [25]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Implementation Details We leverage open-source code bases (PETR [29], BEVFormer [25], and DFA3D [19]) in" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "our experiments. To ensure a fair and stringent comparison, we maintain all original configurations of these methods, making only one modification: the addition of an attentive contextualization module. We conduct qualitative analysis and ablation study on the BEVFormer-base model. We train all models for 24 epochs using 8 NVIDIA Tesla A100 GPUs, following the configurations and settings outlined in previous works [19, 25, 29]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 175, + 545, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 175, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 545, + 198 + ], + "type": "text", + "content": "4.2. The Effectiveness of MvACon across Different Methods" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 205, + 545, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 313 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our proposed MvA-Con method, we first apply it to two typical query-based MV3D object detection paradigms: the perspective-based decoder-only detector (PETR [29]) and the encoder-decoder based detector (BEVFormer [25]). We choose these as our baselines because state-of-the-art (SOTA) query-based MV3D detectors [19, 30, 52, 60] primarily follow these two paradigms. We also test our method on DFA3D [19] to demonstrate its generalizability to SOTA methods." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 314, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 480 + ], + "type": "text", + "content": "On the NuScenes dataset, Table 1 shows that our proposed MvACon consistently improves performance across different detectors. Specifically, for the perspective-based decoder-only detector PETR, it improves the baseline by 0.8 NDS. For the encoder-decoder based detector BEVFormer, our method achieves an improvement of 1.3 in NDS on average. On a more advanced, depth-context enhanced BEVFormer (DFA3D), our method further improves performance by up to 0.5 NDS. Notably, our MvACon achieves significant improvement in location (mAP, mATE), orientation (mAOE), and velocity prediction (mAVE) for encoder-decoder based detectors. It also markedly enhances performance in location (mAP, mATE) and velocity (mAVE) prediction for the perspective-based decoder-only detector." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 481, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 540 + ], + "type": "text", + "content": "On the Waymo dataset, since there are few released codes for MV3D detectors on Waymo except for the BEVFormer, we only test BEVFormer on Waymo-mini following its settings with results shown in Table 2. Our MvACon shows consistent improvement on Waymo metrics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 548, + 477, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 548, + 477, + 559 + ], + "spans": [ + { + "bbox": [ + 305, + 548, + 477, + 559 + ], + "type": "text", + "content": "4.3. How Does Our MvACon Work?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 567, + 545, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 698 + ], + "type": "text", + "content": "We elaborate on the effects of our MvACon by providing detailed analyses during the 2D-to-3D feature lifting process. We first demonstrate what the learned cluster contexts encode, then show how these contexts affect the behavior of deformable points during feature lifting. Lastly, we illustrate how our MvACon improves detection results by presenting a qualitative comparison within a scene. We select BEVFormer-base as our analysis target due to its incorporation of six layers of deformable attention modules in the encoder. More qualitative analysis is provided in the supplementary materials." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "type": "text", + "content": "What do the Learned Cluster Contexts Encode? We vi-" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16692" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 65, + 482, + 218 + ], + "blocks": [ + { + "bbox": [ + 111, + 65, + 482, + 218 + ], + "lines": [ + { + "bbox": [ + 111, + 65, + 482, + 218 + ], + "spans": [ + { + "bbox": [ + 111, + 65, + 482, + 218 + ], + "type": "table", + "html": "
MethodmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
PETR-VovNet-99 [29]37.874.627.248.890.621.242.6
PETR-VovNet-99-MvACon38.2 (+0.5)73.927.050.584.121.343.4 (+0.8)
BEVFormer-t [25]25.290.029.465.565.721.635.4
BEVFormer-t-MvACon25.9 (+0.7)88.428.864.660.522.536.5 (+1.1)
BEVFormer-s [25]37.072.128.040.743.622.047.9
BEVFormer-s-MvACon39.3 (+2.3)71.327.740.142.019.749.6 (+1.7)
BEVFormer-b [25]41.667.327.437.239.419.851.7
BEVFormer-b-MvACon42.6 (+1.0)66.427.635.036.220.052.8 (+1.1)
DFA3D-s [19]40.172.127.941.139.119.650.1
DFA3D-s-MvACon40.171.027.438.337.220.850.6 (+0.5)
DFA3D-b [19]43.065.427.137.434.120.553.1
DFA3D-b-MvACon43.2 (+0.2)66.427.534.432.320.753.5 (+0.4)
", + "image_path": "d9b0f9ec2c20605534ddef99ffe26296ef9d22350491fc62eb85643b4e7e12a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 88, + 260, + 192, + 320 + ], + "blocks": [ + { + "bbox": [ + 88, + 260, + 192, + 320 + ], + "lines": [ + { + "bbox": [ + 88, + 260, + 192, + 320 + ], + "spans": [ + { + "bbox": [ + 88, + 260, + 192, + 320 + ], + "type": "image", + "image_path": "c0aa0f973601d562baf867b1786bce2e4d9992d5b22c24c6d19d5b0565ee4474.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 88, + 321, + 192, + 381 + ], + "blocks": [ + { + "bbox": [ + 88, + 321, + 192, + 381 + ], + "lines": [ + { + "bbox": [ + 88, + 321, + 192, + 381 + ], + "spans": [ + { + "bbox": [ + 88, + 321, + 192, + 381 + ], + "type": "image", + "image_path": "3e9ee985e76f829fcd96abfa656ad52caef07299387e4e36ea72cfec2d6cdf49.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 383, + 546, + 417 + ], + "lines": [ + { + "bbox": [ + 46, + 383, + 546, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 546, + 417 + ], + "type": "text", + "content": "Figure 3. Visualization results of learned cluster contexts in our MvACon on the NuScenes validation set. We sum all the learned clusters along the channel and upsample it to the original image resolution through bilinear interpolation. We observed that the learned cluster context encodes abundant context information in the scene. We provide details with raw images in the supplementary." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 193, + 260, + 294, + 320 + ], + "blocks": [ + { + "bbox": [ + 193, + 260, + 294, + 320 + ], + "lines": [ + { + "bbox": [ + 193, + 260, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 193, + 260, + 294, + 320 + ], + "type": "image", + "image_path": "f24cf33f389162a391e4b02ad7503f079bc6ec580145ec613ecaec4be3c7096c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 193, + 321, + 294, + 381 + ], + "blocks": [ + { + "bbox": [ + 193, + 321, + 294, + 381 + ], + "lines": [ + { + "bbox": [ + 193, + 321, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 193, + 321, + 294, + 381 + ], + "type": "image", + "image_path": "7d78c0edf5d69db9010c2f652450b12b3432ef79dbdd716dfdc75b32c3f18498.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 295, + 260, + 398, + 320 + ], + "blocks": [ + { + "bbox": [ + 295, + 260, + 398, + 320 + ], + "lines": [ + { + "bbox": [ + 295, + 260, + 398, + 320 + ], + "spans": [ + { + "bbox": [ + 295, + 260, + 398, + 320 + ], + "type": "image", + "image_path": "9b0b78e6189103ed8c556c121f5244ecc9bbc0364c5c9eeefb5ea707b212bd27.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 295, + 321, + 398, + 381 + ], + "blocks": [ + { + "bbox": [ + 295, + 321, + 398, + 381 + ], + "lines": [ + { + "bbox": [ + 295, + 321, + 398, + 381 + ], + "spans": [ + { + "bbox": [ + 295, + 321, + 398, + 381 + ], + "type": "image", + "image_path": "0d8d9fc9e6f66c02159cd0eb1ac1fb0333032de245e690a71bb8cdcd17832197.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 399, + 260, + 501, + 320 + ], + "blocks": [ + { + "bbox": [ + 399, + 260, + 501, + 320 + ], + "lines": [ + { + "bbox": [ + 399, + 260, + 501, + 320 + ], + "spans": [ + { + "bbox": [ + 399, + 260, + 501, + 320 + ], + "type": "image", + "image_path": "a56d1c6240209c0b7350f78a1ed48bf598f448dc478a2108e4a56ae991f00ccf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 399, + 321, + 501, + 381 + ], + "blocks": [ + { + "bbox": [ + 399, + 321, + 501, + 381 + ], + "lines": [ + { + "bbox": [ + 399, + 321, + 501, + 381 + ], + "spans": [ + { + "bbox": [ + 399, + 321, + 501, + 381 + ], + "type": "image", + "image_path": "3c589eba87906b82b521e299d80e8cdf1990b8745ee5df20ed56809d052b1352.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 61, + 425, + 273, + 465 + ], + "blocks": [ + { + "bbox": [ + 46, + 222, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 222, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 547, + 257 + ], + "type": "text", + "content": "Table 1. Comparisons of our method with baslines on the NuScenes validation set. BEVFormer-t/s/b refers to BEVFormer-tiny/small/base in the BEVFormer's open-source codes. 'DFA3D' refers to the adaptation of 2D deformable attention into a (depth-weighted) 3D deformable attention within the BEVFormer model, as adopted in [19]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 425, + 273, + 465 + ], + "lines": [ + { + "bbox": [ + 61, + 425, + 273, + 465 + ], + "spans": [ + { + "bbox": [ + 61, + 425, + 273, + 465 + ], + "type": "table", + "html": "
MethodLET-mAPL↑LET-mAPH↑
BEVFormer-ResNet101 [25]34.946.3
BEVFormer-ResNet101-MvACon35.7 (+0.8)47.5 (+1.2)
", + "image_path": "3144b4b6226437684ebd86a546cfc05b6c13aacb33bd6a40331035193a56b2d1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 88, + 468, + 246, + 480 + ], + "lines": [ + { + "bbox": [ + 88, + 468, + 246, + 480 + ], + "spans": [ + { + "bbox": [ + 88, + 468, + 246, + 480 + ], + "type": "text", + "content": "Table 2. Comparisons on the Waymo-mini." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 493, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 287, + 649 + ], + "type": "text", + "content": "sualize the learned cluster context in a heatmap, by summing all clusters along the channel and then upsampling to the original image resolution using bilinear interpolation. The results, shown in Fig. 3, reveal that despite the low resolution of the original summed heatmap, it can accurately locate foreground objects in the scene after upsampling. This suggests that the cluster contexts encode the foreground layouts in a scene. Furthermore, we note that the upsampled heatmap shows a high response to various foreground objects, even when they are spatially close in 2D. This is attributed to the effectiveness of our MvACon in dense scenes, as demonstrated and analyzed later in our experiments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "How do Cluster Contexts Affect the Behavior of Deformable Points? We demonstrate the dynamics of deformable points in Fig. 4. These points originate from one 2D reference point, which is projected from a 3D BEV anchor point in the BEVFormer encoder. As the encoding" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 427, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 606 + ], + "type": "text", + "content": "layers deepen in the vanilla BEVFormer, we observe that most deformable points have low attention weights, suggesting the model's uncertainty about the relevance of the selected context. Consequently, these contexts contribute minimally during the encoding of BEV query features in the feature lifting process. In contrast, deformable points predicted by our method maintain high confidence weights on foreground objects (e.g., cars), as well as on surrounding buildings. We also note that points on buildings remain stable across encoding layers and consecutive frames (see the supplementary). These high-confidence deformed points in our MvACon may foster a local object-context aware coordinate system, enhancing overall performance, including the estimation of velocity and orientation. This observation aligns with our quantitative findings in Table 1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "type": "text", + "content": "Qualitative Comparisons with the Baseline Method. We provide qualitative comparison with BEVFormer in Fig. 5. Our method exhibits superior performance in dense scenarios where objects are crowded and challenging to localize. We attribute this enhanced performance to the rich foreground layout context in the scene, facilitated by our attentive contextualization module. Additional qualitative comparisons are available in the supplementary." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16693" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 76, + 234, + 151 + ], + "blocks": [ + { + "bbox": [ + 86, + 76, + 234, + 151 + ], + "lines": [ + { + "bbox": [ + 86, + 76, + 234, + 151 + ], + "spans": [ + { + "bbox": [ + 86, + 76, + 234, + 151 + ], + "type": "image", + "image_path": "ea8e463d6a162de83c5c59461ee9eff5ca7f6088b46876653e7c8d36107398dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 235, + 76, + 366, + 151 + ], + "blocks": [ + { + "bbox": [ + 235, + 76, + 366, + 151 + ], + "lines": [ + { + "bbox": [ + 235, + 76, + 366, + 151 + ], + "spans": [ + { + "bbox": [ + 235, + 76, + 366, + 151 + ], + "type": "image", + "image_path": "d5ab9c2bd43a57241e3d3282aa56f8c992107df2f89f37056b8180bd12ba3ad4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 366, + 76, + 496, + 151 + ], + "blocks": [ + { + "bbox": [ + 366, + 76, + 496, + 151 + ], + "lines": [ + { + "bbox": [ + 366, + 76, + 496, + 151 + ], + "spans": [ + { + "bbox": [ + 366, + 76, + 496, + 151 + ], + "type": "image", + "image_path": "3f0a40011a876d1f04a9438d4072a5a65c86ad481ced787ba7be0c1124015fb7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 88, + 152, + 234, + 226 + ], + "blocks": [ + { + "bbox": [ + 88, + 152, + 234, + 226 + ], + "lines": [ + { + "bbox": [ + 88, + 152, + 234, + 226 + ], + "spans": [ + { + "bbox": [ + 88, + 152, + 234, + 226 + ], + "type": "image", + "image_path": "82a96e50fa205cc7b1d84d5e613e3453a601b9a90c04d56cbbd711cda24fcb9b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 235, + 152, + 366, + 226 + ], + "blocks": [ + { + "bbox": [ + 235, + 152, + 366, + 226 + ], + "lines": [ + { + "bbox": [ + 235, + 152, + 366, + 226 + ], + "spans": [ + { + "bbox": [ + 235, + 152, + 366, + 226 + ], + "type": "image", + "image_path": "70ad81e688f2c4c764f5e312ef38fba4f1403e3d956f0da28772c7c3abc462ab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 366, + 152, + 496, + 226 + ], + "blocks": [ + { + "bbox": [ + 366, + 152, + 496, + 226 + ], + "lines": [ + { + "bbox": [ + 366, + 152, + 496, + 226 + ], + "spans": [ + { + "bbox": [ + 366, + 152, + 496, + 226 + ], + "type": "image", + "image_path": "ec8d0c83009dcc5348cf026119890974632859e4c47f40946d9dddfac979f401.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 88, + 228, + 234, + 306 + ], + "blocks": [ + { + "bbox": [ + 88, + 228, + 234, + 306 + ], + "lines": [ + { + "bbox": [ + 88, + 228, + 234, + 306 + ], + "spans": [ + { + "bbox": [ + 88, + 228, + 234, + 306 + ], + "type": "image", + "image_path": "a0803f15021e679c4855e4156aae7e86d2f7a06441cad50fd4ff5a9a290c56d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 235, + 230, + 366, + 306 + ], + "blocks": [ + { + "bbox": [ + 235, + 230, + 366, + 306 + ], + "lines": [ + { + "bbox": [ + 235, + 230, + 366, + 306 + ], + "spans": [ + { + "bbox": [ + 235, + 230, + 366, + 306 + ], + "type": "image", + "image_path": "1aa1fb988b350f72349578cfed2dbb8cbcdb03432482af17434e64ff4e79afd0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 366, + 231, + 496, + 306 + ], + "blocks": [ + { + "bbox": [ + 366, + 231, + 496, + 306 + ], + "lines": [ + { + "bbox": [ + 366, + 231, + 496, + 306 + ], + "spans": [ + { + "bbox": [ + 366, + 231, + 496, + 306 + ], + "type": "image", + "image_path": "d307767d3e5d58525008b9b76b67cd6d38f68cb22260f6ee1723f5281a699a24.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 88, + 307, + 234, + 381 + ], + "blocks": [ + { + "bbox": [ + 88, + 307, + 234, + 381 + ], + "lines": [ + { + "bbox": [ + 88, + 307, + 234, + 381 + ], + "spans": [ + { + "bbox": [ + 88, + 307, + 234, + 381 + ], + "type": "image", + "image_path": "246726bcd7652b194da8259cba91b0c8080c7eed480c107abda1c9d0929374c5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 387, + 545, + 419 + ], + "lines": [ + { + "bbox": [ + 46, + 387, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 545, + 419 + ], + "type": "text", + "content": "Figure 4. Visualization results of the deformable points originating from a 2D reference point, which is projected from a 3D BEV anchor point in the BEVFormer encoder, on NuScenes validation set. We utilize the same BEV anchor point as demonstrated in Fig. 1. From left to right and up to bottom, we display the deformable points output from each layer (#1-#6) in the encoder, respectively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 235, + 307, + 366, + 381 + ], + "blocks": [ + { + "bbox": [ + 235, + 307, + 366, + 381 + ], + "lines": [ + { + "bbox": [ + 235, + 307, + 366, + 381 + ], + "spans": [ + { + "bbox": [ + 235, + 307, + 366, + 381 + ], + "type": "image", + "image_path": "b2bfd2901d36a70ac5abb4bc8e03529ea044f310ba5d0d04cb95e4652df426af.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 366, + 307, + 496, + 381 + ], + "blocks": [ + { + "bbox": [ + 366, + 307, + 496, + 381 + ], + "lines": [ + { + "bbox": [ + 366, + 307, + 496, + 381 + ], + "spans": [ + { + "bbox": [ + 366, + 307, + 496, + 381 + ], + "type": "image", + "image_path": "30201d8134e19dc120cc42b9215218fffe89654e81f79cdae7290c978908fc9d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 67, + 424, + 208, + 553 + ], + "blocks": [ + { + "bbox": [ + 67, + 424, + 208, + 553 + ], + "lines": [ + { + "bbox": [ + 67, + 424, + 208, + 553 + ], + "spans": [ + { + "bbox": [ + 67, + 424, + 208, + 553 + ], + "type": "image", + "image_path": "bc9568dc4ce23aa7eaf48d326f7f561c6f32b107e432a17ea178486f864f0cc4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 219, + 424, + 312, + 487 + ], + "blocks": [ + { + "bbox": [ + 219, + 424, + 312, + 487 + ], + "lines": [ + { + "bbox": [ + 219, + 424, + 312, + 487 + ], + "spans": [ + { + "bbox": [ + 219, + 424, + 312, + 487 + ], + "type": "image", + "image_path": "9c03d03ac0013399ac7ba2bcc9ed0dd23491a7bc8945cfbcaba171bea6f82447.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 320, + 426, + 413, + 486 + ], + "blocks": [ + { + "bbox": [ + 320, + 426, + 413, + 486 + ], + "lines": [ + { + "bbox": [ + 320, + 426, + 413, + 486 + ], + "spans": [ + { + "bbox": [ + 320, + 426, + 413, + 486 + ], + "type": "image", + "image_path": "18f5e56b228531861d0e143ce986f9ee3918cc0a9d7bb550387cc57bcb114820.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 421, + 426, + 515, + 486 + ], + "blocks": [ + { + "bbox": [ + 421, + 426, + 515, + 486 + ], + "lines": [ + { + "bbox": [ + 421, + 426, + 515, + 486 + ], + "spans": [ + { + "bbox": [ + 421, + 426, + 515, + 486 + ], + "type": "image", + "image_path": "210aa166ae1c98c05cd71a15d95ed115479a3fb470a8b3b9ba929b58e8561c02.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 219, + 488, + 312, + 548 + ], + "blocks": [ + { + "bbox": [ + 219, + 488, + 312, + 548 + ], + "lines": [ + { + "bbox": [ + 219, + 488, + 312, + 548 + ], + "spans": [ + { + "bbox": [ + 219, + 488, + 312, + 548 + ], + "type": "image", + "image_path": "3d5bccb46a4b166489d57ecfda6f703addcb729d02a0c0f09ee192ada2acc644.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 320, + 488, + 413, + 548 + ], + "blocks": [ + { + "bbox": [ + 320, + 488, + 413, + 548 + ], + "lines": [ + { + "bbox": [ + 320, + 488, + 413, + 548 + ], + "spans": [ + { + "bbox": [ + 320, + 488, + 413, + 548 + ], + "type": "image", + "image_path": "f62eb02409fedcab9e8b5d61e6a5fea1b6f58b59722a222453c9914a11e04e41.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 421, + 488, + 515, + 548 + ], + "blocks": [ + { + "bbox": [ + 421, + 488, + 515, + 548 + ], + "lines": [ + { + "bbox": [ + 421, + 488, + 515, + 548 + ], + "spans": [ + { + "bbox": [ + 421, + 488, + 515, + 548 + ], + "type": "image", + "image_path": "dd64fd307686675c626386891f762e3913f88f24ea5d84a2d1a7c1ad797abc42.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 67, + 554, + 207, + 682 + ], + "blocks": [ + { + "bbox": [ + 67, + 554, + 207, + 682 + ], + "lines": [ + { + "bbox": [ + 67, + 554, + 207, + 682 + ], + "spans": [ + { + "bbox": [ + 67, + 554, + 207, + 682 + ], + "type": "image", + "image_path": "0f13c5f3c184899e1ed6ed39fcba720db8dc763f03494f4ec1a80c4e89429ade.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 687, + 496, + 698 + ], + "lines": [ + { + "bbox": [ + 96, + 687, + 496, + 698 + ], + "spans": [ + { + "bbox": [ + 96, + 687, + 496, + 698 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons between BEVFormer and our MvACon method on NuScenes validation set." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 219, + 554, + 312, + 620 + ], + "blocks": [ + { + "bbox": [ + 219, + 554, + 312, + 620 + ], + "lines": [ + { + "bbox": [ + 219, + 554, + 312, + 620 + ], + "spans": [ + { + "bbox": [ + 219, + 554, + 312, + 620 + ], + "type": "image", + "image_path": "0fd7e1520e178ad8c353679dbaf3e11c555eb934c84a38344a6bdb2626d7b9f4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 320, + 554, + 413, + 620 + ], + "blocks": [ + { + "bbox": [ + 320, + 554, + 413, + 620 + ], + "lines": [ + { + "bbox": [ + 320, + 554, + 413, + 620 + ], + "spans": [ + { + "bbox": [ + 320, + 554, + 413, + 620 + ], + "type": "image", + "image_path": "1470954d2e27c7807a52607f89565c582b53f982c228ca7519f20ad619f11c14.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 421, + 554, + 515, + 620 + ], + "blocks": [ + { + "bbox": [ + 421, + 554, + 515, + 620 + ], + "lines": [ + { + "bbox": [ + 421, + 554, + 515, + 620 + ], + "spans": [ + { + "bbox": [ + 421, + 554, + 515, + 620 + ], + "type": "image", + "image_path": "a6b303c60dbe6837e90ea6ab9c65cca735e7f05470de43c2148e6f8577df7ea2.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 219, + 621, + 312, + 682 + ], + "blocks": [ + { + "bbox": [ + 219, + 621, + 312, + 682 + ], + "lines": [ + { + "bbox": [ + 219, + 621, + 312, + 682 + ], + "spans": [ + { + "bbox": [ + 219, + 621, + 312, + 682 + ], + "type": "image", + "image_path": "7c360e5184c406e4b73793fee739843b2ca95c17dfd4869ec08aa43ec27fb51e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 320, + 621, + 413, + 682 + ], + "blocks": [ + { + "bbox": [ + 320, + 621, + 413, + 682 + ], + "lines": [ + { + "bbox": [ + 320, + 621, + 413, + 682 + ], + "spans": [ + { + "bbox": [ + 320, + 621, + 413, + 682 + ], + "type": "image", + "image_path": "380357fd83a567b48d3def6b532c979c48039e1ae1aed50dce8cb86af7e32276.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 421, + 621, + 515, + 682 + ], + "blocks": [ + { + "bbox": [ + 421, + 621, + 515, + 682 + ], + "lines": [ + { + "bbox": [ + 421, + 621, + 515, + 682 + ], + "spans": [ + { + "bbox": [ + 421, + 621, + 515, + 682 + ], + "type": "image", + "image_path": "94faaa01fa51e26324087c0d894d46685166bde06bf62b964f0b44bd222a57d3.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16694" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 147, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 147, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 147, + 83 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 89, + 286, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 89, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 89, + 286, + 293 + ], + "type": "text", + "content": "Effectiveness of Different Contextualization Methods. Table 3 demonstrates the impact of various contextual methods on detection performance. We selected three representative contexts: local shift window-based context (SWin [32]), global pooling-based context (PVTV2-linear [55]), and global cluster-based context (PaCa [12]). The results show that the local shift window-based context offers minimal improvements in detection performance, which can be attributed to similar local contexts already provided by convolutional backbone networks. Conversely, enhancing the original feature with global contexts, as observed in PVTV2-linear and PaCa experiments, leads to better performance. Notably, orientation and velocity show significant improvements with these global contexts. The inclusion of cluster contexts further enhances improvements in location, orientation, and velocity prediction, as evidenced in the global cluster experiment." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 51, + 298, + 282, + 335 + ], + "blocks": [ + { + "bbox": [ + 51, + 298, + 282, + 335 + ], + "lines": [ + { + "bbox": [ + 51, + 298, + 282, + 335 + ], + "spans": [ + { + "bbox": [ + 51, + 298, + 282, + 335 + ], + "type": "table", + "html": "
MethodmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
BEVFormer-b [25]41.667.327.437.239.419.851.7
Shift Window (SWin [32])41.5 (-0.1)67.127.837.839.620.151.5 (-0.2)
Global Pooling (PVTV2-linear [55])41.6 (+0.0)66.527.536.837.619.652.0 (+0.3)
Global Cluster (PaCa [12])42.6 (+1.0)66.427.635.036.220.052.8 (+1.1)
", + "image_path": "c592b04b2b45ceb3e12ae95a32f75d48e1ad851e9e19018751ca156914d2002a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 51, + 362, + 282, + 396 + ], + "blocks": [ + { + "bbox": [ + 52, + 338, + 281, + 350 + ], + "lines": [ + { + "bbox": [ + 52, + 338, + 281, + 350 + ], + "spans": [ + { + "bbox": [ + 52, + 338, + 281, + 350 + ], + "type": "text", + "content": "Table 3. Ablation study on different contextualization methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 362, + 282, + 396 + ], + "lines": [ + { + "bbox": [ + 51, + 362, + 282, + 396 + ], + "spans": [ + { + "bbox": [ + 51, + 362, + 282, + 396 + ], + "type": "table", + "html": "
ContextmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
Local (BEVFormer-b [25])41.667.327.437.239.419.851.7
Global Cluster41.666.427.438.335.119.552.1
Local + Global Cluster42.6 (+1.0)66.427.635.036.220.052.8 (+1.1)
", + "image_path": "093ff578e6c5cf801e58b2ab3513c30112b7b8f394f3d8f96e5a6a5236cc9d62.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 51, + 434, + 282, + 473 + ], + "blocks": [ + { + "bbox": [ + 47, + 399, + 286, + 422 + ], + "lines": [ + { + "bbox": [ + 47, + 399, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 286, + 422 + ], + "type": "text", + "content": "Table 4. Ablation study on the relationship between local contexts and global cluster-based contexts." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 434, + 282, + 473 + ], + "lines": [ + { + "bbox": [ + 51, + 434, + 282, + 473 + ], + "spans": [ + { + "bbox": [ + 51, + 434, + 282, + 473 + ], + "type": "table", + "html": "
Method#layers#clustersCross-level\\( \\mathbf{{mAP}} \\dagger \\)\\( \\mathbf{{mATE}} \\downarrow \\)\\( \\mathbf{{mASE}} \\downarrow \\)\\( \\mathbf{{mAOE}} \\downarrow \\)\\( \\mathbf{{mAVE}} \\downarrow \\)\\( \\mathbf{{mAAE}} \\downarrow \\)\\( \\mathbf{{NDS}} \\uparrow \\)
BEVFormer-b [25]---41.667.327.437.239.419.851.7
Baseline 13100✓41.165.627.137.736.418.952.0
Baseline 2650✓41.366.327.336.134.118.852.4
Baseline 36100-42.466.827.637.136.319.152.5
MvACon6100✓42.666.427.635.036.220.052.8
", + "image_path": "5121107333bb6260a7d653d6a7b4f5096ef17106f5f17686a93d44c716ddfd0e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 51, + 511, + 282, + 557 + ], + "blocks": [ + { + "bbox": [ + 47, + 477, + 286, + 498 + ], + "lines": [ + { + "bbox": [ + 47, + 477, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 286, + 498 + ], + "type": "text", + "content": "Table 5. Ablation study on the structure of our attentive contextualization module." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 511, + 282, + 557 + ], + "lines": [ + { + "bbox": [ + 51, + 511, + 282, + 557 + ], + "spans": [ + { + "bbox": [ + 51, + 511, + 282, + 557 + ], + "type": "table", + "html": "
Clustering MethodmAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓NDS↑
BEVFormer-b [25]41.667.327.437.239.419.851.7
Linear41.965.627.238.234.819.352.4
MLP42.466.427.338.235.019.252.6
Conv42.666.427.635.036.220.052.8
", + "image_path": "23d1e5e6b6d9d9fa5f275e915d4f42bddc992cdde2d9e172014d46005a2fe010.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 561, + 286, + 582 + ], + "lines": [ + { + "bbox": [ + 47, + 561, + 286, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 286, + 582 + ], + "type": "text", + "content": "Table 6. Ablation study on clustering operations in the attentive contextualization module." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "content": "Relationship between Local Contexts and Global Cluster Contexts. Table 4 reveals the relationship between local and global cluster contexts in enhancing feature learning for view lifting. The exclusive use of global cluster context results in improved velocity prediction, while local attention contributes to better orientation prediction results. Combining these two contexts enhances predictions in location, orientation, and velocity. This highlights the complementary role that global cluster contexts play in feature encoding for view lifting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "Structure of the Attentive Contextualization Module. Table 5 shows the impact of the structure of our attentive contextualization method. Baseline 1 demonstrates that attentive contextualization can efficiently encode feature information. With only three layers, MvACon achieves an improvement of 0.3 NDS compared to the vanilla BEVFormer. Baseline 2 indicates that attentive contextualization requires a sufficient number of clusters to extract abundant cluster contexts in the scene. Baseline 3 suggests that attending to clusters across the feature map aids in improving orientation, velocity, and mAP prediction. Table 6 illustrates that the convolution operation for clustering yields better results in terms of orientation, mAP, and NDS prediction. Conversely, point-based operations (such as a linear layer or multi-layer perceptron) demonstrate superior performance in location and velocity prediction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 274, + 378, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 378, + 286 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 378, + 286 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 294, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 534 + ], + "type": "text", + "content": "This paper presents Multi-View Attentive Contextualization (MvACon) for improving query-based multi-view 3D (MV3D) object detection. It addresses the limitations of two main paradigms of query based MV3D object detector in a unified way: decoder-only dense attention methods like PETR, which lack high-resolution features due to computational constraints, and encoder-decoder sparse 2D/3D deformable attention methods such as BEVFormer and DFA3D. Our MvACon contextualizes the original feature maps extracted from the backbone network using a cluster-attention operation built on the recently proposed patch-to-cluster attention (PaCa). In experiments, we show that our MvACon effectively and consistently enhances query-based MV3D object detectors by encoding more useful contexts, thereby facilitating better 2D-to-3D feature lifting. Rigorously controlled experiments reveal that, for decoder-only detectors, the cluster attention contextualization significantly improves localization and velocity prediction. For encoder-decoder based detectors, it effectively reduces errors in location, orientation, and velocity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 544, + 403, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 403, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 403, + 558 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 564, + 545, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 707 + ], + "type": "text", + "content": "X. Liu and T. Wu were partly supported by NSF IIS-1909644, ARO Grant W911NF1810295, ARO Grant W911NF2210010, NSF CMMI-2024688, NSF IUSE2013451, and a research gift fund from the Innopeak Technology, Inc. (an affiliate of OPPO). The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ARO, NSF, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16695" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "text", + "content": "[2] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In ICCV, pages 9287-9296, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 288, + 225 + ], + "type": "text", + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 281 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 339 + ], + "type": "text", + "content": "[5] Dian Chen, Jie Li, Vitor Guizilini, Rares Andrei Ambrus, and Adrien Gaidon. Viewpoint equivariance for multiview 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9213-9222, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 341, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 287, + 407 + ], + "type": "text", + "content": "[6] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2781-2790, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 408, + 287, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 287, + 451 + ], + "type": "text", + "content": "[7] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. Monopair: Monocular 3d object detection using pairwise spatial relationships. In CVPR, pages 12093-12102, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 454, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 486 + ], + "type": "text", + "content": "[8] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3d object detection in autonomous driving. In CVPR, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 489, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 287, + 532 + ], + "type": "text", + "content": "[9] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3d object detection. In CVPR Workshops, pages 1000-1001, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 534, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 599 + ], + "type": "text", + "content": "[10] 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. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 677 + ], + "type": "text", + "content": "[11] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[12] Ryan Grainger, Thomas Paniagua, Xi Song, Naresh Cuntoor, Mun Wai Lee, and Tianfu Wu. Paca-vit: Learning patch-to-cluster attention in vision transformers. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18568-18578, 2023. 2, 3, 5, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 98, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 98, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 98, + 545, + 140 + ], + "type": "text", + "content": "[13] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems, 34:15908-15919, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 143, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 143, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 308, + 143, + 545, + 175 + ], + "type": "text", + "content": "[14] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 178, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 178, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 308, + 178, + 545, + 220 + ], + "type": "text", + "content": "[15] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 224, + 545, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 224, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 308, + 224, + 545, + 257 + ], + "type": "text", + "content": "[16] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In CVPR, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 258, + 545, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 258, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 545, + 313 + ], + "type": "text", + "content": "[17] Abhinav Kumar, Garrick Brazil, and Xiaoming Liu. Groomed-nms: Grouped mathematically differentiable nms for monocular 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8973-8983, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "type": "text", + "content": "[18] Abhinav Kumar, Garrick Brazil, Enrique Corona, Armin Parchami, and Xiaoming Liu. Deviant: Depth equivariant network for monocular 3d object detection. In ECCV, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 361, + 545, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 416 + ], + "type": "text", + "content": "[19] Hongyang Li, Hao Zhang, Zhaoyang Zeng, Shilong Liu, Feng Li, Tianhe Ren, and Lei Zhang. Dfa3d: 3d deformable attention for 2d-to-3d feature lifting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6684–6693, 2023. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 418, + 545, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 418, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 308, + 418, + 545, + 462 + ], + "type": "text", + "content": "[20] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. Rtm3d: Real-time monocular 3d detection from object keypoints for autonomous driving. In European Conference on Computer Vision, pages 644-660. Springer, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 464, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 464, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 308, + 464, + 545, + 497 + ], + "type": "text", + "content": "[21] Yawei Li, Kai Zhang, Jiezhang Cao, Radu Timofte, and Luc Van Gool. Localvit: Bringing locality to vision transformers. arXiv preprint arXiv:2104.05707, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 499, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 499, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 499, + 545, + 531 + ], + "type": "text", + "content": "[22] Yingyan Li, Yuntao Chen, Jiawei He, and Zhaoxiang Zhang. Densely constrained depth estimator for monocular 3d object detection. ECCV, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 534, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 589 + ], + "type": "text", + "content": "[23] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "type": "text", + "content": "[24] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[25] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16696" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[26] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[27] Xianpeng Liu, Nan Xue, and Tianfu Wu. Learning auxiliary monocular contexts helps monocular 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1810-1818, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "type": "text", + "content": "[28] Xianpeng Liu, Ce Zheng, Kelvin B Cheng, Nan Xue, GuoJun Qi, and Tianfu Wu. Monocular 3d object detection with bounding box denoising in 3d by perceiver. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6436-6446, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "type": "text", + "content": "[29] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022. 2, 3, 4, 5, 6, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 287, + 319 + ], + "type": "text", + "content": "[30] Yingfei Liu, Junjie Yan, Fan Jia, Shuailin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 320, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 287, + 364 + ], + "type": "text", + "content": "[31] Yang Liu, Yao Zhang, Yixin Wang, Feng Hou, Jin Yuan, Jiang Tian, Yang Zhang, Zhongchao Shi, Jianping Fan, and Zhiqiang He. A survey of visual transformers. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 365, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 287, + 420 + ], + "type": "text", + "content": "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 422, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 422, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 49, + 422, + 287, + 475 + ], + "type": "text", + "content": "[33] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15641-15650, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 477, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 477, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 49, + 477, + 287, + 520 + ], + "type": "text", + "content": "[34] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. arXiv:2107.13774, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 522, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 522, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 49, + 522, + 287, + 566 + ], + "type": "text", + "content": "[35] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In ICCV, pages 6851-6860, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 567, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 567, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 49, + 567, + 287, + 600 + ], + "type": "text", + "content": "[36] Xinzhu Ma, Shinan Liu, Zhiyi Xia, Hongwen Zhang, Xingyu Zeng, and Wanli Ouyang. Rethinking pseudo-lidar representation. In ECCV, pages 311-327. Springer, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 287, + 644 + ], + "type": "text", + "content": "[37] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3d object detection. In CVPR, pages 4721-4730, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "text", + "content": "[38] Yuexin Ma, Tai Wang, Xuyang Bai, Huitong Yang, Yuenan Hou, Yaming Wang, Yu Qiao, Ruigang Yang, Dinesh Manocha, and Xinge Zhu. Vision-centric bev perception: A survey. arXiv preprint arXiv:2208.02797, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "text", + "content": "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 106, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 139 + ], + "type": "text", + "content": "[40] Arsalan Mousavian, Dragomir Anguelov, John Flynn, and Jana Kosecka. 3d bounding box estimation using deep learning and geometry. In CVPR, pages 7074-7082, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "type": "text", + "content": "[41] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "type": "text", + "content": "[42] Dennis Park, Jie Li, Dian Chen, Vitor Guizilini, and Adrien Gaidon. Depth is all you need for monocular 3d detection. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7024-7031, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[43] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 284, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 316 + ], + "type": "text", + "content": "[44] Liang Peng, Xiaopei Wu, Zheng Yang, Haifeng Liu, and Deng Cai. Did-m3d: Decoupling instance depth for monocular 3d object detection. In ECCV, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 317, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 371 + ], + "type": "text", + "content": "[45] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 373, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 373, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 373, + 545, + 403 + ], + "type": "text", + "content": "[46] Changyong Shu, Fisher Yu, and Yifan Liu. 3d point positional encoding for multi-camera 3d object detection transformers. arXiv preprint arXiv:2211.14710, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 405, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 448 + ], + "type": "text", + "content": "[47] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In ICCV, pages 1991-1999, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "type": "text", + "content": "[48] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16519-16529, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 505, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 558 + ], + "type": "text", + "content": "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "type": "text", + "content": "[50] 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. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 604, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 667 + ], + "type": "text", + "content": "[51] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12894–12904, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[52] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. arXiv preprint arXiv:2303.11926, 2023. 2, 5" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16697" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[53] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "type": "text", + "content": "[54] Tai Wang, Pang Jiangmiao, and Lin Dahua. Monocular 3d object detection with depth from motion. In ECCV, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 140, + 286, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 286, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 286, + 183 + ], + "type": "text", + "content": "[55] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 185, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 185, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 49, + 185, + 286, + 237 + ], + "type": "text", + "content": "[56] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In CVPR, pages 8445-8453, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 239, + 286, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 239, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 49, + 239, + 286, + 293 + ], + "type": "text", + "content": "[57] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 295, + 286, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 295, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 295, + 286, + 358 + ], + "type": "text", + "content": "[58] Bichen Wu, Chenfeng Xu, Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Zhicheng Yan, Masayoshi Tomizuka, Joseph Gonzalez, Kurt Keutzer, and Peter Vajda. Visual transformers: Token-based image representation and processing for computer vision. arXiv preprint arXiv:2006.03677, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 361, + 286, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 361, + 286, + 415 + ], + "spans": [ + { + "bbox": [ + 49, + 361, + 286, + 415 + ], + "type": "text", + "content": "[59] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerflike representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 416, + 286, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 416, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 49, + 416, + 286, + 491 + ], + "type": "text", + "content": "[60] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 494, + 286, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 494, + 286, + 545 + ], + "spans": [ + { + "bbox": [ + 49, + 494, + 286, + 545 + ], + "type": "text", + "content": "[61] Kun Yuan, Shaopeng Guo, Ziwei Liu, Aojun Zhou, Fengwei Yu, and Wei Wu. Incorporating convolution designs into visual transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 579-588, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 548, + 286, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 548, + 286, + 591 + ], + "spans": [ + { + "bbox": [ + 49, + 548, + 286, + 591 + ], + "type": "text", + "content": "[62] Li Yuan, Qibin Hou, Zihang Jiang, Jiashi Feng, and Shuicheng Yan. Volo: Vision outlooker for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 45(5):6575-6586, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 593, + 286, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 593, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 49, + 593, + 286, + 646 + ], + "type": "text", + "content": "[63] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2998-3008, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 647, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 286, + 678 + ], + "type": "text", + "content": "[64] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In CVPR, pages 3289-3298, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "type": "text", + "content": "[65] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Dalong Du, Jie Zhou, and Jiwen Lu. Dimension embeddings for monocular 3d object detection. In CVPR, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 172 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[66] Ce Zheng, Xianpeng Liu, Guo-Jun Qi, and Chen Chen. Potter: Pooling attention transformer for efficient human mesh recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1620, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[67] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16698" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/8f321c57-13fe-4536-adc3-b4b1a5e6a84e_content_list.json b/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/8f321c57-13fe-4536-adc3-b4b1a5e6a84e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7c8c1ca844b46e26b83f39dcd0223235bd966b33 --- /dev/null +++ b/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/8f321c57-13fe-4536-adc3-b4b1a5e6a84e_content_list.json @@ -0,0 +1,1670 @@ +[ + { + "type": "text", + "text": "Multi-agent Collaborative Perception via Motion-aware Robust Communication Network", + "text_level": 1, + "bbox": [ + 225, + 130, + 743, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shixin Hong $^{1}$ , Yu Liu $^{2\\dagger}$ , Zhi Li $^{1}$ , Shaohui Li $^{1}$ , You He $^{2}$", + "bbox": [ + 264, + 202, + 702, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Shenzhen International Graduate School, Tsinghua University", + "bbox": [ + 241, + 220, + 733, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Electronic Engineering, Tsinghua University", + "bbox": [ + 245, + 239, + 723, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hsx22@mails.tsinghua.edu.cn, liuyu77360132@126.com,", + "bbox": [ + 256, + 258, + 709, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhilizl, lishaohui}@sz.tsinghua.edu.cn, heyou@mail.tsinghua.edu.cn", + "bbox": [ + 183, + 277, + 774, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Collaborative perception allows for information sharing between multiple agents, such as vehicles and infrastructure, to obtain a comprehensive view of the environment through communication and fusion. Current research on multi-agent collaborative perception systems often assumes ideal communication and perception environments and neglects the effect of real-world noise such as pose noise, motion blur, and perception noise. To address this gap, in this paper, we propose a novel motion-aware robust communication network (MRCNet) that mitigates noise interference and achieves accurate and robust collaborative perception. MRCNet consists of two main components: multi-scale robust fusion (MRF) addresses pose noise by developing cross-semantic multi-scale enhanced aggregation to fuse features of different scales, while motion enhanced mechanism (MEM) captures motion context to compensate for information blurring caused by moving objects. Experimental results on popular collaborative 3D object detection datasets demonstrate that MRCNet outperforms competing methods in noisy scenarios with improved perception performance using less bandwidth. Our code will be released at https://github.com/IndigoChildren/collaborative-perception-MRCNet.", + "bbox": [ + 75, + 358, + 472, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Collaborative perception enables information sharing among multiple agents, including vehicles and infrastructure, to provide a comprehensive view of the environment through information communication and fusion [37]. This helps individual autonomous vehicles to tackle challenging perception scenarios, such as physical occlusion or long-range detection [27]. Collaborative perception technology has gained significant attention due to its advantages [20].", + "bbox": [ + 75, + 741, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous works on multi-agent collaborative perception", + "bbox": [ + 96, + 863, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "systems focus on improving perception performance by designing superior fusion modules [2, 14, 29, 35] and effective communication strategies [8, 24, 33]. For example, the CoBEVT [39] framework uses sparse transformers [32] to capture local and global spatial interactions across agents. The UMC [34] framework optimises communication by using a two-stage entropy-based selection mechanism. However, these frameworks assume ideal communication and perception environments [7], while real-world scenarios are affected by noise, such as pose noise in vehicle localization, perception noise in raw data [21], and the motion blur in fast-moving object detection. These noises cannot be ignored for real-world applications of multi-agent collaborative systems. Unfortunately, studies on robust collaborative perception with real-world noise are largely limited.", + "bbox": [ + 500, + 328, + 890, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In fact, different forms of noise in the real world can affect the performance of collaborative perception systems in different ways. For example, when noise interferes with the positioning system, it affects the accuracy of pose information. Pose information is vital for fusing the perception data (e.g., raw data, intermediate features, or detection outputs) of the collaborative agents in different coordinates [40]. Thus, pose noise can lead to information misalignment and inconsistency among agents [26] during data fusion, resulting in a reduced system performance. Perception noise introduces distortions into the sensor data, which affects the final performance of the perception system. In addition, the presence of motion blur due to rapid object movement affects the accuracy of target detection [1,9,15,30].", + "bbox": [ + 496, + 555, + 892, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the development of a robust collaborative perception system against real-world noise is complicated by the following challenges. First, a robust fusion strategy that adequately processes the multi-source perception data is required to deal with the perceptual misalignment caused by pose noise. Second, a feature selection strategy that filters out highly informative features is crucial to reduce bandwidth consumption and improve the efficiency of multi-agent collaboration [28]. Finally, it is desirable to find a sta", + "bbox": [ + 496, + 765, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author: Yu Liu.", + "bbox": [ + 101, + 887, + 272, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15301", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b86db373d2b311de9cce979a13674800279643e0f238b664fd24a7ef6da08f52.jpg", + "image_caption": [ + "Figure 1. The overview of our proposed MRCNet framework. In MRCNet, agents exchange highly informative semantic features in two stages through multiple rounds of communication to mitigate the effect of noise. Meanwhile, the ego agent uses motion context extracted from historical sequences in the memory bank to reduce the effect of motion blur from moving objects." + ], + "image_footnote": [], + "bbox": [ + 122, + 85, + 890, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ble and accurate feature extraction that eliminates the motion blur of rapidly moving objects.", + "bbox": [ + 75, + 409, + 468, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we present MRCNet, a motion-aware robust communication network that mitigates noise interference and achieves more robust and accurate collaborative perception. As shown in Figure 1, MRCNet facilitates effective information collaboration among agents and improves the perception performance of the ego agent through two main components, i.e., multiscale robust fusion (MRF) and motion enhanced mechanism (MEM). Specifically, for the challenge of feature misalignment due to collaborators' pose noise, MRF improves the robustness and performance of collaborative perception through multi-scale semantic feature fusion. For perception noise, we filter out highly informative semantic features for multi-round communication based on channel-wise selection operations, while reducing communication bandwidth consumption. In addition, to address the problem of motion blur due to object movement, we introduce the MEM module, which is based on the gate recurrent unit (GRU) architecture and is designed to capture and incorporate the motion context of the ego agent. This mechanism aggregates historical information to refine the features of the current frame. To validate the effectiveness of our MRCNet framework, we carry out a comprehensive evaluation on three open-source collaborative 3D object detection datasets, including V2XSim [13], OPV2V [41], and V2XSet [40]. The experimental results show that MRCNet outperforms competing methods in noisy scenarios. Our main contributions can be summarised as follows:", + "bbox": [ + 75, + 439, + 470, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We present a focused study of the real-world noise problem in the collaborative perception task.", + "bbox": [ + 94, + 869, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a motion-aware collaborative perception framework, i.e., MRCNet, which presents a novel multi-scale feature communication and fusion strategy to improve the robustness of collaboration.", + "- Collaborative experiments are conducted on three large open-source datasets, proving that our proposed MRCNet can achieve the best perception performance in noisy scenarios with less bandwidth." + ], + "bbox": [ + 517, + 409, + 890, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 550, + 640, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Communication in Collaborative Perception", + "text_level": 1, + "bbox": [ + 500, + 575, + 874, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous works [8, 23, 24] focus on solving the communication problem, as excessive communication bandwidth consumption can lead to communication congestion and latency. To address this challenge, most research considers how to achieve a better trade-off between performance and bandwidth. There are two main methods to address the challenge of bandwidth constraints: compressing the communication volume or reducing the number of communicators. Several works reduce the communication volume by using learnable methods to facilitate efficient collaboration [18]. Where2comm [8] focuses on perceptually critical areas and transmits foreground features with high confidence. UMC [34] designs a two-stage communication strategy using extended entropy theory to filter out low quality regions. As for selecting of effective collaborators, Who2Com [24] proposes a multi-stage handshake communication mechanism that determines which agents to connect with, and Select2Col [19] introduces a collaborator selection method that selects contributing collaborators with low latency. However, there is still a lack of research on", + "bbox": [ + 496, + 598, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15302", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the problem of pose noise in collaborative communication. Therefore, we propose the multi-scale robust fusion module, which selects highly informative features from multiscale features for communication and aggregation, thereby reducing the effect of multi-source noise in collaboration.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Information Fusion in Collaborative Perception", + "text_level": 1, + "bbox": [ + 76, + 174, + 468, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to addressing communication issues, collaborative perception systems focus on designing collaboration modules to fuse information from multiple agents. The goal is to improve perception capabilities and thereby optimise performance in downstream tasks. Based on the stage of data sharing and collaboration, the collaborative perception can be broadly divided into three categories [14, 40], i.e., early fusion [42], intermediate fusion [8, 23, 35, 41], late fusion [26, 31]. Early fusion involves sharing raw data for a comprehensive view, but requires significant bandwidth and overlooks contextual information. Intermediate fusion extracts intermediate features from each agent's observations before transmission, balancing perception performance and communication bandwidth usage. Late fusion shares detection outputs among agents, but produces suboptimal results due to individual inaccuracies or incomplete perceptions. Among them, intermediate fusion methods gain popularity due to their performance-bandwidth trade-off [35, 40, 41]. V2VNet [35] proposes a spatially aware graph neural network to aggregate features from multiple agents. OPV2V [41] uses a global self-attention module for feature fusion. As an extension of intermediate fusion, our study introduces a novel cross-semantic fusion module, which is designed to aggregate features in the presence of feature misalignment caused by pose noise. Furthermore, we integrate motion context into the perception system to improve the understanding of the environment.", + "bbox": [ + 75, + 196, + 472, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 617, + 168, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we focus on improving the robustness and accuracy of collaborative perception in complex collaborative scenarios. To achieve this, we propose a novel motion-aware robust communication network (MRCNet) that integrates multi-scale feature selection, effective information fusion, and motion-aware feature extraction into collaborative perception. As shown in Figure 1, MRCNet consists of four key components: feature extraction and selection, multi-scale robust collaboration, motion enhanced mechanism, and the 3D object detection head.", + "bbox": [ + 75, + 642, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 801, + 279, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At time $t$ , in a collaborative perception task involving $N$ agents, the selected $i$ -th agent is defined as the ego agent, and a communication graph is established among them. Other agents within the communication range serve as collaborators, providing complementary information and", + "bbox": [ + 76, + 824, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "helping the ego agent to make a comprehensive perception. Let $\\mathcal{X}_i^t$ be the point cloud observation and $\\mathcal{Y}_i^t$ be the perception supervision of the ego agent. Each collaborator provides the ego agent with the information extracted from their local observations $\\mathcal{X}_j^t$ . Additionally, considering that we aim at modelling the motion-aware collaborations, we design a memory bank of the ego agent that stores $K$ frames of historical point data, denoted as $\\{\\mathcal{X}_i^{t - k}\\}_{k = 1}^K = \\{\\mathcal{X}_i^{t - 1},\\mathcal{X}_i^{t - 2},\\dots,\\mathcal{X}_i^{t - K}\\}$ .", + "bbox": [ + 496, + 90, + 890, + 231 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus, the motion-aware collaborative perception task can be defined as follows:", + "bbox": [ + 498, + 231, + 890, + 258 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {Y}} _ {i} ^ {t} = \\sum_ {i = 1} ^ {N} \\left\\{\\mathcal {Y} _ {i} ^ {t} | \\Omega \\left(\\mathcal {X} _ {i} ^ {t}, \\left\\{\\mathcal {X} _ {j} ^ {t} \\right\\} _ {j \\neq i} ^ {N}, \\left\\{\\mathcal {X} _ {i} ^ {t - k} \\right\\} _ {k = 1} ^ {K}\\right) \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 268, + 890, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Omega (\\cdot)$ is the proposed MRCNet. In this paper, we use the 3D object detection task to evaluate the performance of collaborative perception. Therefore, $\\hat{\\mathcal{V}}_i^t$ are defined as the predicted detection boxes along this line.", + "bbox": [ + 496, + 321, + 893, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Feature Extraction and Selection", + "text_level": 1, + "bbox": [ + 500, + 393, + 787, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Extraction. After the ego agent constructs a vehicle-to-everything (V2X) collaboration graph connecting neighbouring agents at the current frame $t$ , it exchanges pose data between agents for alignment purposes. The collaborators then transform their raw point cloud data into the ego agent's coordination using the transformation matrix $\\Lambda$ computed based on the agents' unique pose $\\boldsymbol{\\xi}^t$ . Next, each agent uses a shared feature extractor consisting of $S$ layers [41] to obtain multi-scale features from the raw data. This process yields a set of features $\\mathcal{F}_i^t = \\left\\{\\mathcal{F}_i^{s,t}\\right\\}_{s=1}^S$ for the $i$ -th agent, where $i$ ranges from 1 to $N$ .", + "bbox": [ + 496, + 416, + 890, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Selection. In contrast to previous methods that transmit all features, we introduce a novel feature selection technique. Perception noise introduces distortions into the collaborative features, thereby misleading the collaborative system. Therefore, our feature selection technique selects highly informative features from the multi-scale features for communication, effectively reducing the effect of noise interference and conserving communication bandwidth.", + "bbox": [ + 496, + 585, + 890, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The features extracted from the foreground regions are particularly important due to their rich semantic information [43,45]. To select highly informative features for effective communication, we use channel pooling operations to evaluate the semantic density within each feature. Initially, different scales of features are spatially flattened and then concatenated into a feature sequence $\\pmb{E}^{t} \\in \\mathbb{R}^{C \\times L}$ , where $C$ is the number of channels of the feature, $L$ is the length of the multi-scale feature after flatten. Subsequently, we perform channel-wise pooling operations on $\\pmb{E}^{t}$ to generate the semantic confidence map $m$ :", + "bbox": [ + 496, + 705, + 890, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {m} = \\left(\\operatorname {M a x P o o l} \\left(\\boldsymbol {E} ^ {t}\\right) + \\operatorname {A v g P o o l} \\left(\\boldsymbol {E} ^ {t}\\right)\\right) \\in \\mathbb {R} ^ {1 \\times L}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15303", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7af542631a63c4b5fd0b8b0806be1f3d35ba833f5cdd3694d40b427ac7d0bf04.jpg", + "image_caption": [ + "Figure 2. The architecture of the proposed MEA component. (a) shows the improved pyramid sampling aggregation module. (b) shows the multi-scale deformable self-attention module." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 467, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where MaxPool $(\\cdot)$ and AvgPool $(\\cdot)$ refer to the channelwise max pooling and average pooling operations. The binary selection matrix for communication is defined as:", + "bbox": [ + 75, + 332, + 468, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {c} = \\Phi_ {\\text {s e l e c t}} (\\boldsymbol {m}) \\in \\{0, 1 \\} ^ {1 \\times L}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 385, + 468, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Phi_{\\mathrm{select}}(\\cdot)$ is the selection function to select the most informative features for communication. The selected features with high information are then obtained as $D^{t} = (c\\odot E^{t})\\in \\mathbb{R}^{C\\times L}$ , where $\\odot$ is the element-wise multiplication. The selected feature map provides highly informative features with less bandwidth usage.", + "bbox": [ + 76, + 412, + 468, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-Scale Robust Fusion", + "text_level": 1, + "bbox": [ + 76, + 511, + 316, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-agent collaboration improves the visual representation and perception performance of the ego agent by exploiting shared semantic information. Previous attention-based [39-41] approaches overlook to consider the robustness of the collaboration system to noise. Therefore, we introduce a novel multi-scale robust fusion (MRF) module, which is designed for the robust multi-scale feature fusion through two-stage communication. Multi-scale feature fusion combines information of different semantics to improve perception capability and robustness to pose noise. The two-stage communication module consists of multiscale enhanced aggregation (MEA) and collaborative features aggregation (CFA). MEA uses cross-semantic fusion to overcome the challenge of feature misalignment caused by pose noise, and CFA enables the ego agent to maximise the perception information aggregation.", + "bbox": [ + 75, + 534, + 468, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-scale Enhanced Aggregation. In the first stage, the ego agent broadcasts highly informative features to the collaborators. For the sake of clarity, let's consider a communication process involving only two agents. Upon receiving highly informative features, the collaborator first converts these features from sequence to multi-scale representation, then performs a preliminary fusion using the max fusion operation to obtain the multi-scale feature $\\mathcal{U}_j^t = \\{\\mathcal{U}_j^{s,t}\\}_{s=1}^S$ .", + "bbox": [ + 75, + 777, + 468, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the challenge of feature misalignment, MEA incorporates cross-semantic modules to fully exploit the strengths of different semantics [6]. As shown in Figure 2, MEA consists of two components: an improved pyramid sampling aggregation (IPSA) module and a multi-scale deformable self-attention (MDSA) module [46].", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Low-level features capture fine-grained details, while high-level semantic features provide integrated contextual understanding. The pyramid sampling aggregation module [12, 25] uses simple convolution and channel-wise concatenation operations to fuse multi-scale features. This process gathers information through cross-semantic fusion, resulting in a more comprehensive and robust feature representation that effectively mitigates noise interference. We pyramid sample the feature $\\mathcal{U}_j^{s,t}$ to obtain the new multiscale feature $\\left\\{\\mathcal{U}_j^{sl,t}\\right\\}_{l=1}^S$ , where $\\mathcal{U}_j^{sl,t}$ is the transformation of $\\mathcal{U}_j^{s,t}$ from the $s$ -th scale to the $l$ -th scale, and the feature sizes of the new multi-scale feature correspond to $\\mathcal{U}_j^t$ .", + "bbox": [ + 496, + 181, + 892, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After multiple rounds of communication, the fine details in the transmitted features tend to be compromised due to cross-semantic fusion during the pyramid sampling. Therefore, we improve the pyramid sampling aggregation module by adding original multi-scale features from collaborators to compensate for detailed features, which is the proposed IPSA module. The output can be represented as:", + "bbox": [ + 496, + 380, + 890, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {U}} _ {j} ^ {s, t} = \\operatorname {c o n v} _ {1 \\times 1} \\left(\\left[ \\mathcal {U} _ {j} ^ {s 1, t}, \\mathcal {U} _ {j} ^ {s 2, t}, \\mathcal {U} _ {j} ^ {s 3, t} \\right]\\right) + \\mathcal {U} _ {j} ^ {t}, s \\in [ 1, 2, 3 ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 496, + 494, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[\\cdot ,\\cdot ,\\cdot ]$ is the channel-wise concatenation, and $\\mathrm{conv}_{1\\times 1}(\\cdot)$ is the $1\\times 1$ convolution operation.", + "bbox": [ + 496, + 532, + 890, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the IPSA module, the collaborator applies the MDSA [46] module to further fuse multi-scale features and mitigate the effect of pose noise on feature distortion. The MSDA module consists of three linear layers $\\mathbf{W}_a$ , $\\mathbf{W}_v$ , $\\mathbf{W}_o$ , which separately compute the sampling offset $\\Delta q$ , the attention weight $\\mathbf{A}$ and the value $\\mathbf{V}$ at position $q$ , where $q$ is the two-dimensional reference point in $\\hat{\\mathcal{U}}_j^{s,t}$ . The output of MSDA at position $q$ can be expressed as follows:", + "bbox": [ + 496, + 564, + 892, + 685 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M S D A} (q) = \\sum_ {h = 1} ^ {H} \\boldsymbol {W} _ {h}.\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 696, + 655, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\sum_ {s = 1} ^ {S} \\sum_ {p = 1} ^ {P} \\psi \\left(\\boldsymbol {W} _ {a} \\tilde {\\mathcal {U}} _ {j} ^ {s, t} (q)\\right) \\cdot \\boldsymbol {W} _ {v} \\left(\\tilde {\\mathcal {U}} _ {j} ^ {s, t} \\left(q + \\boldsymbol {W} _ {o} \\tilde {\\mathcal {U}} _ {j} ^ {s, t} (q)\\right)\\right)\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 739, + 911, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n= \\sum_ {h = 1} ^ {H} \\boldsymbol {W} _ {h} \\cdot \\left(\\sum_ {s = 1} ^ {S} \\sum_ {p = 1} ^ {P} \\boldsymbol {A} ^ {h s p} (q) \\cdot \\mathcal {V} ^ {h s} \\left(q + \\Delta^ {h s p} (q)\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 785, + 890, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H$ is the number of attention heads, $S$ is the quantity of different feature scales, and $P$ is the number of sampling points utilized in each scale. $\\mathbf{W}_h$ is the projection matrix", + "bbox": [ + 500, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15304", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the softmax function $\\psi (\\cdot)$ is used to determine the attention weight. Ultimately, we output the multi-scale format feature $\\mathcal{C}_j^t = \\{\\mathcal{C}_j^{s,t}\\}_{s = 1}^S$ after MSDA.", + "bbox": [ + 75, + 90, + 468, + 141 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Collaborative Features Aggregation. In the second stage of communication, the $j$ -th collaborator uses the same feature selection technique to send highly informative features $\\pmb{D}_j^t = (\\pmb{m}_j \\odot \\pmb{E}_j^t)$ back to the ego agent, where $\\pmb{E}_j^t$ is the feature sequence representation of $\\mathcal{C}_j^t$ . In order to maximise the aggregation of the semantic information provided by the collaborator, inspired by [29], the ego agent uses the spatial-wise adaptive fusion module on each scale of features. The ego agent first stacks the feature maps of the agents to form $\\mathcal{A}_i^t = \\{\\mathcal{A}_i^{s,t}\\}_{s=1}^S$ , $\\mathcal{A}_i^{s,t} \\in \\mathbb{R}^{n \\times C \\times H_s \\times W_s}$ , where $s$ is the scale index and $n$ is the maximum number of agents. Max and average pooling is applied to $\\mathcal{A}_i^{s,t}$ over the first channel axis to create new feature maps $\\mathcal{A}_{i,\\max}^{s,t} \\in \\mathbb{R}^{1 \\times C \\times H_s \\times W_s}$ and $\\mathcal{A}_{i,\\mathrm{avg}}^{s,t} \\in \\mathbb{R}^{1 \\times C \\times H_s \\times W_s}$ , where $H_s$ and $W_s$ represent the height and width of the feature. These pooled feature maps are concatenated and passed through a 3D convolutional layer with a ReLU activation function to perform further fusion. The aggregated features provide complementary information to the ego agent.", + "bbox": [ + 75, + 141, + 472, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Motion Enhanced Mechanism", + "text_level": 1, + "bbox": [ + 76, + 449, + 344, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned above, the collaborative perception module increases the range of perception and provides compensation for occlusions [41], but it cannot directly address the motion blur introduced by moving objects. To combat this, we introduce the motion enhanced mechanism (MEM), which uses motion context from historical point cloud frames to achieve motion-aware perception.", + "bbox": [ + 75, + 473, + 468, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When the ego agent initiates a collaboration request, it projects the last $K$ frames of the point cloud stored in the memory bank to its current coordinate, producing a series of bird's eye view (BEV) feature maps $\\{\\mathcal{B}^{t - 1},\\dots,\\mathcal{B}^{t - K}\\}$ . Inspired by [4,22], we design a recurrent unit called the gated motion unit (GMU). This unit aggregates motion context to enhance frame-level features. Figure 3 illustrates the operation of GMU at timestamp $t$ , showing that GMU takes three inputs stored in the ego agent's memory bank: the feature $\\mathcal{B}^{t - 1}$ of the previous frame, the feature $\\mathcal{B}^t$ of the current frame, and the enhanced feature $\\mathcal{H}^{t - 1}$ . Both $\\mathcal{B}^0$ and $H^0$ are initialized as $\\mathcal{B}^1$ , and the output $\\mathcal{H}^t$ represents the enhanced feature of the current frame after passing through GMU.", + "bbox": [ + 75, + 580, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Between successive frames, object motion causes spatial transitions of features. Therefore, we use $\\mathcal{M}^t = \\mathcal{B}^t -\\mathcal{B}^{t - 1}$ to capture motion context, which can represent the motion clues of moving objects [17]. To reduce the effect of motion blur and to exploit historical information, we use a deformable convolution network [3]. The deformable convolution network can better capture the features of objects by adaptively adjusting the kernel size and shape. Specifically,", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/70121dcc58cfd5bbe5c0e598d4c04b2fb3a4fd47b1aa789c8a66065db53725be.jpg", + "image_caption": [ + "Figure 3. The architecture of the proposed GMU component." + ], + "image_footnote": [], + "bbox": [ + 513, + 88, + 883, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the $3 \\times 3$ convolutional layer is used on the motion context $\\mathcal{M}_t$ to obtain the spatial offsets $\\Delta p_n^t$ . The convolutional operation is then applied to $\\mathcal{B}^{t-1}$ to extract historical features at position $p$ . The output feature map $\\mathcal{I}^t$ at location $p$ is:", + "bbox": [ + 498, + 261, + 890, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} ^ {t} (p) = \\sum_ {p _ {n} \\in \\mathcal {R}} \\boldsymbol {W} \\left(p _ {n}\\right) \\cdot \\mathcal {B} ^ {t - 1} \\left(p + p _ {n} + \\Delta p _ {n} ^ {t}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 329, + 890, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}$ is the convolution weight, $\\mathcal{R}$ is the kernel grid and $p_n$ enumerates the locations in $\\mathcal{R}$ .", + "bbox": [ + 498, + 369, + 890, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To reduce the motion blur effect caused by the moving object between frames, it is necessary to determine whether the features at each position should be updated. Inspired by [36], we design an update gate $\\mathcal{G}$ using the channel-spatial attention module. The channel-spatial attention module is applied to the stack feature of $\\mathcal{I}^t$ and $\\mathcal{H}^{t - 1}$ . After applying the sigmoid operation, the overall attention map is normalised to a range of values between 0 and 1, as formulated below:", + "bbox": [ + 498, + 400, + 892, + 535 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} = \\sigma (\\Gamma ([ \\mathcal {I} ^ {t}, \\mathcal {H} ^ {t - 1} ])) \\in \\mathbb {R} ^ {C \\times H \\times W}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 542, + 890, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $[\\cdot, \\cdot]$ is the channel-wise concatenation, $\\Gamma(\\cdot)$ is the channel-spatial attention module, and $\\sigma(\\cdot)$ is the sigmoid function for value normalisation. The value in $\\mathcal{G}$ indicates the probability that the corresponding position of $\\mathcal{H}^{t-1}$ is reserved. Then the improved feature $\\mathcal{H}^t$ is refined with the previous improved feature $\\mathcal{H}^{t-1}$ as follows:", + "bbox": [ + 496, + 569, + 890, + 660 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} ^ {t} = \\mathcal {G} \\odot \\mathcal {H} ^ {t - 1} + (1 - \\mathcal {G}) \\odot \\mathcal {I} ^ {t}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 667, + 890, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{H}^t\\in \\mathbb{R}^{C\\times H\\times W}$ is the final output of the GMU.", + "bbox": [ + 498, + 691, + 861, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. 3D Object Detection Head", + "text_level": 1, + "bbox": [ + 500, + 715, + 736, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the fused output after MRF, the multiscale feature is encoded to the same size and concatenated to $\\mathcal{Z}_i^t\\in \\mathbb{R}^{C\\times H\\times W}$ . Then, we concatenate $\\mathcal{Z}_i^t$ and $\\mathcal{H}^t$ along the channel dimension to derive the final representation $\\mathcal{P}_i^t$ . To generate the final object detection results, we then use a dual-branch convolutional module on $\\mathcal{P}_i^t$ to perform classification and regression tasks for each detection anchor. Similar to the previous study [44], the formulation for the object detection loss $L_{\\mathrm{det}}$ is expressed as:", + "bbox": [ + 496, + 739, + 890, + 876 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d e t}} = L _ {\\mathrm {c l s}} + \\beta L _ {\\mathrm {r e g}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 885, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15305", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/be7b78e382707bc3a1666e83d462e580bb328e96c7f998df93c8157c8dfd3538.jpg", + "image_caption": [], + "image_footnote": [ + "MRCNet(Ours) Where2comm V2X-ViT Late Fusion When2com No Fusion Early Fusion" + ], + "bbox": [ + 99, + 88, + 349, + 195 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f7fb61e47650ec619889b7872b1f0e817b1514f4ef5b4413ebac62aca828691a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 89, + 609, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ef1b9ec0cdabf07d5cb048d4c2cb91dc3dd4c5bea92beff0e22b7fd5a596cd2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 90, + 867, + 195 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e7ab0d0f4f4245c11eeb0c26a7e4a05c85ab1b5f894f4a6f09a210e44cfbc33f.jpg", + "image_caption": [ + "Figure 4. Robustness of our proposed method and compared benchmark to varying levels of localization noise." + ], + "image_footnote": [ + "MRC(Ours) Where2comm V2X-ViT Late Fusion When2com No Fusion Early Fusion" + ], + "bbox": [ + 99, + 247, + 348, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5672424319162e1d2ae08c798684f9119b79510bd60dcff835b99be9cbdec885.jpg", + "image_caption": [ + "Figure 5. Robustness of our proposed method and compared benchmark to varying levels of heading noise." + ], + "image_footnote": [], + "bbox": [ + 361, + 250, + 609, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da6309332837a9662bd78cf8ecc429924171a0f50d0863cf24e4535362681146.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 250, + 867, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{\\mathrm{cls}}$ is the cross-entropy loss for object classification, $L_{\\mathrm{reg}}$ is the smooth L1 loss for box regression [16], and $\\beta$ is the balancing hyperparameter.", + "bbox": [ + 76, + 416, + 468, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 477, + 209, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the robustness of our module in noisy collaboration scenarios, we conduct experiments on three large open-source datasets. In particular, our experiments focus on LiDAR-based 3D object detection. We measure detection performance using average precision (AP) at intersection-over-union (IoU) thresholds of 0.50 and 0.70, and note that the AP values are given as percentages. To ensure a fair comparison, we exclude data and feature compression operations in the compared methods.", + "bbox": [ + 75, + 503, + 468, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 650, + 385, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. V2XSim [13] is a V2X collaborative perception dataset co-simulated by SUMO [10] and CARLA [5]. It contains 10000 frames of LiDAR point clouds and 501K annotated 3D boxes. There are 8000/1000/1000 frames for the training/validation/testing, respectively. The detection range of the agent is $x \\in [-32m, 32m]$ , $y \\in [-32m, 32m]$ . OPV2V [41] is a vehicle-to-vehicle collaborative perception dataset co-simulated by OpenCDA [38] and CARLA [5], including 11464 frames of 3D point clouds with 230K annotated 3D boxes, split into training/validation/testing of 6764/1981/2719 frames. The detection range is $x \\in [-140m, 140m]$ , $y \\in [-40m, 40m]$ . V2XSet [40] is the first open dataset that includes both V2X cooperation and realistic noise simulation, which is co-simulated by OpenCDA [38] and CARLA [5]. There", + "bbox": [ + 75, + 672, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are a total of 11447 frames in the dataset, and the train/Validation/testing splits are 6694/1920/2833, respectively. The detection range setting is the same as OPV2V.", + "bbox": [ + 496, + 416, + 890, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. In our implementation, we use PointPillars [11] as the multi-scale feature extractor, with the grid size of $(0.4m, 0.4m)$ for the point cloud discretisation. And we get intermediate features in three scales, which is $S = 3$ as we define in the method. The MSDA module uses 8 attention heads and 4 sampling points per communication round. For V2XSim, we extract information from three historical frames, while for V2XSet and OPV2V, we use two historical frames. To train the detection model, we set the hyperparameter in the loss function to $\\beta = 2$ according to [11]. We use the Adam optimizer with an initial learning rate of $\\{1\\mathrm{e} - 3, 5\\mathrm{e} - 4, 3\\mathrm{e} - 4\\}$ for the OPV2V, V2XSet and V2XSim datasets. The learning rate is gradually reduced at epochs 35 and 45 on the V2XSim dataset using a decay factor of 0.1. For the OPV2V and V2XSet datasets, the decay occurs at epochs 30 and 40. All models are trained on 2 NVIDIA RTX 3090 GPUs with a batch size of 2 for a total of 50 epochs.", + "bbox": [ + 496, + 462, + 892, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark Comparison. Our considering benchmark includes early fusion, intermediate fusion, and late fusion techniques as well as single agent perception for comparison. In addition, we compare with six state-of-the-art (SOTA) intermediate fusion models: When2Com [23], V2VNet [35], AttFuse [41], V2X-ViT [40], DiscoNet [14], CoBEVT [39], and Where2Comm [8].", + "bbox": [ + 496, + 733, + 890, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 500, + 847, + 692, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The pose of the agent can be represented by a 6D vector. Since the agent only has yaw angle to measure rotation,", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15306", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/49b7db73d4c2136222ff3c4fa91b788aa8f00755349144768beda15442753faa.jpg", + "table_caption": [ + "Table 1. Performance comparison on the V2XSim, OPV2V, and V2XSet datasets. The results are reported in AP0.5/0.7." + ], + "table_footnote": [], + "table_body": "
ModelV2XSimOPV2VV2XSet
AP0.5AP0.7AP0.5AP0.7AP0.5AP0.7
No Fusion65.7352.5769.3856.4064.8850.83
Late Fusion71.2256.9982.7661.6476.3153.33
Early Fusion84.9465.9789.2772.9683.4051.71
When2com [23]73.3652.5282.7867.6674.7340.62
V2VNet [35]82.7065.3188.6974.9384.7964.73
AttFuse [41]81.7066.2488.5472.9184.3766.27
V2X-ViT [40]82.3264.4186.7475.7082.4263.14
DiscoNet [14]83.5666.1288.0572.0782.3464.79
CoBEVT [39]81.0065.0688.9972.8084.8465.14
Where2comm [8]77.5757.0286.5868.9780.8358.72
MRCNet85.3369.8289.7776.1285.0066.31
", + "bbox": [ + 80, + 121, + 467, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we simplify the pose to $\\xi = (x,y,\\theta)$ in 2D space [26], where $x, y$ and $\\theta$ represent the 2D center position and the yaw angle of the agent's accurate global position, respectively. To simulate pose noise in real-world scenarios, we add Gaussian noise with a standard deviation of $0.2m$ and $0.2^{\\circ}$ to the collaborators' poses in the datasets. We compare the performance of 3D object detection with other methods on the datasets, as shown in Table 1. According to the experimental results, despite the introduction of pose noise, our proposed MRCNet module shows superior performance compared to the benchmark on all datasets.", + "bbox": [ + 75, + 314, + 468, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, to investigate the effect of different components of pose noise on the collaborative perception system, we conduct experiments to evaluate the robustness of the models to localization noise and heading noise, separately. Experiments are also performed to assess the effect of various levels of localization and heading noise. Specifically, we add Gaussian noise with standard $\\sigma_{xy} \\in [0.2m, 0.6m]$ , $\\sigma_{\\theta} \\in [0.2^{\\circ}, 1.0^{\\circ}]$ separately to the collaborators' poses. As shown in the Figure 4 and the Figure 5, an increase in Gaussian noise leads to a rapid decrease in AP0.7 values for the other methods. This suggests that misaligned spatial features from collaborators can mislead the collaboration system. However, our method maintains superior performance even under significant pose noise, demonstrating its ability to effectively select and fuse multi-level semantic features while achieving robustness to pose noise.", + "bbox": [ + 75, + 483, + 468, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Component Analysis", + "text_level": 1, + "bbox": [ + 76, + 738, + 272, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contribution of Major Components. To evaluate the effectiveness of our proposed component, we progressively remove (i) MEM, (ii) FS (feature selection), (iii) IPSA and (iii) MSDA and present the detection precision. Table 2 shows the results of our ablation studies, with each innovative component contributing to the performance gains.", + "bbox": [ + 75, + 762, + 467, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Varying Number of History Frames. We evaluate the effect of integrating different numbers of history frames from the ego agent. As shown in Figure 6, integrating multiple", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9a078091ac06eb97f95330d1b0657c8dbcd8dbd2004352cedd379e67d07868c9.jpg", + "table_caption": [ + "Table 2. Results of the ablation study of the proposed core components on the V2XSim, OPV2V and V2XSet datasets." + ], + "table_footnote": [], + "table_body": "
MEMFSISPAMSDAV2XSimOPV2VV2XSet
AP@0.5AP@0.7AP@0.5AP@0.7AP@0.5AP@0.7
80.7764.2885.7874.0183.1263.88
✓81.4667.6486.2974.6383.4464.71
✓✓82.8769.1587.3275.3184.2165.31
✓✓✓83.6069.2188.1175.7784.4365.47
✓✓✓✓85.3369.8289.7776.1285.0066.31
", + "bbox": [ + 503, + 121, + 888, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fecf1cfe3b33de799825d1435a00f776df77544625854bab1dda2a0779a6abca.jpg", + "image_caption": [ + "Figure 6. Collaborative Performance with varying number of history frames. The MEM module works best when fusing three history frames on the V2XSim dataset, and two frames on the OPV2V and V2XSet datasets." + ], + "image_footnote": [], + "bbox": [ + 501, + 198, + 696, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78418ebd09de673fd4f9ceef989eb892680da0561553e4c601a460c3319f8524.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 198, + 887, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/397e1ed6a59741a97e4336dde28a6f2f376a2f6af10d414e33d3a74e1af29239.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 501, + 397, + 696, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ea1957539fe975792ecf9605ba19b67c1d731adbdaea33753794c709ebb5494.jpg", + "image_caption": [ + "(b)", + "Figure 7. (a) shows the perception performance comparison when using the proposed multi-scale method versus the single-scale approach. (b) is the comparison of collaborative perception performance of our proposed method and other baseline methods with different communication volume." + ], + "image_footnote": [], + "bbox": [ + 702, + 398, + 888, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frames into the MEM module leads to an improvement in the detection accuracy in the V2XSim [13] dataset. However, once the number of history frames exceeds three, no further improvement is observed, as MEM faces challenges in effectively capturing features with large changes.", + "bbox": [ + 496, + 625, + 890, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Multi-Scale Collaboration. To evaluate the effectiveness of multi-scale feature collaboration in complex scene perception, we separately evaluate the detection accuracy achieved by single-scale and multi-scale feature collaboration. In our implementation, the single-scale feature collaboration is modified by using only a single-scale deformable self-attention module [46] within MEA. The results of this comparison are shown in Figure 7a. By using multi-scale feature fusion, our method exploits the advantages of different semantics for better noise robustness and achieves an improved detection performance.", + "bbox": [ + 496, + 702, + 890, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of Communication Volume. Figure 7b shows the performance-bandwidth trade-off of the proposed", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15307", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26643039f58337bed6ad187b5e1f0988fc59e6c9a2efbcea6eee7f30b7f53d2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 90, + 272, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/64668c3216b00077414f1844d4e8662f728cfa10472e5a2796e41123710d7251.jpg", + "image_caption": [ + "(a) V2X-ViT" + ], + "image_footnote": [], + "bbox": [ + 80, + 239, + 272, + 387 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c8cf16f47167cf1c9048b54e91a8b69924b7c323f9a436d3534ca578a7859c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 90, + 478, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/489cc1c04844189389ef3823922cb66433aae87fdeb3597f72a20513883eded1.jpg", + "image_caption": [ + "(b) DiscoNet" + ], + "image_footnote": [], + "bbox": [ + 287, + 239, + 478, + 387 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d1ffaf89fb32aabcb7d6c0b957da63adde4d34b138df2df07e74af0f90f6a07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 90, + 683, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/352e874738cfbe270d950d8ec8afff1f6fb857ef1f965cf6b3f4dd347a997cc7.jpg", + "image_caption": [ + "(c) CoBEVT", + "Figure 8. In the qualitative comparison on the V2XSim dataset under pose noise, green and red boxes indicate the ground truths and the detection outputs, respectively. Our method outperforms previous SOTA models by achieving more accurate detection results." + ], + "image_footnote": [], + "bbox": [ + 491, + 239, + 683, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b40969ff842e538d313a683affd43588fde959eaf52adec16f293e87283f3ca0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 90, + 887, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b01a784d541602e904c31ec3d8464f52fa907d5b62a6654b13a30dae2ea5052c.jpg", + "image_caption": [ + "(d) MRCNet (Ours)" + ], + "image_footnote": [], + "bbox": [ + 697, + 239, + 887, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MRCNet and the benchmark at different levels of communication bandwidth consumption. No fusion serves as the baseline, as it does not involve any communication. Our analysis reveals several key trends: (i) Where2Comm [8] initially improves detection performance by transmitting the most critical information, but struggles to filter out noisy data as the transmission volume increases. When Where2Comm [8] transmits all feature information, the global attention mechanism captures the spatial information and improves its effectiveness. (ii) Our proposed MRCNet method consistently shows better detection performance at different bandwidth levels, achieving superior results, and multi-round communication can improve the performance of collaborative perception, but excessive misaligned feature transmission can also cause degradation.", + "bbox": [ + 75, + 463, + 472, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Case Study", + "text_level": 1, + "bbox": [ + 76, + 707, + 200, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Detection Results Visualization. Figure 8 provides a visual comparison of the detection results of different models in real-world scenarios. Our proposed MRCNet outperforms SOTA methods, with well-aligned bounding boxes that closely match the ground truth labels. In contrast, other models such as V2X-ViT [40], DiscoNet [14], and CoBEVT [39] fail to accurately detect fast moving objects accurately, resulting in missed objects or falsely-detected predictions. Our proposed MRCNet exploits the multi-level semantic features of collaborative systems and demonstrates improved robustness against noise interference. In", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "addition, the proposed motion-aware fusion model extracts motion information derived from moving objects, which can effectively mitigate the negative effects of motion blur in complex scenes, thus improving detection accuracy.", + "bbox": [ + 498, + 463, + 892, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 544, + 617, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we study the real-world noise problem in multi-agent collaborative perception systems. We propose a novel framework called motion-aware robust communication network (MRCNet) to mitigate noise interference and achieve robust collaborative perception. MRCNet uses two-stage communication to enable the selection, communication and fusion of multi-level semantic features among agents. To address the problem of motion blur caused by fast-moving objects, a GRU-based module is proposed to capture the motion context and to refine the current BEV feature. Comprehensive experiments on three open-source datasets show that the proposed MRCNet outperforms existing competing methods in noisy scenarios. The further consideration of communication issues and model heterogeneity will be a future direction.", + "bbox": [ + 496, + 571, + 892, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was partially supported by grants from the National Natural Science Foundation of China (No.62293544 and 62301299), and China Postdoctoral Science Foundation (No.2022M720077). We would also like to thank Yue Hu from Shanghai Jiao Tong University for sharing the collaborative perception datasets.", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15308", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hongxiang Cai, Zeyuan Zhang, Zhenyu Zhou, Ziyin Li, Wenbo Ding, and Jiuhua Zhao. Bevfusion4d: Learning lidar-camera fusion under bird's-eye-view via cross-modality guidance and temporal aggregation. arXiv preprint arXiv:2303.17099, 2023. 1", + "[2] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1", + "[3] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 5", + "[4] Rahul Dey and Fathi M Salem. Gate-variants of gated recurrent unit (gru) neural networks. In 2017 IEEE 60th international midwest symposium on circuits and systems (MWS-CAS), pages 1597-1600. IEEE, 2017. 5", + "[5] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6", + "[6] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 772-782, 2022. 4", + "[7] Yushan Han, Hui Zhang, Huifang Li, Yi Jin, Congyan Lang, and Yidong Li. Collaborative perception in autonomous driving: Methods, datasets and challenges. arXiv preprint arXiv:2301.06262, 2023. 1", + "[8] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 1, 2, 3, 6, 7, 8", + "[9] Lingdong Kong, Youquan Liu, Xin Li, Runnan Chen, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Robo3d: Towards robust and reliable 3d perception against corruptions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19994-20006, 2023. 1", + "[10] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. Recent development and applications of sumosimulation of urban mobility. International journal on advances in systems and measurements, 5(3&4), 2012. 6", + "[11] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 6", + "[12] Junhyung Lee, Junho Koh, Youngwoo Lee, and Jun Won Choi. D-align: Dual query co-attention network for 3d object detection based on multi-frame point cloud sequence. In" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9238-9244. IEEE, 2023. 4", + "[13] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4):10914-10921, 2022. 2, 6, 7", + "[14] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 1, 3, 6, 7, 8", + "[15] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1", + "[16] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6", + "[17] Hao Liu, Zequn Jie, Karlekar Jayashree, Meibin Qi, Jianguo Jiang, Shuicheng Yan, and Jiashi Feng. Video-based person re-identification with accumulative motion context. IEEE transactions on circuits and systems for video technology, 28(10):2788-2802, 2017. 5", + "[18] Si Liu, Chen Gao, Yuan Chen, Xingyu Peng, Xianghao Kong, Kun Wang, Runsheng Xu, Wentao Jiang, Hao Xiang, Jiaqi Ma, et al. Towards vehicle-to-everything autonomous driving: A survey on collaborative perception. arXiv preprint arXiv:2308.16714, 2023. 2", + "[19] Yuntao Liu, Qian Huang, Rongpeng Li, Xianfu Chen, Zhifeng Zhao, Shuyuan Zhao, Yongdong Zhu, and Honggang Zhang. Rethinking collaborative perception from the spatial-temporal importance of semantic information. arXiv preprint arXiv:2307.16517, 2023. 2", + "[20] Yu Liu, Zhi Li, Zhizhuo Jiang, and You He. Prospects for multi-agent collaboration and gaming: challenge, technology, and application. Frontiers of Information Technology & Electronic Engineering, 23(7):1002-1009, 2022. 1", + "[21] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 1", + "[22] Yiheng Liu, Zhenxun Yuan, Wengang Zhou, and Houqiang Li. Spatial and temporal mutual promotion for video-based person re-identification. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 8786–8793, 2019. 5", + "[23] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106-4115, 2020. 2, 3, 6, 7", + "[24] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15309", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "perception via learnable handshake communication. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6876-6883. IEEE, 2020. 1, 2", + "[25] Zhe Liu, Xin Zhao, Tengteng Huang, Ruolan Hu, Yu Zhou, and Xiang Bai. Tanet: Robust 3d object detection from point clouds with triple attention. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11677-11684, 2020. 4", + "[26] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 1, 3, 7", + "[27] Guiyang Luo, Hui Zhang, Quan Yuan, and Jinglin Li. Complementarity-enhanced and redundancy-minimized collaboration network for multi-agent perception. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3578-3586, 2022. 1", + "[28] Hangyu Mao, Zhibo Gong, Zhengchao Zhang, Zhen Xiao, and Yan Ni. Learning multi-agent communication under limited-bandwidth restriction for internet packet routing. arXiv preprint arXiv:1903.05561, 2019. 1", + "[29] Donghao Qiao and Farhana Zulkernine. Adaptive feature fusion for cooperative perception using lidar point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1186-1195, 2023. 1, 5", + "[30] Yao Rong, Xiangyu Wei, Tianwei Lin, Yueyu Wang, and Enkelejda Kasneci. Dynstatf: An efficient feature fusion strategy for lidar 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3237-3246, 2023. 1", + "[31] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. A cooperative perception system robust to localization errors. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-6. IEEE, 2023. 3", + "[32] 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", + "[33] Junyong Wang, Yuan Zeng, and Yi Gong. Collaborative 3d object detection for automatic vehicle systems via learnable communications. arXiv preprint arXiv:2205.11849, 2022. 1", + "[34] Tianhang Wang, Guang Chen, Kai Chen, Zhengfa Liu, Bo Zhang, Alois Knoll, and Changjun Jiang. Umc: A unified bandwidth-efficient and multi-resolution based collaborative perception framework. arXiv preprint arXiv:2303.12400, 2023. 1, 2", + "[35] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 605–621. Springer, 2020. 1, 3, 6, 7", + "[36] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Hao Xiang, Runsheng Xu, Xin Xia, Zhaoliang Zheng, Bolei Zhou, and Jiaqi Ma. V2xp-asg: Generating adversarial scenes for vehicle-to-everything perception. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3584-3591. IEEE, 2023. 1", + "[38] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6", + "[39] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. arXiv preprint arXiv:2207.02202, 2022. 1, 4, 6, 7, 8", + "[40] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 1, 2, 3, 4, 6, 7, 8", + "[41] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3, 4, 5, 6, 7", + "[42] Yunshuang Yuan, Hao Cheng, and Monika Sester. Keypoints-based deep feature fusion for cooperative vehicle detection of autonomous driving. IEEE Robotics and Automation Letters, 7(2):3054-3061, 2022. 3", + "[43] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3085-3094, 2019. 3", + "[44] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 5", + "[45] Lei Zhu, Zijun Deng, Xiaowei Hu, Chi-Wing Fu, Xuemiao Xu, Jing Qin, and Pheng-Ann Heng. Bidirectional feature pyramid network with recurrent attention residual modules for shadow detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 121–136, 2018. 3", + "[46] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4, 7" + ], + "bbox": [ + 501, + 92, + 890, + 724 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15310", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/8f321c57-13fe-4536-adc3-b4b1a5e6a84e_model.json b/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/8f321c57-13fe-4536-adc3-b4b1a5e6a84e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..45b356abb7f38fa9935d2e4bd83c1c44a79271bd --- /dev/null +++ b/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/8f321c57-13fe-4536-adc3-b4b1a5e6a84e_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.226, + 0.131, + 0.745, + 0.175 + ], + "angle": 0, + "content": "Multi-agent Collaborative Perception via Motion-aware Robust Communication Network" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.203, + 0.704, + 0.221 + ], + "angle": 0, + "content": "Shixin Hong\\(^{1}\\), Yu Liu\\(^{2\\dagger}\\), Zhi Li\\(^{1}\\), Shaohui Li\\(^{1}\\), You He\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.221, + 0.735, + 0.24 + ], + "angle": 0, + "content": "1Shenzhen International Graduate School, Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.24, + 0.725, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)Department of Electronic Engineering, Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.26, + 0.71, + 0.275 + ], + "angle": 0, + "content": "hsx22@mails.tsinghua.edu.cn, liuyu77360132@126.com," + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.278, + 0.776, + 0.293 + ], + "angle": 0, + "content": "{zhilizl, lishaohui}@sz.tsinghua.edu.cn, heyou@mail.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.473, + 0.706 + ], + "angle": 0, + "content": "Collaborative perception allows for information sharing between multiple agents, such as vehicles and infrastructure, to obtain a comprehensive view of the environment through communication and fusion. Current research on multi-agent collaborative perception systems often assumes ideal communication and perception environments and neglects the effect of real-world noise such as pose noise, motion blur, and perception noise. To address this gap, in this paper, we propose a novel motion-aware robust communication network (MRCNet) that mitigates noise interference and achieves accurate and robust collaborative perception. MRCNet consists of two main components: multi-scale robust fusion (MRF) addresses pose noise by developing cross-semantic multi-scale enhanced aggregation to fuse features of different scales, while motion enhanced mechanism (MEM) captures motion context to compensate for information blurring caused by moving objects. Experimental results on popular collaborative 3D object detection datasets demonstrate that MRCNet outperforms competing methods in noisy scenarios with improved perception performance using less bandwidth. Our code will be released at https://github.com/IndigoChildren/collaborative-perception-MRCNet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.717, + 0.21, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Collaborative perception enables information sharing among multiple agents, including vehicles and infrastructure, to provide a comprehensive view of the environment through information communication and fusion [37]. This helps individual autonomous vehicles to tackle challenging perception scenarios, such as physical occlusion or long-range detection [27]. Collaborative perception technology has gained significant attention due to its advantages [20]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.864, + 0.469, + 0.879 + ], + "angle": 0, + "content": "Previous works on multi-agent collaborative perception" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.329, + 0.892, + 0.555 + ], + "angle": 0, + "content": "systems focus on improving perception performance by designing superior fusion modules [2, 14, 29, 35] and effective communication strategies [8, 24, 33]. For example, the CoBEVT [39] framework uses sparse transformers [32] to capture local and global spatial interactions across agents. The UMC [34] framework optimises communication by using a two-stage entropy-based selection mechanism. However, these frameworks assume ideal communication and perception environments [7], while real-world scenarios are affected by noise, such as pose noise in vehicle localization, perception noise in raw data [21], and the motion blur in fast-moving object detection. These noises cannot be ignored for real-world applications of multi-agent collaborative systems. Unfortunately, studies on robust collaborative perception with real-world noise are largely limited." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.893, + 0.766 + ], + "angle": 0, + "content": "In fact, different forms of noise in the real world can affect the performance of collaborative perception systems in different ways. For example, when noise interferes with the positioning system, it affects the accuracy of pose information. Pose information is vital for fusing the perception data (e.g., raw data, intermediate features, or detection outputs) of the collaborative agents in different coordinates [40]. Thus, pose noise can lead to information misalignment and inconsistency among agents [26] during data fusion, resulting in a reduced system performance. Perception noise introduces distortions into the sensor data, which affects the final performance of the perception system. In addition, the presence of motion blur due to rapid object movement affects the accuracy of target detection [1,9,15,30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.895, + 0.902 + ], + "angle": 0, + "content": "However, the development of a robust collaborative perception system against real-world noise is complicated by the following challenges. First, a robust fusion strategy that adequately processes the multi-source perception data is required to deal with the perceptual misalignment caused by pose noise. Second, a feature selection strategy that filters out highly informative features is crucial to reduce bandwidth consumption and improve the efficiency of multi-agent collaboration [28]. Finally, it is desirable to find a sta" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.273, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author: Yu Liu." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15301" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.087, + 0.891, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.359, + 0.895, + 0.403 + ], + "angle": 0, + "content": "Figure 1. The overview of our proposed MRCNet framework. In MRCNet, agents exchange highly informative semantic features in two stages through multiple rounds of communication to mitigate the effect of noise. Meanwhile, the ego agent uses motion context extracted from historical sequences in the memory bank to reduce the effect of motion blur from moving objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.439 + ], + "angle": 0, + "content": "ble and accurate feature extraction that eliminates the motion blur of rapidly moving objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.471, + 0.862 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we present MRCNet, a motion-aware robust communication network that mitigates noise interference and achieves more robust and accurate collaborative perception. As shown in Figure 1, MRCNet facilitates effective information collaboration among agents and improves the perception performance of the ego agent through two main components, i.e., multiscale robust fusion (MRF) and motion enhanced mechanism (MEM). Specifically, for the challenge of feature misalignment due to collaborators' pose noise, MRF improves the robustness and performance of collaborative perception through multi-scale semantic feature fusion. For perception noise, we filter out highly informative semantic features for multi-round communication based on channel-wise selection operations, while reducing communication bandwidth consumption. In addition, to address the problem of motion blur due to object movement, we introduce the MEM module, which is based on the gate recurrent unit (GRU) architecture and is designed to capture and incorporate the motion context of the ego agent. This mechanism aggregates historical information to refine the features of the current frame. To validate the effectiveness of our MRCNet framework, we carry out a comprehensive evaluation on three open-source collaborative 3D object detection datasets, including V2XSim [13], OPV2V [41], and V2XSet [40]. The experimental results show that MRCNet outperforms competing methods in noisy scenarios. Our main contributions can be summarised as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- We present a focused study of the real-world noise problem in the collaborative perception task." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.41, + 0.892, + 0.469 + ], + "angle": 0, + "content": "- We propose a motion-aware collaborative perception framework, i.e., MRCNet, which presents a novel multi-scale feature communication and fusion strategy to improve the robustness of collaboration." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.478, + 0.892, + 0.539 + ], + "angle": 0, + "content": "- Collaborative experiments are conducted on three large open-source datasets, proving that our proposed MRCNet can achieve the best perception performance in noisy scenarios with less bandwidth." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.41, + 0.892, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.551, + 0.642, + 0.567 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.576, + 0.875, + 0.592 + ], + "angle": 0, + "content": "2.1. Communication in Collaborative Perception" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Previous works [8, 23, 24] focus on solving the communication problem, as excessive communication bandwidth consumption can lead to communication congestion and latency. To address this challenge, most research considers how to achieve a better trade-off between performance and bandwidth. There are two main methods to address the challenge of bandwidth constraints: compressing the communication volume or reducing the number of communicators. Several works reduce the communication volume by using learnable methods to facilitate efficient collaboration [18]. Where2comm [8] focuses on perceptually critical areas and transmits foreground features with high confidence. UMC [34] designs a two-stage communication strategy using extended entropy theory to filter out low quality regions. As for selecting of effective collaborators, Who2Com [24] proposes a multi-stage handshake communication mechanism that determines which agents to connect with, and Select2Col [19] introduces a collaborator selection method that selects contributing collaborators with low latency. However, there is still a lack of research on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15302" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "the problem of pose noise in collaborative communication. Therefore, we propose the multi-scale robust fusion module, which selects highly informative features from multiscale features for communication and aggregation, thereby reducing the effect of multi-source noise in collaboration." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.175, + 0.469, + 0.192 + ], + "angle": 0, + "content": "2.2. Information Fusion in Collaborative Perception" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.473, + 0.608 + ], + "angle": 0, + "content": "In addition to addressing communication issues, collaborative perception systems focus on designing collaboration modules to fuse information from multiple agents. The goal is to improve perception capabilities and thereby optimise performance in downstream tasks. Based on the stage of data sharing and collaboration, the collaborative perception can be broadly divided into three categories [14, 40], i.e., early fusion [42], intermediate fusion [8, 23, 35, 41], late fusion [26, 31]. Early fusion involves sharing raw data for a comprehensive view, but requires significant bandwidth and overlooks contextual information. Intermediate fusion extracts intermediate features from each agent's observations before transmission, balancing perception performance and communication bandwidth usage. Late fusion shares detection outputs among agents, but produces suboptimal results due to individual inaccuracies or incomplete perceptions. Among them, intermediate fusion methods gain popularity due to their performance-bandwidth trade-off [35, 40, 41]. V2VNet [35] proposes a spatially aware graph neural network to aggregate features from multiple agents. OPV2V [41] uses a global self-attention module for feature fusion. As an extension of intermediate fusion, our study introduces a novel cross-semantic fusion module, which is designed to aggregate features in the presence of feature misalignment caused by pose noise. Furthermore, we integrate motion context into the perception system to improve the understanding of the environment." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.618, + 0.169, + 0.634 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.794 + ], + "angle": 0, + "content": "In this paper, we focus on improving the robustness and accuracy of collaborative perception in complex collaborative scenarios. To achieve this, we propose a novel motion-aware robust communication network (MRCNet) that integrates multi-scale feature selection, effective information fusion, and motion-aware feature extraction into collaborative perception. As shown in Figure 1, MRCNet consists of four key components: feature extraction and selection, multi-scale robust collaboration, motion enhanced mechanism, and the 3D object detection head." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.802, + 0.281, + 0.817 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "At time \\( t \\), in a collaborative perception task involving \\( N \\) agents, the selected \\( i \\)-th agent is defined as the ego agent, and a communication graph is established among them. Other agents within the communication range serve as collaborators, providing complementary information and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.232 + ], + "angle": 0, + "content": "helping the ego agent to make a comprehensive perception. Let \\(\\mathcal{X}_i^t\\) be the point cloud observation and \\(\\mathcal{Y}_i^t\\) be the perception supervision of the ego agent. Each collaborator provides the ego agent with the information extracted from their local observations \\(\\mathcal{X}_j^t\\). Additionally, considering that we aim at modelling the motion-aware collaborations, we design a memory bank of the ego agent that stores \\(K\\) frames of historical point data, denoted as \\(\\{\\mathcal{X}_i^{t - k}\\}_{k = 1}^K = \\{\\mathcal{X}_i^{t - 1},\\mathcal{X}_i^{t - 2},\\dots,\\mathcal{X}_i^{t - K}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.232, + 0.892, + 0.26 + ], + "angle": 0, + "content": "Thus, the motion-aware collaborative perception task can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.27, + 0.892, + 0.312 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {Y}} _ {i} ^ {t} = \\sum_ {i = 1} ^ {N} \\left\\{\\mathcal {Y} _ {i} ^ {t} | \\Omega \\left(\\mathcal {X} _ {i} ^ {t}, \\left\\{\\mathcal {X} _ {j} ^ {t} \\right\\} _ {j \\neq i} ^ {N}, \\left\\{\\mathcal {X} _ {i} ^ {t - k} \\right\\} _ {k = 1} ^ {K}\\right) \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.894, + 0.385 + ], + "angle": 0, + "content": "where \\(\\Omega (\\cdot)\\) is the proposed MRCNet. In this paper, we use the 3D object detection task to evaluate the performance of collaborative perception. Therefore, \\(\\hat{\\mathcal{V}}_i^t\\) are defined as the predicted detection boxes along this line." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.394, + 0.788, + 0.408 + ], + "angle": 0, + "content": "3.2. Feature Extraction and Selection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.585 + ], + "angle": 0, + "content": "Feature Extraction. After the ego agent constructs a vehicle-to-everything (V2X) collaboration graph connecting neighbouring agents at the current frame \\( t \\), it exchanges pose data between agents for alignment purposes. The collaborators then transform their raw point cloud data into the ego agent's coordination using the transformation matrix \\( \\Lambda \\) computed based on the agents' unique pose \\( \\boldsymbol{\\xi}^t \\). Next, each agent uses a shared feature extractor consisting of \\( S \\) layers [41] to obtain multi-scale features from the raw data. This process yields a set of features \\( \\mathcal{F}_i^t = \\left\\{\\mathcal{F}_i^{s,t}\\right\\}_{s=1}^S \\) for the \\( i \\)-th agent, where \\( i \\) ranges from 1 to \\( N \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.892, + 0.706 + ], + "angle": 0, + "content": "Feature Selection. In contrast to previous methods that transmit all features, we introduce a novel feature selection technique. Perception noise introduces distortions into the collaborative features, thereby misleading the collaborative system. Therefore, our feature selection technique selects highly informative features from the multi-scale features for communication, effectively reducing the effect of noise interference and conserving communication bandwidth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.707, + 0.892, + 0.873 + ], + "angle": 0, + "content": "The features extracted from the foreground regions are particularly important due to their rich semantic information [43,45]. To select highly informative features for effective communication, we use channel pooling operations to evaluate the semantic density within each feature. Initially, different scales of features are spatially flattened and then concatenated into a feature sequence \\( \\pmb{E}^{t} \\in \\mathbb{R}^{C \\times L} \\), where \\( C \\) is the number of channels of the feature, \\( L \\) is the length of the multi-scale feature after flatten. Subsequently, we perform channel-wise pooling operations on \\( \\pmb{E}^{t} \\) to generate the semantic confidence map \\( m \\):" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.884, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {m} = \\left(\\operatorname {M a x P o o l} \\left(\\boldsymbol {E} ^ {t}\\right) + \\operatorname {A v g P o o l} \\left(\\boldsymbol {E} ^ {t}\\right)\\right) \\in \\mathbb {R} ^ {1 \\times L}, \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15303" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.089, + 0.468, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.276, + 0.47, + 0.318 + ], + "angle": 0, + "content": "Figure 2. The architecture of the proposed MEA component. (a) shows the improved pyramid sampling aggregation module. (b) shows the multi-scale deformable self-attention module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.469, + 0.379 + ], + "angle": 0, + "content": "where MaxPool \\((\\cdot)\\) and AvgPool \\((\\cdot)\\) refer to the channelwise max pooling and average pooling operations. The binary selection matrix for communication is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.386, + 0.469, + 0.405 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {c} = \\Phi_ {\\text {s e l e c t}} (\\boldsymbol {m}) \\in \\{0, 1 \\} ^ {1 \\times L}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.413, + 0.469, + 0.504 + ], + "angle": 0, + "content": "where \\(\\Phi_{\\mathrm{select}}(\\cdot)\\) is the selection function to select the most informative features for communication. The selected features with high information are then obtained as \\(D^{t} = (c\\odot E^{t})\\in \\mathbb{R}^{C\\times L}\\), where \\(\\odot\\) is the element-wise multiplication. The selected feature map provides highly informative features with less bandwidth usage." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.512, + 0.318, + 0.527 + ], + "angle": 0, + "content": "3.3. Multi-Scale Robust Fusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.469, + 0.777 + ], + "angle": 0, + "content": "Cross-agent collaboration improves the visual representation and perception performance of the ego agent by exploiting shared semantic information. Previous attention-based [39-41] approaches overlook to consider the robustness of the collaboration system to noise. Therefore, we introduce a novel multi-scale robust fusion (MRF) module, which is designed for the robust multi-scale feature fusion through two-stage communication. Multi-scale feature fusion combines information of different semantics to improve perception capability and robustness to pose noise. The two-stage communication module consists of multiscale enhanced aggregation (MEA) and collaborative features aggregation (CFA). MEA uses cross-semantic fusion to overcome the challenge of feature misalignment caused by pose noise, and CFA enables the ego agent to maximise the perception information aggregation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.778, + 0.469, + 0.906 + ], + "angle": 0, + "content": "Multi-scale Enhanced Aggregation. In the first stage, the ego agent broadcasts highly informative features to the collaborators. For the sake of clarity, let's consider a communication process involving only two agents. Upon receiving highly informative features, the collaborator first converts these features from sequence to multi-scale representation, then performs a preliminary fusion using the max fusion operation to obtain the multi-scale feature \\(\\mathcal{U}_j^t = \\{\\mathcal{U}_j^{s,t}\\}_{s=1}^S\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "To address the challenge of feature misalignment, MEA incorporates cross-semantic modules to fully exploit the strengths of different semantics [6]. As shown in Figure 2, MEA consists of two components: an improved pyramid sampling aggregation (IPSA) module and a multi-scale deformable self-attention (MDSA) module [46]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.893, + 0.381 + ], + "angle": 0, + "content": "Low-level features capture fine-grained details, while high-level semantic features provide integrated contextual understanding. The pyramid sampling aggregation module [12, 25] uses simple convolution and channel-wise concatenation operations to fuse multi-scale features. This process gathers information through cross-semantic fusion, resulting in a more comprehensive and robust feature representation that effectively mitigates noise interference. We pyramid sample the feature \\(\\mathcal{U}_j^{s,t}\\) to obtain the new multiscale feature \\(\\left\\{\\mathcal{U}_j^{sl,t}\\right\\}_{l=1}^S\\), where \\(\\mathcal{U}_j^{sl,t}\\) is the transformation of \\(\\mathcal{U}_j^{s,t}\\) from the \\(s\\)-th scale to the \\(l\\)-th scale, and the feature sizes of the new multi-scale feature correspond to \\(\\mathcal{U}_j^t\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.485 + ], + "angle": 0, + "content": "After multiple rounds of communication, the fine details in the transmitted features tend to be compromised due to cross-semantic fusion during the pyramid sampling. Therefore, we improve the pyramid sampling aggregation module by adding original multi-scale features from collaborators to compensate for detailed features, which is the proposed IPSA module. The output can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.498, + 0.496, + 0.891, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {U}} _ {j} ^ {s, t} = \\operatorname {c o n v} _ {1 \\times 1} \\left(\\left[ \\mathcal {U} _ {j} ^ {s 1, t}, \\mathcal {U} _ {j} ^ {s 2, t}, \\mathcal {U} _ {j} ^ {s 3, t} \\right]\\right) + \\mathcal {U} _ {j} ^ {t}, s \\in [ 1, 2, 3 ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.565 + ], + "angle": 0, + "content": "where \\([\\cdot ,\\cdot ,\\cdot ]\\) is the channel-wise concatenation, and \\(\\mathrm{conv}_{1\\times 1}(\\cdot)\\) is the \\(1\\times 1\\) convolution operation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.893, + 0.686 + ], + "angle": 0, + "content": "Following the IPSA module, the collaborator applies the MDSA [46] module to further fuse multi-scale features and mitigate the effect of pose noise on feature distortion. The MSDA module consists of three linear layers \\( \\mathbf{W}_a \\), \\( \\mathbf{W}_v \\), \\( \\mathbf{W}_o \\), which separately compute the sampling offset \\( \\Delta q \\), the attention weight \\( \\mathbf{A} \\) and the value \\( \\mathbf{V} \\) at position \\( q \\), where \\( q \\) is the two-dimensional reference point in \\( \\hat{\\mathcal{U}}_j^{s,t} \\). The output of MSDA at position \\( q \\) can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.697, + 0.656, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M S D A} (q) = \\sum_ {h = 1} ^ {H} \\boldsymbol {W} _ {h}.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.741, + 0.913, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\left(\\sum_ {s = 1} ^ {S} \\sum_ {p = 1} ^ {P} \\psi \\left(\\boldsymbol {W} _ {a} \\tilde {\\mathcal {U}} _ {j} ^ {s, t} (q)\\right) \\cdot \\boldsymbol {W} _ {v} \\left(\\tilde {\\mathcal {U}} _ {j} ^ {s, t} \\left(q + \\boldsymbol {W} _ {o} \\tilde {\\mathcal {U}} _ {j} ^ {s, t} (q)\\right)\\right)\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.786, + 0.892, + 0.844 + ], + "angle": 0, + "content": "\\[\n= \\sum_ {h = 1} ^ {H} \\boldsymbol {W} _ {h} \\cdot \\left(\\sum_ {s = 1} ^ {S} \\sum_ {p = 1} ^ {P} \\boldsymbol {A} ^ {h s p} (q) \\cdot \\mathcal {V} ^ {h s} \\left(q + \\Delta^ {h s p} (q)\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(H\\) is the number of attention heads, \\(S\\) is the quantity of different feature scales, and \\(P\\) is the number of sampling points utilized in each scale. \\(\\mathbf{W}_h\\) is the projection matrix" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15304" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.142 + ], + "angle": 0, + "content": "and the softmax function \\(\\psi (\\cdot)\\) is used to determine the attention weight. Ultimately, we output the multi-scale format feature \\(\\mathcal{C}_j^t = \\{\\mathcal{C}_j^{s,t}\\}_{s = 1}^S\\) after MSDA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.142, + 0.473, + 0.436 + ], + "angle": 0, + "content": "Collaborative Features Aggregation. In the second stage of communication, the \\(j\\)-th collaborator uses the same feature selection technique to send highly informative features \\(\\pmb{D}_j^t = (\\pmb{m}_j \\odot \\pmb{E}_j^t)\\) back to the ego agent, where \\(\\pmb{E}_j^t\\) is the feature sequence representation of \\(\\mathcal{C}_j^t\\). In order to maximise the aggregation of the semantic information provided by the collaborator, inspired by [29], the ego agent uses the spatial-wise adaptive fusion module on each scale of features. The ego agent first stacks the feature maps of the agents to form \\(\\mathcal{A}_i^t = \\{\\mathcal{A}_i^{s,t}\\}_{s=1}^S\\), \\(\\mathcal{A}_i^{s,t} \\in \\mathbb{R}^{n \\times C \\times H_s \\times W_s}\\), where \\(s\\) is the scale index and \\(n\\) is the maximum number of agents. Max and average pooling is applied to \\(\\mathcal{A}_i^{s,t}\\) over the first channel axis to create new feature maps \\(\\mathcal{A}_{i,\\max}^{s,t} \\in \\mathbb{R}^{1 \\times C \\times H_s \\times W_s}\\) and \\(\\mathcal{A}_{i,\\mathrm{avg}}^{s,t} \\in \\mathbb{R}^{1 \\times C \\times H_s \\times W_s}\\), where \\(H_s\\) and \\(W_s\\) represent the height and width of the feature. These pooled feature maps are concatenated and passed through a 3D convolutional layer with a ReLU activation function to perform further fusion. The aggregated features provide complementary information to the ego agent." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.45, + 0.345, + 0.465 + ], + "angle": 0, + "content": "3.4. Motion Enhanced Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.47, + 0.581 + ], + "angle": 0, + "content": "As mentioned above, the collaborative perception module increases the range of perception and provides compensation for occlusions [41], but it cannot directly address the motion blur introduced by moving objects. To combat this, we introduce the motion enhanced mechanism (MEM), which uses motion context from historical point cloud frames to achieve motion-aware perception." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.779 + ], + "angle": 0, + "content": "When the ego agent initiates a collaboration request, it projects the last \\(K\\) frames of the point cloud stored in the memory bank to its current coordinate, producing a series of bird's eye view (BEV) feature maps \\(\\{\\mathcal{B}^{t - 1},\\dots,\\mathcal{B}^{t - K}\\}\\). Inspired by [4,22], we design a recurrent unit called the gated motion unit (GMU). This unit aggregates motion context to enhance frame-level features. Figure 3 illustrates the operation of GMU at timestamp \\(t\\), showing that GMU takes three inputs stored in the ego agent's memory bank: the feature \\(\\mathcal{B}^{t - 1}\\) of the previous frame, the feature \\(\\mathcal{B}^t\\) of the current frame, and the enhanced feature \\(\\mathcal{H}^{t - 1}\\). Both \\(\\mathcal{B}^0\\) and \\(H^0\\) are initialized as \\(\\mathcal{B}^1\\), and the output \\(\\mathcal{H}^t\\) represents the enhanced feature of the current frame after passing through GMU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Between successive frames, object motion causes spatial transitions of features. Therefore, we use \\(\\mathcal{M}^t = \\mathcal{B}^t -\\mathcal{B}^{t - 1}\\) to capture motion context, which can represent the motion clues of moving objects [17]. To reduce the effect of motion blur and to exploit historical information, we use a deformable convolution network [3]. The deformable convolution network can better capture the features of objects by adaptively adjusting the kernel size and shape. Specifically," + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.089, + 0.885, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.232, + 0.878, + 0.247 + ], + "angle": 0, + "content": "Figure 3. The architecture of the proposed GMU component." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.262, + 0.892, + 0.323 + ], + "angle": 0, + "content": "the \\(3 \\times 3\\) convolutional layer is used on the motion context \\(\\mathcal{M}_t\\) to obtain the spatial offsets \\(\\Delta p_n^t\\). The convolutional operation is then applied to \\(\\mathcal{B}^{t-1}\\) to extract historical features at position \\(p\\). The output feature map \\(\\mathcal{I}^t\\) at location \\(p\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.33, + 0.892, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} ^ {t} (p) = \\sum_ {p _ {n} \\in \\mathcal {R}} \\boldsymbol {W} \\left(p _ {n}\\right) \\cdot \\mathcal {B} ^ {t - 1} \\left(p + p _ {n} + \\Delta p _ {n} ^ {t}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.37, + 0.892, + 0.4 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}\\) is the convolution weight, \\(\\mathcal{R}\\) is the kernel grid and \\(p_n\\) enumerates the locations in \\(\\mathcal{R}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.401, + 0.893, + 0.536 + ], + "angle": 0, + "content": "To reduce the motion blur effect caused by the moving object between frames, it is necessary to determine whether the features at each position should be updated. Inspired by [36], we design an update gate \\(\\mathcal{G}\\) using the channel-spatial attention module. The channel-spatial attention module is applied to the stack feature of \\(\\mathcal{I}^t\\) and \\(\\mathcal{H}^{t - 1}\\). After applying the sigmoid operation, the overall attention map is normalised to a range of values between 0 and 1, as formulated below:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.543, + 0.892, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} = \\sigma (\\Gamma ([ \\mathcal {I} ^ {t}, \\mathcal {H} ^ {t - 1} ])) \\in \\mathbb {R} ^ {C \\times H \\times W}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.661 + ], + "angle": 0, + "content": "where \\([\\cdot, \\cdot]\\) is the channel-wise concatenation, \\(\\Gamma(\\cdot)\\) is the channel-spatial attention module, and \\(\\sigma(\\cdot)\\) is the sigmoid function for value normalisation. The value in \\(\\mathcal{G}\\) indicates the probability that the corresponding position of \\(\\mathcal{H}^{t-1}\\) is reserved. Then the improved feature \\(\\mathcal{H}^t\\) is refined with the previous improved feature \\(\\mathcal{H}^{t-1}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.669, + 0.892, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} ^ {t} = \\mathcal {G} \\odot \\mathcal {H} ^ {t - 1} + (1 - \\mathcal {G}) \\odot \\mathcal {I} ^ {t}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.693, + 0.862, + 0.71 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}^t\\in \\mathbb{R}^{C\\times H\\times W}\\) is the final output of the GMU." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.738, + 0.733 + ], + "angle": 0, + "content": "3.5. 3D Object Detection Head" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.741, + 0.892, + 0.877 + ], + "angle": 0, + "content": "After obtaining the fused output after MRF, the multiscale feature is encoded to the same size and concatenated to \\(\\mathcal{Z}_i^t\\in \\mathbb{R}^{C\\times H\\times W}\\). Then, we concatenate \\(\\mathcal{Z}_i^t\\) and \\(\\mathcal{H}^t\\) along the channel dimension to derive the final representation \\(\\mathcal{P}_i^t\\). To generate the final object detection results, we then use a dual-branch convolutional module on \\(\\mathcal{P}_i^t\\) to perform classification and regression tasks for each detection anchor. Similar to the previous study [44], the formulation for the object detection loss \\(L_{\\mathrm{det}}\\) is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.886, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d e t}} = L _ {\\mathrm {c l s}} + \\beta L _ {\\mathrm {r e g}}, \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15305" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.35, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.09, + 0.61, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.091, + 0.868, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.204, + 0.2, + 0.776, + 0.21 + ], + "angle": 0, + "content": "MRCNet(Ours) Where2comm V2X-ViT Late Fusion When2com No Fusion Early Fusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.224, + 0.812, + 0.239 + ], + "angle": 0, + "content": "Figure 4. Robustness of our proposed method and compared benchmark to varying levels of localization noise." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.248, + 0.349, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.25, + 0.61, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.251, + 0.869, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.212, + 0.359, + 0.769, + 0.37 + ], + "angle": 0, + "content": "MRC(Ours) Where2comm V2X-ViT Late Fusion When2com No Fusion Early Fusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.385, + 0.8, + 0.4 + ], + "angle": 0, + "content": "Figure 5. Robustness of our proposed method and compared benchmark to varying levels of heading noise." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.417, + 0.47, + 0.464 + ], + "angle": 0, + "content": "where \\( L_{\\mathrm{cls}} \\) is the cross-entropy loss for object classification, \\( L_{\\mathrm{reg}} \\) is the smooth L1 loss for box regression [16], and \\( \\beta \\) is the balancing hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.478, + 0.21, + 0.495 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.64 + ], + "angle": 0, + "content": "To assess the robustness of our module in noisy collaboration scenarios, we conduct experiments on three large open-source datasets. In particular, our experiments focus on LiDAR-based 3D object detection. We measure detection performance using average precision (AP) at intersection-over-union (IoU) thresholds of 0.50 and 0.70, and note that the AP values are given as percentages. To ensure a fair comparison, we exclude data and feature compression operations in the compared methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.387, + 0.667 + ], + "angle": 0, + "content": "4.1. Datasets and Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Datasets. V2XSim [13] is a V2X collaborative perception dataset co-simulated by SUMO [10] and CARLA [5]. It contains 10000 frames of LiDAR point clouds and 501K annotated 3D boxes. There are 8000/1000/1000 frames for the training/validation/testing, respectively. The detection range of the agent is \\( x \\in [-32m, 32m] \\), \\( y \\in [-32m, 32m] \\). OPV2V [41] is a vehicle-to-vehicle collaborative perception dataset co-simulated by OpenCDA [38] and CARLA [5], including 11464 frames of 3D point clouds with 230K annotated 3D boxes, split into training/validation/testing of 6764/1981/2719 frames. The detection range is \\( x \\in [-140m, 140m] \\), \\( y \\in [-40m, 40m] \\). V2XSet [40] is the first open dataset that includes both V2X cooperation and realistic noise simulation, which is co-simulated by OpenCDA [38] and CARLA [5]. There" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.462 + ], + "angle": 0, + "content": "are a total of 11447 frames in the dataset, and the train/Validation/testing splits are 6694/1920/2833, respectively. The detection range setting is the same as OPV2V." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.893, + 0.734 + ], + "angle": 0, + "content": "Implementation Details. In our implementation, we use PointPillars [11] as the multi-scale feature extractor, with the grid size of \\((0.4m, 0.4m)\\) for the point cloud discretisation. And we get intermediate features in three scales, which is \\(S = 3\\) as we define in the method. The MSDA module uses 8 attention heads and 4 sampling points per communication round. For V2XSim, we extract information from three historical frames, while for V2XSet and OPV2V, we use two historical frames. To train the detection model, we set the hyperparameter in the loss function to \\(\\beta = 2\\) according to [11]. We use the Adam optimizer with an initial learning rate of \\(\\{1\\mathrm{e} - 3, 5\\mathrm{e} - 4, 3\\mathrm{e} - 4\\}\\) for the OPV2V, V2XSet and V2XSim datasets. The learning rate is gradually reduced at epochs 35 and 45 on the V2XSim dataset using a decay factor of 0.1. For the OPV2V and V2XSet datasets, the decay occurs at epochs 30 and 40. All models are trained on 2 NVIDIA RTX 3090 GPUs with a batch size of 2 for a total of 50 epochs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Benchmark Comparison. Our considering benchmark includes early fusion, intermediate fusion, and late fusion techniques as well as single agent perception for comparison. In addition, we compare with six state-of-the-art (SOTA) intermediate fusion models: When2Com [23], V2VNet [35], AttFuse [41], V2X-ViT [40], DiscoNet [14], CoBEVT [39], and Where2Comm [8]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.694, + 0.864 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The pose of the agent can be represented by a 6D vector. Since the agent only has yaw angle to measure rotation," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15306" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.089, + 0.47, + 0.117 + ], + "angle": 0, + "content": "Table 1. Performance comparison on the V2XSim, OPV2V, and V2XSet datasets. The results are reported in AP0.5/0.7." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.122, + 0.468, + 0.299 + ], + "angle": 0, + "content": "
ModelV2XSimOPV2VV2XSet
AP0.5AP0.7AP0.5AP0.7AP0.5AP0.7
No Fusion65.7352.5769.3856.4064.8850.83
Late Fusion71.2256.9982.7661.6476.3153.33
Early Fusion84.9465.9789.2772.9683.4051.71
When2com [23]73.3652.5282.7867.6674.7340.62
V2VNet [35]82.7065.3188.6974.9384.7964.73
AttFuse [41]81.7066.2488.5472.9184.3766.27
V2X-ViT [40]82.3264.4186.7475.7082.4263.14
DiscoNet [14]83.5666.1288.0572.0782.3464.79
CoBEVT [39]81.0065.0688.9972.8084.8465.14
Where2comm [8]77.5757.0286.5868.9780.8358.72
MRCNet85.3369.8289.7776.1285.0066.31
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.47, + 0.481 + ], + "angle": 0, + "content": "we simplify the pose to \\(\\xi = (x,y,\\theta)\\) in 2D space [26], where \\(x, y\\) and \\(\\theta\\) represent the 2D center position and the yaw angle of the agent's accurate global position, respectively. To simulate pose noise in real-world scenarios, we add Gaussian noise with a standard deviation of \\(0.2m\\) and \\(0.2^{\\circ}\\) to the collaborators' poses in the datasets. We compare the performance of 3D object detection with other methods on the datasets, as shown in Table 1. According to the experimental results, despite the introduction of pose noise, our proposed MRCNet module shows superior performance compared to the benchmark on all datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.484, + 0.47, + 0.725 + ], + "angle": 0, + "content": "In addition, to investigate the effect of different components of pose noise on the collaborative perception system, we conduct experiments to evaluate the robustness of the models to localization noise and heading noise, separately. Experiments are also performed to assess the effect of various levels of localization and heading noise. Specifically, we add Gaussian noise with standard \\(\\sigma_{xy} \\in [0.2m, 0.6m]\\), \\(\\sigma_{\\theta} \\in [0.2^{\\circ}, 1.0^{\\circ}]\\) separately to the collaborators' poses. As shown in the Figure 4 and the Figure 5, an increase in Gaussian noise leads to a rapid decrease in AP0.7 values for the other methods. This suggests that misaligned spatial features from collaborators can mislead the collaboration system. However, our method maintains superior performance even under significant pose noise, demonstrating its ability to effectively select and fuse multi-level semantic features while achieving robustness to pose noise." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.739, + 0.273, + 0.755 + ], + "angle": 0, + "content": "4.3. Component Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.468, + 0.854 + ], + "angle": 0, + "content": "Contribution of Major Components. To evaluate the effectiveness of our proposed component, we progressively remove (i) MEM, (ii) FS (feature selection), (iii) IPSA and (iii) MSDA and present the detection precision. Table 2 shows the results of our ablation studies, with each innovative component contributing to the performance gains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Varying Number of History Frames. We evaluate the effect of integrating different numbers of history frames from the ego agent. As shown in Figure 6, integrating multiple" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.89, + 0.117 + ], + "angle": 0, + "content": "Table 2. Results of the ablation study of the proposed core components on the V2XSim, OPV2V and V2XSet datasets." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.122, + 0.89, + 0.19 + ], + "angle": 0, + "content": "
MEMFSISPAMSDAV2XSimOPV2VV2XSet
AP@0.5AP@0.7AP@0.5AP@0.7AP@0.5AP@0.7
80.7764.2885.7874.0183.1263.88
✓81.4667.6486.2974.6383.4464.71
✓✓82.8769.1587.3275.3184.2165.31
✓✓✓83.6069.2188.1175.7784.4365.47
✓✓✓✓85.3369.8289.7776.1285.0066.31
" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.199, + 0.697, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.199, + 0.888, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "Figure 6. Collaborative Performance with varying number of history frames. The MEM module works best when fusing three history frames on the V2XSim dataset, and two frames on the OPV2V and V2XSet datasets." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.398, + 0.697, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.52, + 0.609, + 0.53 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.399, + 0.889, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.52, + 0.805, + 0.53 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.544, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Figure 7. (a) shows the perception performance comparison when using the proposed multi-scale method versus the single-scale approach. (b) is the comparison of collaborative perception performance of our proposed method and other baseline methods with different communication volume." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.701 + ], + "angle": 0, + "content": "frames into the MEM module leads to an improvement in the detection accuracy in the V2XSim [13] dataset. However, once the number of history frames exceeds three, no further improvement is observed, as MEM faces challenges in effectively capturing features with large changes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Effect of Multi-Scale Collaboration. To evaluate the effectiveness of multi-scale feature collaboration in complex scene perception, we separately evaluate the detection accuracy achieved by single-scale and multi-scale feature collaboration. In our implementation, the single-scale feature collaboration is modified by using only a single-scale deformable self-attention module [46] within MEA. The results of this comparison are shown in Figure 7a. By using multi-scale feature fusion, our method exploits the advantages of different semantics for better noise robustness and achieves an improved detection performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparison of Communication Volume. Figure 7b shows the performance-bandwidth trade-off of the proposed" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "15307" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.273, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.241, + 0.274, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.393, + 0.213, + 0.405 + ], + "angle": 0, + "content": "(a) V2X-ViT" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.091, + 0.48, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.241, + 0.48, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.393, + 0.418, + 0.405 + ], + "angle": 0, + "content": "(b) DiscoNet" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.091, + 0.684, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.241, + 0.684, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.393, + 0.623, + 0.405 + ], + "angle": 0, + "content": "(c) CoBEVT" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.092, + 0.888, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.241, + 0.888, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.393, + 0.846, + 0.405 + ], + "angle": 0, + "content": "(d) MRCNet (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.418, + 0.893, + 0.446 + ], + "angle": 0, + "content": "Figure 8. In the qualitative comparison on the V2XSim dataset under pose noise, green and red boxes indicate the ground truths and the detection outputs, respectively. Our method outperforms previous SOTA models by achieving more accurate detection results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.473, + 0.691 + ], + "angle": 0, + "content": "MRCNet and the benchmark at different levels of communication bandwidth consumption. No fusion serves as the baseline, as it does not involve any communication. Our analysis reveals several key trends: (i) Where2Comm [8] initially improves detection performance by transmitting the most critical information, but struggles to filter out noisy data as the transmission volume increases. When Where2Comm [8] transmits all feature information, the global attention mechanism captures the spatial information and improves its effectiveness. (ii) Our proposed MRCNet method consistently shows better detection performance at different bandwidth levels, achieving superior results, and multi-round communication can improve the performance of collaborative perception, but excessive misaligned feature transmission can also cause degradation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.708, + 0.201, + 0.725 + ], + "angle": 0, + "content": "4.4. Case Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Detection Results Visualization. Figure 8 provides a visual comparison of the detection results of different models in real-world scenarios. Our proposed MRCNet outperforms SOTA methods, with well-aligned bounding boxes that closely match the ground truth labels. In contrast, other models such as V2X-ViT [40], DiscoNet [14], and CoBEVT [39] fail to accurately detect fast moving objects accurately, resulting in missed objects or falsely-detected predictions. Our proposed MRCNet exploits the multi-level semantic features of collaborative systems and demonstrates improved robustness against noise interference. In" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.464, + 0.893, + 0.525 + ], + "angle": 0, + "content": "addition, the proposed motion-aware fusion model extracts motion information derived from moving objects, which can effectively mitigate the negative effects of motion blur in complex scenes, thus improving detection accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.619, + 0.56 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.573, + 0.893, + 0.799 + ], + "angle": 0, + "content": "In this paper, we study the real-world noise problem in multi-agent collaborative perception systems. We propose a novel framework called motion-aware robust communication network (MRCNet) to mitigate noise interference and achieve robust collaborative perception. MRCNet uses two-stage communication to enable the selection, communication and fusion of multi-level semantic features among agents. To address the problem of motion blur caused by fast-moving objects, a GRU-based module is proposed to capture the motion context and to refine the current BEV feature. Comprehensive experiments on three open-source datasets show that the proposed MRCNet outperforms existing competing methods in noisy scenarios. The further consideration of communication issues and model heterogeneity will be a future direction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was partially supported by grants from the National Natural Science Foundation of China (No.62293544 and 62301299), and China Postdoctoral Science Foundation (No.2022M720077). We would also like to thank Yue Hu from Shanghai Jiao Tong University for sharing the collaborative perception datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15308" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Hongxiang Cai, Zeyuan Zhang, Zhenyu Zhou, Ziyin Li, Wenbo Ding, and Jiuhua Zhao. Bevfusion4d: Learning lidar-camera fusion under bird's-eye-view via cross-modality guidance and temporal aggregation. arXiv preprint arXiv:2303.17099, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.315 + ], + "angle": 0, + "content": "[3] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[4] Rahul Dey and Fathi M Salem. Gate-variants of gated recurrent unit (gru) neural networks. In 2017 IEEE 60th international midwest symposium on circuits and systems (MWS-CAS), pages 1597-1600. IEEE, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.47, + 0.429 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 772-782, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[7] Yushan Han, Hui Zhang, Huifang Li, Yi Jin, Congyan Lang, and Yidong Li. Collaborative perception in autonomous driving: Methods, datasets and challenges. arXiv preprint arXiv:2301.06262, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[8] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[9] Lingdong Kong, Youquan Liu, Xin Li, Runnan Chen, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Robo3d: Towards robust and reliable 3d perception against corruptions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19994-20006, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[10] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. Recent development and applications of sumosimulation of urban mobility. International journal on advances in systems and measurements, 5(3&4), 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[11] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] Junhyung Lee, Junho Koh, Youngwoo Lee, and Jun Won Choi. D-align: Dual query co-attention network for 3d object detection based on multi-frame point cloud sequence. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9238-9244. IEEE, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[13] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4):10914-10921, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[14] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.894, + 0.333 + ], + "angle": 0, + "content": "[15] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.894, + 0.39 + ], + "angle": 0, + "content": "[16] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.894, + 0.459 + ], + "angle": 0, + "content": "[17] Hao Liu, Zequn Jie, Karlekar Jayashree, Meibin Qi, Jianguo Jiang, Shuicheng Yan, and Jiashi Feng. Video-based person re-identification with accumulative motion context. IEEE transactions on circuits and systems for video technology, 28(10):2788-2802, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.894, + 0.53 + ], + "angle": 0, + "content": "[18] Si Liu, Chen Gao, Yuan Chen, Xingyu Peng, Xianghao Kong, Kun Wang, Runsheng Xu, Wentao Jiang, Hao Xiang, Jiaqi Ma, et al. Towards vehicle-to-everything autonomous driving: A survey on collaborative perception. arXiv preprint arXiv:2308.16714, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.894, + 0.602 + ], + "angle": 0, + "content": "[19] Yuntao Liu, Qian Huang, Rongpeng Li, Xianfu Chen, Zhifeng Zhao, Shuyuan Zhao, Yongdong Zhu, and Honggang Zhang. Rethinking collaborative perception from the spatial-temporal importance of semantic information. arXiv preprint arXiv:2307.16517, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.894, + 0.659 + ], + "angle": 0, + "content": "[20] Yu Liu, Zhi Li, Zhizhuo Jiang, and You He. Prospects for multi-agent collaboration and gaming: challenge, technology, and application. Frontiers of Information Technology & Electronic Engineering, 23(7):1002-1009, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[21] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.894, + 0.8 + ], + "angle": 0, + "content": "[22] Yiheng Liu, Zhenxun Yuan, Wengang Zhou, and Houqiang Li. Spatial and temporal mutual promotion for video-based person re-identification. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 8786–8793, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[23] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106-4115, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[24] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15309" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "perception via learnable handshake communication. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6876-6883. IEEE, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[25] Zhe Liu, Xin Zhao, Tengteng Huang, Ruolan Hu, Yu Zhou, and Xiang Bai. Tanet: Robust 3d object detection from point clouds with triple attention. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11677-11684, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[26] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[27] Guiyang Luo, Hui Zhang, Quan Yuan, and Jinglin Li. Complementarity-enhanced and redundancy-minimized collaboration network for multi-agent perception. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3578-3586, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[28] Hangyu Mao, Zhibo Gong, Zhengchao Zhang, Zhen Xiao, and Yan Ni. Learning multi-agent communication under limited-bandwidth restriction for internet packet routing. arXiv preprint arXiv:1903.05561, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[29] Donghao Qiao and Farhana Zulkernine. Adaptive feature fusion for cooperative perception using lidar point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1186-1195, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.471, + 0.524 + ], + "angle": 0, + "content": "[30] Yao Rong, Xiangyu Wei, Tianwei Lin, Yueyu Wang, and Enkelejda Kasneci. Dynstatf: An efficient feature fusion strategy for lidar 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3237-3246, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[31] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. A cooperative perception system robust to localization errors. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-6. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.471, + 0.636 + ], + "angle": 0, + "content": "[32] 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" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.471, + 0.678 + ], + "angle": 0, + "content": "[33] Junyong Wang, Yuan Zeng, and Yi Gong. Collaborative 3d object detection for automatic vehicle systems via learnable communications. arXiv preprint arXiv:2205.11849, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.471, + 0.746 + ], + "angle": 0, + "content": "[34] Tianhang Wang, Guang Chen, Kai Chen, Zhengfa Liu, Bo Zhang, Alois Knoll, and Changjun Jiang. Umc: A unified bandwidth-efficient and multi-resolution based collaborative perception framework. arXiv preprint arXiv:2303.12400, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[35] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 605–621. Springer, 2020. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[36] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[37] Hao Xiang, Runsheng Xu, Xin Xia, Zhaoliang Zheng, Bolei Zhou, and Jiaqi Ma. V2xp-asg: Generating adversarial scenes for vehicle-to-everything perception. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3584-3591. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[38] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[39] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. arXiv preprint arXiv:2207.02202, 2022. 1, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[40] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.36, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[41] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.445, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[42] Yunshuang Yuan, Hao Cheng, and Monika Sester. Keypoints-based deep feature fusion for cooperative vehicle detection of autonomous driving. IEEE Robotics and Automation Letters, 7(2):3054-3061, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.501, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[43] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3085-3094, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.558, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[44] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[45] Lei Zhu, Zijun Deng, Xiaowei Hu, Chi-Wing Fu, Xuemiao Xu, Jing Qin, and Pheng-Ann Heng. Bidirectional feature pyramid network with recurrent attention residual modules for shadow detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 121–136, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[46] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15310" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/8f321c57-13fe-4536-adc3-b4b1a5e6a84e_origin.pdf b/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/8f321c57-13fe-4536-adc3-b4b1a5e6a84e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..79c330ca9001eeeba09d164c2203869adb6040d9 --- /dev/null +++ b/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/8f321c57-13fe-4536-adc3-b4b1a5e6a84e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d99a73bb4fc008b81b8a5384ccc02e63fda02fa1df9188230330f89a59063df4 +size 2272907 diff --git a/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/full.md b/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/full.md new file mode 100644 index 0000000000000000000000000000000000000000..50754865ba470e6bcea768871d8c89e74ff491ad --- /dev/null +++ b/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/full.md @@ -0,0 +1,332 @@ +# Multi-agent Collaborative Perception via Motion-aware Robust Communication Network + +Shixin Hong $^{1}$ , Yu Liu $^{2\dagger}$ , Zhi Li $^{1}$ , Shaohui Li $^{1}$ , You He $^{2}$ + +1Shenzhen International Graduate School, Tsinghua University + +$^{2}$ Department of Electronic Engineering, Tsinghua University + +hsx22@mails.tsinghua.edu.cn, liuyu77360132@126.com, + +{zhilizl, lishaohui}@sz.tsinghua.edu.cn, heyou@mail.tsinghua.edu.cn + +# Abstract + +Collaborative perception allows for information sharing between multiple agents, such as vehicles and infrastructure, to obtain a comprehensive view of the environment through communication and fusion. Current research on multi-agent collaborative perception systems often assumes ideal communication and perception environments and neglects the effect of real-world noise such as pose noise, motion blur, and perception noise. To address this gap, in this paper, we propose a novel motion-aware robust communication network (MRCNet) that mitigates noise interference and achieves accurate and robust collaborative perception. MRCNet consists of two main components: multi-scale robust fusion (MRF) addresses pose noise by developing cross-semantic multi-scale enhanced aggregation to fuse features of different scales, while motion enhanced mechanism (MEM) captures motion context to compensate for information blurring caused by moving objects. Experimental results on popular collaborative 3D object detection datasets demonstrate that MRCNet outperforms competing methods in noisy scenarios with improved perception performance using less bandwidth. Our code will be released at https://github.com/IndigoChildren/collaborative-perception-MRCNet. + +# 1. Introduction + +Collaborative perception enables information sharing among multiple agents, including vehicles and infrastructure, to provide a comprehensive view of the environment through information communication and fusion [37]. This helps individual autonomous vehicles to tackle challenging perception scenarios, such as physical occlusion or long-range detection [27]. Collaborative perception technology has gained significant attention due to its advantages [20]. + +Previous works on multi-agent collaborative perception + +systems focus on improving perception performance by designing superior fusion modules [2, 14, 29, 35] and effective communication strategies [8, 24, 33]. For example, the CoBEVT [39] framework uses sparse transformers [32] to capture local and global spatial interactions across agents. The UMC [34] framework optimises communication by using a two-stage entropy-based selection mechanism. However, these frameworks assume ideal communication and perception environments [7], while real-world scenarios are affected by noise, such as pose noise in vehicle localization, perception noise in raw data [21], and the motion blur in fast-moving object detection. These noises cannot be ignored for real-world applications of multi-agent collaborative systems. Unfortunately, studies on robust collaborative perception with real-world noise are largely limited. + +In fact, different forms of noise in the real world can affect the performance of collaborative perception systems in different ways. For example, when noise interferes with the positioning system, it affects the accuracy of pose information. Pose information is vital for fusing the perception data (e.g., raw data, intermediate features, or detection outputs) of the collaborative agents in different coordinates [40]. Thus, pose noise can lead to information misalignment and inconsistency among agents [26] during data fusion, resulting in a reduced system performance. Perception noise introduces distortions into the sensor data, which affects the final performance of the perception system. In addition, the presence of motion blur due to rapid object movement affects the accuracy of target detection [1,9,15,30]. + +However, the development of a robust collaborative perception system against real-world noise is complicated by the following challenges. First, a robust fusion strategy that adequately processes the multi-source perception data is required to deal with the perceptual misalignment caused by pose noise. Second, a feature selection strategy that filters out highly informative features is crucial to reduce bandwidth consumption and improve the efficiency of multi-agent collaboration [28]. Finally, it is desirable to find a sta + +![](images/b86db373d2b311de9cce979a13674800279643e0f238b664fd24a7ef6da08f52.jpg) +Figure 1. The overview of our proposed MRCNet framework. In MRCNet, agents exchange highly informative semantic features in two stages through multiple rounds of communication to mitigate the effect of noise. Meanwhile, the ego agent uses motion context extracted from historical sequences in the memory bank to reduce the effect of motion blur from moving objects. + +ble and accurate feature extraction that eliminates the motion blur of rapidly moving objects. + +To address the aforementioned challenges, we present MRCNet, a motion-aware robust communication network that mitigates noise interference and achieves more robust and accurate collaborative perception. As shown in Figure 1, MRCNet facilitates effective information collaboration among agents and improves the perception performance of the ego agent through two main components, i.e., multiscale robust fusion (MRF) and motion enhanced mechanism (MEM). Specifically, for the challenge of feature misalignment due to collaborators' pose noise, MRF improves the robustness and performance of collaborative perception through multi-scale semantic feature fusion. For perception noise, we filter out highly informative semantic features for multi-round communication based on channel-wise selection operations, while reducing communication bandwidth consumption. In addition, to address the problem of motion blur due to object movement, we introduce the MEM module, which is based on the gate recurrent unit (GRU) architecture and is designed to capture and incorporate the motion context of the ego agent. This mechanism aggregates historical information to refine the features of the current frame. To validate the effectiveness of our MRCNet framework, we carry out a comprehensive evaluation on three open-source collaborative 3D object detection datasets, including V2XSim [13], OPV2V [41], and V2XSet [40]. The experimental results show that MRCNet outperforms competing methods in noisy scenarios. Our main contributions can be summarised as follows: + +- We present a focused study of the real-world noise problem in the collaborative perception task. + +- We propose a motion-aware collaborative perception framework, i.e., MRCNet, which presents a novel multi-scale feature communication and fusion strategy to improve the robustness of collaboration. +- Collaborative experiments are conducted on three large open-source datasets, proving that our proposed MRCNet can achieve the best perception performance in noisy scenarios with less bandwidth. + +# 2. Related Work + +# 2.1. Communication in Collaborative Perception + +Previous works [8, 23, 24] focus on solving the communication problem, as excessive communication bandwidth consumption can lead to communication congestion and latency. To address this challenge, most research considers how to achieve a better trade-off between performance and bandwidth. There are two main methods to address the challenge of bandwidth constraints: compressing the communication volume or reducing the number of communicators. Several works reduce the communication volume by using learnable methods to facilitate efficient collaboration [18]. Where2comm [8] focuses on perceptually critical areas and transmits foreground features with high confidence. UMC [34] designs a two-stage communication strategy using extended entropy theory to filter out low quality regions. As for selecting of effective collaborators, Who2Com [24] proposes a multi-stage handshake communication mechanism that determines which agents to connect with, and Select2Col [19] introduces a collaborator selection method that selects contributing collaborators with low latency. However, there is still a lack of research on + +the problem of pose noise in collaborative communication. Therefore, we propose the multi-scale robust fusion module, which selects highly informative features from multiscale features for communication and aggregation, thereby reducing the effect of multi-source noise in collaboration. + +# 2.2. Information Fusion in Collaborative Perception + +In addition to addressing communication issues, collaborative perception systems focus on designing collaboration modules to fuse information from multiple agents. The goal is to improve perception capabilities and thereby optimise performance in downstream tasks. Based on the stage of data sharing and collaboration, the collaborative perception can be broadly divided into three categories [14, 40], i.e., early fusion [42], intermediate fusion [8, 23, 35, 41], late fusion [26, 31]. Early fusion involves sharing raw data for a comprehensive view, but requires significant bandwidth and overlooks contextual information. Intermediate fusion extracts intermediate features from each agent's observations before transmission, balancing perception performance and communication bandwidth usage. Late fusion shares detection outputs among agents, but produces suboptimal results due to individual inaccuracies or incomplete perceptions. Among them, intermediate fusion methods gain popularity due to their performance-bandwidth trade-off [35, 40, 41]. V2VNet [35] proposes a spatially aware graph neural network to aggregate features from multiple agents. OPV2V [41] uses a global self-attention module for feature fusion. As an extension of intermediate fusion, our study introduces a novel cross-semantic fusion module, which is designed to aggregate features in the presence of feature misalignment caused by pose noise. Furthermore, we integrate motion context into the perception system to improve the understanding of the environment. + +# 3. Method + +In this paper, we focus on improving the robustness and accuracy of collaborative perception in complex collaborative scenarios. To achieve this, we propose a novel motion-aware robust communication network (MRCNet) that integrates multi-scale feature selection, effective information fusion, and motion-aware feature extraction into collaborative perception. As shown in Figure 1, MRCNet consists of four key components: feature extraction and selection, multi-scale robust collaboration, motion enhanced mechanism, and the 3D object detection head. + +# 3.1. Problem Formulation + +At time $t$ , in a collaborative perception task involving $N$ agents, the selected $i$ -th agent is defined as the ego agent, and a communication graph is established among them. Other agents within the communication range serve as collaborators, providing complementary information and + +helping the ego agent to make a comprehensive perception. Let $\mathcal{X}_i^t$ be the point cloud observation and $\mathcal{Y}_i^t$ be the perception supervision of the ego agent. Each collaborator provides the ego agent with the information extracted from their local observations $\mathcal{X}_j^t$ . Additionally, considering that we aim at modelling the motion-aware collaborations, we design a memory bank of the ego agent that stores $K$ frames of historical point data, denoted as $\{\mathcal{X}_i^{t - k}\}_{k = 1}^K = \{\mathcal{X}_i^{t - 1},\mathcal{X}_i^{t - 2},\dots,\mathcal{X}_i^{t - K}\}$ . + +Thus, the motion-aware collaborative perception task can be defined as follows: + +$$ +\hat {\mathcal {Y}} _ {i} ^ {t} = \sum_ {i = 1} ^ {N} \left\{\mathcal {Y} _ {i} ^ {t} | \Omega \left(\mathcal {X} _ {i} ^ {t}, \left\{\mathcal {X} _ {j} ^ {t} \right\} _ {j \neq i} ^ {N}, \left\{\mathcal {X} _ {i} ^ {t - k} \right\} _ {k = 1} ^ {K}\right) \right\}, \tag {1} +$$ + +where $\Omega (\cdot)$ is the proposed MRCNet. In this paper, we use the 3D object detection task to evaluate the performance of collaborative perception. Therefore, $\hat{\mathcal{V}}_i^t$ are defined as the predicted detection boxes along this line. + +# 3.2. Feature Extraction and Selection + +Feature Extraction. After the ego agent constructs a vehicle-to-everything (V2X) collaboration graph connecting neighbouring agents at the current frame $t$ , it exchanges pose data between agents for alignment purposes. The collaborators then transform their raw point cloud data into the ego agent's coordination using the transformation matrix $\Lambda$ computed based on the agents' unique pose $\boldsymbol{\xi}^t$ . Next, each agent uses a shared feature extractor consisting of $S$ layers [41] to obtain multi-scale features from the raw data. This process yields a set of features $\mathcal{F}_i^t = \left\{\mathcal{F}_i^{s,t}\right\}_{s=1}^S$ for the $i$ -th agent, where $i$ ranges from 1 to $N$ . + +Feature Selection. In contrast to previous methods that transmit all features, we introduce a novel feature selection technique. Perception noise introduces distortions into the collaborative features, thereby misleading the collaborative system. Therefore, our feature selection technique selects highly informative features from the multi-scale features for communication, effectively reducing the effect of noise interference and conserving communication bandwidth. + +The features extracted from the foreground regions are particularly important due to their rich semantic information [43,45]. To select highly informative features for effective communication, we use channel pooling operations to evaluate the semantic density within each feature. Initially, different scales of features are spatially flattened and then concatenated into a feature sequence $\pmb{E}^{t} \in \mathbb{R}^{C \times L}$ , where $C$ is the number of channels of the feature, $L$ is the length of the multi-scale feature after flatten. Subsequently, we perform channel-wise pooling operations on $\pmb{E}^{t}$ to generate the semantic confidence map $m$ : + +$$ +\boldsymbol {m} = \left(\operatorname {M a x P o o l} \left(\boldsymbol {E} ^ {t}\right) + \operatorname {A v g P o o l} \left(\boldsymbol {E} ^ {t}\right)\right) \in \mathbb {R} ^ {1 \times L}, \tag {2} +$$ + +![](images/7af542631a63c4b5fd0b8b0806be1f3d35ba833f5cdd3694d40b427ac7d0bf04.jpg) +Figure 2. The architecture of the proposed MEA component. (a) shows the improved pyramid sampling aggregation module. (b) shows the multi-scale deformable self-attention module. + +where MaxPool $(\cdot)$ and AvgPool $(\cdot)$ refer to the channelwise max pooling and average pooling operations. The binary selection matrix for communication is defined as: + +$$ +\boldsymbol {c} = \Phi_ {\text {s e l e c t}} (\boldsymbol {m}) \in \{0, 1 \} ^ {1 \times L}, \tag {3} +$$ + +where $\Phi_{\mathrm{select}}(\cdot)$ is the selection function to select the most informative features for communication. The selected features with high information are then obtained as $D^{t} = (c\odot E^{t})\in \mathbb{R}^{C\times L}$ , where $\odot$ is the element-wise multiplication. The selected feature map provides highly informative features with less bandwidth usage. + +# 3.3. Multi-Scale Robust Fusion + +Cross-agent collaboration improves the visual representation and perception performance of the ego agent by exploiting shared semantic information. Previous attention-based [39-41] approaches overlook to consider the robustness of the collaboration system to noise. Therefore, we introduce a novel multi-scale robust fusion (MRF) module, which is designed for the robust multi-scale feature fusion through two-stage communication. Multi-scale feature fusion combines information of different semantics to improve perception capability and robustness to pose noise. The two-stage communication module consists of multiscale enhanced aggregation (MEA) and collaborative features aggregation (CFA). MEA uses cross-semantic fusion to overcome the challenge of feature misalignment caused by pose noise, and CFA enables the ego agent to maximise the perception information aggregation. + +Multi-scale Enhanced Aggregation. In the first stage, the ego agent broadcasts highly informative features to the collaborators. For the sake of clarity, let's consider a communication process involving only two agents. Upon receiving highly informative features, the collaborator first converts these features from sequence to multi-scale representation, then performs a preliminary fusion using the max fusion operation to obtain the multi-scale feature $\mathcal{U}_j^t = \{\mathcal{U}_j^{s,t}\}_{s=1}^S$ . + +To address the challenge of feature misalignment, MEA incorporates cross-semantic modules to fully exploit the strengths of different semantics [6]. As shown in Figure 2, MEA consists of two components: an improved pyramid sampling aggregation (IPSA) module and a multi-scale deformable self-attention (MDSA) module [46]. + +Low-level features capture fine-grained details, while high-level semantic features provide integrated contextual understanding. The pyramid sampling aggregation module [12, 25] uses simple convolution and channel-wise concatenation operations to fuse multi-scale features. This process gathers information through cross-semantic fusion, resulting in a more comprehensive and robust feature representation that effectively mitigates noise interference. We pyramid sample the feature $\mathcal{U}_j^{s,t}$ to obtain the new multiscale feature $\left\{\mathcal{U}_j^{sl,t}\right\}_{l=1}^S$ , where $\mathcal{U}_j^{sl,t}$ is the transformation of $\mathcal{U}_j^{s,t}$ from the $s$ -th scale to the $l$ -th scale, and the feature sizes of the new multi-scale feature correspond to $\mathcal{U}_j^t$ . + +After multiple rounds of communication, the fine details in the transmitted features tend to be compromised due to cross-semantic fusion during the pyramid sampling. Therefore, we improve the pyramid sampling aggregation module by adding original multi-scale features from collaborators to compensate for detailed features, which is the proposed IPSA module. The output can be represented as: + +$$ +\tilde {\mathcal {U}} _ {j} ^ {s, t} = \operatorname {c o n v} _ {1 \times 1} \left(\left[ \mathcal {U} _ {j} ^ {s 1, t}, \mathcal {U} _ {j} ^ {s 2, t}, \mathcal {U} _ {j} ^ {s 3, t} \right]\right) + \mathcal {U} _ {j} ^ {t}, s \in [ 1, 2, 3 ], \tag {4} +$$ + +where $[\cdot ,\cdot ,\cdot ]$ is the channel-wise concatenation, and $\mathrm{conv}_{1\times 1}(\cdot)$ is the $1\times 1$ convolution operation. + +Following the IPSA module, the collaborator applies the MDSA [46] module to further fuse multi-scale features and mitigate the effect of pose noise on feature distortion. The MSDA module consists of three linear layers $\mathbf{W}_a$ , $\mathbf{W}_v$ , $\mathbf{W}_o$ , which separately compute the sampling offset $\Delta q$ , the attention weight $\mathbf{A}$ and the value $\mathbf{V}$ at position $q$ , where $q$ is the two-dimensional reference point in $\hat{\mathcal{U}}_j^{s,t}$ . The output of MSDA at position $q$ can be expressed as follows: + +$$ +\operatorname {M S D A} (q) = \sum_ {h = 1} ^ {H} \boldsymbol {W} _ {h}. +$$ + +$$ +\left(\sum_ {s = 1} ^ {S} \sum_ {p = 1} ^ {P} \psi \left(\boldsymbol {W} _ {a} \tilde {\mathcal {U}} _ {j} ^ {s, t} (q)\right) \cdot \boldsymbol {W} _ {v} \left(\tilde {\mathcal {U}} _ {j} ^ {s, t} \left(q + \boldsymbol {W} _ {o} \tilde {\mathcal {U}} _ {j} ^ {s, t} (q)\right)\right)\right) +$$ + +$$ += \sum_ {h = 1} ^ {H} \boldsymbol {W} _ {h} \cdot \left(\sum_ {s = 1} ^ {S} \sum_ {p = 1} ^ {P} \boldsymbol {A} ^ {h s p} (q) \cdot \mathcal {V} ^ {h s} \left(q + \Delta^ {h s p} (q)\right)\right), \tag {5} +$$ + +where $H$ is the number of attention heads, $S$ is the quantity of different feature scales, and $P$ is the number of sampling points utilized in each scale. $\mathbf{W}_h$ is the projection matrix + +and the softmax function $\psi (\cdot)$ is used to determine the attention weight. Ultimately, we output the multi-scale format feature $\mathcal{C}_j^t = \{\mathcal{C}_j^{s,t}\}_{s = 1}^S$ after MSDA. + +Collaborative Features Aggregation. In the second stage of communication, the $j$ -th collaborator uses the same feature selection technique to send highly informative features $\pmb{D}_j^t = (\pmb{m}_j \odot \pmb{E}_j^t)$ back to the ego agent, where $\pmb{E}_j^t$ is the feature sequence representation of $\mathcal{C}_j^t$ . In order to maximise the aggregation of the semantic information provided by the collaborator, inspired by [29], the ego agent uses the spatial-wise adaptive fusion module on each scale of features. The ego agent first stacks the feature maps of the agents to form $\mathcal{A}_i^t = \{\mathcal{A}_i^{s,t}\}_{s=1}^S$ , $\mathcal{A}_i^{s,t} \in \mathbb{R}^{n \times C \times H_s \times W_s}$ , where $s$ is the scale index and $n$ is the maximum number of agents. Max and average pooling is applied to $\mathcal{A}_i^{s,t}$ over the first channel axis to create new feature maps $\mathcal{A}_{i,\max}^{s,t} \in \mathbb{R}^{1 \times C \times H_s \times W_s}$ and $\mathcal{A}_{i,\mathrm{avg}}^{s,t} \in \mathbb{R}^{1 \times C \times H_s \times W_s}$ , where $H_s$ and $W_s$ represent the height and width of the feature. These pooled feature maps are concatenated and passed through a 3D convolutional layer with a ReLU activation function to perform further fusion. The aggregated features provide complementary information to the ego agent. + +# 3.4. Motion Enhanced Mechanism + +As mentioned above, the collaborative perception module increases the range of perception and provides compensation for occlusions [41], but it cannot directly address the motion blur introduced by moving objects. To combat this, we introduce the motion enhanced mechanism (MEM), which uses motion context from historical point cloud frames to achieve motion-aware perception. + +When the ego agent initiates a collaboration request, it projects the last $K$ frames of the point cloud stored in the memory bank to its current coordinate, producing a series of bird's eye view (BEV) feature maps $\{\mathcal{B}^{t - 1},\dots,\mathcal{B}^{t - K}\}$ . Inspired by [4,22], we design a recurrent unit called the gated motion unit (GMU). This unit aggregates motion context to enhance frame-level features. Figure 3 illustrates the operation of GMU at timestamp $t$ , showing that GMU takes three inputs stored in the ego agent's memory bank: the feature $\mathcal{B}^{t - 1}$ of the previous frame, the feature $\mathcal{B}^t$ of the current frame, and the enhanced feature $\mathcal{H}^{t - 1}$ . Both $\mathcal{B}^0$ and $H^0$ are initialized as $\mathcal{B}^1$ , and the output $\mathcal{H}^t$ represents the enhanced feature of the current frame after passing through GMU. + +Between successive frames, object motion causes spatial transitions of features. Therefore, we use $\mathcal{M}^t = \mathcal{B}^t -\mathcal{B}^{t - 1}$ to capture motion context, which can represent the motion clues of moving objects [17]. To reduce the effect of motion blur and to exploit historical information, we use a deformable convolution network [3]. The deformable convolution network can better capture the features of objects by adaptively adjusting the kernel size and shape. Specifically, + +![](images/70121dcc58cfd5bbe5c0e598d4c04b2fb3a4fd47b1aa789c8a66065db53725be.jpg) +Figure 3. The architecture of the proposed GMU component. + +the $3 \times 3$ convolutional layer is used on the motion context $\mathcal{M}_t$ to obtain the spatial offsets $\Delta p_n^t$ . The convolutional operation is then applied to $\mathcal{B}^{t-1}$ to extract historical features at position $p$ . The output feature map $\mathcal{I}^t$ at location $p$ is: + +$$ +\mathcal {I} ^ {t} (p) = \sum_ {p _ {n} \in \mathcal {R}} \boldsymbol {W} \left(p _ {n}\right) \cdot \mathcal {B} ^ {t - 1} \left(p + p _ {n} + \Delta p _ {n} ^ {t}\right), \tag {6} +$$ + +where $\mathbf{W}$ is the convolution weight, $\mathcal{R}$ is the kernel grid and $p_n$ enumerates the locations in $\mathcal{R}$ . + +To reduce the motion blur effect caused by the moving object between frames, it is necessary to determine whether the features at each position should be updated. Inspired by [36], we design an update gate $\mathcal{G}$ using the channel-spatial attention module. The channel-spatial attention module is applied to the stack feature of $\mathcal{I}^t$ and $\mathcal{H}^{t - 1}$ . After applying the sigmoid operation, the overall attention map is normalised to a range of values between 0 and 1, as formulated below: + +$$ +\mathcal {G} = \sigma (\Gamma ([ \mathcal {I} ^ {t}, \mathcal {H} ^ {t - 1} ])) \in \mathbb {R} ^ {C \times H \times W}, \tag {7} +$$ + +where $[\cdot, \cdot]$ is the channel-wise concatenation, $\Gamma(\cdot)$ is the channel-spatial attention module, and $\sigma(\cdot)$ is the sigmoid function for value normalisation. The value in $\mathcal{G}$ indicates the probability that the corresponding position of $\mathcal{H}^{t-1}$ is reserved. Then the improved feature $\mathcal{H}^t$ is refined with the previous improved feature $\mathcal{H}^{t-1}$ as follows: + +$$ +\mathcal {H} ^ {t} = \mathcal {G} \odot \mathcal {H} ^ {t - 1} + (1 - \mathcal {G}) \odot \mathcal {I} ^ {t}, \tag {8} +$$ + +where $\mathcal{H}^t\in \mathbb{R}^{C\times H\times W}$ is the final output of the GMU. + +# 3.5. 3D Object Detection Head + +After obtaining the fused output after MRF, the multiscale feature is encoded to the same size and concatenated to $\mathcal{Z}_i^t\in \mathbb{R}^{C\times H\times W}$ . Then, we concatenate $\mathcal{Z}_i^t$ and $\mathcal{H}^t$ along the channel dimension to derive the final representation $\mathcal{P}_i^t$ . To generate the final object detection results, we then use a dual-branch convolutional module on $\mathcal{P}_i^t$ to perform classification and regression tasks for each detection anchor. Similar to the previous study [44], the formulation for the object detection loss $L_{\mathrm{det}}$ is expressed as: + +$$ +L _ {\text {d e t}} = L _ {\mathrm {c l s}} + \beta L _ {\mathrm {r e g}}, \tag {9} +$$ + +![](images/be7b78e382707bc3a1666e83d462e580bb328e96c7f998df93c8157c8dfd3538.jpg) +MRCNet(Ours) Where2comm V2X-ViT Late Fusion When2com No Fusion Early Fusion + +![](images/f7fb61e47650ec619889b7872b1f0e817b1514f4ef5b4413ebac62aca828691a.jpg) + +![](images/1ef1b9ec0cdabf07d5cb048d4c2cb91dc3dd4c5bea92beff0e22b7fd5a596cd2.jpg) + +Figure 4. Robustness of our proposed method and compared benchmark to varying levels of localization noise. +![](images/e7ab0d0f4f4245c11eeb0c26a7e4a05c85ab1b5f894f4a6f09a210e44cfbc33f.jpg) +MRC(Ours) Where2comm V2X-ViT Late Fusion When2com No Fusion Early Fusion + +![](images/5672424319162e1d2ae08c798684f9119b79510bd60dcff835b99be9cbdec885.jpg) +Figure 5. Robustness of our proposed method and compared benchmark to varying levels of heading noise. + +![](images/da6309332837a9662bd78cf8ecc429924171a0f50d0863cf24e4535362681146.jpg) + +where $L_{\mathrm{cls}}$ is the cross-entropy loss for object classification, $L_{\mathrm{reg}}$ is the smooth L1 loss for box regression [16], and $\beta$ is the balancing hyperparameter. + +# 4. Experiments + +To assess the robustness of our module in noisy collaboration scenarios, we conduct experiments on three large open-source datasets. In particular, our experiments focus on LiDAR-based 3D object detection. We measure detection performance using average precision (AP) at intersection-over-union (IoU) thresholds of 0.50 and 0.70, and note that the AP values are given as percentages. To ensure a fair comparison, we exclude data and feature compression operations in the compared methods. + +# 4.1. Datasets and Experimental Settings + +Datasets. V2XSim [13] is a V2X collaborative perception dataset co-simulated by SUMO [10] and CARLA [5]. It contains 10000 frames of LiDAR point clouds and 501K annotated 3D boxes. There are 8000/1000/1000 frames for the training/validation/testing, respectively. The detection range of the agent is $x \in [-32m, 32m]$ , $y \in [-32m, 32m]$ . OPV2V [41] is a vehicle-to-vehicle collaborative perception dataset co-simulated by OpenCDA [38] and CARLA [5], including 11464 frames of 3D point clouds with 230K annotated 3D boxes, split into training/validation/testing of 6764/1981/2719 frames. The detection range is $x \in [-140m, 140m]$ , $y \in [-40m, 40m]$ . V2XSet [40] is the first open dataset that includes both V2X cooperation and realistic noise simulation, which is co-simulated by OpenCDA [38] and CARLA [5]. There + +are a total of 11447 frames in the dataset, and the train/Validation/testing splits are 6694/1920/2833, respectively. The detection range setting is the same as OPV2V. + +Implementation Details. In our implementation, we use PointPillars [11] as the multi-scale feature extractor, with the grid size of $(0.4m, 0.4m)$ for the point cloud discretisation. And we get intermediate features in three scales, which is $S = 3$ as we define in the method. The MSDA module uses 8 attention heads and 4 sampling points per communication round. For V2XSim, we extract information from three historical frames, while for V2XSet and OPV2V, we use two historical frames. To train the detection model, we set the hyperparameter in the loss function to $\beta = 2$ according to [11]. We use the Adam optimizer with an initial learning rate of $\{1\mathrm{e} - 3, 5\mathrm{e} - 4, 3\mathrm{e} - 4\}$ for the OPV2V, V2XSet and V2XSim datasets. The learning rate is gradually reduced at epochs 35 and 45 on the V2XSim dataset using a decay factor of 0.1. For the OPV2V and V2XSet datasets, the decay occurs at epochs 30 and 40. All models are trained on 2 NVIDIA RTX 3090 GPUs with a batch size of 2 for a total of 50 epochs. + +Benchmark Comparison. Our considering benchmark includes early fusion, intermediate fusion, and late fusion techniques as well as single agent perception for comparison. In addition, we compare with six state-of-the-art (SOTA) intermediate fusion models: When2Com [23], V2VNet [35], AttFuse [41], V2X-ViT [40], DiscoNet [14], CoBEVT [39], and Where2Comm [8]. + +# 4.2. Quantitative Results + +The pose of the agent can be represented by a 6D vector. Since the agent only has yaw angle to measure rotation, + +Table 1. Performance comparison on the V2XSim, OPV2V, and V2XSet datasets. The results are reported in AP0.5/0.7. + +
ModelV2XSimOPV2VV2XSet
AP0.5AP0.7AP0.5AP0.7AP0.5AP0.7
No Fusion65.7352.5769.3856.4064.8850.83
Late Fusion71.2256.9982.7661.6476.3153.33
Early Fusion84.9465.9789.2772.9683.4051.71
When2com [23]73.3652.5282.7867.6674.7340.62
V2VNet [35]82.7065.3188.6974.9384.7964.73
AttFuse [41]81.7066.2488.5472.9184.3766.27
V2X-ViT [40]82.3264.4186.7475.7082.4263.14
DiscoNet [14]83.5666.1288.0572.0782.3464.79
CoBEVT [39]81.0065.0688.9972.8084.8465.14
Where2comm [8]77.5757.0286.5868.9780.8358.72
MRCNet85.3369.8289.7776.1285.0066.31
+ +we simplify the pose to $\xi = (x,y,\theta)$ in 2D space [26], where $x, y$ and $\theta$ represent the 2D center position and the yaw angle of the agent's accurate global position, respectively. To simulate pose noise in real-world scenarios, we add Gaussian noise with a standard deviation of $0.2m$ and $0.2^{\circ}$ to the collaborators' poses in the datasets. We compare the performance of 3D object detection with other methods on the datasets, as shown in Table 1. According to the experimental results, despite the introduction of pose noise, our proposed MRCNet module shows superior performance compared to the benchmark on all datasets. + +In addition, to investigate the effect of different components of pose noise on the collaborative perception system, we conduct experiments to evaluate the robustness of the models to localization noise and heading noise, separately. Experiments are also performed to assess the effect of various levels of localization and heading noise. Specifically, we add Gaussian noise with standard $\sigma_{xy} \in [0.2m, 0.6m]$ , $\sigma_{\theta} \in [0.2^{\circ}, 1.0^{\circ}]$ separately to the collaborators' poses. As shown in the Figure 4 and the Figure 5, an increase in Gaussian noise leads to a rapid decrease in AP0.7 values for the other methods. This suggests that misaligned spatial features from collaborators can mislead the collaboration system. However, our method maintains superior performance even under significant pose noise, demonstrating its ability to effectively select and fuse multi-level semantic features while achieving robustness to pose noise. + +# 4.3. Component Analysis + +Contribution of Major Components. To evaluate the effectiveness of our proposed component, we progressively remove (i) MEM, (ii) FS (feature selection), (iii) IPSA and (iii) MSDA and present the detection precision. Table 2 shows the results of our ablation studies, with each innovative component contributing to the performance gains. + +Varying Number of History Frames. We evaluate the effect of integrating different numbers of history frames from the ego agent. As shown in Figure 6, integrating multiple + +Table 2. Results of the ablation study of the proposed core components on the V2XSim, OPV2V and V2XSet datasets. + +
MEMFSISPAMSDAV2XSimOPV2VV2XSet
AP@0.5AP@0.7AP@0.5AP@0.7AP@0.5AP@0.7
80.7764.2885.7874.0183.1263.88
✓81.4667.6486.2974.6383.4464.71
✓✓82.8769.1587.3275.3184.2165.31
✓✓✓83.6069.2188.1175.7784.4365.47
✓✓✓✓85.3369.8289.7776.1285.0066.31
+ +![](images/fecf1cfe3b33de799825d1435a00f776df77544625854bab1dda2a0779a6abca.jpg) +Figure 6. Collaborative Performance with varying number of history frames. The MEM module works best when fusing three history frames on the V2XSim dataset, and two frames on the OPV2V and V2XSet datasets. + +![](images/78418ebd09de673fd4f9ceef989eb892680da0561553e4c601a460c3319f8524.jpg) + +![](images/397e1ed6a59741a97e4336dde28a6f2f376a2f6af10d414e33d3a74e1af29239.jpg) +(a) + +![](images/9ea1957539fe975792ecf9605ba19b67c1d731adbdaea33753794c709ebb5494.jpg) +(b) +Figure 7. (a) shows the perception performance comparison when using the proposed multi-scale method versus the single-scale approach. (b) is the comparison of collaborative perception performance of our proposed method and other baseline methods with different communication volume. + +frames into the MEM module leads to an improvement in the detection accuracy in the V2XSim [13] dataset. However, once the number of history frames exceeds three, no further improvement is observed, as MEM faces challenges in effectively capturing features with large changes. + +Effect of Multi-Scale Collaboration. To evaluate the effectiveness of multi-scale feature collaboration in complex scene perception, we separately evaluate the detection accuracy achieved by single-scale and multi-scale feature collaboration. In our implementation, the single-scale feature collaboration is modified by using only a single-scale deformable self-attention module [46] within MEA. The results of this comparison are shown in Figure 7a. By using multi-scale feature fusion, our method exploits the advantages of different semantics for better noise robustness and achieves an improved detection performance. + +Comparison of Communication Volume. Figure 7b shows the performance-bandwidth trade-off of the proposed + +![](images/26643039f58337bed6ad187b5e1f0988fc59e6c9a2efbcea6eee7f30b7f53d2f.jpg) + +![](images/64668c3216b00077414f1844d4e8662f728cfa10472e5a2796e41123710d7251.jpg) +(a) V2X-ViT + +![](images/1c8cf16f47167cf1c9048b54e91a8b69924b7c323f9a436d3534ca578a7859c3.jpg) + +![](images/489cc1c04844189389ef3823922cb66433aae87fdeb3597f72a20513883eded1.jpg) +(b) DiscoNet + +![](images/4d1ffaf89fb32aabcb7d6c0b957da63adde4d34b138df2df07e74af0f90f6a07.jpg) + +![](images/352e874738cfbe270d950d8ec8afff1f6fb857ef1f965cf6b3f4dd347a997cc7.jpg) +(c) CoBEVT +Figure 8. In the qualitative comparison on the V2XSim dataset under pose noise, green and red boxes indicate the ground truths and the detection outputs, respectively. Our method outperforms previous SOTA models by achieving more accurate detection results. + +![](images/b40969ff842e538d313a683affd43588fde959eaf52adec16f293e87283f3ca0.jpg) + +![](images/b01a784d541602e904c31ec3d8464f52fa907d5b62a6654b13a30dae2ea5052c.jpg) +(d) MRCNet (Ours) + +MRCNet and the benchmark at different levels of communication bandwidth consumption. No fusion serves as the baseline, as it does not involve any communication. Our analysis reveals several key trends: (i) Where2Comm [8] initially improves detection performance by transmitting the most critical information, but struggles to filter out noisy data as the transmission volume increases. When Where2Comm [8] transmits all feature information, the global attention mechanism captures the spatial information and improves its effectiveness. (ii) Our proposed MRCNet method consistently shows better detection performance at different bandwidth levels, achieving superior results, and multi-round communication can improve the performance of collaborative perception, but excessive misaligned feature transmission can also cause degradation. + +# 4.4. Case Study + +Detection Results Visualization. Figure 8 provides a visual comparison of the detection results of different models in real-world scenarios. Our proposed MRCNet outperforms SOTA methods, with well-aligned bounding boxes that closely match the ground truth labels. In contrast, other models such as V2X-ViT [40], DiscoNet [14], and CoBEVT [39] fail to accurately detect fast moving objects accurately, resulting in missed objects or falsely-detected predictions. Our proposed MRCNet exploits the multi-level semantic features of collaborative systems and demonstrates improved robustness against noise interference. In + +addition, the proposed motion-aware fusion model extracts motion information derived from moving objects, which can effectively mitigate the negative effects of motion blur in complex scenes, thus improving detection accuracy. + +# 5. Conclusion + +In this paper, we study the real-world noise problem in multi-agent collaborative perception systems. We propose a novel framework called motion-aware robust communication network (MRCNet) to mitigate noise interference and achieve robust collaborative perception. MRCNet uses two-stage communication to enable the selection, communication and fusion of multi-level semantic features among agents. To address the problem of motion blur caused by fast-moving objects, a GRU-based module is proposed to capture the motion context and to refine the current BEV feature. Comprehensive experiments on three open-source datasets show that the proposed MRCNet outperforms existing competing methods in noisy scenarios. The further consideration of communication issues and model heterogeneity will be a future direction. + +Acknowledgements. This work was partially supported by grants from the National Natural Science Foundation of China (No.62293544 and 62301299), and China Postdoctoral Science Foundation (No.2022M720077). We would also like to thank Yue Hu from Shanghai Jiao Tong University for sharing the collaborative perception datasets. + +# References + +[1] Hongxiang Cai, Zeyuan Zhang, Zhenyu Zhou, Ziyin Li, Wenbo Ding, and Jiuhua Zhao. Bevfusion4d: Learning lidar-camera fusion under bird's-eye-view via cross-modality guidance and temporal aggregation. arXiv preprint arXiv:2303.17099, 2023. 1 +[2] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1 +[3] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 5 +[4] Rahul Dey and Fathi M Salem. Gate-variants of gated recurrent unit (gru) neural networks. In 2017 IEEE 60th international midwest symposium on circuits and systems (MWS-CAS), pages 1597-1600. IEEE, 2017. 5 +[5] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6 +[6] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 772-782, 2022. 4 +[7] Yushan Han, Hui Zhang, Huifang Li, Yi Jin, Congyan Lang, and Yidong Li. Collaborative perception in autonomous driving: Methods, datasets and challenges. arXiv preprint arXiv:2301.06262, 2023. 1 +[8] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 1, 2, 3, 6, 7, 8 +[9] Lingdong Kong, Youquan Liu, Xin Li, Runnan Chen, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Robo3d: Towards robust and reliable 3d perception against corruptions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19994-20006, 2023. 1 +[10] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. Recent development and applications of sumosimulation of urban mobility. International journal on advances in systems and measurements, 5(3&4), 2012. 6 +[11] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 6 +[12] Junhyung Lee, Junho Koh, Youngwoo Lee, and Jun Won Choi. D-align: Dual query co-attention network for 3d object detection based on multi-frame point cloud sequence. In + +2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9238-9244. IEEE, 2023. 4 +[13] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4):10914-10921, 2022. 2, 6, 7 +[14] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 1, 3, 6, 7, 8 +[15] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1 +[16] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6 +[17] Hao Liu, Zequn Jie, Karlekar Jayashree, Meibin Qi, Jianguo Jiang, Shuicheng Yan, and Jiashi Feng. Video-based person re-identification with accumulative motion context. IEEE transactions on circuits and systems for video technology, 28(10):2788-2802, 2017. 5 +[18] Si Liu, Chen Gao, Yuan Chen, Xingyu Peng, Xianghao Kong, Kun Wang, Runsheng Xu, Wentao Jiang, Hao Xiang, Jiaqi Ma, et al. Towards vehicle-to-everything autonomous driving: A survey on collaborative perception. arXiv preprint arXiv:2308.16714, 2023. 2 +[19] Yuntao Liu, Qian Huang, Rongpeng Li, Xianfu Chen, Zhifeng Zhao, Shuyuan Zhao, Yongdong Zhu, and Honggang Zhang. Rethinking collaborative perception from the spatial-temporal importance of semantic information. arXiv preprint arXiv:2307.16517, 2023. 2 +[20] Yu Liu, Zhi Li, Zhizhuo Jiang, and You He. Prospects for multi-agent collaboration and gaming: challenge, technology, and application. Frontiers of Information Technology & Electronic Engineering, 23(7):1002-1009, 2022. 1 +[21] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 1 +[22] Yiheng Liu, Zhenxun Yuan, Wengang Zhou, and Houqiang Li. Spatial and temporal mutual promotion for video-based person re-identification. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 8786–8793, 2019. 5 +[23] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106-4115, 2020. 2, 3, 6, 7 +[24] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative + +perception via learnable handshake communication. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6876-6883. IEEE, 2020. 1, 2 +[25] Zhe Liu, Xin Zhao, Tengteng Huang, Ruolan Hu, Yu Zhou, and Xiang Bai. Tanet: Robust 3d object detection from point clouds with triple attention. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11677-11684, 2020. 4 +[26] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 1, 3, 7 +[27] Guiyang Luo, Hui Zhang, Quan Yuan, and Jinglin Li. Complementarity-enhanced and redundancy-minimized collaboration network for multi-agent perception. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3578-3586, 2022. 1 +[28] Hangyu Mao, Zhibo Gong, Zhengchao Zhang, Zhen Xiao, and Yan Ni. Learning multi-agent communication under limited-bandwidth restriction for internet packet routing. arXiv preprint arXiv:1903.05561, 2019. 1 +[29] Donghao Qiao and Farhana Zulkernine. Adaptive feature fusion for cooperative perception using lidar point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1186-1195, 2023. 1, 5 +[30] Yao Rong, Xiangyu Wei, Tianwei Lin, Yueyu Wang, and Enkelejda Kasneci. Dynstatf: An efficient feature fusion strategy for lidar 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3237-3246, 2023. 1 +[31] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. A cooperative perception system robust to localization errors. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-6. IEEE, 2023. 3 +[32] 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 +[33] Junyong Wang, Yuan Zeng, and Yi Gong. Collaborative 3d object detection for automatic vehicle systems via learnable communications. arXiv preprint arXiv:2205.11849, 2022. 1 +[34] Tianhang Wang, Guang Chen, Kai Chen, Zhengfa Liu, Bo Zhang, Alois Knoll, and Changjun Jiang. Umc: A unified bandwidth-efficient and multi-resolution based collaborative perception framework. arXiv preprint arXiv:2303.12400, 2023. 1, 2 +[35] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 605–621. Springer, 2020. 1, 3, 6, 7 +[36] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5 + +[37] Hao Xiang, Runsheng Xu, Xin Xia, Zhaoliang Zheng, Bolei Zhou, and Jiaqi Ma. V2xp-asg: Generating adversarial scenes for vehicle-to-everything perception. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3584-3591. IEEE, 2023. 1 +[38] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6 +[39] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. arXiv preprint arXiv:2207.02202, 2022. 1, 4, 6, 7, 8 +[40] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 1, 2, 3, 4, 6, 7, 8 +[41] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3, 4, 5, 6, 7 +[42] Yunshuang Yuan, Hao Cheng, and Monika Sester. Keypoints-based deep feature fusion for cooperative vehicle detection of autonomous driving. IEEE Robotics and Automation Letters, 7(2):3054-3061, 2022. 3 +[43] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3085-3094, 2019. 3 +[44] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 5 +[45] Lei Zhu, Zijun Deng, Xiaowei Hu, Chi-Wing Fu, Xuemiao Xu, Jing Qin, and Pheng-Ann Heng. Bidirectional feature pyramid network with recurrent attention residual modules for shadow detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 121–136, 2018. 3 +[46] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4, 7 \ No newline at end of file diff --git a/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/images.zip b/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d7d1fd1ddb4ded03df49de35f4103ff2555b5439 --- /dev/null +++ b/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71689dd8ecee513588fcdc19589b112968bfcf4443b891c70f31f55b38e5ef86 +size 677219 diff --git a/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/layout.json b/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e0021a11df4b1999c32620ed43aa8a5fb0a2e953 --- /dev/null +++ b/2024/Multi-agent Collaborative Perception via Motion-aware Robust Communication Network/layout.json @@ -0,0 +1,9642 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 103, + 455, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 103, + 455, + 138 + ], + "spans": [ + { + "bbox": [ + 138, + 103, + 455, + 138 + ], + "type": "text", + "content": "Multi-agent Collaborative Perception via Motion-aware Robust Communication Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "spans": [ + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "text", + "content": "Shixin Hong" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "text", + "content": ", Yu Liu" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "text", + "content": ", Zhi Li" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "text", + "content": ", Shaohui Li" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "text", + "content": ", You He" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 175, + 449, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 175, + 449, + 190 + ], + "spans": [ + { + "bbox": [ + 148, + 175, + 449, + 190 + ], + "type": "text", + "content": "1Shenzhen International Graduate School, Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 150, + 190, + 443, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 190, + 443, + 204 + ], + "spans": [ + { + "bbox": [ + 150, + 190, + 443, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 150, + 190, + 443, + 204 + ], + "type": "text", + "content": "Department of Electronic Engineering, Tsinghua University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 157, + 205, + 434, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 205, + 434, + 217 + ], + "spans": [ + { + "bbox": [ + 157, + 205, + 434, + 217 + ], + "type": "text", + "content": "hsx22@mails.tsinghua.edu.cn, liuyu77360132@126.com," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 112, + 220, + 474, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 220, + 474, + 232 + ], + "spans": [ + { + "bbox": [ + 112, + 220, + 474, + 232 + ], + "type": "text", + "content": "{zhilizl, lishaohui}@sz.tsinghua.edu.cn, heyou@mail.tsinghua.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 284, + 289, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 289, + 559 + ], + "type": "text", + "content": "Collaborative perception allows for information sharing between multiple agents, such as vehicles and infrastructure, to obtain a comprehensive view of the environment through communication and fusion. Current research on multi-agent collaborative perception systems often assumes ideal communication and perception environments and neglects the effect of real-world noise such as pose noise, motion blur, and perception noise. To address this gap, in this paper, we propose a novel motion-aware robust communication network (MRCNet) that mitigates noise interference and achieves accurate and robust collaborative perception. MRCNet consists of two main components: multi-scale robust fusion (MRF) addresses pose noise by developing cross-semantic multi-scale enhanced aggregation to fuse features of different scales, while motion enhanced mechanism (MEM) captures motion context to compensate for information blurring caused by moving objects. Experimental results on popular collaborative 3D object detection datasets demonstrate that MRCNet outperforms competing methods in noisy scenarios with improved perception performance using less bandwidth. Our code will be released at https://github.com/IndigoChildren/collaborative-perception-MRCNet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 587, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 684 + ], + "type": "text", + "content": "Collaborative perception enables information sharing among multiple agents, including vehicles and infrastructure, to provide a comprehensive view of the environment through information communication and fusion [37]. This helps individual autonomous vehicles to tackle challenging perception scenarios, such as physical occlusion or long-range detection [27]. Collaborative perception technology has gained significant attention due to its advantages [20]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 684, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 684, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 59, + 684, + 287, + 696 + ], + "type": "text", + "content": "Previous works on multi-agent collaborative perception" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 260, + 545, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 260, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 306, + 260, + 545, + 439 + ], + "type": "text", + "content": "systems focus on improving perception performance by designing superior fusion modules [2, 14, 29, 35] and effective communication strategies [8, 24, 33]. For example, the CoBEVT [39] framework uses sparse transformers [32] to capture local and global spatial interactions across agents. The UMC [34] framework optimises communication by using a two-stage entropy-based selection mechanism. However, these frameworks assume ideal communication and perception environments [7], while real-world scenarios are affected by noise, such as pose noise in vehicle localization, perception noise in raw data [21], and the motion blur in fast-moving object detection. These noises cannot be ignored for real-world applications of multi-agent collaborative systems. Unfortunately, studies on robust collaborative perception with real-world noise are largely limited." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 440, + 546, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 546, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 546, + 606 + ], + "type": "text", + "content": "In fact, different forms of noise in the real world can affect the performance of collaborative perception systems in different ways. For example, when noise interferes with the positioning system, it affects the accuracy of pose information. Pose information is vital for fusing the perception data (e.g., raw data, intermediate features, or detection outputs) of the collaborative agents in different coordinates [40]. Thus, pose noise can lead to information misalignment and inconsistency among agents [26] during data fusion, resulting in a reduced system performance. Perception noise introduces distortions into the sensor data, which affects the final performance of the perception system. In addition, the presence of motion blur due to rapid object movement affects the accuracy of target detection [1,9,15,30]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "content": "However, the development of a robust collaborative perception system against real-world noise is complicated by the following challenges. First, a robust fusion strategy that adequately processes the multi-source perception data is required to deal with the perceptual misalignment caused by pose noise. Second, a feature selection strategy that filters out highly informative features is crucial to reduce bandwidth consumption and improve the efficiency of multi-agent collaboration [28]. Finally, it is desirable to find a sta" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 167, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 167, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 167, + 712 + ], + "type": "text", + "content": "†Corresponding author: Yu Liu." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15301" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 68, + 545, + 275 + ], + "blocks": [ + { + "bbox": [ + 75, + 68, + 545, + 275 + ], + "lines": [ + { + "bbox": [ + 75, + 68, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 75, + 68, + 545, + 275 + ], + "type": "image", + "image_path": "b86db373d2b311de9cce979a13674800279643e0f238b664fd24a7ef6da08f52.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 284, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 284, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 547, + 319 + ], + "type": "text", + "content": "Figure 1. The overview of our proposed MRCNet framework. In MRCNet, agents exchange highly informative semantic features in two stages through multiple rounds of communication to mitigate the effect of noise. Meanwhile, the ego agent uses motion context extracted from historical sequences in the memory bank to reduce the effect of motion blur from moving objects." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 324, + 287, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 347 + ], + "type": "text", + "content": "ble and accurate feature extraction that eliminates the motion blur of rapidly moving objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 348, + 288, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 288, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 288, + 682 + ], + "type": "text", + "content": "To address the aforementioned challenges, we present MRCNet, a motion-aware robust communication network that mitigates noise interference and achieves more robust and accurate collaborative perception. As shown in Figure 1, MRCNet facilitates effective information collaboration among agents and improves the perception performance of the ego agent through two main components, i.e., multiscale robust fusion (MRF) and motion enhanced mechanism (MEM). Specifically, for the challenge of feature misalignment due to collaborators' pose noise, MRF improves the robustness and performance of collaborative perception through multi-scale semantic feature fusion. For perception noise, we filter out highly informative semantic features for multi-round communication based on channel-wise selection operations, while reducing communication bandwidth consumption. In addition, to address the problem of motion blur due to object movement, we introduce the MEM module, which is based on the gate recurrent unit (GRU) architecture and is designed to capture and incorporate the motion context of the ego agent. This mechanism aggregates historical information to refine the features of the current frame. To validate the effectiveness of our MRCNet framework, we carry out a comprehensive evaluation on three open-source collaborative 3D object detection datasets, including V2XSim [13], OPV2V [41], and V2XSet [40]. The experimental results show that MRCNet outperforms competing methods in noisy scenarios. Our main contributions can be summarised as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 288, + 714 + ], + "type": "text", + "content": "- We present a focused study of the real-world noise problem in the collaborative perception task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 324, + 545, + 426 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 317, + 324, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 324, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 317, + 324, + 545, + 371 + ], + "type": "text", + "content": "- We propose a motion-aware collaborative perception framework, i.e., MRCNet, which presents a novel multi-scale feature communication and fusion strategy to improve the robustness of collaboration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 378, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 378, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 378, + 545, + 426 + ], + "type": "text", + "content": "- Collaborative experiments are conducted on three large open-source datasets, proving that our proposed MRCNet can achieve the best perception performance in noisy scenarios with less bandwidth." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 436, + 392, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 436, + 392, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 436, + 392, + 449 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 456, + 535, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 535, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 535, + 468 + ], + "type": "text", + "content": "2.1. Communication in Collaborative Perception" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": "Previous works [8, 23, 24] focus on solving the communication problem, as excessive communication bandwidth consumption can lead to communication congestion and latency. To address this challenge, most research considers how to achieve a better trade-off between performance and bandwidth. There are two main methods to address the challenge of bandwidth constraints: compressing the communication volume or reducing the number of communicators. Several works reduce the communication volume by using learnable methods to facilitate efficient collaboration [18]. Where2comm [8] focuses on perceptually critical areas and transmits foreground features with high confidence. UMC [34] designs a two-stage communication strategy using extended entropy theory to filter out low quality regions. As for selecting of effective collaborators, Who2Com [24] proposes a multi-stage handshake communication mechanism that determines which agents to connect with, and Select2Col [19] introduces a collaborator selection method that selects contributing collaborators with low latency. However, there is still a lack of research on" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15302" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "the problem of pose noise in collaborative communication. Therefore, we propose the multi-scale robust fusion module, which selects highly informative features from multiscale features for communication and aggregation, thereby reducing the effect of multi-source noise in collaboration." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 138, + 287, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 138, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 138, + 287, + 152 + ], + "type": "text", + "content": "2.2. Information Fusion in Collaborative Perception" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 156, + 289, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 289, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 289, + 481 + ], + "type": "text", + "content": "In addition to addressing communication issues, collaborative perception systems focus on designing collaboration modules to fuse information from multiple agents. The goal is to improve perception capabilities and thereby optimise performance in downstream tasks. Based on the stage of data sharing and collaboration, the collaborative perception can be broadly divided into three categories [14, 40], i.e., early fusion [42], intermediate fusion [8, 23, 35, 41], late fusion [26, 31]. Early fusion involves sharing raw data for a comprehensive view, but requires significant bandwidth and overlooks contextual information. Intermediate fusion extracts intermediate features from each agent's observations before transmission, balancing perception performance and communication bandwidth usage. Late fusion shares detection outputs among agents, but produces suboptimal results due to individual inaccuracies or incomplete perceptions. Among them, intermediate fusion methods gain popularity due to their performance-bandwidth trade-off [35, 40, 41]. V2VNet [35] proposes a spatially aware graph neural network to aggregate features from multiple agents. OPV2V [41] uses a global self-attention module for feature fusion. As an extension of intermediate fusion, our study introduces a novel cross-semantic fusion module, which is designed to aggregate features in the presence of feature misalignment caused by pose noise. Furthermore, we integrate motion context into the perception system to improve the understanding of the environment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 489, + 103, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 103, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 103, + 502 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "text", + "content": "In this paper, we focus on improving the robustness and accuracy of collaborative perception in complex collaborative scenarios. To achieve this, we propose a novel motion-aware robust communication network (MRCNet) that integrates multi-scale feature selection, effective information fusion, and motion-aware feature extraction into collaborative perception. As shown in Figure 1, MRCNet consists of four key components: feature extraction and selection, multi-scale robust collaboration, motion enhanced mechanism, and the 3D object detection head." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 635, + 171, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 171, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 171, + 647 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "text", + "content": "At time " + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "text", + "content": ", in a collaborative perception task involving " + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "text", + "content": " agents, the selected " + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "text", + "content": "-th agent is defined as the ego agent, and a communication graph is established among them. Other agents within the communication range serve as collaborators, providing complementary information and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 183 + ], + "type": "text", + "content": "helping the ego agent to make a comprehensive perception. Let " + }, + { + "bbox": [ + 304, + 72, + 545, + 183 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i^t" + }, + { + "bbox": [ + 304, + 72, + 545, + 183 + ], + "type": "text", + "content": " be the point cloud observation and " + }, + { + "bbox": [ + 304, + 72, + 545, + 183 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_i^t" + }, + { + "bbox": [ + 304, + 72, + 545, + 183 + ], + "type": "text", + "content": " be the perception supervision of the ego agent. Each collaborator provides the ego agent with the information extracted from their local observations " + }, + { + "bbox": [ + 304, + 72, + 545, + 183 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_j^t" + }, + { + "bbox": [ + 304, + 72, + 545, + 183 + ], + "type": "text", + "content": ". Additionally, considering that we aim at modelling the motion-aware collaborations, we design a memory bank of the ego agent that stores " + }, + { + "bbox": [ + 304, + 72, + 545, + 183 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 545, + 183 + ], + "type": "text", + "content": " frames of historical point data, denoted as " + }, + { + "bbox": [ + 304, + 72, + 545, + 183 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{X}_i^{t - k}\\}_{k = 1}^K = \\{\\mathcal{X}_i^{t - 1},\\mathcal{X}_i^{t - 2},\\dots,\\mathcal{X}_i^{t - K}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 183 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 183, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 183, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 183, + 545, + 205 + ], + "type": "text", + "content": "Thus, the motion-aware collaborative perception task can be defined as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 213, + 545, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 213, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 545, + 247 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {Y}} _ {i} ^ {t} = \\sum_ {i = 1} ^ {N} \\left\\{\\mathcal {Y} _ {i} ^ {t} | \\Omega \\left(\\mathcal {X} _ {i} ^ {t}, \\left\\{\\mathcal {X} _ {j} ^ {t} \\right\\} _ {j \\neq i} ^ {N}, \\left\\{\\mathcal {X} _ {i} ^ {t - k} \\right\\} _ {k = 1} ^ {K}\\right) \\right\\}, \\tag {1}", + "image_path": "269dbcad6cffa55c59d1ee4bc05f0ad683f4725fb3126ef5f9f790d9897d241d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 255, + 547, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 255, + 547, + 304 + ], + "type": "inline_equation", + "content": "\\Omega (\\cdot)" + }, + { + "bbox": [ + 304, + 255, + 547, + 304 + ], + "type": "text", + "content": " is the proposed MRCNet. In this paper, we use the 3D object detection task to evaluate the performance of collaborative perception. Therefore, " + }, + { + "bbox": [ + 304, + 255, + 547, + 304 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}_i^t" + }, + { + "bbox": [ + 304, + 255, + 547, + 304 + ], + "type": "text", + "content": " are defined as the predicted detection boxes along this line." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 312, + 482, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 482, + 323 + ], + "type": "text", + "content": "3.2. Feature Extraction and Selection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "text", + "content": "Feature Extraction. After the ego agent constructs a vehicle-to-everything (V2X) collaboration graph connecting neighbouring agents at the current frame " + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "text", + "content": ", it exchanges pose data between agents for alignment purposes. The collaborators then transform their raw point cloud data into the ego agent's coordination using the transformation matrix " + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "text", + "content": " computed based on the agents' unique pose " + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\xi}^t" + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "text", + "content": ". Next, each agent uses a shared feature extractor consisting of " + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "text", + "content": " layers [41] to obtain multi-scale features from the raw data. This process yields a set of features " + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i^t = \\left\\{\\mathcal{F}_i^{s,t}\\right\\}_{s=1}^S" + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "text", + "content": "-th agent, where " + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "text", + "content": " ranges from 1 to " + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 330, + 545, + 463 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 464, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 559 + ], + "type": "text", + "content": "Feature Selection. In contrast to previous methods that transmit all features, we introduce a novel feature selection technique. Perception noise introduces distortions into the collaborative features, thereby misleading the collaborative system. Therefore, our feature selection technique selects highly informative features from the multi-scale features for communication, effectively reducing the effect of noise interference and conserving communication bandwidth." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 559, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 545, + 691 + ], + "type": "text", + "content": "The features extracted from the foreground regions are particularly important due to their rich semantic information [43,45]. To select highly informative features for effective communication, we use channel pooling operations to evaluate the semantic density within each feature. Initially, different scales of features are spatially flattened and then concatenated into a feature sequence " + }, + { + "bbox": [ + 304, + 559, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\pmb{E}^{t} \\in \\mathbb{R}^{C \\times L}" + }, + { + "bbox": [ + 304, + 559, + 545, + 691 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 559, + 545, + 691 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 559, + 545, + 691 + ], + "type": "text", + "content": " is the number of channels of the feature, " + }, + { + "bbox": [ + 304, + 559, + 545, + 691 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 559, + 545, + 691 + ], + "type": "text", + "content": " is the length of the multi-scale feature after flatten. Subsequently, we perform channel-wise pooling operations on " + }, + { + "bbox": [ + 304, + 559, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\pmb{E}^{t}" + }, + { + "bbox": [ + 304, + 559, + 545, + 691 + ], + "type": "text", + "content": " to generate the semantic confidence map " + }, + { + "bbox": [ + 304, + 559, + 545, + 691 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 559, + 545, + 691 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 318, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {m} = \\left(\\operatorname {M a x P o o l} \\left(\\boldsymbol {E} ^ {t}\\right) + \\operatorname {A v g P o o l} \\left(\\boldsymbol {E} ^ {t}\\right)\\right) \\in \\mathbb {R} ^ {1 \\times L}, \\tag {2}", + "image_path": "a063a5f6fccfbaf0efa9f73c7d4d2b3df6f9374b9689d311d801e1157e539f56.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15303" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 286, + 209 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 286, + 209 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 286, + 209 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 286, + 209 + ], + "type": "image", + "image_path": "7af542631a63c4b5fd0b8b0806be1f3d35ba833f5cdd3694d40b427ac7d0bf04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 218, + 287, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 218, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 287, + 251 + ], + "type": "text", + "content": "Figure 2. The architecture of the proposed MEA component. (a) shows the improved pyramid sampling aggregation module. (b) shows the multi-scale deformable self-attention module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 263, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 300 + ], + "type": "text", + "content": "where MaxPool " + }, + { + "bbox": [ + 46, + 263, + 287, + 300 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 46, + 263, + 287, + 300 + ], + "type": "text", + "content": " and AvgPool " + }, + { + "bbox": [ + 46, + 263, + 287, + 300 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 46, + 263, + 287, + 300 + ], + "type": "text", + "content": " refer to the channelwise max pooling and average pooling operations. The binary selection matrix for communication is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 305, + 287, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 305, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 287, + 320 + ], + "type": "interline_equation", + "content": "\\boldsymbol {c} = \\Phi_ {\\text {s e l e c t}} (\\boldsymbol {m}) \\in \\{0, 1 \\} ^ {1 \\times L}, \\tag {3}", + "image_path": "773eec740532ead9820b117336356d8a30d5dc294f03cf0c4954420718259cdb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 327, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 287, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 327, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{select}}(\\cdot)" + }, + { + "bbox": [ + 47, + 327, + 287, + 399 + ], + "type": "text", + "content": " is the selection function to select the most informative features for communication. The selected features with high information are then obtained as " + }, + { + "bbox": [ + 47, + 327, + 287, + 399 + ], + "type": "inline_equation", + "content": "D^{t} = (c\\odot E^{t})\\in \\mathbb{R}^{C\\times L}" + }, + { + "bbox": [ + 47, + 327, + 287, + 399 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 327, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 327, + 287, + 399 + ], + "type": "text", + "content": " is the element-wise multiplication. The selected feature map provides highly informative features with less bandwidth usage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 405, + 194, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 405, + 194, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 194, + 417 + ], + "type": "text", + "content": "3.3. Multi-Scale Robust Fusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 423, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 615 + ], + "type": "text", + "content": "Cross-agent collaboration improves the visual representation and perception performance of the ego agent by exploiting shared semantic information. Previous attention-based [39-41] approaches overlook to consider the robustness of the collaboration system to noise. Therefore, we introduce a novel multi-scale robust fusion (MRF) module, which is designed for the robust multi-scale feature fusion through two-stage communication. Multi-scale feature fusion combines information of different semantics to improve perception capability and robustness to pose noise. The two-stage communication module consists of multiscale enhanced aggregation (MEA) and collaborative features aggregation (CFA). MEA uses cross-semantic fusion to overcome the challenge of feature misalignment caused by pose noise, and CFA enables the ego agent to maximise the perception information aggregation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 616, + 287, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 717 + ], + "type": "text", + "content": "Multi-scale Enhanced Aggregation. In the first stage, the ego agent broadcasts highly informative features to the collaborators. For the sake of clarity, let's consider a communication process involving only two agents. Upon receiving highly informative features, the collaborator first converts these features from sequence to multi-scale representation, then performs a preliminary fusion using the max fusion operation to obtain the multi-scale feature " + }, + { + "bbox": [ + 46, + 616, + 287, + 717 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_j^t = \\{\\mathcal{U}_j^{s,t}\\}_{s=1}^S" + }, + { + "bbox": [ + 46, + 616, + 287, + 717 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "To address the challenge of feature misalignment, MEA incorporates cross-semantic modules to fully exploit the strengths of different semantics [6]. As shown in Figure 2, MEA consists of two components: an improved pyramid sampling aggregation (IPSA) module and a multi-scale deformable self-attention (MDSA) module [46]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "text", + "content": "Low-level features capture fine-grained details, while high-level semantic features provide integrated contextual understanding. The pyramid sampling aggregation module [12, 25] uses simple convolution and channel-wise concatenation operations to fuse multi-scale features. This process gathers information through cross-semantic fusion, resulting in a more comprehensive and robust feature representation that effectively mitigates noise interference. We pyramid sample the feature " + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_j^{s,t}" + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "text", + "content": " to obtain the new multiscale feature " + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "inline_equation", + "content": "\\left\\{\\mathcal{U}_j^{sl,t}\\right\\}_{l=1}^S" + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_j^{sl,t}" + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "text", + "content": " is the transformation of " + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_j^{s,t}" + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "text", + "content": "-th scale to the " + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "text", + "content": "-th scale, and the feature sizes of the new multi-scale feature correspond to " + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_j^t" + }, + { + "bbox": [ + 304, + 144, + 546, + 301 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 301, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 384 + ], + "type": "text", + "content": "After multiple rounds of communication, the fine details in the transmitted features tend to be compromised due to cross-semantic fusion during the pyramid sampling. Therefore, we improve the pyramid sampling aggregation module by adding original multi-scale features from collaborators to compensate for detailed features, which is the proposed IPSA module. The output can be represented as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 392, + 545, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 422 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {U}} _ {j} ^ {s, t} = \\operatorname {c o n v} _ {1 \\times 1} \\left(\\left[ \\mathcal {U} _ {j} ^ {s 1, t}, \\mathcal {U} _ {j} ^ {s 2, t}, \\mathcal {U} _ {j} ^ {s 3, t} \\right]\\right) + \\mathcal {U} _ {j} ^ {t}, s \\in [ 1, 2, 3 ], \\tag {4}", + "image_path": "ac0c61ed31cd4762a6f5b5b257c09db14f8d2e8129b9172074295c618652d4e0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 422, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 447 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 422, + 545, + 447 + ], + "type": "inline_equation", + "content": "[\\cdot ,\\cdot ,\\cdot ]" + }, + { + "bbox": [ + 304, + 422, + 545, + 447 + ], + "type": "text", + "content": " is the channel-wise concatenation, and " + }, + { + "bbox": [ + 304, + 422, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\mathrm{conv}_{1\\times 1}(\\cdot)" + }, + { + "bbox": [ + 304, + 422, + 545, + 447 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 422, + 545, + 447 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 304, + 422, + 545, + 447 + ], + "type": "text", + "content": " convolution operation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": "Following the IPSA module, the collaborator applies the MDSA [46] module to further fuse multi-scale features and mitigate the effect of pose noise on feature distortion. The MSDA module consists of three linear layers " + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_a" + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v" + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o" + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": ", which separately compute the sampling offset " + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "inline_equation", + "content": "\\Delta q" + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": ", the attention weight " + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": " and the value " + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": " at position " + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": " is the two-dimensional reference point in " + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{U}}_j^{s,t}" + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": ". The output of MSDA at position " + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": " can be expressed as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 552, + 401, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 401, + 583 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 401, + 583 + ], + "type": "interline_equation", + "content": "\\operatorname {M S D A} (q) = \\sum_ {h = 1} ^ {H} \\boldsymbol {W} _ {h}.", + "image_path": "184aeed90b7433bcdb84d20c83fced21725c860127ad4a7c762efe5189c3ae96.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 586, + 558, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 586, + 558, + 620 + ], + "spans": [ + { + "bbox": [ + 310, + 586, + 558, + 620 + ], + "type": "interline_equation", + "content": "\\left(\\sum_ {s = 1} ^ {S} \\sum_ {p = 1} ^ {P} \\psi \\left(\\boldsymbol {W} _ {a} \\tilde {\\mathcal {U}} _ {j} ^ {s, t} (q)\\right) \\cdot \\boldsymbol {W} _ {v} \\left(\\tilde {\\mathcal {U}} _ {j} ^ {s, t} \\left(q + \\boldsymbol {W} _ {o} \\tilde {\\mathcal {U}} _ {j} ^ {s, t} (q)\\right)\\right)\\right)", + "image_path": "4c741128acf55de35611ef6c27e8c87047d9e00690262ce8f4e524cbd93e2cb5.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 310, + 622, + 545, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 622, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 310, + 622, + 545, + 668 + ], + "type": "interline_equation", + "content": "= \\sum_ {h = 1} ^ {H} \\boldsymbol {W} _ {h} \\cdot \\left(\\sum_ {s = 1} ^ {S} \\sum_ {p = 1} ^ {P} \\boldsymbol {A} ^ {h s p} (q) \\cdot \\mathcal {V} ^ {h s} \\left(q + \\Delta^ {h s p} (q)\\right)\\right), \\tag {5}", + "image_path": "bbedfd0f512960d562b5524ef7bbfdf6fac1e7366f80f54ff502d45a30d3b485.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": " is the number of attention heads, " + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": " is the quantity of different feature scales, and " + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": " is the number of sampling points utilized in each scale. " + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_h" + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": " is the projection matrix" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15304" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 112 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 112 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 112 + ], + "type": "text", + "content": "and the softmax function " + }, + { + "bbox": [ + 46, + 72, + 287, + 112 + ], + "type": "inline_equation", + "content": "\\psi (\\cdot)" + }, + { + "bbox": [ + 46, + 72, + 287, + 112 + ], + "type": "text", + "content": " is used to determine the attention weight. Ultimately, we output the multi-scale format feature " + }, + { + "bbox": [ + 46, + 72, + 287, + 112 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_j^t = \\{\\mathcal{C}_j^{s,t}\\}_{s = 1}^S" + }, + { + "bbox": [ + 46, + 72, + 287, + 112 + ], + "type": "text", + "content": " after MSDA." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "content": "Collaborative Features Aggregation. In the second stage of communication, the " + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "content": "-th collaborator uses the same feature selection technique to send highly informative features " + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "inline_equation", + "content": "\\pmb{D}_j^t = (\\pmb{m}_j \\odot \\pmb{E}_j^t)" + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "content": " back to the ego agent, where " + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "inline_equation", + "content": "\\pmb{E}_j^t" + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "content": " is the feature sequence representation of " + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_j^t" + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "content": ". In order to maximise the aggregation of the semantic information provided by the collaborator, inspired by [29], the ego agent uses the spatial-wise adaptive fusion module on each scale of features. The ego agent first stacks the feature maps of the agents to form " + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i^t = \\{\\mathcal{A}_i^{s,t}\\}_{s=1}^S" + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i^{s,t} \\in \\mathbb{R}^{n \\times C \\times H_s \\times W_s}" + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "content": " is the scale index and " + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "content": " is the maximum number of agents. Max and average pooling is applied to " + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i^{s,t}" + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "content": " over the first channel axis to create new feature maps " + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{i,\\max}^{s,t} \\in \\mathbb{R}^{1 \\times C \\times H_s \\times W_s}" + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{i,\\mathrm{avg}}^{s,t} \\in \\mathbb{R}^{1 \\times C \\times H_s \\times W_s}" + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "inline_equation", + "content": "H_s" + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "inline_equation", + "content": "W_s" + }, + { + "bbox": [ + 46, + 112, + 289, + 345 + ], + "type": "text", + "content": " represent the height and width of the feature. These pooled feature maps are concatenated and passed through a 3D convolutional layer with a ReLU activation function to perform further fusion. The aggregated features provide complementary information to the ego agent." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 356, + 211, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 211, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 211, + 368 + ], + "type": "text", + "content": "3.4. Motion Enhanced Mechanism" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 375, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 460 + ], + "type": "text", + "content": "As mentioned above, the collaborative perception module increases the range of perception and provides compensation for occlusions [41], but it cannot directly address the motion blur introduced by moving objects. To combat this, we introduce the motion enhanced mechanism (MEM), which uses motion context from historical point cloud frames to achieve motion-aware perception." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": "When the ego agent initiates a collaboration request, it projects the last " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": " frames of the point cloud stored in the memory bank to its current coordinate, producing a series of bird's eye view (BEV) feature maps " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}^{t - 1},\\dots,\\mathcal{B}^{t - K}\\}" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": ". Inspired by [4,22], we design a recurrent unit called the gated motion unit (GMU). This unit aggregates motion context to enhance frame-level features. Figure 3 illustrates the operation of GMU at timestamp " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": ", showing that GMU takes three inputs stored in the ego agent's memory bank: the feature " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{t - 1}" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": " of the previous frame, the feature " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^t" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": " of the current frame, and the enhanced feature " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{t - 1}" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": ". Both " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^0" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "H^0" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": " are initialized as " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^1" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": ", and the output " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^t" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": " represents the enhanced feature of the current frame after passing through GMU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Between successive frames, object motion causes spatial transitions of features. Therefore, we use " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^t = \\mathcal{B}^t -\\mathcal{B}^{t - 1}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " to capture motion context, which can represent the motion clues of moving objects [17]. To reduce the effect of motion blur and to exploit historical information, we use a deformable convolution network [3]. The deformable convolution network can better capture the features of objects by adaptively adjusting the kernel size and shape. Specifically," + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 314, + 70, + 541, + 175 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 541, + 175 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 541, + 175 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 541, + 175 + ], + "type": "image", + "image_path": "70121dcc58cfd5bbe5c0e598d4c04b2fb3a4fd47b1aa789c8a66065db53725be.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 183, + 537, + 195 + ], + "lines": [ + { + "bbox": [ + 313, + 183, + 537, + 195 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 537, + 195 + ], + "type": "text", + "content": "Figure 3. The architecture of the proposed GMU component." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "text", + "content": "the " + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "text", + "content": " convolutional layer is used on the motion context " + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "text", + "content": " to obtain the spatial offsets " + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\Delta p_n^t" + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "text", + "content": ". The convolutional operation is then applied to " + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{t-1}" + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "text", + "content": " to extract historical features at position " + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "text", + "content": ". The output feature map " + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^t" + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "text", + "content": " at location " + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 261, + 545, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 261, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 319, + 261, + 545, + 287 + ], + "type": "interline_equation", + "content": "\\mathcal {I} ^ {t} (p) = \\sum_ {p _ {n} \\in \\mathcal {R}} \\boldsymbol {W} \\left(p _ {n}\\right) \\cdot \\mathcal {B} ^ {t - 1} \\left(p + p _ {n} + \\Delta p _ {n} ^ {t}\\right), \\tag {6}", + "image_path": "f17a7bd42b8c6e80fc620e7e39d0792002673277c07c66c0872482ccc91782e3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "text", + "content": " is the convolution weight, " + }, + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "text", + "content": " is the kernel grid and " + }, + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "inline_equation", + "content": "p_n" + }, + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "text", + "content": " enumerates the locations in " + }, + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 317, + 546, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 317, + 546, + 424 + ], + "spans": [ + { + "bbox": [ + 305, + 317, + 546, + 424 + ], + "type": "text", + "content": "To reduce the motion blur effect caused by the moving object between frames, it is necessary to determine whether the features at each position should be updated. Inspired by [36], we design an update gate " + }, + { + "bbox": [ + 305, + 317, + 546, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 305, + 317, + 546, + 424 + ], + "type": "text", + "content": " using the channel-spatial attention module. The channel-spatial attention module is applied to the stack feature of " + }, + { + "bbox": [ + 305, + 317, + 546, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^t" + }, + { + "bbox": [ + 305, + 317, + 546, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 317, + 546, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{t - 1}" + }, + { + "bbox": [ + 305, + 317, + 546, + 424 + ], + "type": "text", + "content": ". After applying the sigmoid operation, the overall attention map is normalised to a range of values between 0 and 1, as formulated below:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 430, + 545, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 430, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 345, + 430, + 545, + 445 + ], + "type": "interline_equation", + "content": "\\mathcal {G} = \\sigma (\\Gamma ([ \\mathcal {I} ^ {t}, \\mathcal {H} ^ {t - 1} ])) \\in \\mathbb {R} ^ {C \\times H \\times W}, \\tag {7}", + "image_path": "251c5c22dc71b842d2bcba31902f2dec37ffcd4960941a220365c9a8613d8cb0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "text", + "content": " is the channel-wise concatenation, " + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\Gamma(\\cdot)" + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "text", + "content": " is the channel-spatial attention module, and " + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "text", + "content": " is the sigmoid function for value normalisation. The value in " + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "text", + "content": " indicates the probability that the corresponding position of " + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{t-1}" + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "text", + "content": " is reserved. Then the improved feature " + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^t" + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "text", + "content": " is refined with the previous improved feature " + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{t-1}" + }, + { + "bbox": [ + 304, + 451, + 545, + 523 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 357, + 529, + 545, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 529, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 357, + 529, + 545, + 543 + ], + "type": "interline_equation", + "content": "\\mathcal {H} ^ {t} = \\mathcal {G} \\odot \\mathcal {H} ^ {t - 1} + (1 - \\mathcal {G}) \\odot \\mathcal {I} ^ {t}, \\tag {8}", + "image_path": "a1ef69a6f395666c8f7285cb9592a4d83c59e4a51f3b118584eaa133d542c4cb.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 548, + 527, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 548, + 527, + 562 + ], + "spans": [ + { + "bbox": [ + 305, + 548, + 527, + 562 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 548, + 527, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^t\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 305, + 548, + 527, + 562 + ], + "type": "text", + "content": " is the final output of the GMU." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 567, + 451, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 451, + 580 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 451, + 580 + ], + "type": "text", + "content": "3.5. 3D Object Detection Head" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "type": "text", + "content": "After obtaining the fused output after MRF, the multiscale feature is encoded to the same size and concatenated to " + }, + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_i^t\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "type": "text", + "content": ". Then, we concatenate " + }, + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_i^t" + }, + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^t" + }, + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "type": "text", + "content": " along the channel dimension to derive the final representation " + }, + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^t" + }, + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "type": "text", + "content": ". To generate the final object detection results, we then use a dual-branch convolutional module on " + }, + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^t" + }, + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "type": "text", + "content": " to perform classification and regression tasks for each detection anchor. Similar to the previous study [44], the formulation for the object detection loss " + }, + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{det}}" + }, + { + "bbox": [ + 304, + 586, + 545, + 694 + ], + "type": "text", + "content": " is expressed as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 383, + 701, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 701, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 383, + 701, + 545, + 715 + ], + "type": "interline_equation", + "content": "L _ {\\text {d e t}} = L _ {\\mathrm {c l s}} + \\beta L _ {\\mathrm {r e g}}, \\tag {9}", + "image_path": "c29b3e3803d1cec6fd7282ffe6b7f4ccf91c5b5313948fc71f057a02e681e84a.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15305" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 214, + 155 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 214, + 155 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 214, + 155 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 214, + 155 + ], + "type": "image", + "image_path": "be7b78e382707bc3a1666e83d462e580bb328e96c7f998df93c8157c8dfd3538.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 158, + 474, + 166 + ], + "lines": [ + { + "bbox": [ + 124, + 158, + 474, + 166 + ], + "spans": [ + { + "bbox": [ + 124, + 158, + 474, + 166 + ], + "type": "text", + "content": "MRCNet(Ours) Where2comm V2X-ViT Late Fusion When2com No Fusion Early Fusion" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 71, + 373, + 156 + ], + "blocks": [ + { + "bbox": [ + 220, + 71, + 373, + 156 + ], + "lines": [ + { + "bbox": [ + 220, + 71, + 373, + 156 + ], + "spans": [ + { + "bbox": [ + 220, + 71, + 373, + 156 + ], + "type": "image", + "image_path": "f7fb61e47650ec619889b7872b1f0e817b1514f4ef5b4413ebac62aca828691a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 380, + 72, + 531, + 155 + ], + "blocks": [ + { + "bbox": [ + 380, + 72, + 531, + 155 + ], + "lines": [ + { + "bbox": [ + 380, + 72, + 531, + 155 + ], + "spans": [ + { + "bbox": [ + 380, + 72, + 531, + 155 + ], + "type": "image", + "image_path": "1ef1b9ec0cdabf07d5cb048d4c2cb91dc3dd4c5bea92beff0e22b7fd5a596cd2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 61, + 196, + 213, + 281 + ], + "blocks": [ + { + "bbox": [ + 95, + 177, + 496, + 189 + ], + "lines": [ + { + "bbox": [ + 95, + 177, + 496, + 189 + ], + "spans": [ + { + "bbox": [ + 95, + 177, + 496, + 189 + ], + "type": "text", + "content": "Figure 4. Robustness of our proposed method and compared benchmark to varying levels of localization noise." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 196, + 213, + 281 + ], + "lines": [ + { + "bbox": [ + 61, + 196, + 213, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 213, + 281 + ], + "type": "image", + "image_path": "e7ab0d0f4f4245c11eeb0c26a7e4a05c85ab1b5f894f4a6f09a210e44cfbc33f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 284, + 470, + 293 + ], + "lines": [ + { + "bbox": [ + 129, + 284, + 470, + 293 + ], + "spans": [ + { + "bbox": [ + 129, + 284, + 470, + 293 + ], + "type": "text", + "content": "MRC(Ours) Where2comm V2X-ViT Late Fusion When2com No Fusion Early Fusion" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 221, + 198, + 373, + 281 + ], + "blocks": [ + { + "bbox": [ + 221, + 198, + 373, + 281 + ], + "lines": [ + { + "bbox": [ + 221, + 198, + 373, + 281 + ], + "spans": [ + { + "bbox": [ + 221, + 198, + 373, + 281 + ], + "type": "image", + "image_path": "5672424319162e1d2ae08c798684f9119b79510bd60dcff835b99be9cbdec885.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 304, + 489, + 316 + ], + "lines": [ + { + "bbox": [ + 102, + 304, + 489, + 316 + ], + "spans": [ + { + "bbox": [ + 102, + 304, + 489, + 316 + ], + "type": "text", + "content": "Figure 5. Robustness of our proposed method and compared benchmark to varying levels of heading noise." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 380, + 198, + 531, + 281 + ], + "blocks": [ + { + "bbox": [ + 380, + 198, + 531, + 281 + ], + "lines": [ + { + "bbox": [ + 380, + 198, + 531, + 281 + ], + "spans": [ + { + "bbox": [ + 380, + 198, + 531, + 281 + ], + "type": "image", + "image_path": "da6309332837a9662bd78cf8ecc429924171a0f50d0863cf24e4535362681146.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 330, + 287, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 330, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 287, + 367 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 330, + 287, + 367 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cls}}" + }, + { + "bbox": [ + 47, + 330, + 287, + 367 + ], + "type": "text", + "content": " is the cross-entropy loss for object classification, " + }, + { + "bbox": [ + 47, + 330, + 287, + 367 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{reg}}" + }, + { + "bbox": [ + 47, + 330, + 287, + 367 + ], + "type": "text", + "content": " is the smooth L1 loss for box regression [16], and " + }, + { + "bbox": [ + 47, + 330, + 287, + 367 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 330, + 287, + 367 + ], + "type": "text", + "content": " is the balancing hyperparameter." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 378, + 128, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 128, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 128, + 392 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 399, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 506 + ], + "type": "text", + "content": "To assess the robustness of our module in noisy collaboration scenarios, we conduct experiments on three large open-source datasets. In particular, our experiments focus on LiDAR-based 3D object detection. We measure detection performance using average precision (AP) at intersection-over-union (IoU) thresholds of 0.50 and 0.70, and note that the AP values are given as percentages. To ensure a fair comparison, we exclude data and feature compression operations in the compared methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 515, + 236, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 236, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 236, + 528 + ], + "type": "text", + "content": "4.1. Datasets and Experimental Settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": "Datasets. V2XSim [13] is a V2X collaborative perception dataset co-simulated by SUMO [10] and CARLA [5]. It contains 10000 frames of LiDAR point clouds and 501K annotated 3D boxes. There are 8000/1000/1000 frames for the training/validation/testing, respectively. The detection range of the agent is " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "x \\in [-32m, 32m]" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "y \\in [-32m, 32m]" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": ". OPV2V [41] is a vehicle-to-vehicle collaborative perception dataset co-simulated by OpenCDA [38] and CARLA [5], including 11464 frames of 3D point clouds with 230K annotated 3D boxes, split into training/validation/testing of 6764/1981/2719 frames. The detection range is " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "x \\in [-140m, 140m]" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "y \\in [-40m, 40m]" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": ". V2XSet [40] is the first open dataset that includes both V2X cooperation and realistic noise simulation, which is co-simulated by OpenCDA [38] and CARLA [5]. There" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 330, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 365 + ], + "type": "text", + "content": "are a total of 11447 frames in the dataset, and the train/Validation/testing splits are 6694/1920/2833, respectively. The detection range setting is the same as OPV2V." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 366, + 546, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 546, + 581 + ], + "type": "text", + "content": "Implementation Details. In our implementation, we use PointPillars [11] as the multi-scale feature extractor, with the grid size of " + }, + { + "bbox": [ + 304, + 366, + 546, + 581 + ], + "type": "inline_equation", + "content": "(0.4m, 0.4m)" + }, + { + "bbox": [ + 304, + 366, + 546, + 581 + ], + "type": "text", + "content": " for the point cloud discretisation. And we get intermediate features in three scales, which is " + }, + { + "bbox": [ + 304, + 366, + 546, + 581 + ], + "type": "inline_equation", + "content": "S = 3" + }, + { + "bbox": [ + 304, + 366, + 546, + 581 + ], + "type": "text", + "content": " as we define in the method. The MSDA module uses 8 attention heads and 4 sampling points per communication round. For V2XSim, we extract information from three historical frames, while for V2XSet and OPV2V, we use two historical frames. To train the detection model, we set the hyperparameter in the loss function to " + }, + { + "bbox": [ + 304, + 366, + 546, + 581 + ], + "type": "inline_equation", + "content": "\\beta = 2" + }, + { + "bbox": [ + 304, + 366, + 546, + 581 + ], + "type": "text", + "content": " according to [11]. We use the Adam optimizer with an initial learning rate of " + }, + { + "bbox": [ + 304, + 366, + 546, + 581 + ], + "type": "inline_equation", + "content": "\\{1\\mathrm{e} - 3, 5\\mathrm{e} - 4, 3\\mathrm{e} - 4\\}" + }, + { + "bbox": [ + 304, + 366, + 546, + 581 + ], + "type": "text", + "content": " for the OPV2V, V2XSet and V2XSim datasets. The learning rate is gradually reduced at epochs 35 and 45 on the V2XSim dataset using a decay factor of 0.1. For the OPV2V and V2XSet datasets, the decay occurs at epochs 30 and 40. All models are trained on 2 NVIDIA RTX 3090 GPUs with a batch size of 2 for a total of 50 epochs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "type": "text", + "content": "Benchmark Comparison. Our considering benchmark includes early fusion, intermediate fusion, and late fusion techniques as well as single agent perception for comparison. In addition, we compare with six state-of-the-art (SOTA) intermediate fusion models: When2Com [23], V2VNet [35], AttFuse [41], V2X-ViT [40], DiscoNet [14], CoBEVT [39], and Where2Comm [8]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 671, + 424, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 424, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 424, + 684 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "The pose of the agent can be represented by a 6D vector. Since the agent only has yaw angle to measure rotation," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15306" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 96, + 286, + 236 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "type": "text", + "content": "Table 1. Performance comparison on the V2XSim, OPV2V, and V2XSet datasets. The results are reported in AP0.5/0.7." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 96, + 286, + 236 + ], + "lines": [ + { + "bbox": [ + 49, + 96, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 49, + 96, + 286, + 236 + ], + "type": "table", + "html": "
ModelV2XSimOPV2VV2XSet
AP0.5AP0.7AP0.5AP0.7AP0.5AP0.7
No Fusion65.7352.5769.3856.4064.8850.83
Late Fusion71.2256.9982.7661.6476.3153.33
Early Fusion84.9465.9789.2772.9683.4051.71
When2com [23]73.3652.5282.7867.6674.7340.62
V2VNet [35]82.7065.3188.6974.9384.7964.73
AttFuse [41]81.7066.2488.5472.9184.3766.27
V2X-ViT [40]82.3264.4186.7475.7082.4263.14
DiscoNet [14]83.5666.1288.0572.0782.3464.79
CoBEVT [39]81.0065.0688.9972.8084.8465.14
Where2comm [8]77.5757.0286.5868.9780.8358.72
MRCNet85.3369.8289.7776.1285.0066.31
", + "image_path": "49b7db73d4c2136222ff3c4fa91b788aa8f00755349144768beda15442753faa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 249, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 287, + 380 + ], + "type": "text", + "content": "we simplify the pose to " + }, + { + "bbox": [ + 46, + 249, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\xi = (x,y,\\theta)" + }, + { + "bbox": [ + 46, + 249, + 287, + 380 + ], + "type": "text", + "content": " in 2D space [26], where " + }, + { + "bbox": [ + 46, + 249, + 287, + 380 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 46, + 249, + 287, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 249, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 249, + 287, + 380 + ], + "type": "text", + "content": " represent the 2D center position and the yaw angle of the agent's accurate global position, respectively. To simulate pose noise in real-world scenarios, we add Gaussian noise with a standard deviation of " + }, + { + "bbox": [ + 46, + 249, + 287, + 380 + ], + "type": "inline_equation", + "content": "0.2m" + }, + { + "bbox": [ + 46, + 249, + 287, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 249, + 287, + 380 + ], + "type": "inline_equation", + "content": "0.2^{\\circ}" + }, + { + "bbox": [ + 46, + 249, + 287, + 380 + ], + "type": "text", + "content": " to the collaborators' poses in the datasets. We compare the performance of 3D object detection with other methods on the datasets, as shown in Table 1. According to the experimental results, despite the introduction of pose noise, our proposed MRCNet module shows superior performance compared to the benchmark on all datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 383, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 383, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 287, + 574 + ], + "type": "text", + "content": "In addition, to investigate the effect of different components of pose noise on the collaborative perception system, we conduct experiments to evaluate the robustness of the models to localization noise and heading noise, separately. Experiments are also performed to assess the effect of various levels of localization and heading noise. Specifically, we add Gaussian noise with standard " + }, + { + "bbox": [ + 46, + 383, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\sigma_{xy} \\in [0.2m, 0.6m]" + }, + { + "bbox": [ + 46, + 383, + 287, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 383, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\sigma_{\\theta} \\in [0.2^{\\circ}, 1.0^{\\circ}]" + }, + { + "bbox": [ + 46, + 383, + 287, + 574 + ], + "type": "text", + "content": " separately to the collaborators' poses. As shown in the Figure 4 and the Figure 5, an increase in Gaussian noise leads to a rapid decrease in AP0.7 values for the other methods. This suggests that misaligned spatial features from collaborators can mislead the collaboration system. However, our method maintains superior performance even under significant pose noise, demonstrating its ability to effectively select and fuse multi-level semantic features while achieving robustness to pose noise." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 585, + 167, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 167, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 167, + 597 + ], + "type": "text", + "content": "4.3. Component Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 604, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 286, + 676 + ], + "type": "text", + "content": "Contribution of Major Components. To evaluate the effectiveness of our proposed component, we progressively remove (i) MEM, (ii) FS (feature selection), (iii) IPSA and (iii) MSDA and present the detection precision. Table 2 shows the results of our ablation studies, with each innovative component contributing to the performance gains." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "Varying Number of History Frames. We evaluate the effect of integrating different numbers of history frames from the ego agent. As shown in Figure 6, integrating multiple" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 96, + 544, + 150 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 544, + 92 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 544, + 92 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 544, + 92 + ], + "type": "text", + "content": "Table 2. Results of the ablation study of the proposed core components on the V2XSim, OPV2V and V2XSet datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 96, + 544, + 150 + ], + "lines": [ + { + "bbox": [ + 308, + 96, + 544, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 544, + 150 + ], + "type": "table", + "html": "
MEMFSISPAMSDAV2XSimOPV2VV2XSet
AP@0.5AP@0.7AP@0.5AP@0.7AP@0.5AP@0.7
80.7764.2885.7874.0183.1263.88
✓81.4667.6486.2974.6383.4464.71
✓✓82.8769.1587.3275.3184.2165.31
✓✓✓83.6069.2188.1175.7784.4365.47
✓✓✓✓85.3369.8289.7776.1285.0066.31
", + "image_path": "9a078091ac06eb97f95330d1b0657c8dbcd8dbd2004352cedd379e67d07868c9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 157, + 426, + 254 + ], + "blocks": [ + { + "bbox": [ + 307, + 157, + 426, + 254 + ], + "lines": [ + { + "bbox": [ + 307, + 157, + 426, + 254 + ], + "spans": [ + { + "bbox": [ + 307, + 157, + 426, + 254 + ], + "type": "image", + "image_path": "fecf1cfe3b33de799825d1435a00f776df77544625854bab1dda2a0779a6abca.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 264, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 308 + ], + "type": "text", + "content": "Figure 6. Collaborative Performance with varying number of history frames. The MEM module works best when fusing three history frames on the V2XSim dataset, and two frames on the OPV2V and V2XSet datasets." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 426, + 157, + 543, + 254 + ], + "blocks": [ + { + "bbox": [ + 426, + 157, + 543, + 254 + ], + "lines": [ + { + "bbox": [ + 426, + 157, + 543, + 254 + ], + "spans": [ + { + "bbox": [ + 426, + 157, + 543, + 254 + ], + "type": "image", + "image_path": "78418ebd09de673fd4f9ceef989eb892680da0561553e4c601a460c3319f8524.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 315, + 426, + 407 + ], + "blocks": [ + { + "bbox": [ + 307, + 315, + 426, + 407 + ], + "lines": [ + { + "bbox": [ + 307, + 315, + 426, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 315, + 426, + 407 + ], + "type": "image", + "image_path": "397e1ed6a59741a97e4336dde28a6f2f376a2f6af10d414e33d3a74e1af29239.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 411, + 372, + 419 + ], + "lines": [ + { + "bbox": [ + 362, + 411, + 372, + 419 + ], + "spans": [ + { + "bbox": [ + 362, + 411, + 372, + 419 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 430, + 316, + 544, + 407 + ], + "blocks": [ + { + "bbox": [ + 430, + 316, + 544, + 407 + ], + "lines": [ + { + "bbox": [ + 430, + 316, + 544, + 407 + ], + "spans": [ + { + "bbox": [ + 430, + 316, + 544, + 407 + ], + "type": "image", + "image_path": "9ea1957539fe975792ecf9605ba19b67c1d731adbdaea33753794c709ebb5494.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 411, + 492, + 419 + ], + "lines": [ + { + "bbox": [ + 482, + 411, + 492, + 419 + ], + "spans": [ + { + "bbox": [ + 482, + 411, + 492, + 419 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 430, + 545, + 484 + ], + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 484 + ], + "type": "text", + "content": "Figure 7. (a) shows the perception performance comparison when using the proposed multi-scale method versus the single-scale approach. (b) is the comparison of collaborative perception performance of our proposed method and other baseline methods with different communication volume." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 495, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 555 + ], + "type": "text", + "content": "frames into the MEM module leads to an improvement in the detection accuracy in the V2XSim [13] dataset. However, once the number of history frames exceeds three, no further improvement is observed, as MEM faces challenges in effectively capturing features with large changes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 556, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 688 + ], + "type": "text", + "content": "Effect of Multi-Scale Collaboration. To evaluate the effectiveness of multi-scale feature collaboration in complex scene perception, we separately evaluate the detection accuracy achieved by single-scale and multi-scale feature collaboration. In our implementation, the single-scale feature collaboration is modified by using only a single-scale deformable self-attention module [46] within MEA. The results of this comparison are shown in Figure 7a. By using multi-scale feature fusion, our method exploits the advantages of different semantics for better noise robustness and achieves an improved detection performance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Comparison of Communication Volume. Figure 7b shows the performance-bandwidth trade-off of the proposed" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15307" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 167, + 189 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 167, + 189 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 167, + 189 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 167, + 189 + ], + "type": "image", + "image_path": "26643039f58337bed6ad187b5e1f0988fc59e6c9a2efbcea6eee7f30b7f53d2f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 190, + 167, + 307 + ], + "blocks": [ + { + "bbox": [ + 49, + 190, + 167, + 307 + ], + "lines": [ + { + "bbox": [ + 49, + 190, + 167, + 307 + ], + "spans": [ + { + "bbox": [ + 49, + 190, + 167, + 307 + ], + "type": "image", + "image_path": "64668c3216b00077414f1844d4e8662f728cfa10472e5a2796e41123710d7251.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 311, + 130, + 320 + ], + "lines": [ + { + "bbox": [ + 87, + 311, + 130, + 320 + ], + "spans": [ + { + "bbox": [ + 87, + 311, + 130, + 320 + ], + "type": "text", + "content": "(a) V2X-ViT" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 176, + 72, + 293, + 189 + ], + "blocks": [ + { + "bbox": [ + 176, + 72, + 293, + 189 + ], + "lines": [ + { + "bbox": [ + 176, + 72, + 293, + 189 + ], + "spans": [ + { + "bbox": [ + 176, + 72, + 293, + 189 + ], + "type": "image", + "image_path": "1c8cf16f47167cf1c9048b54e91a8b69924b7c323f9a436d3534ca578a7859c3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 176, + 190, + 293, + 307 + ], + "blocks": [ + { + "bbox": [ + 176, + 190, + 293, + 307 + ], + "lines": [ + { + "bbox": [ + 176, + 190, + 293, + 307 + ], + "spans": [ + { + "bbox": [ + 176, + 190, + 293, + 307 + ], + "type": "image", + "image_path": "489cc1c04844189389ef3823922cb66433aae87fdeb3597f72a20513883eded1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 311, + 255, + 320 + ], + "lines": [ + { + "bbox": [ + 212, + 311, + 255, + 320 + ], + "spans": [ + { + "bbox": [ + 212, + 311, + 255, + 320 + ], + "type": "text", + "content": "(b) DiscoNet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 301, + 72, + 418, + 188 + ], + "blocks": [ + { + "bbox": [ + 301, + 72, + 418, + 188 + ], + "lines": [ + { + "bbox": [ + 301, + 72, + 418, + 188 + ], + "spans": [ + { + "bbox": [ + 301, + 72, + 418, + 188 + ], + "type": "image", + "image_path": "4d1ffaf89fb32aabcb7d6c0b957da63adde4d34b138df2df07e74af0f90f6a07.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 301, + 190, + 418, + 306 + ], + "blocks": [ + { + "bbox": [ + 301, + 190, + 418, + 306 + ], + "lines": [ + { + "bbox": [ + 301, + 190, + 418, + 306 + ], + "spans": [ + { + "bbox": [ + 301, + 190, + 418, + 306 + ], + "type": "image", + "image_path": "352e874738cfbe270d950d8ec8afff1f6fb857ef1f965cf6b3f4dd347a997cc7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 311, + 381, + 320 + ], + "lines": [ + { + "bbox": [ + 339, + 311, + 381, + 320 + ], + "spans": [ + { + "bbox": [ + 339, + 311, + 381, + 320 + ], + "type": "text", + "content": "(c) CoBEVT" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 331, + 546, + 353 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 546, + 353 + ], + "type": "text", + "content": "Figure 8. In the qualitative comparison on the V2XSim dataset under pose noise, green and red boxes indicate the ground truths and the detection outputs, respectively. Our method outperforms previous SOTA models by achieving more accurate detection results." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 543, + 188 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 543, + 188 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 543, + 188 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 543, + 188 + ], + "type": "image", + "image_path": "b40969ff842e538d313a683affd43588fde959eaf52adec16f293e87283f3ca0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 427, + 190, + 543, + 306 + ], + "blocks": [ + { + "bbox": [ + 427, + 190, + 543, + 306 + ], + "lines": [ + { + "bbox": [ + 427, + 190, + 543, + 306 + ], + "spans": [ + { + "bbox": [ + 427, + 190, + 543, + 306 + ], + "type": "image", + "image_path": "b01a784d541602e904c31ec3d8464f52fa907d5b62a6654b13a30dae2ea5052c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 311, + 517, + 320 + ], + "lines": [ + { + "bbox": [ + 452, + 311, + 517, + 320 + ], + "spans": [ + { + "bbox": [ + 452, + 311, + 517, + 320 + ], + "type": "text", + "content": "(d) MRCNet (Ours)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 367, + 289, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 289, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 289, + 547 + ], + "type": "text", + "content": "MRCNet and the benchmark at different levels of communication bandwidth consumption. No fusion serves as the baseline, as it does not involve any communication. Our analysis reveals several key trends: (i) Where2Comm [8] initially improves detection performance by transmitting the most critical information, but struggles to filter out noisy data as the transmission volume increases. When Where2Comm [8] transmits all feature information, the global attention mechanism captures the spatial information and improves its effectiveness. (ii) Our proposed MRCNet method consistently shows better detection performance at different bandwidth levels, achieving superior results, and multi-round communication can improve the performance of collaborative perception, but excessive misaligned feature transmission can also cause degradation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 560, + 123, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 123, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 123, + 574 + ], + "type": "text", + "content": "4.4. Case Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "Detection Results Visualization. Figure 8 provides a visual comparison of the detection results of different models in real-world scenarios. Our proposed MRCNet outperforms SOTA methods, with well-aligned bounding boxes that closely match the ground truth labels. In contrast, other models such as V2X-ViT [40], DiscoNet [14], and CoBEVT [39] fail to accurately detect fast moving objects accurately, resulting in missed objects or falsely-detected predictions. Our proposed MRCNet exploits the multi-level semantic features of collaborative systems and demonstrates improved robustness against noise interference. In" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 367, + 546, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 367, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 367, + 546, + 415 + ], + "type": "text", + "content": "addition, the proposed motion-aware fusion model extracts motion information derived from moving objects, which can effectively mitigate the negative effects of motion blur in complex scenes, thus improving detection accuracy." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 431, + 378, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 378, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 378, + 443 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 453, + 546, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 546, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 546, + 632 + ], + "type": "text", + "content": "In this paper, we study the real-world noise problem in multi-agent collaborative perception systems. We propose a novel framework called motion-aware robust communication network (MRCNet) to mitigate noise interference and achieve robust collaborative perception. MRCNet uses two-stage communication to enable the selection, communication and fusion of multi-level semantic features among agents. To address the problem of motion blur caused by fast-moving objects, a GRU-based module is proposed to capture the motion context and to refine the current BEV feature. Comprehensive experiments on three open-source datasets show that the proposed MRCNet outperforms existing competing methods in noisy scenarios. The further consideration of communication issues and model heterogeneity will be a future direction." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was partially supported by grants from the National Natural Science Foundation of China (No.62293544 and 62301299), and China Postdoctoral Science Foundation (No.2022M720077). We would also like to thank Yue Hu from Shanghai Jiao Tong University for sharing the collaborative perception datasets." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15308" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Hongxiang Cai, Zeyuan Zhang, Zhenyu Zhou, Ziyin Li, Wenbo Ding, and Jiuhua Zhao. Bevfusion4d: Learning lidar-camera fusion under bird's-eye-view via cross-modality guidance and temporal aggregation. arXiv preprint arXiv:2303.17099, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[2] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "type": "text", + "content": "[3] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 251, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 294 + ], + "type": "text", + "content": "[4] Rahul Dey and Fathi M Salem. Gate-variants of gated recurrent unit (gru) neural networks. In 2017 IEEE 60th international midwest symposium on circuits and systems (MWS-CAS), pages 1597-1600. IEEE, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 296, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 339 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 341, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 287, + 407 + ], + "type": "text", + "content": "[6] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 772-782, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 409, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 453 + ], + "type": "text", + "content": "[7] Yushan Han, Hui Zhang, Huifang Li, Yi Jin, Congyan Lang, and Yidong Li. Collaborative perception in autonomous driving: Methods, datasets and challenges. arXiv preprint arXiv:2301.06262, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "type": "text", + "content": "[8] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 576 + ], + "type": "text", + "content": "[9] Lingdong Kong, Youquan Liu, Xin Li, Runnan Chen, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Robo3d: Towards robust and reliable 3d perception against corruptions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19994-20006, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "text", + "content": "[10] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. Recent development and applications of sumosimulation of urban mobility. International journal on advances in systems and measurements, 5(3&4), 2012. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "text", + "content": "[11] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[12] Junhyung Lee, Junho Koh, Youngwoo Lee, and Jun Won Choi. D-align: Dual query co-attention network for 3d object detection based on multi-frame point cloud sequence. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9238-9244. IEEE, 2023. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "type": "text", + "content": "[13] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4):10914-10921, 2022. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "type": "text", + "content": "[14] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 547, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 547, + 263 + ], + "type": "text", + "content": "[15] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 265, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 547, + 308 + ], + "type": "text", + "content": "[16] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 310, + 547, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 547, + 363 + ], + "type": "text", + "content": "[17] Hao Liu, Zequn Jie, Karlekar Jayashree, Meibin Qi, Jianguo Jiang, Shuicheng Yan, and Jiashi Feng. Video-based person re-identification with accumulative motion context. IEEE transactions on circuits and systems for video technology, 28(10):2788-2802, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 365, + 547, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 547, + 419 + ], + "type": "text", + "content": "[18] Si Liu, Chen Gao, Yuan Chen, Xingyu Peng, Xianghao Kong, Kun Wang, Runsheng Xu, Wentao Jiang, Hao Xiang, Jiaqi Ma, et al. Towards vehicle-to-everything autonomous driving: A survey on collaborative perception. arXiv preprint arXiv:2308.16714, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 422, + 547, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 547, + 476 + ], + "type": "text", + "content": "[19] Yuntao Liu, Qian Huang, Rongpeng Li, Xianfu Chen, Zhifeng Zhao, Shuyuan Zhao, Yongdong Zhu, and Honggang Zhang. Rethinking collaborative perception from the spatial-temporal importance of semantic information. arXiv preprint arXiv:2307.16517, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 478, + 547, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 547, + 521 + ], + "type": "text", + "content": "[20] Yu Liu, Zhi Li, Zhizhuo Jiang, and You He. Prospects for multi-agent collaboration and gaming: challenge, technology, and application. Frontiers of Information Technology & Electronic Engineering, 23(7):1002-1009, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 523, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 547, + 578 + ], + "type": "text", + "content": "[21] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 579, + 547, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 547, + 633 + ], + "type": "text", + "content": "[22] Yiheng Liu, Zhenxun Yuan, Wengang Zhou, and Houqiang Li. Spatial and temporal mutual promotion for video-based person re-identification. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 8786–8793, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 635, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 547, + 689 + ], + "type": "text", + "content": "[23] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106-4115, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "text", + "content": "[24] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15309" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "perception via learnable handshake communication. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6876-6883. IEEE, 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "text", + "content": "[25] Zhe Liu, Xin Zhao, Tengteng Huang, Ruolan Hu, Yu Zhou, and Xiang Bai. Tanet: Robust 3d object detection from point clouds with triple attention. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11677-11684, 2020. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 216 + ], + "type": "text", + "content": "[26] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 1, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 271 + ], + "type": "text", + "content": "[27] Guiyang Luo, Hui Zhang, Quan Yuan, and Jinglin Li. Complementarity-enhanced and redundancy-minimized collaboration network for multi-agent perception. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3578-3586, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "text", + "content": "[28] Hangyu Mao, Zhibo Gong, Zhengchao Zhang, Zhen Xiao, and Yan Ni. Learning multi-agent communication under limited-bandwidth restriction for internet packet routing. arXiv preprint arXiv:1903.05561, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "text", + "content": "[29] Donghao Qiao and Farhana Zulkernine. Adaptive feature fusion for cooperative perception using lidar point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1186-1195, 2023. 1, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "type": "text", + "content": "[30] Yao Rong, Xiangyu Wei, Tianwei Lin, Yueyu Wang, and Enkelejda Kasneci. Dynstatf: An efficient feature fusion strategy for lidar 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3237-3246, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "text", + "content": "[31] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. A cooperative perception system robust to localization errors. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-6. IEEE, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "type": "text", + "content": "[32] 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" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 504, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 288, + 536 + ], + "type": "text", + "content": "[33] Junyong Wang, Yuan Zeng, and Yi Gong. Collaborative 3d object detection for automatic vehicle systems via learnable communications. arXiv preprint arXiv:2205.11849, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 590 + ], + "type": "text", + "content": "[34] Tianhang Wang, Guang Chen, Kai Chen, Zhengfa Liu, Bo Zhang, Alois Knoll, and Changjun Jiang. Umc: A unified bandwidth-efficient and multi-resolution based collaborative perception framework. arXiv preprint arXiv:2303.12400, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 667 + ], + "type": "text", + "content": "[35] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 605–621. Springer, 2020. 1, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "text", + "content": "[36] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 574 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[37] Hao Xiang, Runsheng Xu, Xin Xia, Zhaoliang Zheng, Bolei Zhou, and Jiaqi Ma. V2xp-asg: Generating adversarial scenes for vehicle-to-everything perception. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3584-3591. IEEE, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[38] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 184, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 228 + ], + "type": "text", + "content": "[39] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. arXiv preprint arXiv:2207.02202, 2022. 1, 4, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 229, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 545, + 283 + ], + "type": "text", + "content": "[40] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 285, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 350 + ], + "type": "text", + "content": "[41] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 352, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 352, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 352, + 545, + 396 + ], + "type": "text", + "content": "[42] Yunshuang Yuan, Hao Cheng, and Monika Sester. Keypoints-based deep feature fusion for cooperative vehicle detection of autonomous driving. IEEE Robotics and Automation Letters, 7(2):3054-3061, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 396, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 396, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 396, + 545, + 441 + ], + "type": "text", + "content": "[43] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3085-3094, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 441, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 464 + ], + "type": "text", + "content": "[44] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "type": "text", + "content": "[45] Lei Zhu, Zijun Deng, Xiaowei Hu, Chi-Wing Fu, Xuemiao Xu, Jing Qin, and Pheng-Ann Heng. Bidirectional feature pyramid network with recurrent attention residual modules for shadow detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 121–136, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 531, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 574 + ], + "type": "text", + "content": "[46] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4, 7" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15310" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/12609e19-efed-4425-b5ad-32eb26a81347_content_list.json b/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/12609e19-efed-4425-b5ad-32eb26a81347_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6ab479078b1a5fd3e9445d9a5a80bdce21680af3 --- /dev/null +++ b/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/12609e19-efed-4425-b5ad-32eb26a81347_content_list.json @@ -0,0 +1,1739 @@ +[ + { + "type": "text", + "text": "Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning", + "text_level": 1, + "bbox": [ + 109, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaewoo Jeong*, Daehee Park*, and Kuk-Jin Yoon KAIST", + "bbox": [ + 289, + 203, + 679, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jeong207,bag2824,kjyoon}@kaist.ac.kr", + "bbox": [ + 321, + 241, + 650, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human pose forecasting garners attention for its diverse applications. However, challenges in modeling the multi-modal nature of human motion and intricate interactions among agents persist, particularly with longer timescales and more agents. In this paper, we propose an interaction-aware trajectory-conditioned long-term multi-agent human pose forecasting model, utilizing a coarse-to-fine prediction approach: multi-modal global trajectories are initially forecasted, followed by respective local pose forecasts conditioned on each mode. In doing so, our Trajectory2Pose model introduces a graph-based agent-wise interaction module for a reciprocal forecast of local motion-conditioned global trajectory and trajectory-conditioned local pose. Our model effectively handles the multi-modality of human motion and the complexity of long-term multi-agent interactions, improving performance in complex environments. Furthermore, we address the lack of long-term $(6s+)$ multi-agent $(5+)$ datasets by constructing a new dataset from real-world images and 2D annotations, enabling a comprehensive evaluation of our proposed model. State-of-the-art prediction performance on both complex and simpler datasets confirms the generalized effectiveness of our method. The code is available at https://github.com/Jaewoo97/T2P.", + "bbox": [ + 76, + 323, + 473, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human pose forecasting aims to predict future human motion based on observed past motion [20, 31, 32, 35, 37, 53, 86]. Humans instinctively perform such tasks, allowing them to naturally navigate in crowded areas or identify and circumvent potential dangers. For this reason, human pose forecasting plays an important role in various computer vision tasks [21, 23, 27, 54, 85, 91]. Indeed, recent years have seen a proliferation of work on multi-agent motion forecasting which aim towards modeling complex multi-agent inter", + "bbox": [ + 75, + 739, + 468, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a55f1d7b76b2c2e415c064861262d024e6c7ec4a7e55e3fbc6e46e53ca2d746.jpg", + "image_caption": [ + "Figure 1. Human motion is goal-directed and influenced by other entities. Therefore, global intention contains hints for local intention, allowing us to infer local pose from global trajectories. Our method first forecasts global trajectories, upon which local poses are conditioned for subsequent forecasts. Pose and trajectory-wise inter-agent interactions are considered for both predictions." + ], + "image_footnote": [], + "bbox": [ + 501, + 287, + 890, + 364 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/3f7c3ec175f3756e35f2530be803e0cdb77f6ecfb0c264decb0f66dd2b118f2f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Global Intent
ABC
Walk side-by-side\nAvoid CKneel
", + "bbox": [ + 503, + 368, + 679, + 421 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/6b18546f0257669459a9bba4f8ef1e76bf67c65ca930d9e3126525d9e431a813.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Local Intent
ABC
Walking motion\nTalking gesturesTie shoe
", + "bbox": [ + 712, + 368, + 890, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "action [20, 47, 53, 71, 74].", + "bbox": [ + 500, + 534, + 679, + 550 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although various methods have been proposed, they share two major limitations. The first is a limitation on long-term predictions, as previous studies predicted up to 3 seconds at most [4, 47, 74, 75]. However, a sufficiently long forecast horizon is essential to fully leverage human pose forecasting for diverse downstream tasks in the scope of identifying potential danger or understanding human behavior. The second is that multi-person interactions are not proficiently learned. Existing methods consider the joints of multiple people all at once as objects of interaction [47, 65, 74], resulting in an excessive complexity with respect to the number of joints. Due to such inefficient modeling, these approaches are found to be incompetent in long-term $(3s+)$ multi-agent $(6+)$ settings, limiting their practicality on complex real-world environments.", + "bbox": [ + 496, + 551, + 890, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Moreover, these challenges are also due to the limitations of datasets. Existing pose forecasting datasets have limited sequence length ( $\\sim 3\\mathrm{s}$ ) and number of agents ( $\\sim 2$ ). Therefore, previous works [47, 69, 75] have randomly blended disparate datasets to model multi-agent interaction with up to 10 agents. Yet, such naively merged data lacks authentic interaction as agents from different scenes remain uninfluenced. As such, there was no opportunity to develop and", + "bbox": [ + 496, + 779, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Denotes equal contribution.", + "bbox": [ + 94, + 886, + 250, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1617", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "evaluate a model in a long-term multi-agent environment.", + "bbox": [ + 76, + 90, + 455, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we present a solution from both model and dataset perspectives to tackle long-term multi-agent human pose forecasting. First, from a model perspective, we propose an interaction-aware trajectory-conditioned pose forecasting method. We point out that the limitations of existing methods on long-term multi-agent environments lead to poor performance in handling the multi-modal nature of human motion and correspondingly complex interactions. To improve upon handling multi-modality in these complex settings, we use a coarse-to-fine approach to enjoy effective interaction modeling by propagating agent-wise coarse representations. Agent-wise pose and trajectory embeddings are obtained in their respective local coordinates, followed by a holistic interaction modeling via our proposed Traj-pose module. Interaction-aware forecasts are then made by initial coarse global hip joint trajectory forecast followed by fine local pose forecasts in its hip joint coordinates, conditioned on the global trajectory as shown in Fig. 1. As discovered in previous research [2, 54], learning an agent-wise global intention as coarse trajectories is less challenging than predicting every joint-wise motion. We leverage these hints from global trajectories, which are further conditioned towards forecasting local motion that embodies the interaction-aware spatio-temporal context.", + "bbox": [ + 75, + 108, + 468, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From a dataset perspective, we parsed a novel real-world dataset for long-term multi-agent human pose forecasting. We utilize JRDB dataset [66] which consists of multi-view video and collected in various environments. Since 3D pose annotations are not provided in the original JRDB, we extracted sequences of 3D human pose from visible agents in omnidirectional images using the latest algorithm for 3D pose extraction from image [59]. We then ensure the reliability of 3D pose information by filtering and adjusting the extracted 3D poses based on 2D pose and 3D bounding box annotations. As a result, we construct a real-world 3D human pose forecasting dataset, JRDB-GlobMultiPose (JRDB-GMP), where up to 24 agents exist for up to 5 seconds. The proposed pose forecasting model is validated on both previous datasets and newly created JRDB-GMP dataset. Our method shows state-of-the-art forecasting performance in both global and local accuracy metrics, not only on JRDB-GMP but also on all previous datasets. Therefore, our contributions are as follows:", + "bbox": [ + 75, + 472, + 468, + 758 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a interaction-aware trajectory-conditioned pose forecasting method $(T2P)$ for long-term multi-agent 3D human pose forecasting.", + "- We propose a long-term, multi-agent real-world 3D human pose forecasting dataset which contains up to 24 persons and forecasts up to 5 seconds.", + "- We validate our $T2P$ model on both previous datasets and our new JRDB-GMP dataset. Our method achieves state-of-the-art forecasting performance on all datasets." + ], + "bbox": [ + 76, + 761, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 500, + 90, + 643, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Human pose forecasting", + "text_level": 1, + "bbox": [ + 500, + 111, + 720, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Pose Forecasting involves predicting a future pose sequence with temporal length of a prediction horizon, given a historical pose sequence [5, 8, 11, 16, 36, 45, 49, 52, 68, 71]. In the early stage, methods were developed to forecast single person motion within a short timeframe ( $\\sim$ 1s) [9, 34, 58, 73]. However, to improve applicability on diverse downstream computer vision tasks, forecasts are to be made on multi-person poses [1, 2, 63] for longer prediction horizons [6, 62]. Forecasting future inherently involves a stochastic nature, and handling such multi-modality has been attempted by forecasting multiple future poses of a single agent [4]. However, comparatively marginal efforts have been employed in the more complex long-term multi-agent scenes [75]. Such absence is mostly due to the lack of a proper dataset. The commonly used evaluation datasets are CMU-Mocap [13], 3DPW [67], UMPM [64], MuPoTS-3D [39], all of which contain 2 agents at most in a given scene and have short prediction horizons within 3 seconds. Most recent research arbitrarily combines individual scenes to create datasets with more than three individuals [47, 69, 75]. However, such a synthetic approach does not account for authentic agent interactions.", + "bbox": [ + 496, + 135, + 890, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Trajectory prediction", + "text_level": 1, + "bbox": [ + 500, + 479, + 700, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Trajectory prediction involves predicting the future path of an object given its past trajectory [3, 10, 25, 38, 40, 41, 51, 76, 77, 89, 90]. Unlike human pose forecasting which aims to predict every joint position, trajectory prediction regards each agent as a point mass, typically the center of mass or center point of a detected bounding box. Research in trajectory forecasting is interested in not only vehicles but also many types of agents including humans, cyclists, and so on [29, 43, 72, 84]. One substantial direction in research within this field is the Goal-conditioned prediction approach [18, 33, 82]. Goal-conditioned prediction approach first predicts the final destination within the prediction horizon with multiple goal proposals [28, 70]. Then, a thorough future path is conditioned on each mode of the multi-modal proposals. Compared to directly predicting full trajectories, the goal-conditioned approach follows a coarse-to-fine prediction and is effective in learning highly stochastic multi-modality of complex scenes [19, 42, 88].", + "bbox": [ + 496, + 502, + 890, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Human pose estimation from image", + "text_level": 1, + "bbox": [ + 500, + 786, + 808, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human pose estimation is the task of inferring the pose of a person from an image or a video [22, 24, 26, 30, 57, 60, 78, 80, 81]. Initial deep learning-based methods first utilized convolutional neural networks to estimate 2D and 3D poses from single or multiple images [12, 14, 61, 79, 83]. Recent approaches engage in more challenging tasks such as", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1618", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1341b7a715ef1637d68c209a56efc4686d3ca20c761d52245beea7e57a33c25d.jpg", + "image_caption": [ + "Figure 2. Illustration of our T2P framework. We decompose global motion into global trajectory and local pose. Multi-modal global trajectory proposals are predicted from past global trajectory and local pose embeddings. Then, future local poses are conditioned and forecasted on each trajectory proposal to compose the final human pose prediction. Predicted local poses are added to their mode-specific global trajectories in a joint-wise manner, obtaining the global human poses as the final output." + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 893, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "estimating 3D poses from monocular videos [7, 46, 48, 50, 55, 56] using self-supervised learning and generative methods [15, 17]. Most recent methods estimate multi-person poses in a crowded environment with considerable occlusions [44, 87]. We account for the aforementioned need of a complex dataset by extracting 3D pose from images using these methods. Specifically, we use a monocular 3D pose estimation method $BEV$ [59] to construct a 3D human motion forecasting dataset with long-term multi-agent characteristics from real-world image sequences. $BEV$ robustly estimates human pose in a scale-ambiguous and crowded environment, reliably extracting 3D poses from the omnidirectional image sequences of JRDB dataset [66].", + "bbox": [ + 75, + 421, + 472, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 637, + 169, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem definition", + "text_level": 1, + "bbox": [ + 76, + 664, + 256, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-agent human pose forecasting aims to learn a mapping function between the observed 3D pose of $N_{A}$ agents composed of $J$ joints, $\\mathbf{X}:\\left\\{\\mathbf{x}_t^{n,j}\\right\\}_{-T_p:0}^{N_A,J}$ , and future pose $\\mathbf{Y}:F\\times \\left\\{\\mathbf{x}_t^{n,j}\\right\\}_{0:T_f}^{N_A,J}$ in global coordinates where $F$ denotes the number of modes. Here, $T_{p}$ and $T_{f}$ are history length and prediction horizon while $\\mathbf{x}_t^{n,j} = (x_t^{n,j},y_t^{n,j},z_t^{n,j})$ is 3D global coordinate of joint $j$ of agent $n$ at time $t$ . While the global position of joint is represented in $\\mathbf{x}$ , we additionally define local position $\\mathbf{p}$ . The local position is defined in local coordinate of each agent, calculated by subtracting the global position of the hip joint of each agent. Therefore, local position of joint is defined as $\\mathbf{p}^{n,j} = \\mathbf{x}^{n,j} - \\mathbf{x}^{n,\\mathrm{hip}}$ . We", + "bbox": [ + 75, + 690, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "define the trajectory of global hip joint position as global trajectory, $Tr:\\{\\mathbf{x}^{n,\\mathrm{hip}}\\}^{N_A}$ . We also define local pose as local position of all joints, $Po:\\{\\mathbf{p}^{n,\\mathrm{j}}\\}^{N_A,J}$ . We denote past and future timesteps of global trajectory and local motion as $Tr_{\\mathcal{P}},Tr_{\\mathcal{F}}\\in Tr$ and $Po_{\\mathcal{P}},Po_{\\mathcal{F}}\\in Po$ , where $\\mathcal{P}$ and $\\mathcal{F}$ respectively denotes past and future.", + "bbox": [ + 496, + 421, + 893, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overall framework", + "text_level": 1, + "bbox": [ + 498, + 530, + 684, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We disentangle the overall human motion into global trajectories and local poses, as depicted in top left of Fig. 2. Following a coarse-to-fine strategy, multiple global trajectories are first forecasted to model the coarse modes of global intentions. Based on these forecasts, local pose predictions are conditioned on each mode to jointly constitute a thorough motion. In doing so, our model is widely divided into two portions: Trajectory predictor consists of trajectory encoder and decoder and pose predictor consists of pose encoder and decoder. Both predictors engage in the reciprocal exchange of both trajectory and pose information, facilitating the inference of cues between global and local motion. The detailed methods of each stage are described below:", + "bbox": [ + 496, + 553, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Model structure", + "text_level": 1, + "bbox": [ + 500, + 761, + 661, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.1 Pose encoder", + "text_level": 1, + "bbox": [ + 500, + 785, + 643, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike the holistic approach of previous works that encode and decode all agents' joint motions in global coordinates, our pose encoder encodes the pose dynamics in local coordinates. In addition, our pose encoder only considers intra-agent joint interaction. As a result, the encoded pose embedding represents agent-specific local mo", + "bbox": [ + 496, + 809, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1619", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion, containing insights on global intent. We follow our baseline [47] and construct the encoder with Multi-Person Body-Part (MPBP) module and transformer networks. As depicted in Fig. 2, body part sequences are constructed in frequency domain, followed by intra-agent attention-based encoding of the body parts to acquire pose embedding $Z_{Po}$ .", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Trajectory module", + "text_level": 1, + "bbox": [ + 76, + 205, + 261, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Trajectory module aims to extract embeddings from the agents' past global trajectory. Using an encoder structure from [88], multi-agent interaction-based trajectory embedding $Z_{Tr}$ is extracted which contains rudimentary insight on global intent. Interaction between agent trajectories is represented based on the reference agent $i$ 's global trajectory segment vector $\\mathbf{v}_t^i = \\mathbf{x}_t^{i,\\mathrm{hip}} - \\mathbf{x}_{t-1}^{i,\\mathrm{hip}}$ . For rotational invariance, neighbor actor $j$ 's vector is normalized by the reference vector's orientation at latest timestep $t=0$ . Separate MLP layers then compute the reference agent and neighboring agent embeddings $z_{Tr_i}^t, z_{Tr_j}^t$ as follows:", + "bbox": [ + 75, + 231, + 472, + 400 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z _ {T r _ {i}} ^ {t} = \\phi_ {r e f} (R _ {i} ^ {T} \\mathbf {v} _ {t} ^ {i}) \\\\ z _ {T r _ {j}} ^ {t} = \\phi_ {n b r} ([ R _ {i} ^ {T} (\\mathbf {v} _ {t} ^ {j}), R _ {i} ^ {T} (\\mathbf {v} _ {t} ^ {i}) ]) \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 414, + 467, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi_{ref}$ and $\\phi_{nbr}$ are different MLP blocks, $R_i \\in \\mathbb{R}^{3 \\times 3}$ is the rotation matrix of agent $j$ against agent $i$ , $[\\cdot, \\cdot]$ is concatenation. The resulting agent-specific reference and neighbor embeddings constitute trajectory embedding $z_{Tr}^t$ .", + "bbox": [ + 76, + 469, + 468, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.3 Traj-posed module", + "text_level": 1, + "bbox": [ + 76, + 554, + 253, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human maneuver contains various dynamic activities characterized by the agent's multi-modal intents. Auxiliary human motion such as arm gesture, rotational orientation of upper body and head implies the agent's intent in global motion. In that sense, harvesting meaningful insights from past local joint motion helps proficient modeling of coarse multimodality as future trajectory proposals. Therefore, we propose Traj-Pose Module that fuses agent-wise embeddings of both trajectory and pose to fully utilize these information in modeling global intentions.", + "bbox": [ + 75, + 580, + 468, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, MLP is used to match the temporal domain of pose embedding $Z_{Po}$ to that of $Z_{Tr}$ , after which both are concatenated as agent-wise traj-posed embedding $Z$ .", + "bbox": [ + 75, + 732, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ = \\left[ Z _ {T r}, \\phi_ {M L P} \\left(Z _ {P o}\\right) \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 791, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The resulting $Z$ is comprised of agent and timestep-respective trajectory and pose embeddings: $z_{i}^{t},z_{j}^{t}\\in z^{t}\\in Z$ Then, $\\widetilde{Z}$ is acquired from the graph attention with an agent-wise update where each agent embedding $z_{i}^{t}$ and its neighbor embedding $z_{j}^{t}$ are used as query and key/value. Similar", + "bbox": [ + 76, + 821, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to trajectory interaction encoder of HiVT [88], the graph attention operation is operated as follows:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\alpha_ {i} ^ {t} = \\operatorname {s o f t m a x} \\left(\\frac {q _ {i} ^ {t \\top}}{\\sqrt {d _ {k}}} \\cdot \\left[ \\left\\{k _ {j} ^ {t} \\right\\} _ {j \\in N _ {i}} \\right]\\right), \\\\ m _ {i} ^ {t} = \\sum_ {j \\in N _ {i}} \\alpha_ {i} ^ {t} v _ {j} ^ {t}, \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 131, + 890, + 200 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i} ^ {t} = \\operatorname {s i g m o i d} \\left(W ^ {\\text {g a t e}} \\left[ z _ {i} ^ {t}, m _ {i} ^ {t} \\right]\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 204, + 813, + 220 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {z} _ {i} ^ {t} = g _ {i} ^ {t} \\odot W ^ {\\text {s e l f}} z _ {i} ^ {t} + \\left(1 - g _ {i} ^ {t}\\right) \\odot m _ {i} ^ {t}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 223, + 815, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{i}$ is a set of agent $i$ 's neighbors, $W^{gate}$ and $W^{self}$ are learnable matrices, and $\\odot$ is element-wise product.", + "bbox": [ + 500, + 253, + 890, + 285 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.4 Trajectory decoder", + "text_level": 1, + "bbox": [ + 500, + 301, + 687, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Trajectory is forecasted from the output of trajectory encoder which encodes both past global trajectory and local pose information. Since its graph operation is operated by each timestep, a temporal encoder is used as a temporal encoder to integrate $\\widehat{Z}$ in the temporal dimension. A multihead self-attention temporal encoder is used as the temporal encoder. Aggregator then takes into account variations in local coordinate frames to accurately represent geometric relationships within the global coordinate system via a graph operation. MLP is subsequently applied to span embedding $F$ times for multi-modal prediction, which is residually added to the $\\times F$ repeated embedding before the aggregator. Finally, another MLP is used to extract multi-modal future global trajectory proposals of hip joint $Tr_{\\mathcal{F}} \\in \\mathbb{R}^{F \\times T_f \\times 3}$ . The multi-modal embedding is also passed onto the Pose Decoder to forecast local poses.", + "bbox": [ + 498, + 325, + 890, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.5 Pose decoder", + "text_level": 1, + "bbox": [ + 500, + 585, + 643, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Future human pose depends on past human poses and global intention. The pose decoder is designed to consider these factors while generating local poses via mode-specific trajectory conditioning. A transformer (TRM) decoder is used to decode local motions, where pose embedding $Z_{Po}$ is used as key/value and concatenation of trajectory and pose query.", + "bbox": [ + 498, + 609, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\phi_ {M L P} ([ Q _ {P o}, Q _ {T r} ]), K = Z _ {P o}, V = Z _ {P o} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 714, + 890, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using both pose and trajectory queries, past pose embedding $Z_{Po}$ is conditioned on both MPBP sequence at $t = 0$ and the multi-modal trajectory proposals which contain global intent. Subsequently, inverse discrete cosine transform (idct) is applied to convert the future pose proposals from frequency domain to local coordinate domain, $P_{o_{\\mathcal{F}}}$ . The final multi-modal future pose in global coordinates is acquired as Eq. 5 where $\\oplus$ is a joint-wise addition operation.", + "bbox": [ + 496, + 737, + 890, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} = T r _ {\\mathcal {F}} \\oplus P o _ {\\mathcal {F}}, \\quad \\mathbf {Y} \\in \\mathbb {R} ^ {F \\times N _ {A} \\times T _ {f} \\times 3} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 883, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1620", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/02c21887687e40ab21d1c50bdc7c2f9e7df8636a629ffde9346a16fdf0453a20.jpg", + "table_caption": [ + "Table 1. Comparison of statistics between existing human pose forecasting datasets and newly proposed JRDB-GMP dataset." + ], + "table_footnote": [], + "table_body": "
Dataset
CMU-Mocap (UMPM)MuPoTs -3D3DPWJRDB-GMP 1s/2s2s/5s
Duration (s)400026717001863
Location #-20-27
Sample #1300019243211534593
avg. agent3326.86.8
med. agent #33255
max agent #3322422
avg. vel. (m/s)0.30.260.570.460.38
avg. disp.(m)0.630.551.130.640.79
max. disp.(m)4.622.4510.718.4411.0
", + "bbox": [ + 81, + 121, + 468, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fbebee859bddfbb4d3d177925974b79304b1a13b0ffdfc0cf8b14adb1eab21cf.jpg", + "image_caption": [ + "Figure 3. Example scenes from the JRDB-GMP dataset, illustrating its long-term, multi-agent nature." + ], + "image_footnote": [], + "bbox": [ + 83, + 282, + 483, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training objective", + "text_level": 1, + "bbox": [ + 76, + 503, + 253, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both objectives of global trajectory and local pose forecasting are trained jointly. For both global trajectory and local pose prediction, $L2$ loss is propagated to the mode with minimal $L2$ distance with the ground truth.", + "bbox": [ + 75, + 527, + 468, + 588 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {T r} = \\sum_ {n = 1} ^ {N _ {A}} \\sum_ {t = 1} ^ {T _ {f}} \\| \\widetilde {y} _ {T r, n} ^ {t} - \\hat {y} _ {T r, n} ^ {t} \\|\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 598, + 400, + 640 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {P o} = \\sum_ {n = 1} ^ {N _ {A}} \\sum_ {t = 1} ^ {T _ {f}} \\sum_ {j = 1} ^ {J - 1} \\| \\left(\\widetilde {y} _ {P o, n} ^ {t, j} - \\hat {y} _ {P o, n} ^ {t, j}\\right) \\| \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 643, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {T r} + L _ {P o}\n$$\n", + "text_format": "latex", + "bbox": [ + 289, + 691, + 400, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5.JRDB-GMP dataset", + "text_level": 1, + "bbox": [ + 76, + 724, + 267, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the absence of existing long-term $(3s+)$ multi-agent $(6+)$ dataset, we compose a unique 3D human pose forecasting dataset in a real-world environment from JRDB [66]. The original JRDB dataset is constructed by a moving robot that records human activity around a school campus using 5 omnidirectional cameras and LiDAR. Image sequences along with 2D pose annotation and 3D bounding box annotations are provided in the original dataset. However, since 3D human pose annotations are unavailable, we separately parse accurate 3D human pose from provided inputs and", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/507fc764ff5e37ab86f96fc8431046ff42ffa209d9888a29300d9bf19aa962c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 85, + 903, + 247 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a7dce04996df690df884af1d9a2c03334fe254bd55eadb8f4382bf933af9807b.jpg", + "image_caption": [ + "Figure 4. Various motions from the JRDB-MultiGlobPose dataset, providing rich motion queues for inter-agent interaction inference." + ], + "image_footnote": [], + "bbox": [ + 506, + 250, + 903, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "annotations. First, a SOTA monocular 3D pose extraction method [59] is used to extract raw 3D joint positions from image sequences. Then, 2D pose and 3D bounding box annotations are used to refine the raw joint positions and minimize noise. We use 2D pose annotations to initially filter out the 3D poses with noise. With camera parameters and refined 3D pose, we project it on 2D image plane, then L2 distance between projected 2D pose and GT 2D pose annotation is calculated. If the mean L2 distance per each agent at a time stamp is over a threshold, that instance is filtered out. 2D pose annotations are also projected in 3D space to refine the remaining 3D poses, ensuring the accuracy of the 3D pose information of our JRDB-GMP dataset. Further details are elaborated in the supplementary materials.", + "bbox": [ + 496, + 446, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3 visualizes some scenes of the constructed dataset. Accurate extraction of 3D poses has been made even with considerable occlusion via the use of 2D poses. The dataset includes agents with both long and short traverse distances and rich inter-agent interactions in both trajectory and local pose aspects. Figure 4 illustrates diverse local poses included in the dataset, which serve as motion cues of inter-agent interaction. Both figures confirm our method's accuracy in extracting 3D multi-human pose, even in crowded environments. Table 1 scrutinizes the statistics compared to previously used datasets. Compared to earlier datasets, the average number of agents is more than twice as high. In addition, comparing JRDB-GMP 1s/2s to CMU-Mocap and MuPoTs datasets, JRDB contains more diverse and longer motion as shown by a similar magnitude of average displacement but longer maximum displacement.", + "bbox": [ + 496, + 659, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1621", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/64422717c5f9b02cc9d8c810df24b468e44699a5df6f66ab64673ae7626420d7.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of our method to previous methods on CMU-mocap (UMPM), 3DPW, and JRDB-GlobMultiPose datasets with number of prediction modes $(F)$ as 6. Lower is better for all metrics. The best results are marked in bold." + ], + "table_footnote": [], + "table_body": "
DatasetCMU-mocap (UMPM)3DPWJRDB-GlobMultiPose
In/out length (s)1/20.8/1.61/22/5
Evaluation time (s)120.81.6122.55
JPEMRT [69]164.7280.1159.1251.2259.3349.3438.4474.0
JRT [74]168.5316.9181.9287.3237.9373.1351.9538.8
TBIFormer [47]170.0290.9153.9265.8257.1339.3443.2481.3
Ours152.4262.7142.6236.2224.0301.4341.6390.4
APEMRT [69]127.0164.4117.9153.272.387.388.5101.9
JRT [74]121.2181.6133.4178.0112.6154.396.7120.2
TBIFormer [47]125.1160.8115.4152.770.683.388.2102.9
Ours114.4151.7114.6150.070.883.382.294.7
FDEMRT [69]99.6204.7102.7185.3235.2325.2418.2454.8
JRT [74]117.7250.8133.7235.4211.4337.4318.5497.2
TBIFormer [47]112.1228.5106.7215.9232.4314.6423.9458.8
Ours88.7188.974.1158.2194.7271.5313.9361.0
", + "bbox": [ + 153, + 117, + 813, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 351, + 200, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 76, + 377, + 171, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We test our model on three datasets: CMU-Mocap (UMPM) [13, 64], 3DPW [67], and our JRDB-GMP. Although our model is designed to forecast human poses in a long-term multi-agent environment, we also report experimental results on previous benchmark datasets with simpler scenes. Mocap-UMPM is a mixed dataset of Mocap and UMPM containing synthesized human interaction between three agents [47]. 3DPW is a dataset with 2 agents traversing a real-world environment. We report the test results on each after separate training on respective datasets.", + "bbox": [ + 75, + 402, + 468, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Metrics", + "text_level": 1, + "bbox": [ + 76, + 566, + 173, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the following widely-used metrics. For a detailed definition, please refer to the supplementary material.", + "bbox": [ + 75, + 590, + 468, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "APE: Aligned mean per joint Position Error is used as a metric to evaluate the forecasted local motion. $L2$ distance of each joint in the hip joint coordinate is averaged over all joints for a given timestep.", + "bbox": [ + 75, + 621, + 468, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FDE: Final Distance Error evaluates the forecasted global trajectory by calculating the $L2$ distance of a given timestep.", + "bbox": [ + 75, + 681, + 468, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "JPE: Joint Precision Error evaluates both global and local predictions by mean $L2$ distance of all joints for a timestep.", + "bbox": [ + 75, + 710, + 467, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 755, + 290, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our model on a single A6000 GPU. 2 layers of pose encoder transformer are stacked, followed by 2 layers of transformer in pose decoder. Embedding dimensions of 96 and 128 are used for trajectory and pose embeddings, respectively. The transformed key, value dimension of 64 is used for all transformer architectures. A learning rate of 0.003 is used with an AdamW optimizer with weight decay. Further details can be found in the supplementary materials.", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ec862d630b93d64b62f7b919c299579825ae63720035418019c972b381231aa1.jpg", + "table_caption": [ + "Table 3. Short-term prediction results on CMU-Mocap (UMPM) dataset, where 1s of poses are forecasted given 2s of poses." + ], + "table_footnote": [], + "table_body": "
MetricJPEAPEFDE
Time (s)0.20.61.00.20.61.00.20.61.0
MRT64.515221749.811014039.497.9153
JRT31.510417328.785.912517.763.9120
TBIFormer37.410415832.885.811923.363.7104
Ours37.810215833.884.411614.949.192.6
", + "bbox": [ + 504, + 378, + 888, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Baselines", + "text_level": 1, + "bbox": [ + 500, + 468, + 607, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method against the latest SOTA methods for multi-agent pose forecasting [47, 69, 74]. To compare the multi-modal predictions of these three methods, we extend their prediction modes by spanning embedding $K$ times in the same manner as ours. All baselines are trained and evaluated on CMU-mocap (UMPM) and 3DPW datasets. CMU-Mocap (UMPM) dataset predicts 2 seconds from 1 second of poses, and 3DPW predicts 1.6 seconds from 0.8 seconds of poses, both from 6 modes. For 3DPW dataset, we slightly lengthen the forecast horizon to evaluate long term predictions. For JRDB-GMP dataset, both short (1s/2s) and long term (2s/5s) predictions are evaluated for all models. Lastly, We use HiVT [88] as the baseline For global trajectory prediction of our model.", + "bbox": [ + 496, + 492, + 890, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 500, + 715, + 586, + 732 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.Quantitative results", + "text_level": 1, + "bbox": [ + 500, + 741, + 687, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 compares the quantitative performances on three datasets. Our method exhibits considerable performance gain against all previous SOTA methods, not only on the proposed long-term multi-agent dataset but also on the existing two datasets. Such generalized competence demonstrates the applicability of our trajectory-conditioned pose forecasting method to various real-world scenarios. In detail, our approach achieves over $10\\%$ gain of FDE on all datasets. This improvement on forecasting global locomo", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1622", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/10bbec035091361641fd3dfb6ac0716dfbc5352256d96f352c218c3ce2aae723.jpg", + "image_caption": [ + "t \n2.5s \n5s" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 276, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a0d733b417b0691157108f7eda42efe5ca5e599fe5f92e2043110e954c4d6a3.jpg", + "image_caption": [ + "Figure 5. Visualization of a long-term forecasting scene from JRDB-GMP (2/5) dataset. Past poses for input are shown on the leftmost column, GT future poses on the next, and forecasts by ours, MRT, TBIFormer, and JRT, respectively." + ], + "image_footnote": [], + "bbox": [ + 310, + 89, + 426, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2a9882d1651207055f03c33321a9385ce57ff85ec9b2489d2431b9c9a4cf832.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 90, + 544, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b44c99c37110e82209b08963524d194ca4c4042e416e7bbf7e289a55409aec50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 90, + 642, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a6003607be61c44db3936bb5304853d7277c6be956e5bea3528142594565f3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 90, + 766, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6148c95f486c1457ca9e5c06749adb80bd4715c870cbfa4143d5448d43d8daef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 90, + 888, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f23d1a7548a187491fa3133a38a67f1db919411589fe758e0fcd2a8693da24d.jpg", + "image_caption": [ + "Figure 6. Visualization of a CMU-Mocap (UMPM) scene. Past poses are shown on the upper row, GT future poses on the next, and forecasts by TBIFormer and ours on the latter two rows. To visualize motion, we stack several frames around the target time stamp. Black/red/blue arrows refer to the direction of the global trajectory, and yellow arrows refer to the direction of foot motion." + ], + "image_footnote": [], + "bbox": [ + 78, + 266, + 467, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion could be accredited to the decoupled forecasting of global trajectory and local pose. Previous methods holistically predict both global and local movements, limiting both performances due to superfluous interactions to consider between all joints. Conversely, our approach can extract accurate global intent by decoupling past motion into global and local representations. Moreover, our effective interaction modeling of global and local pose also helps to predict a more accurate global trajectory under multi-agent environment as shown in the latter ablation study.", + "bbox": [ + 75, + 580, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For APE metric, our method also surpasses previous SOTA models on all datasets, highlighting the accurate extraction of local pose intent. Such improvement shows that our approach generates plausible local motion due to its proficient sampling from coarse global intents. Our method simplifies the task by learning multi-modality in a coarse-to-fine approach. Its subsequent local motion forecasting is inferred from coarsely modeled multi-modality, a greatly simplified task compared to extracting intent from entangled multi-modality as well as multi-agent interaction.", + "bbox": [ + 75, + 733, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These improvements on both global and local scales", + "bbox": [ + 96, + 885, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "jointly contribute toward lowering the JPE metric, demonstrating proficiency of our method in forecasting overall human motion. Based on such competence, our method which aimed towards improving on forecasting long-term multiagent environments also exhibits similar or better performances on short timeframes as shown in Tab. 3. In addition, our approach also excels even on sole local motion with minimal global displacement, as elaborated in section 7.2 of supplementary materials.", + "bbox": [ + 496, + 268, + 893, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Qualitative results", + "text_level": 1, + "bbox": [ + 500, + 417, + 676, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method forecasts a more plausible global pose in longer timescales ( $\\sim 5\\mathrm{s}$ ) as shown in the interacting scene of five agents in Fig. 5. Looking at the input and GT sequences, the leftmost person avoids the traversing couple from right to left. The two people in front are stationary while talking to each other. MRT and TBIFormer forecasts implausible overlapped poses at the final prediction horizon ( $t = 5\\mathrm{s}$ ). JRT fails to learn the global locomotion of agents due to the high complexity of its attention mechanism and is stuck in the local minimum of predicting the inactivity of all agents. On the other hand, our model forecasts plausible poses where the closely interacting two agents walk side-by-side.", + "bbox": [ + 496, + 441, + 892, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 6 illustrates exemplary sequences where more natural local motion has been forecasted by our method. Comparing forecasts on a scene of walking agents, our method generates a much more plausible sequence where the stepping foot remains stationary. On the other hand, the previous SOTA method, TBIFormer, struggles to learn the natural walking mechanism of human legs and a parallel translation of both feet is exhibited. Such discrepancy shows that trajectory-conditioning for inferring local motion from global intent generates more proficient details in human motion than SOTA methods. More visualizations could be found in the supplementary materials.", + "bbox": [ + 496, + 623, + 893, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation studies", + "text_level": 1, + "bbox": [ + 500, + 816, + 661, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different number of modes. The main quantitative results report prediction results with $F$ as 6 to compare the ability to address the multi-modal nature of human motion during pose forecasting. Table. 4 additionally compares forecast", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1623", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/64970b331bc7a6b1f21847f375f9cd2dbb57b93508af6d744446afb5b3c212f9.jpg", + "table_caption": [ + "Table 4. Comparison of performance with different number of modes in CMU-Mocap (UMPM) dataset." + ], + "table_footnote": [], + "table_body": "
F16
Metric @ 2sAPEJPEAPEJPE
MRT163.9366.4164.4280.1
JRT176.7367.4181.6316.9
TBIFormer160.1374.3160.8290.9
Ours154.4366.4151.7262.7
", + "bbox": [ + 117, + 122, + 426, + 215 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/90cf8cf8052668432edfe65e6e5ab7da69fb50d116ca64af34f49ad32cf5aad9.jpg", + "table_caption": [ + "Table 5. Ablation studies on core components of model structures. Experiments are done with JRDB-GMP dataset to evaluate multiagent long-term performance." + ], + "table_footnote": [], + "table_body": "
Exp. #Trajectory encoderPose decoderMetrics
Local pose embeddingAgent interactionTrajectory -conditioningJPE @5sAPE @5sFDE @5s
-471.4101.7457.9
1✓400.595.1370.9
2✓403.394.7374.2
3✓401.293.0372.8
4✓✓395.693.8366.8
5✓✓392.795.2363.4
6✓✓✓391.291.4363.3
", + "bbox": [ + 81, + 275, + 468, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "results with $F$ as 1. Our method again achieves noticeable improvement in APE over the baseline on single-modal forecasts. With $F = 1$ , although our method barely enjoys improvement in forecasting global motion due to the absence of multi-modality, its superiority in APE shows the validity of our coarse-to-fine forecasting strategy that also effectively captures agent interaction. Our method improves with multi-modal predictions, demonstrating the proficiency of a coarse-to-fine approach in interpreting the stochastic nature of human motion and its intent. Note that our method improves in APE along with an increase in $F$ unlike previous methods, indicating a unique aptitude in addressing the multi-modal nature of not only global locomotion but also local pose intent via trajectory-conditioning.", + "bbox": [ + 75, + 412, + 468, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importance of each architecture component. Table. 5 reports the influence of core components of our model. For the trajectory encoder, we evaluate the importance of using local pose embedding and modeling agent interaction. Comparing experiments 1, 5 and 3, 4, both show improvements in JPE and FDE metrics with the use of local pose embedding. Our method has taken advantage of detailed local pose cues to infer an agent's global intention. For interaction modeling, its use is beneficial for both global and local forecasts as compared by experiments 4 and 6. These joint improvements demonstrate the importance of considering local and global motion interactions for their respective forecasts. As for the pose decoder, comparisons of experiments 2,4 and 5,6 both show improvements in APE metric. Such consistent improvement verifies the effectiveness of the trajectory-conditioned local motion forecast approach in generating plausible local motion from global intention.", + "bbox": [ + 75, + 626, + 468, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importance of interaction modeling. Accurate model", + "bbox": [ + 76, + 885, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5381b8d451966346c75439fb30ea8b4350813992a9c09099f1f9e742849f00a4.jpg", + "table_caption": [ + "Table 6. Ablation studies on agent interaction cutoff distance on JRDB-GMP." + ], + "table_footnote": [], + "table_body": "
JPE @ 5s
TBIFormerOurs
w/o interaction483.8406.0
w/ interaction < 2m-403.5
w/ interaction < 4m-400.5
w/ interaction all481.3390.4
", + "bbox": [ + 553, + 122, + 836, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing of inter-agent interaction becomes more pivotal to forecast in more complex environments. Indeed, its complexity grows in a long-term multi-agent scene. When holistically considering joint-wise interaction for all timesteps, the computation complexity is acquired as $O(T^2 \\cdot N^2 \\cdot J^2)$ , where $T$ is the number of timesteps, $N$ the number of agents, and $J$ the number of joints. On the contrary, with interaction modeling in global trajectory scale, our method reduces the computation cost by $TJ^2$ into $O(T \\cdot N^2)$ . This enables efficient and proficient modeling of intra (pose) and inter (trajectory)-agent interactions as shown by Tab. 6. While body part-wise interaction modeling only improved by $0.52\\%$ for TBIFormer, ours improves up to $3.84\\%$ with interaction modeling. This demonstrates the proficiency of our efficient interaction modeling-based method in inferring global and local intents from complex interactions. In addition, the gradual improvement of JPE according to a wider interaction range confirms the importance of interaction modeling of more agents, which cannot be learned from the arbitrarily mixed previous datasets.", + "bbox": [ + 496, + 224, + 890, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 542, + 617, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel interaction-aware trajectory-conditioned approach to handle long-term multi-agent motion forecasting, along with a new dataset suited for such scope. Our proposed model utilizes a coarse-to-fine approach and decouples overall motion prediction into global and local components. Multi-modality of human motion is proficiently modeled via inferring fine local intents from coarse global intents, along with efficient agent-wise interaction modeling. As for the dataset, our JRDB-GMP dataset contains unprecedented long-term $(5s+)$ multi-agent $(6+)$ interactions in a real-world setting. Our method achieves state-of-the-art performance on all previous datasets and JRDB-GMP dataset, offering generalized practical implications in real-world applications.", + "bbox": [ + 496, + 568, + 890, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601).", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1624", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vida Adeli, Ehsan Adeli, Ian Reid, Juan Carlos Niebles, and Hamid Rezatofighi. Socially and contextually aware human motion and pose forecasting. IEEE Robotics and Automation Letters, 5(4):6033-6040, 2020. 2", + "[2] Vida Adeli, Mahsa Ehsanpour, Ian Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13390-13400, 2021. 2", + "[3] Gorkay Aydemir, Adil Kaan Akan, and Fatma Güney. Adapt: Efficient multi-agent trajectory prediction with adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8295-8305, 2023. 2", + "[4] German Barquero, Sergio Escalera, and Cristina Palmero. Belfusion: Latent diffusion for behavior-driven human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2317-2327, 2023. 1, 2", + "[5] Arij Bouazizi, Adrian Holzbock, Ulrich Kressel, Klaus Dietmayer, and Vasileios Belagiannis. Motionmixer: Mlp-based 3d human body pose forecasting. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 791-798. International Joint Conferences on Artificial Intelligence Organization, 2022. Main Track. 2", + "[6] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 387–404. Springer, 2020. 2", + "[7] Wenhao Chai, Zhongyu Jiang, Jenq-Neng Hwang, and Gaoang Wang. Global adaptation meets local generalization: Unsupervised domain adaptation for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14655-14665, 2023. 3", + "[8] Ling-Hao Chen, JiaWei Zhang, Yewen Li, Yiren Pang, Xiaobo Xia, and Tongliang Liu. Humanmac: Masked motion completion for human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9544-9555, 2023. 2", + "[9] Hsu-kuang Chiu, Ehsan Adeli, Borui Wang, De-An Huang, and Juan Carlos Niebles. Action-agnostic human pose forecasting. In 2019 IEEE winter conference on applications of computer vision (WACV), pages 1423–1432. IEEE, 2019. 2", + "[10] Sehwan Choi, Jungho Kim, Junyong Yun, and Jun Won Choi. R-pred: Two-stage motion prediction via tube-query attention-based trajectory refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8525-8535, 2023. 2", + "[11] Rohan Choudhury, Kris M. Kitani, and László A. Jeni. Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In Proceedings of the IEEE/CVF International" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision (ICCV), pages 14750-14760, 2023. 2", + "[12] Andrea Porfiri Dal Cin, Giacomo Boracchi, and Luca Magri. Multi-body depth and camera pose estimation from multiple views. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17804-17814, 2023. 2", + "[13] CMU-Graphics-Lab. Cmu graphics lab motion capture database. http://mocap.cs.cmu.edu/, 2003. 2, 6", + "[14] Runyang Feng, Yixing Gao, Xueqing Ma, Tze Ho Elden Tse, and Hyung Jin Chang. Mutual information-based temporal difference learning for human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17131-17141, 2023. 2", + "[15] Runyang Feng, Yixing Gao, Tze Ho Elden Tse, Xueqing Ma, and Hyung Jin Chang. Diffpose: Spatiotemporal diffusion model for video-based human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14861-14872, 2023. 3", + "[16] Xuehao Gao, Shaoyi Du, Yang Wu, and Yang Yang. Decompose more and aggregate better: Two closer looks at frequency representation learning for human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6451-6460, 2023. 2", + "[17] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13041-13051, 2023. 3", + "[18] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15303-15312, 2021. 2", + "[19] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15303-15312, 2021. 2", + "[20] Wen Guo, Xiaoyu Bie, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Multi-person extreme motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13053-13064, 2022. 1", + "[21] Je-Seok Ham, Dae Hoe Kim, NamKyo Jung, and Jinyoung Moon. Cipf: Crossing intention prediction network based on feature fusion modules for improving pedestrian safety. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3665-3674, 2023. 1", + "[22] Karl Holmquist and Bastian Wandt. Diffpose: Multi-hypothesis human pose estimation using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15977-15987, 2023. 2", + "[23] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 1" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1625", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Boyuan Jiang, Lei Hu, and Shihong Xia. Probabilistic triangulation for uncalibrated multi-view 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14850-14860, 2023. 2", + "[25] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9644-9653, 2023. 2", + "[26] Zhehan Kan, Shuoshuo Chen, Ce Zhang, Yushun Tang, and Zhihai He. Self-correctable and adaptable inference for generalizable human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5537-5546, 2023. 2", + "[27] Jeongho Kim, Wooksu Shin, Hancheol Park, and Jongwon Baek. Addressing the occlusion problem in multi-camera people tracking with human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5462-5468, 2023. 1", + "[28] Mihee Lee, Samuel S Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. Musevae: multi-scale vae for environment-aware long term trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2221-2230, 2022. 2", + "[29] Lihuan Li, Maurice Pagnucco, and Yang Song. Graph-based spatial transformer with memory replay for multifuture pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2241, 2022. 2", + "[30] Huan Liu, Qiang Chen, Zichang Tan, Jiang-Jiang Liu, Jian Wang, Xiangbo Su, Xiaolong Li, Kun Yao, Junyu Han, Errui Ding, Yao Zhao, and Jingdong Wang. Group pose: A simple baseline for end-to-end multi-person pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15029-15038, 2023. 2", + "[31] Tiezheng Ma, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Progressively generating better initial guesses towards next stages for high-quality human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6437-6446, 2022. 1", + "[32] Takahiro Maeda and Norimichi Ukita. Motionaug: Augmentation with physical correction for human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6427-6436, 2022. 1", + "[33] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15233-15242, 2021. 2", + "[34] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9489-9497, 2019. 2", + "[35] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Generating smooth pose sequences for diverse human motion pre" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "diction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13309-13318, 2021. 1", + "[36] Wei Mao, Richard I Hartley, Mathieu Salzmann, et al. Contact-aware human motion forecasting. Advances in Neural Information Processing Systems, 35:7356-7367, 2022. 2", + "[37] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Weakly-supervised action transition learning for stochastic human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8151-8160, 2022. 1", + "[38] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5517-5526, 2023. 2", + "[39] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 2018 International Conference on 3D Vision (3DV), pages 120-130, 2018. 2", + "[40] Jiquan Ngiam, Vijay Vasudevan, Benjamin Caine, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, David J Weiss, Benjamin Sapp, Zhifeng Chen, and Jonathon Shlens. Scene transformer: A unified architecture for predicting future trajectories of multiple agents. In International Conference on Learning Representations, 2022. 2", + "[41] Daehee Park, Jaewoo Jeong, and Kuk-Jin Yoon. Improving transferability for cross-domain trajectory prediction via neural stochastic differential equation. arXiv preprint arXiv:2312.15906, 2023. 2", + "[42] Daehee Park, Hobin Ryu, Yunseo Yang, Jegyeong Cho, Jiwon Kim, and Kuk-Jin Yoon. Leveraging future relationship reasoning for vehicle trajectory prediction. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[43] Daehee Park, Jaeseok Jeong, Sung-Hoon Yoon, Jaewoo Jeong, and Kuk-Jin Yoon. T4p: Test-time training of trajectory prediction via masked autoencoder and actor-specific token memory, 2024. 2", + "[44] Sungchan Park, Eunyi You, Inhoe Lee, and Joonseok Lee. Towards robust and smooth 3d multi-person pose estimation from monocular videos in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14772-14782, 2023. 3", + "[45] Behnam Parsaeifard, Saeed Saadatnejad, Yuejiang Liu, Taylor Mordan, and Alexandre Alahi. Learning decoupled representations for human pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2294-2303, 2021. 2", + "[46] Qucheng Peng, Ce Zheng, and Chen Chen. Source-free domain adaptive human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4826-4836, 2023. 3", + "[47] Xiaogang Peng, Siyuan Mao, and Zizhao Wu. Trajectory-aware body interaction transformer for multi-person pose forecasting. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 503, + 92, + 890, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1626", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 17121- 17130, 2023. 1, 2, 4, 6", + "[48] Zhongwei Qiu, Qiansheng Yang, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Chang Xu, Dongmei Fu, and Jingdong Wang. Psvt: End-to-end multi-person 3d pose and shape estimation with progressive video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21254-21263, 2023. 3", + "[49] Muhammad Rameez Ur Rahman, Luca Scofano, Edoardo De Matteis, Alessandro Flaborea, Alessio Sampieri, and Fabio Galasso. Best practices for 2-body pose forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3613-3623, 2023. 2", + "[50] Dripta S. Raychaudhuri, Calvin-Khang Ta, Arindam Dutta, Rohit Lal, and Amit K. Roy-Chowdhury. Prior-guided source-free domain adaptation for human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14996-15006, 2023. 3", + "[51] Luke Rowe, Martin Ethier, Eli-Henry Dykhne, and Krzysztof Czarnecki. Fjmp: Factorized joint multi-agent motion prediction over learned directed acyclic interaction graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13745-13755, 2023. 2", + "[52] Saeed Saadatnejad, Ali Rasekh, Mohammadreza Mofayezi, Yasamin Medghalchi, Sara Rajabzadeh, Taylor Mordan, and Alexandre Alahi. A generic diffusion-based approach for 3d human pose prediction in the wild. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8246-8253. IEEE, 2023. 2", + "[53] Tim Salzmann, Marco Pavone, and Markus Ryll. Motron: Multimodal probabilistic human motion forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6457-6466, 2022. 1", + "[54] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. IEEE Robotics and Automation Letters, 2023. 1, 2", + "[55] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14761-14771, 2023. 3", + "[56] Xiaolong Shen, Zongxin Yang, Xiaohan Wang, Jianxin Ma, Chang Zhou, and Yi Yang. Global-to-local modeling for video-based 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8887-8896, 2023. 3", + "[57] Mingyi Shi, Sebastian Starke, Yuting Ye, Taku Komura, and Jungdam Won. Phasemp: Robust 3d pose estimation via phase-conditioned human motion prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14725-14737, 2023. 2", + "[58] Theodoros Sofianos, Alessio Sampieri, Luca Franco, and Fabio Galasso. Space-time-separable graph convolutional" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "network for pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11209-11218, 2021. 2", + "[59] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13243-13252, 2022. 2, 3, 5", + "[60] Yuran Sun, Alan William Dougherty, Zhuoying Zhang, Yi King Choi, and Chuan Wu. Mixsynthformer: A transformer encoder-like structure with mixed synthetic self-attention for efficient human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14884–14893, 2023. 2", + "[61] Zhenhua Tang, Zhaofan Qiu, Yanbin Hao, Richang Hong, and Ting Yao. 3d human pose estimation with spatiotemporal criss-cross attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4790-4799, 2023. 2", + "[62] Julian Tanke, Chintan Zaveri, and Juergen Gall. Intention-based long-term human motion anticipation. In 2021 International Conference on 3D Vision (3DV), pages 596-605. IEEE, 2021. 2", + "[63] Julian Tanke, Linguang Zhang, Amy Zhao, Chengcheng Tang, Yujun Cai, Lezi Wang, Po-Chen Wu, Juergen Gall, and Cem Keskin. Social diffusion: Long-term multiple human motion anticipation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9601-9611, 2023. 2", + "[64] N.P. van der Aa, X. Luo, G.J. Giezeman, R.T. Tan, and R.C. Veltkamp. Umpm benchmark: A multi-person dataset with synchronized video and motion capture data for evaluation of articulated human motion and interaction. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 1264-1269, 2011. 2, 6", + "[65] Edward Vendrow, Satyajit Kumar, Ehsan Adeli, and Hamid Rezatofighi. Somoformer: Multi-person pose forecasting with transformers. arXiv preprint arXiv:2208.14023, 2022. 1", + "[66] Edward Vendrow, Duy Tho Le, Jianfei Cai, and Hamid Rezatofighi. Jrdb-pose: A large-scale dataset for multiperson pose estimation and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4811–4820, 2023. 2, 3, 5", + "[67] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2, 6", + "[68] Chenxi Wang, Yunfeng Wang, Zixuan Huang, and Zhiwen Chen. Simple baseline for single human motion forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2260-2265, 2021. 2", + "[69] Jiashun Wang, Huazhe Xu, Medhini Narasimhan, and Xiaolong Wang. Multi-person 3d motion prediction with multirange transformers. Advances in Neural Information Processing Systems, 34:6036-6049, 2021. 1, 2, 6" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1627", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Mingkun Wang, Xinge Zhu, Changqian Yu, Wei Li, Yuexin Ma, Ruochun Jin, Xiaoguang Ren, Dongchun Ren, Mingxu Wang, and Wenjing Yang. Ganet: Goal area network for motion forecasting. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1609-1615. IEEE, 2023. 2", + "[71] Yucheng Xing and Xin Wang. Hdg-ode: A hierarchical continuous-time model for human pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14700–14712, 2023. 1, 2", + "[72] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6488-6497, 2022. 2", + "[73] Chenxin Xu, Robby T. Tan, Yuhong Tan, Siheng Chen, Xin-chao Wang, and Yanfeng Wang. Auxiliary tasks benefit 3d skeleton-based human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9509–9520, 2023. 2", + "[74] Qingyao Xu, Weibo Mao, Jingze Gong, Chenxin Xu, Siheng Chen, Weidi Xie, Ya Zhang, and Yanfeng Wang. Joint-relation transformer for multi-person motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9816-9826, 2023. 1, 6", + "[75] Sirui Xu, Yu-Xiong Wang, and Liangyan Gui. Stochastic multi-person 3d motion forecasting. In The Eleventh International Conference on Learning Representations, 2023. 1, 2", + "[76] Yi Xu, Armin Bazarjani, Hyung-gun Chi, Chiho Choi, and Yun Fu. Uncovering the missing pattern: Unified framework towards trajectory imputation and prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9632-9643, 2023. 2", + "[77] Maosheng Ye, Jiamiao Xu, Xunnong Xu, Tengfei Wang, Tongyi Cao, and Qifeng Chen. Bootstrap motion forecasting with self-consistent constraints. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8504-8514, 2023. 2", + "[78] Yingxuan You, Hong Liu, Ti Wang, Wenhao Li, Runwei Ding, and Xia Li. Co-evolution of pose and mesh for 3d human body estimation from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14963-14973, 2023. 2", + "[79] Bruce X.B. Yu, Zhi Zhang, Yongxu Liu, Sheng-hua Zhong, Yan Liu, and Chang Wen Chen. Gla-gen: Global-local adaptive graph convolutional network for 3d human pose estimation from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8818-8829, 2023. 2", + "[80] Kai Zhai, Qiang Nie, Bo Ouyang, Xiang Li, and Shanlin Yang. Hopfir: Hop-wise graphformer with intragroup joint refinement for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14985-14995, 2023. 2", + "[81] Yi Zhang, Pengliang Ji, Angtian Wang, Jieru Mei, Adam Kortylewski, and Alan Yuille. 3d-aware neural body fitting for" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "occlusion robust 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9399-9410, 2023. 2", + "[82] He Zhao and Richard P. Wildes. Where are you heading? dynamic trajectory prediction with expert goal examples. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7629-7638, 2021. 2", + "[83] Qitao Zhao, Ce Zheng, Mengyuan Liu, Pichao Wang, and Chen Chen. Poseformerv2: Exploring frequency domain for efficient and robust 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8877-8886, 2023. 2", + "[84] Fang Zheng, Le Wang, Sanping Zhou, Wei Tang, Zhenxing Niu, Nanning Zheng, and Gang Hua. Unlimited neighborhood interaction for heterogeneous trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13168-13177, 2021. 2", + "[85] Jingxiao Zheng, Xinwei Shi, Alexander Gorban, Junhua Mao, Yang Song, Charles R Qi, Ting Liu, Visesh Chari, Andre Cornman, Yin Zhou, et al. Multi-modal 3d human pose estimation with 2d weak supervision in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4478-4487, 2022. 1", + "[86] Chongyang Zhong, Lei Hu, Zihao Zhang, Yongjing Ye, and Shihong Xia. Spatio-temporal gating-adjacency gen for human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6447–6456, 2022. 1", + "[87] Mu Zhou, Lucas Stoffl, Mackenzie Weygandt Mathis, and Alexander Mathis. Rethinking pose estimation in crowds: Overcoming the detection information bottleneck and ambiguity. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14689-14699, 2023. 3", + "[88] Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Kejie Lu. Hivt: Hierarchical vector transformer for multi-agent motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8823-8833, 2022. 2, 4, 6", + "[89] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 2", + "[90] Dekai Zhu, Guangyao Zhai, Yan Di, Fabian Manhardt, Hendrik Berkemeyer, Tuan Tran, Nassir Navab, Federico Tombari, and Benjamin Busam. Ipcc-tp: Utilizing incremental pearson correlation coefficient for joint multi-agent trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5507-5516, 2023. 2", + "[91] Tao Zhuo, Zhiyong Cheng, Peng Zhang, Yongkang Wong, and Mohan Kankanhalli. Unsupervised online video object segmentation with motion property understanding. IEEE Transactions on Image Processing, 29:237-249, 2019. 1" + ], + "bbox": [ + 503, + 92, + 890, + 852 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1628", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/12609e19-efed-4425-b5ad-32eb26a81347_model.json b/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/12609e19-efed-4425-b5ad-32eb26a81347_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c390131ea66b894e6b36de7e3ecba0d0ce29b955 --- /dev/null +++ b/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/12609e19-efed-4425-b5ad-32eb26a81347_model.json @@ -0,0 +1,2765 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.13, + 0.861, + 0.177 + ], + "angle": 0, + "content": "Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.204, + 0.68, + 0.237 + ], + "angle": 0, + "content": "Jaewoo Jeong*, Daehee Park*, and Kuk-Jin Yoon KAIST" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.242, + 0.651, + 0.258 + ], + "angle": 0, + "content": "{jeong207,bag2824,kjyoon}@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.324, + 0.474, + 0.687 + ], + "angle": 0, + "content": "Human pose forecasting garners attention for its diverse applications. However, challenges in modeling the multi-modal nature of human motion and intricate interactions among agents persist, particularly with longer timescales and more agents. In this paper, we propose an interaction-aware trajectory-conditioned long-term multi-agent human pose forecasting model, utilizing a coarse-to-fine prediction approach: multi-modal global trajectories are initially forecasted, followed by respective local pose forecasts conditioned on each mode. In doing so, our Trajectory2Pose model introduces a graph-based agent-wise interaction module for a reciprocal forecast of local motion-conditioned global trajectory and trajectory-conditioned local pose. Our model effectively handles the multi-modality of human motion and the complexity of long-term multi-agent interactions, improving performance in complex environments. Furthermore, we address the lack of long-term \\((6s+)\\) multi-agent \\((5+)\\) datasets by constructing a new dataset from real-world images and 2D annotations, enabling a comprehensive evaluation of our proposed model. State-of-the-art prediction performance on both complex and simpler datasets confirms the generalized effectiveness of our method. The code is available at https://github.com/Jaewoo97/T2P." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.716, + 0.21, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.741, + 0.47, + 0.878 + ], + "angle": 0, + "content": "Human pose forecasting aims to predict future human motion based on observed past motion [20, 31, 32, 35, 37, 53, 86]. Humans instinctively perform such tasks, allowing them to naturally navigate in crowded areas or identify and circumvent potential dangers. For this reason, human pose forecasting plays an important role in various computer vision tasks [21, 23, 27, 54, 85, 91]. Indeed, recent years have seen a proliferation of work on multi-agent motion forecasting which aim towards modeling complex multi-agent inter" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.369, + 0.68, + 0.422 + ], + "angle": 0, + "content": "
Global Intent
ABC
Walk side-by-side\nAvoid CKneel
" + }, + { + "type": "table", + "bbox": [ + 0.713, + 0.369, + 0.891, + 0.422 + ], + "angle": 0, + "content": "
Local Intent
ABC
Walking motion\nTalking gesturesTie shoe
" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.428, + 0.895, + 0.512 + ], + "angle": 0, + "content": "Figure 1. Human motion is goal-directed and influenced by other entities. Therefore, global intention contains hints for local intention, allowing us to infer local pose from global trajectories. Our method first forecasts global trajectories, upon which local poses are conditioned for subsequent forecasts. Pose and trajectory-wise inter-agent interactions are considered for both predictions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.535, + 0.68, + 0.551 + ], + "angle": 0, + "content": "action [20, 47, 53, 71, 74]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Although various methods have been proposed, they share two major limitations. The first is a limitation on long-term predictions, as previous studies predicted up to 3 seconds at most [4, 47, 74, 75]. However, a sufficiently long forecast horizon is essential to fully leverage human pose forecasting for diverse downstream tasks in the scope of identifying potential danger or understanding human behavior. The second is that multi-person interactions are not proficiently learned. Existing methods consider the joints of multiple people all at once as objects of interaction [47, 65, 74], resulting in an excessive complexity with respect to the number of joints. Due to such inefficient modeling, these approaches are found to be incompetent in long-term \\((3s+)\\) multi-agent \\((6+)\\) settings, limiting their practicality on complex real-world environments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Moreover, these challenges are also due to the limitations of datasets. Existing pose forecasting datasets have limited sequence length (\\(\\sim 3\\mathrm{s}\\)) and number of agents (\\(\\sim 2\\)). Therefore, previous works [47, 69, 75] have randomly blended disparate datasets to model multi-agent interaction with up to 10 agents. Yet, such naively merged data lacks authentic interaction as agents from different scenes remain uninfluenced. As such, there was no opportunity to develop and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.251, + 0.9 + ], + "angle": 0, + "content": "*Denotes equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1617" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.457, + 0.106 + ], + "angle": 0, + "content": "evaluate a model in a long-term multi-agent environment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.109, + 0.47, + 0.472 + ], + "angle": 0, + "content": "To this end, we present a solution from both model and dataset perspectives to tackle long-term multi-agent human pose forecasting. First, from a model perspective, we propose an interaction-aware trajectory-conditioned pose forecasting method. We point out that the limitations of existing methods on long-term multi-agent environments lead to poor performance in handling the multi-modal nature of human motion and correspondingly complex interactions. To improve upon handling multi-modality in these complex settings, we use a coarse-to-fine approach to enjoy effective interaction modeling by propagating agent-wise coarse representations. Agent-wise pose and trajectory embeddings are obtained in their respective local coordinates, followed by a holistic interaction modeling via our proposed Traj-pose module. Interaction-aware forecasts are then made by initial coarse global hip joint trajectory forecast followed by fine local pose forecasts in its hip joint coordinates, conditioned on the global trajectory as shown in Fig. 1. As discovered in previous research [2, 54], learning an agent-wise global intention as coarse trajectories is less challenging than predicting every joint-wise motion. We leverage these hints from global trajectories, which are further conditioned towards forecasting local motion that embodies the interaction-aware spatio-temporal context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.759 + ], + "angle": 0, + "content": "From a dataset perspective, we parsed a novel real-world dataset for long-term multi-agent human pose forecasting. We utilize JRDB dataset [66] which consists of multi-view video and collected in various environments. Since 3D pose annotations are not provided in the original JRDB, we extracted sequences of 3D human pose from visible agents in omnidirectional images using the latest algorithm for 3D pose extraction from image [59]. We then ensure the reliability of 3D pose information by filtering and adjusting the extracted 3D poses based on 2D pose and 3D bounding box annotations. As a result, we construct a real-world 3D human pose forecasting dataset, JRDB-GlobMultiPose (JRDB-GMP), where up to 24 agents exist for up to 5 seconds. The proposed pose forecasting model is validated on both previous datasets and newly created JRDB-GMP dataset. Our method shows state-of-the-art forecasting performance in both global and local accuracy metrics, not only on JRDB-GMP but also on all previous datasets. Therefore, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.762, + 0.469, + 0.807 + ], + "angle": 0, + "content": "- We propose a interaction-aware trajectory-conditioned pose forecasting method \\((T2P)\\) for long-term multi-agent 3D human pose forecasting." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.809, + 0.469, + 0.853 + ], + "angle": 0, + "content": "- We propose a long-term, multi-agent real-world 3D human pose forecasting dataset which contains up to 24 persons and forecasts up to 5 seconds." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- We validate our \\( T2P \\) model on both previous datasets and our new JRDB-GMP dataset. Our method achieves state-of-the-art forecasting performance on all datasets." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.762, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.645, + 0.105 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.112, + 0.722, + 0.129 + ], + "angle": 0, + "content": "2.1. Human pose forecasting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.136, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Human Pose Forecasting involves predicting a future pose sequence with temporal length of a prediction horizon, given a historical pose sequence [5, 8, 11, 16, 36, 45, 49, 52, 68, 71]. In the early stage, methods were developed to forecast single person motion within a short timeframe (\\(\\sim\\) 1s) [9, 34, 58, 73]. However, to improve applicability on diverse downstream computer vision tasks, forecasts are to be made on multi-person poses [1, 2, 63] for longer prediction horizons [6, 62]. Forecasting future inherently involves a stochastic nature, and handling such multi-modality has been attempted by forecasting multiple future poses of a single agent [4]. However, comparatively marginal efforts have been employed in the more complex long-term multi-agent scenes [75]. Such absence is mostly due to the lack of a proper dataset. The commonly used evaluation datasets are CMU-Mocap [13], 3DPW [67], UMPM [64], MuPoTS-3D [39], all of which contain 2 agents at most in a given scene and have short prediction horizons within 3 seconds. Most recent research arbitrarily combines individual scenes to create datasets with more than three individuals [47, 69, 75]. However, such a synthetic approach does not account for authentic agent interactions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.48, + 0.702, + 0.496 + ], + "angle": 0, + "content": "2.2. Trajectory prediction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.892, + 0.775 + ], + "angle": 0, + "content": "Trajectory prediction involves predicting the future path of an object given its past trajectory [3, 10, 25, 38, 40, 41, 51, 76, 77, 89, 90]. Unlike human pose forecasting which aims to predict every joint position, trajectory prediction regards each agent as a point mass, typically the center of mass or center point of a detected bounding box. Research in trajectory forecasting is interested in not only vehicles but also many types of agents including humans, cyclists, and so on [29, 43, 72, 84]. One substantial direction in research within this field is the Goal-conditioned prediction approach [18, 33, 82]. Goal-conditioned prediction approach first predicts the final destination within the prediction horizon with multiple goal proposals [28, 70]. Then, a thorough future path is conditioned on each mode of the multi-modal proposals. Compared to directly predicting full trajectories, the goal-conditioned approach follows a coarse-to-fine prediction and is effective in learning highly stochastic multi-modality of complex scenes [19, 42, 88]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.809, + 0.803 + ], + "angle": 0, + "content": "2.3. Human pose estimation from image" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Human pose estimation is the task of inferring the pose of a person from an image or a video [22, 24, 26, 30, 57, 60, 78, 80, 81]. Initial deep learning-based methods first utilized convolutional neural networks to estimate 2D and 3D poses from single or multiple images [12, 14, 61, 79, 83]. Recent approaches engage in more challenging tasks such as" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1618" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.088, + 0.895, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.351, + 0.895, + 0.409 + ], + "angle": 0, + "content": "Figure 2. Illustration of our T2P framework. We decompose global motion into global trajectory and local pose. Multi-modal global trajectory proposals are predicted from past global trajectory and local pose embeddings. Then, future local poses are conditioned and forecasted on each trajectory proposal to compose the final human pose prediction. Predicted local poses are added to their mode-specific global trajectories in a joint-wise manner, obtaining the global human poses as the final output." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.473, + 0.619 + ], + "angle": 0, + "content": "estimating 3D poses from monocular videos [7, 46, 48, 50, 55, 56] using self-supervised learning and generative methods [15, 17]. Most recent methods estimate multi-person poses in a crowded environment with considerable occlusions [44, 87]. We account for the aforementioned need of a complex dataset by extracting 3D pose from images using these methods. Specifically, we use a monocular 3D pose estimation method \\(BEV\\) [59] to construct a 3D human motion forecasting dataset with long-term multi-agent characteristics from real-world image sequences. \\(BEV\\) robustly estimates human pose in a scale-ambiguous and crowded environment, reliably extracting 3D poses from the omnidirectional image sequences of JRDB dataset [66]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.638, + 0.17, + 0.654 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.665, + 0.258, + 0.68 + ], + "angle": 0, + "content": "3.1. Problem definition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.691, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Multi-agent human pose forecasting aims to learn a mapping function between the observed 3D pose of \\(N_{A}\\) agents composed of \\(J\\) joints, \\(\\mathbf{X}:\\left\\{\\mathbf{x}_t^{n,j}\\right\\}_{-T_p:0}^{N_A,J}\\), and future pose \\(\\mathbf{Y}:F\\times \\left\\{\\mathbf{x}_t^{n,j}\\right\\}_{0:T_f}^{N_A,J}\\) in global coordinates where \\(F\\) denotes the number of modes. Here, \\(T_{p}\\) and \\(T_{f}\\) are history length and prediction horizon while \\(\\mathbf{x}_t^{n,j} = (x_t^{n,j},y_t^{n,j},z_t^{n,j})\\) is 3D global coordinate of joint \\(j\\) of agent \\(n\\) at time \\(t\\). While the global position of joint is represented in \\(\\mathbf{x}\\), we additionally define local position \\(\\mathbf{p}\\). The local position is defined in local coordinate of each agent, calculated by subtracting the global position of the hip joint of each agent. Therefore, local position of joint is defined as \\(\\mathbf{p}^{n,j} = \\mathbf{x}^{n,j} - \\mathbf{x}^{n,\\mathrm{hip}}\\). We" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.895, + 0.52 + ], + "angle": 0, + "content": "define the trajectory of global hip joint position as global trajectory, \\(Tr:\\{\\mathbf{x}^{n,\\mathrm{hip}}\\}^{N_A}\\). We also define local pose as local position of all joints, \\(Po:\\{\\mathbf{p}^{n,\\mathrm{j}}\\}^{N_A,J}\\). We denote past and future timesteps of global trajectory and local motion as \\(Tr_{\\mathcal{P}},Tr_{\\mathcal{F}}\\in Tr\\) and \\(Po_{\\mathcal{P}},Po_{\\mathcal{F}}\\in Po\\), where \\(\\mathcal{P}\\) and \\(\\mathcal{F}\\) respectively denotes past and future." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.531, + 0.686, + 0.545 + ], + "angle": 0, + "content": "3.2. Overall framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.751 + ], + "angle": 0, + "content": "We disentangle the overall human motion into global trajectories and local poses, as depicted in top left of Fig. 2. Following a coarse-to-fine strategy, multiple global trajectories are first forecasted to model the coarse modes of global intentions. Based on these forecasts, local pose predictions are conditioned on each mode to jointly constitute a thorough motion. In doing so, our model is widely divided into two portions: Trajectory predictor consists of trajectory encoder and decoder and pose predictor consists of pose encoder and decoder. Both predictors engage in the reciprocal exchange of both trajectory and pose information, facilitating the inference of cues between global and local motion. The detailed methods of each stage are described below:" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.762, + 0.662, + 0.777 + ], + "angle": 0, + "content": "3.3. Model structure" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.645, + 0.799 + ], + "angle": 0, + "content": "3.3.1 Pose encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Unlike the holistic approach of previous works that encode and decode all agents' joint motions in global coordinates, our pose encoder encodes the pose dynamics in local coordinates. In addition, our pose encoder only considers intra-agent joint interaction. As a result, the encoded pose embedding represents agent-specific local mo" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1619" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "tion, containing insights on global intent. We follow our baseline [47] and construct the encoder with Multi-Person Body-Part (MPBP) module and transformer networks. As depicted in Fig. 2, body part sequences are constructed in frequency domain, followed by intra-agent attention-based encoding of the body parts to acquire pose embedding \\( Z_{Po} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.207, + 0.262, + 0.222 + ], + "angle": 0, + "content": "3.3.2 Trajectory module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.473, + 0.401 + ], + "angle": 0, + "content": "Trajectory module aims to extract embeddings from the agents' past global trajectory. Using an encoder structure from [88], multi-agent interaction-based trajectory embedding \\( Z_{Tr} \\) is extracted which contains rudimentary insight on global intent. Interaction between agent trajectories is represented based on the reference agent \\( i \\)'s global trajectory segment vector \\( \\mathbf{v}_t^i = \\mathbf{x}_t^{i,\\mathrm{hip}} - \\mathbf{x}_{t-1}^{i,\\mathrm{hip}} \\). For rotational invariance, neighbor actor \\( j \\)'s vector is normalized by the reference vector's orientation at latest timestep \\( t=0 \\). Separate MLP layers then compute the reference agent and neighboring agent embeddings \\( z_{Tr_i}^t, z_{Tr_j}^t \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.415, + 0.468, + 0.456 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z _ {T r _ {i}} ^ {t} = \\phi_ {r e f} (R _ {i} ^ {T} \\mathbf {v} _ {t} ^ {i}) \\\\ z _ {T r _ {j}} ^ {t} = \\phi_ {n b r} ([ R _ {i} ^ {T} (\\mathbf {v} _ {t} ^ {j}), R _ {i} ^ {T} (\\mathbf {v} _ {t} ^ {i}) ]) \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.47, + 0.47, + 0.533 + ], + "angle": 0, + "content": "where \\(\\phi_{ref}\\) and \\(\\phi_{nbr}\\) are different MLP blocks, \\(R_i \\in \\mathbb{R}^{3 \\times 3}\\) is the rotation matrix of agent \\(j\\) against agent \\(i\\), \\([\\cdot, \\cdot]\\) is concatenation. The resulting agent-specific reference and neighbor embeddings constitute trajectory embedding \\(z_{Tr}^t\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.555, + 0.254, + 0.57 + ], + "angle": 0, + "content": "3.3.3 Traj-posed module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Human maneuver contains various dynamic activities characterized by the agent's multi-modal intents. Auxiliary human motion such as arm gesture, rotational orientation of upper body and head implies the agent's intent in global motion. In that sense, harvesting meaningful insights from past local joint motion helps proficient modeling of coarse multimodality as future trajectory proposals. Therefore, we propose Traj-Pose Module that fuses agent-wise embeddings of both trajectory and pose to fully utilize these information in modeling global intentions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.779 + ], + "angle": 0, + "content": "First, MLP is used to match the temporal domain of pose embedding \\( Z_{Po} \\) to that of \\( Z_{Tr} \\), after which both are concatenated as agent-wise traj-posed embedding \\( Z \\)." + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.792, + 0.469, + 0.809 + ], + "angle": 0, + "content": "\\[\nZ = \\left[ Z _ {T r}, \\phi_ {M L P} \\left(Z _ {P o}\\right) \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.822, + 0.47, + 0.904 + ], + "angle": 0, + "content": "The resulting \\( Z \\) is comprised of agent and timestep-respective trajectory and pose embeddings: \\( z_{i}^{t},z_{j}^{t}\\in z^{t}\\in Z \\) Then, \\( \\widetilde{Z} \\) is acquired from the graph attention with an agent-wise update where each agent embedding \\( z_{i}^{t} \\) and its neighbor embedding \\( z_{j}^{t} \\) are used as query and key/value. Similar" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "to trajectory interaction encoder of HiVT [88], the graph attention operation is operated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.132, + 0.891, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\alpha_ {i} ^ {t} = \\operatorname {s o f t m a x} \\left(\\frac {q _ {i} ^ {t \\top}}{\\sqrt {d _ {k}}} \\cdot \\left[ \\left\\{k _ {j} ^ {t} \\right\\} _ {j \\in N _ {i}} \\right]\\right), \\\\ m _ {i} ^ {t} = \\sum_ {j \\in N _ {i}} \\alpha_ {i} ^ {t} v _ {j} ^ {t}, \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.205, + 0.815, + 0.222 + ], + "angle": 0, + "content": "\\[\ng _ {i} ^ {t} = \\operatorname {s i g m o i d} \\left(W ^ {\\text {g a t e}} \\left[ z _ {i} ^ {t}, m _ {i} ^ {t} \\right]\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.224, + 0.816, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\widetilde {z} _ {i} ^ {t} = g _ {i} ^ {t} \\odot W ^ {\\text {s e l f}} z _ {i} ^ {t} + \\left(1 - g _ {i} ^ {t}\\right) \\odot m _ {i} ^ {t}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.254, + 0.892, + 0.286 + ], + "angle": 0, + "content": "where \\(N_{i}\\) is a set of agent \\(i\\)'s neighbors, \\(W^{gate}\\) and \\(W^{self}\\) are learnable matrices, and \\(\\odot\\) is element-wise product." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.303, + 0.688, + 0.318 + ], + "angle": 0, + "content": "3.3.4 Trajectory decoder" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.327, + 0.892, + 0.569 + ], + "angle": 0, + "content": "Trajectory is forecasted from the output of trajectory encoder which encodes both past global trajectory and local pose information. Since its graph operation is operated by each timestep, a temporal encoder is used as a temporal encoder to integrate \\(\\widehat{Z}\\) in the temporal dimension. A multihead self-attention temporal encoder is used as the temporal encoder. Aggregator then takes into account variations in local coordinate frames to accurately represent geometric relationships within the global coordinate system via a graph operation. MLP is subsequently applied to span embedding \\(F\\) times for multi-modal prediction, which is residually added to the \\(\\times F\\) repeated embedding before the aggregator. Finally, another MLP is used to extract multi-modal future global trajectory proposals of hip joint \\(Tr_{\\mathcal{F}} \\in \\mathbb{R}^{F \\times T_f \\times 3}\\). The multi-modal embedding is also passed onto the Pose Decoder to forecast local poses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.587, + 0.645, + 0.6 + ], + "angle": 0, + "content": "3.3.5 Pose decoder" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.61, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Future human pose depends on past human poses and global intention. The pose decoder is designed to consider these factors while generating local poses via mode-specific trajectory conditioning. A transformer (TRM) decoder is used to decode local motions, where pose embedding \\( Z_{Po} \\) is used as key/value and concatenation of trajectory and pose query." + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.715, + 0.892, + 0.732 + ], + "angle": 0, + "content": "\\[\nQ = \\phi_ {M L P} ([ Q _ {P o}, Q _ {T r} ]), K = Z _ {P o}, V = Z _ {P o} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.738, + 0.892, + 0.872 + ], + "angle": 0, + "content": "Using both pose and trajectory queries, past pose embedding \\( Z_{Po} \\) is conditioned on both MPBP sequence at \\( t = 0 \\) and the multi-modal trajectory proposals which contain global intent. Subsequently, inverse discrete cosine transform (idct) is applied to convert the future pose proposals from frequency domain to local coordinate domain, \\( P_{o_{\\mathcal{F}}} \\). The final multi-modal future pose in global coordinates is acquired as Eq. 5 where \\( \\oplus \\) is a joint-wise addition operation." + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.884, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} = T r _ {\\mathcal {F}} \\oplus P o _ {\\mathcal {F}}, \\quad \\mathbf {Y} \\in \\mathbb {R} ^ {F \\times N _ {A} \\times T _ {f} \\times 3} \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1620" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.118 + ], + "angle": 0, + "content": "Table 1. Comparison of statistics between existing human pose forecasting datasets and newly proposed JRDB-GMP dataset." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.122, + 0.47, + 0.274 + ], + "angle": 0, + "content": "
Dataset
CMU-Mocap (UMPM)MuPoTs -3D3DPWJRDB-GMP 1s/2s2s/5s
Duration (s)400026717001863
Location #-20-27
Sample #1300019243211534593
avg. agent3326.86.8
med. agent #33255
max agent #3322422
avg. vel. (m/s)0.30.260.570.460.38
avg. disp.(m)0.630.551.130.640.79
max. disp.(m)4.622.4510.718.4411.0
" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.283, + 0.485, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.464, + 0.469, + 0.491 + ], + "angle": 0, + "content": "Figure 3. Example scenes from the JRDB-GMP dataset, illustrating its long-term, multi-agent nature." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.505, + 0.254, + 0.521 + ], + "angle": 0, + "content": "3.4. Training objective" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.47, + 0.589 + ], + "angle": 0, + "content": "Both objectives of global trajectory and local pose forecasting are trained jointly. For both global trajectory and local pose prediction, \\( L2 \\) loss is propagated to the mode with minimal \\( L2 \\) distance with the ground truth." + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.599, + 0.401, + 0.641 + ], + "angle": 0, + "content": "\\[\nL _ {T r} = \\sum_ {n = 1} ^ {N _ {A}} \\sum_ {t = 1} ^ {T _ {f}} \\| \\widetilde {y} _ {T r, n} ^ {t} - \\hat {y} _ {T r, n} ^ {t} \\|\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.645, + 0.469, + 0.687 + ], + "angle": 0, + "content": "\\[\nL _ {P o} = \\sum_ {n = 1} ^ {N _ {A}} \\sum_ {t = 1} ^ {T _ {f}} \\sum_ {j = 1} ^ {J - 1} \\| \\left(\\widetilde {y} _ {P o, n} ^ {t, j} - \\hat {y} _ {P o, n} ^ {t, j}\\right) \\| \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.29, + 0.692, + 0.401, + 0.707 + ], + "angle": 0, + "content": "\\[\nL = L _ {T r} + L _ {P o}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.268, + 0.74 + ], + "angle": 0, + "content": "3.5.JRDB-GMP dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Due to the absence of existing long-term \\((3s+)\\) multi-agent \\((6+)\\) dataset, we compose a unique 3D human pose forecasting dataset in a real-world environment from JRDB [66]. The original JRDB dataset is constructed by a moving robot that records human activity around a school campus using 5 omnidirectional cameras and LiDAR. Image sequences along with 2D pose annotation and 3D bounding box annotations are provided in the original dataset. However, since 3D human pose annotations are unavailable, we separately parse accurate 3D human pose from provided inputs and" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.087, + 0.905, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.25, + 0.905, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.408, + 0.892, + 0.436 + ], + "angle": 0, + "content": "Figure 4. Various motions from the JRDB-MultiGlobPose dataset, providing rich motion queues for inter-agent interaction inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.658 + ], + "angle": 0, + "content": "annotations. First, a SOTA monocular 3D pose extraction method [59] is used to extract raw 3D joint positions from image sequences. Then, 2D pose and 3D bounding box annotations are used to refine the raw joint positions and minimize noise. We use 2D pose annotations to initially filter out the 3D poses with noise. With camera parameters and refined 3D pose, we project it on 2D image plane, then L2 distance between projected 2D pose and GT 2D pose annotation is calculated. If the mean L2 distance per each agent at a time stamp is over a threshold, that instance is filtered out. 2D pose annotations are also projected in 3D space to refine the remaining 3D poses, ensuring the accuracy of the 3D pose information of our JRDB-GMP dataset. Further details are elaborated in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Figure 3 visualizes some scenes of the constructed dataset. Accurate extraction of 3D poses has been made even with considerable occlusion via the use of 2D poses. The dataset includes agents with both long and short traverse distances and rich inter-agent interactions in both trajectory and local pose aspects. Figure 4 illustrates diverse local poses included in the dataset, which serve as motion cues of inter-agent interaction. Both figures confirm our method's accuracy in extracting 3D multi-human pose, even in crowded environments. Table 1 scrutinizes the statistics compared to previously used datasets. Compared to earlier datasets, the average number of agents is more than twice as high. In addition, comparing JRDB-GMP 1s/2s to CMU-Mocap and MuPoTs datasets, JRDB contains more diverse and longer motion as shown by a similar magnitude of average displacement but longer maximum displacement." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1621" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.117 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of our method to previous methods on CMU-mocap (UMPM), 3DPW, and JRDB-GlobMultiPose datasets with number of prediction modes \\((F)\\) as 6. Lower is better for all metrics. The best results are marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.154, + 0.118, + 0.814, + 0.343 + ], + "angle": 0, + "content": "
DatasetCMU-mocap (UMPM)3DPWJRDB-GlobMultiPose
In/out length (s)1/20.8/1.61/22/5
Evaluation time (s)120.81.6122.55
JPEMRT [69]164.7280.1159.1251.2259.3349.3438.4474.0
JRT [74]168.5316.9181.9287.3237.9373.1351.9538.8
TBIFormer [47]170.0290.9153.9265.8257.1339.3443.2481.3
Ours152.4262.7142.6236.2224.0301.4341.6390.4
APEMRT [69]127.0164.4117.9153.272.387.388.5101.9
JRT [74]121.2181.6133.4178.0112.6154.396.7120.2
TBIFormer [47]125.1160.8115.4152.770.683.388.2102.9
Ours114.4151.7114.6150.070.883.382.294.7
FDEMRT [69]99.6204.7102.7185.3235.2325.2418.2454.8
JRT [74]117.7250.8133.7235.4211.4337.4318.5497.2
TBIFormer [47]112.1228.5106.7215.9232.4314.6423.9458.8
Ours88.7188.974.1158.2194.7271.5313.9361.0
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.352, + 0.201, + 0.37 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.378, + 0.173, + 0.393 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.554 + ], + "angle": 0, + "content": "We test our model on three datasets: CMU-Mocap (UMPM) [13, 64], 3DPW [67], and our JRDB-GMP. Although our model is designed to forecast human poses in a long-term multi-agent environment, we also report experimental results on previous benchmark datasets with simpler scenes. Mocap-UMPM is a mixed dataset of Mocap and UMPM containing synthesized human interaction between three agents [47]. 3DPW is a dataset with 2 agents traversing a real-world environment. We report the test results on each after separate training on respective datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.567, + 0.174, + 0.582 + ], + "angle": 0, + "content": "4.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.469, + 0.621 + ], + "angle": 0, + "content": "We use the following widely-used metrics. For a detailed definition, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.469, + 0.681 + ], + "angle": 0, + "content": "APE: Aligned mean per joint Position Error is used as a metric to evaluate the forecasted local motion. \\( L2 \\) distance of each joint in the hip joint coordinate is averaged over all joints for a given timestep." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.469, + 0.712 + ], + "angle": 0, + "content": "FDE: Final Distance Error evaluates the forecasted global trajectory by calculating the \\( L2 \\) distance of a given timestep." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.468, + 0.743 + ], + "angle": 0, + "content": "JPE: Joint Precision Error evaluates both global and local predictions by mean \\( L2 \\) distance of all joints for a timestep." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.292, + 0.772 + ], + "angle": 0, + "content": "4.3. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We train our model on a single A6000 GPU. 2 layers of pose encoder transformer are stacked, followed by 2 layers of transformer in pose decoder. Embedding dimensions of 96 and 128 are used for trajectory and pose embeddings, respectively. The transformed key, value dimension of 64 is used for all transformer architectures. A learning rate of 0.003 is used with an AdamW optimizer with weight decay. Further details can be found in the supplementary materials." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.351, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Table 3. Short-term prediction results on CMU-Mocap (UMPM) dataset, where 1s of poses are forecasted given 2s of poses." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.38, + 0.89, + 0.463 + ], + "angle": 0, + "content": "
MetricJPEAPEFDE
Time (s)0.20.61.00.20.61.00.20.61.0
MRT64.515221749.811014039.497.9153
JRT31.510417328.785.912517.763.9120
TBIFormer37.410415832.885.811923.363.7104
Ours37.810215833.884.411614.949.192.6
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.608, + 0.485 + ], + "angle": 0, + "content": "4.4. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.705 + ], + "angle": 0, + "content": "We compare our method against the latest SOTA methods for multi-agent pose forecasting [47, 69, 74]. To compare the multi-modal predictions of these three methods, we extend their prediction modes by spanning embedding \\( K \\) times in the same manner as ours. All baselines are trained and evaluated on CMU-mocap (UMPM) and 3DPW datasets. CMU-Mocap (UMPM) dataset predicts 2 seconds from 1 second of poses, and 3DPW predicts 1.6 seconds from 0.8 seconds of poses, both from 6 modes. For 3DPW dataset, we slightly lengthen the forecast horizon to evaluate long term predictions. For JRDB-GMP dataset, both short (1s/2s) and long term (2s/5s) predictions are evaluated for all models. Lastly, We use HiVT [88] as the baseline For global trajectory prediction of our model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.587, + 0.733 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.688, + 0.758 + ], + "angle": 0, + "content": "5.1.Quantitative results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Table 2 compares the quantitative performances on three datasets. Our method exhibits considerable performance gain against all previous SOTA methods, not only on the proposed long-term multi-agent dataset but also on the existing two datasets. Such generalized competence demonstrates the applicability of our trajectory-conditioned pose forecasting method to various real-world scenarios. In detail, our approach achieves over \\(10\\%\\) gain of FDE on all datasets. This improvement on forecasting global locomo" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1622" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.277, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.09, + 0.308, + 0.209 + ], + "angle": 0, + "content": "t \n2.5s \n5s" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.09, + 0.428, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.091, + 0.545, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.091, + 0.643, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.091, + 0.767, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.091, + 0.889, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.234, + 0.893, + 0.262 + ], + "angle": 0, + "content": "Figure 5. Visualization of a long-term forecasting scene from JRDB-GMP (2/5) dataset. Past poses for input are shown on the leftmost column, GT future poses on the next, and forecasts by ours, MRT, TBIFormer, and JRT, respectively." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.267, + 0.468, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.487, + 0.471, + 0.571 + ], + "angle": 0, + "content": "Figure 6. Visualization of a CMU-Mocap (UMPM) scene. Past poses are shown on the upper row, GT future poses on the next, and forecasts by TBIFormer and ours on the latter two rows. To visualize motion, we stack several frames around the target time stamp. Black/red/blue arrows refer to the direction of the global trajectory, and yellow arrows refer to the direction of foot motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.734 + ], + "angle": 0, + "content": "tion could be accredited to the decoupled forecasting of global trajectory and local pose. Previous methods holistically predict both global and local movements, limiting both performances due to superfluous interactions to consider between all joints. Conversely, our approach can extract accurate global intent by decoupling past motion into global and local representations. Moreover, our effective interaction modeling of global and local pose also helps to predict a more accurate global trajectory under multi-agent environment as shown in the latter ablation study." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.884 + ], + "angle": 0, + "content": "For APE metric, our method also surpasses previous SOTA models on all datasets, highlighting the accurate extraction of local pose intent. Such improvement shows that our approach generates plausible local motion due to its proficient sampling from coarse global intents. Our method simplifies the task by learning multi-modality in a coarse-to-fine approach. Its subsequent local motion forecasting is inferred from coarsely modeled multi-modality, a greatly simplified task compared to extracting intent from entangled multi-modality as well as multi-agent interaction." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.901 + ], + "angle": 0, + "content": "These improvements on both global and local scales" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.406 + ], + "angle": 0, + "content": "jointly contribute toward lowering the JPE metric, demonstrating proficiency of our method in forecasting overall human motion. Based on such competence, our method which aimed towards improving on forecasting long-term multiagent environments also exhibits similar or better performances on short timeframes as shown in Tab. 3. In addition, our approach also excels even on sole local motion with minimal global displacement, as elaborated in section 7.2 of supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.418, + 0.678, + 0.433 + ], + "angle": 0, + "content": "5.2. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.893, + 0.623 + ], + "angle": 0, + "content": "Our method forecasts a more plausible global pose in longer timescales (\\(\\sim 5\\mathrm{s}\\)) as shown in the interacting scene of five agents in Fig. 5. Looking at the input and GT sequences, the leftmost person avoids the traversing couple from right to left. The two people in front are stationary while talking to each other. MRT and TBIFormer forecasts implausible overlapped poses at the final prediction horizon (\\(t = 5\\mathrm{s}\\)). JRT fails to learn the global locomotion of agents due to the high complexity of its attention mechanism and is stuck in the local minimum of predicting the inactivity of all agents. On the other hand, our model forecasts plausible poses where the closely interacting two agents walk side-by-side." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.895, + 0.806 + ], + "angle": 0, + "content": "Figure 6 illustrates exemplary sequences where more natural local motion has been forecasted by our method. Comparing forecasts on a scene of walking agents, our method generates a much more plausible sequence where the stepping foot remains stationary. On the other hand, the previous SOTA method, TBIFormer, struggles to learn the natural walking mechanism of human legs and a parallel translation of both feet is exhibited. Such discrepancy shows that trajectory-conditioning for inferring local motion from global intent generates more proficient details in human motion than SOTA methods. More visualizations could be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.662, + 0.832 + ], + "angle": 0, + "content": "5.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Different number of modes. The main quantitative results report prediction results with \\( F \\) as 6 to compare the ability to address the multi-modal nature of human motion during pose forecasting. Table. 4 additionally compares forecast" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1623" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.118 + ], + "angle": 0, + "content": "Table 4. Comparison of performance with different number of modes in CMU-Mocap (UMPM) dataset." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.123, + 0.427, + 0.217 + ], + "angle": 0, + "content": "
F16
Metric @ 2sAPEJPEAPEJPE
MRT163.9366.4164.4280.1
JRT176.7367.4181.6316.9
TBIFormer160.1374.3160.8290.9
Ours154.4366.4151.7262.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.229, + 0.47, + 0.271 + ], + "angle": 0, + "content": "Table 5. Ablation studies on core components of model structures. Experiments are done with JRDB-GMP dataset to evaluate multiagent long-term performance." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.276, + 0.47, + 0.398 + ], + "angle": 0, + "content": "
Exp. #Trajectory encoderPose decoderMetrics
Local pose embeddingAgent interactionTrajectory -conditioningJPE @5sAPE @5sFDE @5s
-471.4101.7457.9
1✓400.595.1370.9
2✓403.394.7374.2
3✓401.293.0372.8
4✓✓395.693.8366.8
5✓✓392.795.2363.4
6✓✓✓391.291.4363.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.625 + ], + "angle": 0, + "content": "results with \\( F \\) as 1. Our method again achieves noticeable improvement in APE over the baseline on single-modal forecasts. With \\( F = 1 \\), although our method barely enjoys improvement in forecasting global motion due to the absence of multi-modality, its superiority in APE shows the validity of our coarse-to-fine forecasting strategy that also effectively captures agent interaction. Our method improves with multi-modal predictions, demonstrating the proficiency of a coarse-to-fine approach in interpreting the stochastic nature of human motion and its intent. Note that our method improves in APE along with an increase in \\( F \\) unlike previous methods, indicating a unique aptitude in addressing the multi-modal nature of not only global locomotion but also local pose intent via trajectory-conditioning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Importance of each architecture component. Table. 5 reports the influence of core components of our model. For the trajectory encoder, we evaluate the importance of using local pose embedding and modeling agent interaction. Comparing experiments 1, 5 and 3, 4, both show improvements in JPE and FDE metrics with the use of local pose embedding. Our method has taken advantage of detailed local pose cues to infer an agent's global intention. For interaction modeling, its use is beneficial for both global and local forecasts as compared by experiments 4 and 6. These joint improvements demonstrate the importance of considering local and global motion interactions for their respective forecasts. As for the pose decoder, comparisons of experiments 2,4 and 5,6 both show improvements in APE metric. Such consistent improvement verifies the effectiveness of the trajectory-conditioned local motion forecast approach in generating plausible local motion from global intention." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Importance of interaction modeling. Accurate model" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.116 + ], + "angle": 0, + "content": "Table 6. Ablation studies on agent interaction cutoff distance on JRDB-GMP." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.123, + 0.837, + 0.217 + ], + "angle": 0, + "content": "
JPE @ 5s
TBIFormerOurs
w/o interaction483.8406.0
w/ interaction < 2m-403.5
w/ interaction < 4m-400.5
w/ interaction all481.3390.4
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.226, + 0.892, + 0.528 + ], + "angle": 0, + "content": "ing of inter-agent interaction becomes more pivotal to forecast in more complex environments. Indeed, its complexity grows in a long-term multi-agent scene. When holistically considering joint-wise interaction for all timesteps, the computation complexity is acquired as \\( O(T^2 \\cdot N^2 \\cdot J^2) \\), where \\( T \\) is the number of timesteps, \\( N \\) the number of agents, and \\( J \\) the number of joints. On the contrary, with interaction modeling in global trajectory scale, our method reduces the computation cost by \\( TJ^2 \\) into \\( O(T \\cdot N^2) \\). This enables efficient and proficient modeling of intra (pose) and inter (trajectory)-agent interactions as shown by Tab. 6. While body part-wise interaction modeling only improved by \\( 0.52\\% \\) for TBIFormer, ours improves up to \\( 3.84\\% \\) with interaction modeling. This demonstrates the proficiency of our efficient interaction modeling-based method in inferring global and local intents from complex interactions. In addition, the gradual improvement of JPE according to a wider interaction range confirms the importance of interaction modeling of more agents, which cannot be learned from the arbitrarily mixed previous datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.543, + 0.619, + 0.559 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.78 + ], + "angle": 0, + "content": "In this work, we propose a novel interaction-aware trajectory-conditioned approach to handle long-term multi-agent motion forecasting, along with a new dataset suited for such scope. Our proposed model utilizes a coarse-to-fine approach and decouples overall motion prediction into global and local components. Multi-modality of human motion is proficiently modeled via inferring fine local intents from coarse global intents, along with efficient agent-wise interaction modeling. As for the dataset, our JRDB-GMP dataset contains unprecedented long-term \\((5s+)\\) multi-agent \\((6+)\\) interactions in a real-world setting. Our method achieves state-of-the-art performance on all previous datasets and JRDB-GMP dataset, offering generalized practical implications in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1624" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Vida Adeli, Ehsan Adeli, Ian Reid, Juan Carlos Niebles, and Hamid Rezatofighi. Socially and contextually aware human motion and pose forecasting. IEEE Robotics and Automation Letters, 5(4):6033-6040, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[2] Vida Adeli, Mahsa Ehsanpour, Ian Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13390-13400, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[3] Gorkay Aydemir, Adil Kaan Akan, and Fatma Güney. Adapt: Efficient multi-agent trajectory prediction with adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8295-8305, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.471, + 0.386 + ], + "angle": 0, + "content": "[4] German Barquero, Sergio Escalera, and Cristina Palmero. Belfusion: Latent diffusion for behavior-driven human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2317-2327, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.472, + 0.485 + ], + "angle": 0, + "content": "[5] Arij Bouazizi, Adrian Holzbock, Ulrich Kressel, Klaus Dietmayer, and Vasileios Belagiannis. Motionmixer: Mlp-based 3d human body pose forecasting. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 791-798. International Joint Conferences on Artificial Intelligence Organization, 2022. Main Track. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.471, + 0.569 + ], + "angle": 0, + "content": "[6] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 387–404. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.573, + 0.47, + 0.655 + ], + "angle": 0, + "content": "[7] Wenhao Chai, Zhongyu Jiang, Jenq-Neng Hwang, and Gaoang Wang. Global adaptation meets local generalization: Unsupervised domain adaptation for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14655-14665, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.659, + 0.471, + 0.728 + ], + "angle": 0, + "content": "[8] Ling-Hao Chen, JiaWei Zhang, Yewen Li, Yiren Pang, Xiaobo Xia, and Tongliang Liu. Humanmac: Masked motion completion for human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9544-9555, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.73, + 0.471, + 0.786 + ], + "angle": 0, + "content": "[9] Hsu-kuang Chiu, Ehsan Adeli, Borui Wang, De-An Huang, and Juan Carlos Niebles. Action-agnostic human pose forecasting. In 2019 IEEE winter conference on applications of computer vision (WACV), pages 1423–1432. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[10] Sehwan Choi, Jungho Kim, Junyong Yun, and Jun Won Choi. R-pred: Two-stage motion prediction via tube-query attention-based trajectory refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8525-8535, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[11] Rohan Choudhury, Kris M. Kitani, and László A. Jeni. Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In Proceedings of the IEEE/CVF International" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision (ICCV), pages 14750-14760, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[12] Andrea Porfiri Dal Cin, Giacomo Boracchi, and Luca Magri. Multi-body depth and camera pose estimation from multiple views. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17804-17814, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[13] CMU-Graphics-Lab. Cmu graphics lab motion capture database. http://mocap.cs.cmu.edu/, 2003. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.22, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[14] Runyang Feng, Yixing Gao, Xueqing Ma, Tze Ho Elden Tse, and Hyung Jin Chang. Mutual information-based temporal difference learning for human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17131-17141, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[15] Runyang Feng, Yixing Gao, Tze Ho Elden Tse, Xueqing Ma, and Hyung Jin Chang. Diffpose: Spatiotemporal diffusion model for video-based human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14861-14872, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.373, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[16] Xuehao Gao, Shaoyi Du, Yang Wu, and Yang Yang. Decompose more and aggregate better: Two closer looks at frequency representation learning for human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6451-6460, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.512 + ], + "angle": 0, + "content": "[17] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13041-13051, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[18] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15303-15312, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[19] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15303-15312, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[20] Wen Guo, Xiaoyu Bie, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Multi-person extreme motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13053-13064, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[21] Je-Seok Ham, Dae Hoe Kim, NamKyo Jung, and Jinyoung Moon. Cipf: Crossing intention prediction network based on feature fusion modules for improving pedestrian safety. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3665-3674, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[22] Karl Holmquist and Bastian Wandt. Diffpose: Multi-hypothesis human pose estimation using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15977-15987, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[23] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1625" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[24] Boyuan Jiang, Lei Hu, and Shihong Xia. Probabilistic triangulation for uncalibrated multi-view 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14850-14860, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[25] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9644-9653, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[26] Zhehan Kan, Shuoshuo Chen, Ce Zhang, Yushun Tang, and Zhihai He. Self-correctable and adaptable inference for generalizable human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5537-5546, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.371 + ], + "angle": 0, + "content": "[27] Jeongho Kim, Wooksu Shin, Hancheol Park, and Jongwon Baek. Addressing the occlusion problem in multi-camera people tracking with human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5462-5468, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.471, + 0.453 + ], + "angle": 0, + "content": "[28] Mihee Lee, Samuel S Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. Musevae: multi-scale vae for environment-aware long term trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2221-2230, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.471, + 0.524 + ], + "angle": 0, + "content": "[29] Lihuan Li, Maurice Pagnucco, and Yang Song. Graph-based spatial transformer with memory replay for multifuture pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2241, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.608 + ], + "angle": 0, + "content": "[30] Huan Liu, Qiang Chen, Zichang Tan, Jiang-Jiang Liu, Jian Wang, Xiangbo Su, Xiaolong Li, Kun Yao, Junyu Han, Errui Ding, Yao Zhao, and Jingdong Wang. Group pose: A simple baseline for end-to-end multi-person pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15029-15038, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.471, + 0.691 + ], + "angle": 0, + "content": "[31] Tiezheng Ma, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Progressively generating better initial guesses towards next stages for high-quality human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6437-6446, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.471, + 0.748 + ], + "angle": 0, + "content": "[32] Takahiro Maeda and Norimichi Ukita. Motionaug: Augmentation with physical correction for human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6427-6436, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.816 + ], + "angle": 0, + "content": "[33] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15233-15242, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[34] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9489-9497, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[35] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Generating smooth pose sequences for diverse human motion pre" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "diction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13309-13318, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[36] Wei Mao, Richard I Hartley, Mathieu Salzmann, et al. Contact-aware human motion forecasting. Advances in Neural Information Processing Systems, 35:7356-7367, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[37] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Weakly-supervised action transition learning for stochastic human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8151-8160, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[38] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5517-5526, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[39] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 2018 International Conference on 3D Vision (3DV), pages 120-130, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[40] Jiquan Ngiam, Vijay Vasudevan, Benjamin Caine, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, David J Weiss, Benjamin Sapp, Zhifeng Chen, and Jonathon Shlens. Scene transformer: A unified architecture for predicting future trajectories of multiple agents. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[41] Daehee Park, Jaewoo Jeong, and Kuk-Jin Yoon. Improving transferability for cross-domain trajectory prediction via neural stochastic differential equation. arXiv preprint arXiv:2312.15906, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[42] Daehee Park, Hobin Ryu, Yunseo Yang, Jegyeong Cho, Jiwon Kim, and Kuk-Jin Yoon. Leveraging future relationship reasoning for vehicle trajectory prediction. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.603, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[43] Daehee Park, Jaeseok Jeong, Sung-Hoon Yoon, Jaewoo Jeong, and Kuk-Jin Yoon. T4p: Test-time training of trajectory prediction via masked autoencoder and actor-specific token memory, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[44] Sungchan Park, Eunyi You, Inhoe Lee, and Joonseok Lee. Towards robust and smooth 3d multi-person pose estimation from monocular videos in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14772-14782, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[45] Behnam Parsaeifard, Saeed Saadatnejad, Yuejiang Liu, Taylor Mordan, and Alexandre Alahi. Learning decoupled representations for human pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2294-2303, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[46] Qucheng Peng, Ce Zheng, and Chen Chen. Source-free domain adaptive human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4826-4836, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[47] Xiaogang Peng, Siyuan Mao, and Zizhao Wu. Trajectory-aware body interaction transformer for multi-person pose forecasting. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1626" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 17121- 17130, 2023. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.471, + 0.219 + ], + "angle": 0, + "content": "[48] Zhongwei Qiu, Qiansheng Yang, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Chang Xu, Dongmei Fu, and Jingdong Wang. Psvt: End-to-end multi-person 3d pose and shape estimation with progressive video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21254-21263, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[49] Muhammad Rameez Ur Rahman, Luca Scofano, Edoardo De Matteis, Alessandro Flaborea, Alessio Sampieri, and Fabio Galasso. Best practices for 2-body pose forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3613-3623, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.471, + 0.361 + ], + "angle": 0, + "content": "[50] Dripta S. Raychaudhuri, Calvin-Khang Ta, Arindam Dutta, Rohit Lal, and Amit K. Roy-Chowdhury. Prior-guided source-free domain adaptation for human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14996-15006, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.471, + 0.445 + ], + "angle": 0, + "content": "[51] Luke Rowe, Martin Ethier, Eli-Henry Dykhne, and Krzysztof Czarnecki. Fjmp: Factorized joint multi-agent motion prediction over learned directed acyclic interaction graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13745-13755, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.471, + 0.531 + ], + "angle": 0, + "content": "[52] Saeed Saadatnejad, Ali Rasekh, Mohammadreza Mofayezi, Yasamin Medghalchi, Sara Rajabzadeh, Taylor Mordan, and Alexandre Alahi. A generic diffusion-based approach for 3d human pose prediction in the wild. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8246-8253. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.471, + 0.588 + ], + "angle": 0, + "content": "[53] Tim Salzmann, Marco Pavone, and Markus Ryll. Motron: Multimodal probabilistic human motion forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6457-6466, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.471, + 0.645 + ], + "angle": 0, + "content": "[54] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. IEEE Robotics and Automation Letters, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.471, + 0.729 + ], + "angle": 0, + "content": "[55] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14761-14771, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[56] Xiaolong Shen, Zongxin Yang, Xiaohan Wang, Jianxin Ma, Chang Zhou, and Yi Yang. Global-to-local modeling for video-based 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8887-8896, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[57] Mingyi Shi, Sebastian Starke, Yuting Ye, Taku Komura, and Jungdam Won. Phasemp: Robust 3d pose estimation via phase-conditioned human motion prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14725-14737, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[58] Theodoros Sofianos, Alessio Sampieri, Luca Franco, and Fabio Galasso. Space-time-separable graph convolutional" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "network for pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11209-11218, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[59] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13243-13252, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[60] Yuran Sun, Alan William Dougherty, Zhuoying Zhang, Yi King Choi, and Chuan Wu. Mixsynthformer: A transformer encoder-like structure with mixed synthetic self-attention for efficient human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14884–14893, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[61] Zhenhua Tang, Zhaofan Qiu, Yanbin Hao, Richang Hong, and Ting Yao. 3d human pose estimation with spatiotemporal criss-cross attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4790-4799, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[62] Julian Tanke, Chintan Zaveri, and Juergen Gall. Intention-based long-term human motion anticipation. In 2021 International Conference on 3D Vision (3DV), pages 596-605. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[63] Julian Tanke, Linguang Zhang, Amy Zhao, Chengcheng Tang, Yujun Cai, Lezi Wang, Po-Chen Wu, Juergen Gall, and Cem Keskin. Social diffusion: Long-term multiple human motion anticipation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9601-9611, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[64] N.P. van der Aa, X. Luo, G.J. Giezeman, R.T. Tan, and R.C. Veltkamp. Umpm benchmark: A multi-person dataset with synchronized video and motion capture data for evaluation of articulated human motion and interaction. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 1264-1269, 2011. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[65] Edward Vendrow, Satyajit Kumar, Ehsan Adeli, and Hamid Rezatofighi. Somoformer: Multi-person pose forecasting with transformers. arXiv preprint arXiv:2208.14023, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[66] Edward Vendrow, Duy Tho Le, Jianfei Cai, and Hamid Rezatofighi. Jrdb-pose: A large-scale dataset for multiperson pose estimation and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4811–4820, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[67] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[68] Chenxi Wang, Yunfeng Wang, Zixuan Huang, and Zhiwen Chen. Simple baseline for single human motion forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2260-2265, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[69] Jiashun Wang, Huazhe Xu, Medhini Narasimhan, and Xiaolong Wang. Multi-person 3d motion prediction with multirange transformers. Advances in Neural Information Processing Systems, 34:6036-6049, 2021. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1627" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[70] Mingkun Wang, Xinge Zhu, Changqian Yu, Wei Li, Yuexin Ma, Ruochun Jin, Xiaoguang Ren, Dongchun Ren, Mingxu Wang, and Wenjing Yang. Ganet: Goal area network for motion forecasting. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1609-1615. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[71] Yucheng Xing and Xin Wang. Hdg-ode: A hierarchical continuous-time model for human pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14700–14712, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[72] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6488-6497, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.472, + 0.375 + ], + "angle": 0, + "content": "[73] Chenxin Xu, Robby T. Tan, Yuhong Tan, Siheng Chen, Xin-chao Wang, and Yanfeng Wang. Auxiliary tasks benefit 3d skeleton-based human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9509–9520, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[74] Qingyao Xu, Weibo Mao, Jingze Gong, Chenxin Xu, Siheng Chen, Weidi Xie, Ya Zhang, and Yanfeng Wang. Joint-relation transformer for multi-person motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9816-9826, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[75] Sirui Xu, Yu-Xiong Wang, and Liangyan Gui. Stochastic multi-person 3d motion forecasting. In The Eleventh International Conference on Learning Representations, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.47, + 0.574 + ], + "angle": 0, + "content": "[76] Yi Xu, Armin Bazarjani, Hyung-gun Chi, Chiho Choi, and Yun Fu. Uncovering the missing pattern: Unified framework towards trajectory imputation and prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9632-9643, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[77] Maosheng Ye, Jiamiao Xu, Xunnong Xu, Tengfei Wang, Tongyi Cao, and Qifeng Chen. Bootstrap motion forecasting with self-consistent constraints. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8504-8514, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[78] Yingxuan You, Hong Liu, Ti Wang, Wenhao Li, Runwei Ding, and Xia Li. Co-evolution of pose and mesh for 3d human body estimation from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14963-14973, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[79] Bruce X.B. Yu, Zhi Zhang, Yongxu Liu, Sheng-hua Zhong, Yan Liu, and Chang Wen Chen. Gla-gen: Global-local adaptive graph convolutional network for 3d human pose estimation from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8818-8829, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[80] Kai Zhai, Qiang Nie, Bo Ouyang, Xiang Li, and Shanlin Yang. Hopfir: Hop-wise graphformer with intragroup joint refinement for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14985-14995, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[81] Yi Zhang, Pengliang Ji, Angtian Wang, Jieru Mei, Adam Kortylewski, and Alan Yuille. 3d-aware neural body fitting for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "occlusion robust 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9399-9410, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[82] He Zhao and Richard P. Wildes. Where are you heading? dynamic trajectory prediction with expert goal examples. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7629-7638, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[83] Qitao Zhao, Ce Zheng, Mengyuan Liu, Pichao Wang, and Chen Chen. Poseformerv2: Exploring frequency domain for efficient and robust 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8877-8886, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[84] Fang Zheng, Le Wang, Sanping Zhou, Wei Tang, Zhenxing Niu, Nanning Zheng, and Gang Hua. Unlimited neighborhood interaction for heterogeneous trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13168-13177, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[85] Jingxiao Zheng, Xinwei Shi, Alexander Gorban, Junhua Mao, Yang Song, Charles R Qi, Ting Liu, Visesh Chari, Andre Cornman, Yin Zhou, et al. Multi-modal 3d human pose estimation with 2d weak supervision in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4478-4487, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.892, + 0.486 + ], + "angle": 0, + "content": "[86] Chongyang Zhong, Lei Hu, Zihao Zhang, Yongjing Ye, and Shihong Xia. Spatio-temporal gating-adjacency gen for human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6447–6456, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.488, + 0.892, + 0.569 + ], + "angle": 0, + "content": "[87] Mu Zhou, Lucas Stoffl, Mackenzie Weygandt Mathis, and Alexander Mathis. Rethinking pose estimation in crowds: Overcoming the detection information bottleneck and ambiguity. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14689-14699, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.572, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[88] Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Kejie Lu. Hivt: Hierarchical vector transformer for multi-agent motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8823-8833, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.643, + 0.892, + 0.698 + ], + "angle": 0, + "content": "[89] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.892, + 0.795 + ], + "angle": 0, + "content": "[90] Dekai Zhu, Guangyao Zhai, Yan Di, Fabian Manhardt, Hendrik Berkemeyer, Tuan Tran, Nassir Navab, Federico Tombari, and Benjamin Busam. Ipcc-tp: Utilizing incremental pearson correlation coefficient for joint multi-agent trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5507-5516, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.798, + 0.892, + 0.853 + ], + "angle": 0, + "content": "[91] Tao Zhuo, Zhiyong Cheng, Peng Zhang, Yongkang Wong, and Mohan Kankanhalli. Unsupervised online video object segmentation with motion property understanding. IEEE Transactions on Image Processing, 29:237-249, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1628" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/12609e19-efed-4425-b5ad-32eb26a81347_origin.pdf b/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/12609e19-efed-4425-b5ad-32eb26a81347_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6caafb9fb85f3f0e49a7faa865e112ce7c86fe7f --- /dev/null +++ b/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/12609e19-efed-4425-b5ad-32eb26a81347_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:733301f867897ef0d5cfc4788dbb0111c4515364b2fafe6a97be407c744f61d3 +size 9262860 diff --git a/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/full.md b/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1804eb58a658a49bfe01d3f64dfffab7b0b739e4 --- /dev/null +++ b/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/full.md @@ -0,0 +1,376 @@ +# Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning + +Jaewoo Jeong*, Daehee Park*, and Kuk-Jin Yoon KAIST + +{jeong207,bag2824,kjyoon}@kaist.ac.kr + +# Abstract + +Human pose forecasting garners attention for its diverse applications. However, challenges in modeling the multi-modal nature of human motion and intricate interactions among agents persist, particularly with longer timescales and more agents. In this paper, we propose an interaction-aware trajectory-conditioned long-term multi-agent human pose forecasting model, utilizing a coarse-to-fine prediction approach: multi-modal global trajectories are initially forecasted, followed by respective local pose forecasts conditioned on each mode. In doing so, our Trajectory2Pose model introduces a graph-based agent-wise interaction module for a reciprocal forecast of local motion-conditioned global trajectory and trajectory-conditioned local pose. Our model effectively handles the multi-modality of human motion and the complexity of long-term multi-agent interactions, improving performance in complex environments. Furthermore, we address the lack of long-term $(6s+)$ multi-agent $(5+)$ datasets by constructing a new dataset from real-world images and 2D annotations, enabling a comprehensive evaluation of our proposed model. State-of-the-art prediction performance on both complex and simpler datasets confirms the generalized effectiveness of our method. The code is available at https://github.com/Jaewoo97/T2P. + +# 1. Introduction + +Human pose forecasting aims to predict future human motion based on observed past motion [20, 31, 32, 35, 37, 53, 86]. Humans instinctively perform such tasks, allowing them to naturally navigate in crowded areas or identify and circumvent potential dangers. For this reason, human pose forecasting plays an important role in various computer vision tasks [21, 23, 27, 54, 85, 91]. Indeed, recent years have seen a proliferation of work on multi-agent motion forecasting which aim towards modeling complex multi-agent inter + +![](images/1a55f1d7b76b2c2e415c064861262d024e6c7ec4a7e55e3fbc6e46e53ca2d746.jpg) +Figure 1. Human motion is goal-directed and influenced by other entities. Therefore, global intention contains hints for local intention, allowing us to infer local pose from global trajectories. Our method first forecasts global trajectories, upon which local poses are conditioned for subsequent forecasts. Pose and trajectory-wise inter-agent interactions are considered for both predictions. + +
Global Intent
ABC
Walk side-by-side +Avoid CKneel
+ +
Local Intent
ABC
Walking motion +Talking gesturesTie shoe
+ +action [20, 47, 53, 71, 74]. + +Although various methods have been proposed, they share two major limitations. The first is a limitation on long-term predictions, as previous studies predicted up to 3 seconds at most [4, 47, 74, 75]. However, a sufficiently long forecast horizon is essential to fully leverage human pose forecasting for diverse downstream tasks in the scope of identifying potential danger or understanding human behavior. The second is that multi-person interactions are not proficiently learned. Existing methods consider the joints of multiple people all at once as objects of interaction [47, 65, 74], resulting in an excessive complexity with respect to the number of joints. Due to such inefficient modeling, these approaches are found to be incompetent in long-term $(3s+)$ multi-agent $(6+)$ settings, limiting their practicality on complex real-world environments. + +Moreover, these challenges are also due to the limitations of datasets. Existing pose forecasting datasets have limited sequence length ( $\sim 3\mathrm{s}$ ) and number of agents ( $\sim 2$ ). Therefore, previous works [47, 69, 75] have randomly blended disparate datasets to model multi-agent interaction with up to 10 agents. Yet, such naively merged data lacks authentic interaction as agents from different scenes remain uninfluenced. As such, there was no opportunity to develop and + +evaluate a model in a long-term multi-agent environment. + +To this end, we present a solution from both model and dataset perspectives to tackle long-term multi-agent human pose forecasting. First, from a model perspective, we propose an interaction-aware trajectory-conditioned pose forecasting method. We point out that the limitations of existing methods on long-term multi-agent environments lead to poor performance in handling the multi-modal nature of human motion and correspondingly complex interactions. To improve upon handling multi-modality in these complex settings, we use a coarse-to-fine approach to enjoy effective interaction modeling by propagating agent-wise coarse representations. Agent-wise pose and trajectory embeddings are obtained in their respective local coordinates, followed by a holistic interaction modeling via our proposed Traj-pose module. Interaction-aware forecasts are then made by initial coarse global hip joint trajectory forecast followed by fine local pose forecasts in its hip joint coordinates, conditioned on the global trajectory as shown in Fig. 1. As discovered in previous research [2, 54], learning an agent-wise global intention as coarse trajectories is less challenging than predicting every joint-wise motion. We leverage these hints from global trajectories, which are further conditioned towards forecasting local motion that embodies the interaction-aware spatio-temporal context. + +From a dataset perspective, we parsed a novel real-world dataset for long-term multi-agent human pose forecasting. We utilize JRDB dataset [66] which consists of multi-view video and collected in various environments. Since 3D pose annotations are not provided in the original JRDB, we extracted sequences of 3D human pose from visible agents in omnidirectional images using the latest algorithm for 3D pose extraction from image [59]. We then ensure the reliability of 3D pose information by filtering and adjusting the extracted 3D poses based on 2D pose and 3D bounding box annotations. As a result, we construct a real-world 3D human pose forecasting dataset, JRDB-GlobMultiPose (JRDB-GMP), where up to 24 agents exist for up to 5 seconds. The proposed pose forecasting model is validated on both previous datasets and newly created JRDB-GMP dataset. Our method shows state-of-the-art forecasting performance in both global and local accuracy metrics, not only on JRDB-GMP but also on all previous datasets. Therefore, our contributions are as follows: + +- We propose a interaction-aware trajectory-conditioned pose forecasting method $(T2P)$ for long-term multi-agent 3D human pose forecasting. +- We propose a long-term, multi-agent real-world 3D human pose forecasting dataset which contains up to 24 persons and forecasts up to 5 seconds. +- We validate our $T2P$ model on both previous datasets and our new JRDB-GMP dataset. Our method achieves state-of-the-art forecasting performance on all datasets. + +# 2. Related works + +# 2.1. Human pose forecasting + +Human Pose Forecasting involves predicting a future pose sequence with temporal length of a prediction horizon, given a historical pose sequence [5, 8, 11, 16, 36, 45, 49, 52, 68, 71]. In the early stage, methods were developed to forecast single person motion within a short timeframe ( $\sim$ 1s) [9, 34, 58, 73]. However, to improve applicability on diverse downstream computer vision tasks, forecasts are to be made on multi-person poses [1, 2, 63] for longer prediction horizons [6, 62]. Forecasting future inherently involves a stochastic nature, and handling such multi-modality has been attempted by forecasting multiple future poses of a single agent [4]. However, comparatively marginal efforts have been employed in the more complex long-term multi-agent scenes [75]. Such absence is mostly due to the lack of a proper dataset. The commonly used evaluation datasets are CMU-Mocap [13], 3DPW [67], UMPM [64], MuPoTS-3D [39], all of which contain 2 agents at most in a given scene and have short prediction horizons within 3 seconds. Most recent research arbitrarily combines individual scenes to create datasets with more than three individuals [47, 69, 75]. However, such a synthetic approach does not account for authentic agent interactions. + +# 2.2. Trajectory prediction + +Trajectory prediction involves predicting the future path of an object given its past trajectory [3, 10, 25, 38, 40, 41, 51, 76, 77, 89, 90]. Unlike human pose forecasting which aims to predict every joint position, trajectory prediction regards each agent as a point mass, typically the center of mass or center point of a detected bounding box. Research in trajectory forecasting is interested in not only vehicles but also many types of agents including humans, cyclists, and so on [29, 43, 72, 84]. One substantial direction in research within this field is the Goal-conditioned prediction approach [18, 33, 82]. Goal-conditioned prediction approach first predicts the final destination within the prediction horizon with multiple goal proposals [28, 70]. Then, a thorough future path is conditioned on each mode of the multi-modal proposals. Compared to directly predicting full trajectories, the goal-conditioned approach follows a coarse-to-fine prediction and is effective in learning highly stochastic multi-modality of complex scenes [19, 42, 88]. + +# 2.3. Human pose estimation from image + +Human pose estimation is the task of inferring the pose of a person from an image or a video [22, 24, 26, 30, 57, 60, 78, 80, 81]. Initial deep learning-based methods first utilized convolutional neural networks to estimate 2D and 3D poses from single or multiple images [12, 14, 61, 79, 83]. Recent approaches engage in more challenging tasks such as + +![](images/1341b7a715ef1637d68c209a56efc4686d3ca20c761d52245beea7e57a33c25d.jpg) +Figure 2. Illustration of our T2P framework. We decompose global motion into global trajectory and local pose. Multi-modal global trajectory proposals are predicted from past global trajectory and local pose embeddings. Then, future local poses are conditioned and forecasted on each trajectory proposal to compose the final human pose prediction. Predicted local poses are added to their mode-specific global trajectories in a joint-wise manner, obtaining the global human poses as the final output. + +estimating 3D poses from monocular videos [7, 46, 48, 50, 55, 56] using self-supervised learning and generative methods [15, 17]. Most recent methods estimate multi-person poses in a crowded environment with considerable occlusions [44, 87]. We account for the aforementioned need of a complex dataset by extracting 3D pose from images using these methods. Specifically, we use a monocular 3D pose estimation method $BEV$ [59] to construct a 3D human motion forecasting dataset with long-term multi-agent characteristics from real-world image sequences. $BEV$ robustly estimates human pose in a scale-ambiguous and crowded environment, reliably extracting 3D poses from the omnidirectional image sequences of JRDB dataset [66]. + +# 3. Method + +# 3.1. Problem definition + +Multi-agent human pose forecasting aims to learn a mapping function between the observed 3D pose of $N_{A}$ agents composed of $J$ joints, $\mathbf{X}:\left\{\mathbf{x}_t^{n,j}\right\}_{-T_p:0}^{N_A,J}$ , and future pose $\mathbf{Y}:F\times \left\{\mathbf{x}_t^{n,j}\right\}_{0:T_f}^{N_A,J}$ in global coordinates where $F$ denotes the number of modes. Here, $T_{p}$ and $T_{f}$ are history length and prediction horizon while $\mathbf{x}_t^{n,j} = (x_t^{n,j},y_t^{n,j},z_t^{n,j})$ is 3D global coordinate of joint $j$ of agent $n$ at time $t$ . While the global position of joint is represented in $\mathbf{x}$ , we additionally define local position $\mathbf{p}$ . The local position is defined in local coordinate of each agent, calculated by subtracting the global position of the hip joint of each agent. Therefore, local position of joint is defined as $\mathbf{p}^{n,j} = \mathbf{x}^{n,j} - \mathbf{x}^{n,\mathrm{hip}}$ . We + +define the trajectory of global hip joint position as global trajectory, $Tr:\{\mathbf{x}^{n,\mathrm{hip}}\}^{N_A}$ . We also define local pose as local position of all joints, $Po:\{\mathbf{p}^{n,\mathrm{j}}\}^{N_A,J}$ . We denote past and future timesteps of global trajectory and local motion as $Tr_{\mathcal{P}},Tr_{\mathcal{F}}\in Tr$ and $Po_{\mathcal{P}},Po_{\mathcal{F}}\in Po$ , where $\mathcal{P}$ and $\mathcal{F}$ respectively denotes past and future. + +# 3.2. Overall framework + +We disentangle the overall human motion into global trajectories and local poses, as depicted in top left of Fig. 2. Following a coarse-to-fine strategy, multiple global trajectories are first forecasted to model the coarse modes of global intentions. Based on these forecasts, local pose predictions are conditioned on each mode to jointly constitute a thorough motion. In doing so, our model is widely divided into two portions: Trajectory predictor consists of trajectory encoder and decoder and pose predictor consists of pose encoder and decoder. Both predictors engage in the reciprocal exchange of both trajectory and pose information, facilitating the inference of cues between global and local motion. The detailed methods of each stage are described below: + +# 3.3. Model structure + +# 3.3.1 Pose encoder + +Unlike the holistic approach of previous works that encode and decode all agents' joint motions in global coordinates, our pose encoder encodes the pose dynamics in local coordinates. In addition, our pose encoder only considers intra-agent joint interaction. As a result, the encoded pose embedding represents agent-specific local mo + +tion, containing insights on global intent. We follow our baseline [47] and construct the encoder with Multi-Person Body-Part (MPBP) module and transformer networks. As depicted in Fig. 2, body part sequences are constructed in frequency domain, followed by intra-agent attention-based encoding of the body parts to acquire pose embedding $Z_{Po}$ . + +# 3.3.2 Trajectory module + +Trajectory module aims to extract embeddings from the agents' past global trajectory. Using an encoder structure from [88], multi-agent interaction-based trajectory embedding $Z_{Tr}$ is extracted which contains rudimentary insight on global intent. Interaction between agent trajectories is represented based on the reference agent $i$ 's global trajectory segment vector $\mathbf{v}_t^i = \mathbf{x}_t^{i,\mathrm{hip}} - \mathbf{x}_{t-1}^{i,\mathrm{hip}}$ . For rotational invariance, neighbor actor $j$ 's vector is normalized by the reference vector's orientation at latest timestep $t=0$ . Separate MLP layers then compute the reference agent and neighboring agent embeddings $z_{Tr_i}^t, z_{Tr_j}^t$ as follows: + +$$ +\begin{array}{l} z _ {T r _ {i}} ^ {t} = \phi_ {r e f} (R _ {i} ^ {T} \mathbf {v} _ {t} ^ {i}) \\ z _ {T r _ {j}} ^ {t} = \phi_ {n b r} ([ R _ {i} ^ {T} (\mathbf {v} _ {t} ^ {j}), R _ {i} ^ {T} (\mathbf {v} _ {t} ^ {i}) ]) \tag {1} \\ \end{array} +$$ + +where $\phi_{ref}$ and $\phi_{nbr}$ are different MLP blocks, $R_i \in \mathbb{R}^{3 \times 3}$ is the rotation matrix of agent $j$ against agent $i$ , $[\cdot, \cdot]$ is concatenation. The resulting agent-specific reference and neighbor embeddings constitute trajectory embedding $z_{Tr}^t$ . + +# 3.3.3 Traj-posed module + +Human maneuver contains various dynamic activities characterized by the agent's multi-modal intents. Auxiliary human motion such as arm gesture, rotational orientation of upper body and head implies the agent's intent in global motion. In that sense, harvesting meaningful insights from past local joint motion helps proficient modeling of coarse multimodality as future trajectory proposals. Therefore, we propose Traj-Pose Module that fuses agent-wise embeddings of both trajectory and pose to fully utilize these information in modeling global intentions. + +First, MLP is used to match the temporal domain of pose embedding $Z_{Po}$ to that of $Z_{Tr}$ , after which both are concatenated as agent-wise traj-posed embedding $Z$ . + +$$ +Z = \left[ Z _ {T r}, \phi_ {M L P} \left(Z _ {P o}\right) \right] \tag {2} +$$ + +The resulting $Z$ is comprised of agent and timestep-respective trajectory and pose embeddings: $z_{i}^{t},z_{j}^{t}\in z^{t}\in Z$ Then, $\widetilde{Z}$ is acquired from the graph attention with an agent-wise update where each agent embedding $z_{i}^{t}$ and its neighbor embedding $z_{j}^{t}$ are used as query and key/value. Similar + +to trajectory interaction encoder of HiVT [88], the graph attention operation is operated as follows: + +$$ +\begin{array}{l} \alpha_ {i} ^ {t} = \operatorname {s o f t m a x} \left(\frac {q _ {i} ^ {t \top}}{\sqrt {d _ {k}}} \cdot \left[ \left\{k _ {j} ^ {t} \right\} _ {j \in N _ {i}} \right]\right), \\ m _ {i} ^ {t} = \sum_ {j \in N _ {i}} \alpha_ {i} ^ {t} v _ {j} ^ {t}, \tag {3} \\ \end{array} +$$ + +$$ +g _ {i} ^ {t} = \operatorname {s i g m o i d} \left(W ^ {\text {g a t e}} \left[ z _ {i} ^ {t}, m _ {i} ^ {t} \right]\right), +$$ + +$$ +\widetilde {z} _ {i} ^ {t} = g _ {i} ^ {t} \odot W ^ {\text {s e l f}} z _ {i} ^ {t} + \left(1 - g _ {i} ^ {t}\right) \odot m _ {i} ^ {t} +$$ + +where $N_{i}$ is a set of agent $i$ 's neighbors, $W^{gate}$ and $W^{self}$ are learnable matrices, and $\odot$ is element-wise product. + +# 3.3.4 Trajectory decoder + +Trajectory is forecasted from the output of trajectory encoder which encodes both past global trajectory and local pose information. Since its graph operation is operated by each timestep, a temporal encoder is used as a temporal encoder to integrate $\widehat{Z}$ in the temporal dimension. A multihead self-attention temporal encoder is used as the temporal encoder. Aggregator then takes into account variations in local coordinate frames to accurately represent geometric relationships within the global coordinate system via a graph operation. MLP is subsequently applied to span embedding $F$ times for multi-modal prediction, which is residually added to the $\times F$ repeated embedding before the aggregator. Finally, another MLP is used to extract multi-modal future global trajectory proposals of hip joint $Tr_{\mathcal{F}} \in \mathbb{R}^{F \times T_f \times 3}$ . The multi-modal embedding is also passed onto the Pose Decoder to forecast local poses. + +# 3.3.5 Pose decoder + +Future human pose depends on past human poses and global intention. The pose decoder is designed to consider these factors while generating local poses via mode-specific trajectory conditioning. A transformer (TRM) decoder is used to decode local motions, where pose embedding $Z_{Po}$ is used as key/value and concatenation of trajectory and pose query. + +$$ +Q = \phi_ {M L P} ([ Q _ {P o}, Q _ {T r} ]), K = Z _ {P o}, V = Z _ {P o} \tag {4} +$$ + +Using both pose and trajectory queries, past pose embedding $Z_{Po}$ is conditioned on both MPBP sequence at $t = 0$ and the multi-modal trajectory proposals which contain global intent. Subsequently, inverse discrete cosine transform (idct) is applied to convert the future pose proposals from frequency domain to local coordinate domain, $P_{o_{\mathcal{F}}}$ . The final multi-modal future pose in global coordinates is acquired as Eq. 5 where $\oplus$ is a joint-wise addition operation. + +$$ +\mathbf {Y} = T r _ {\mathcal {F}} \oplus P o _ {\mathcal {F}}, \quad \mathbf {Y} \in \mathbb {R} ^ {F \times N _ {A} \times T _ {f} \times 3} \tag {5} +$$ + +Table 1. Comparison of statistics between existing human pose forecasting datasets and newly proposed JRDB-GMP dataset. + +
Dataset
CMU-Mocap (UMPM)MuPoTs -3D3DPWJRDB-GMP 1s/2s2s/5s
Duration (s)400026717001863
Location #-20-27
Sample #1300019243211534593
avg. agent3326.86.8
med. agent #33255
max agent #3322422
avg. vel. (m/s)0.30.260.570.460.38
avg. disp.(m)0.630.551.130.640.79
max. disp.(m)4.622.4510.718.4411.0
+ +![](images/fbebee859bddfbb4d3d177925974b79304b1a13b0ffdfc0cf8b14adb1eab21cf.jpg) +Figure 3. Example scenes from the JRDB-GMP dataset, illustrating its long-term, multi-agent nature. + +# 3.4. Training objective + +Both objectives of global trajectory and local pose forecasting are trained jointly. For both global trajectory and local pose prediction, $L2$ loss is propagated to the mode with minimal $L2$ distance with the ground truth. + +$$ +L _ {T r} = \sum_ {n = 1} ^ {N _ {A}} \sum_ {t = 1} ^ {T _ {f}} \| \widetilde {y} _ {T r, n} ^ {t} - \hat {y} _ {T r, n} ^ {t} \| +$$ + +$$ +L _ {P o} = \sum_ {n = 1} ^ {N _ {A}} \sum_ {t = 1} ^ {T _ {f}} \sum_ {j = 1} ^ {J - 1} \| \left(\widetilde {y} _ {P o, n} ^ {t, j} - \hat {y} _ {P o, n} ^ {t, j}\right) \| \tag {6} +$$ + +$$ +L = L _ {T r} + L _ {P o} +$$ + +# 3.5.JRDB-GMP dataset + +Due to the absence of existing long-term $(3s+)$ multi-agent $(6+)$ dataset, we compose a unique 3D human pose forecasting dataset in a real-world environment from JRDB [66]. The original JRDB dataset is constructed by a moving robot that records human activity around a school campus using 5 omnidirectional cameras and LiDAR. Image sequences along with 2D pose annotation and 3D bounding box annotations are provided in the original dataset. However, since 3D human pose annotations are unavailable, we separately parse accurate 3D human pose from provided inputs and + +![](images/507fc764ff5e37ab86f96fc8431046ff42ffa209d9888a29300d9bf19aa962c2.jpg) + +![](images/a7dce04996df690df884af1d9a2c03334fe254bd55eadb8f4382bf933af9807b.jpg) +Figure 4. Various motions from the JRDB-MultiGlobPose dataset, providing rich motion queues for inter-agent interaction inference. + +annotations. First, a SOTA monocular 3D pose extraction method [59] is used to extract raw 3D joint positions from image sequences. Then, 2D pose and 3D bounding box annotations are used to refine the raw joint positions and minimize noise. We use 2D pose annotations to initially filter out the 3D poses with noise. With camera parameters and refined 3D pose, we project it on 2D image plane, then L2 distance between projected 2D pose and GT 2D pose annotation is calculated. If the mean L2 distance per each agent at a time stamp is over a threshold, that instance is filtered out. 2D pose annotations are also projected in 3D space to refine the remaining 3D poses, ensuring the accuracy of the 3D pose information of our JRDB-GMP dataset. Further details are elaborated in the supplementary materials. + +Figure 3 visualizes some scenes of the constructed dataset. Accurate extraction of 3D poses has been made even with considerable occlusion via the use of 2D poses. The dataset includes agents with both long and short traverse distances and rich inter-agent interactions in both trajectory and local pose aspects. Figure 4 illustrates diverse local poses included in the dataset, which serve as motion cues of inter-agent interaction. Both figures confirm our method's accuracy in extracting 3D multi-human pose, even in crowded environments. Table 1 scrutinizes the statistics compared to previously used datasets. Compared to earlier datasets, the average number of agents is more than twice as high. In addition, comparing JRDB-GMP 1s/2s to CMU-Mocap and MuPoTs datasets, JRDB contains more diverse and longer motion as shown by a similar magnitude of average displacement but longer maximum displacement. + +Table 2. Quantitative comparison of our method to previous methods on CMU-mocap (UMPM), 3DPW, and JRDB-GlobMultiPose datasets with number of prediction modes $(F)$ as 6. Lower is better for all metrics. The best results are marked in bold. + +
DatasetCMU-mocap (UMPM)3DPWJRDB-GlobMultiPose
In/out length (s)1/20.8/1.61/22/5
Evaluation time (s)120.81.6122.55
JPEMRT [69]164.7280.1159.1251.2259.3349.3438.4474.0
JRT [74]168.5316.9181.9287.3237.9373.1351.9538.8
TBIFormer [47]170.0290.9153.9265.8257.1339.3443.2481.3
Ours152.4262.7142.6236.2224.0301.4341.6390.4
APEMRT [69]127.0164.4117.9153.272.387.388.5101.9
JRT [74]121.2181.6133.4178.0112.6154.396.7120.2
TBIFormer [47]125.1160.8115.4152.770.683.388.2102.9
Ours114.4151.7114.6150.070.883.382.294.7
FDEMRT [69]99.6204.7102.7185.3235.2325.2418.2454.8
JRT [74]117.7250.8133.7235.4211.4337.4318.5497.2
TBIFormer [47]112.1228.5106.7215.9232.4314.6423.9458.8
Ours88.7188.974.1158.2194.7271.5313.9361.0
+ +# 4. Experiment + +# 4.1. Dataset + +We test our model on three datasets: CMU-Mocap (UMPM) [13, 64], 3DPW [67], and our JRDB-GMP. Although our model is designed to forecast human poses in a long-term multi-agent environment, we also report experimental results on previous benchmark datasets with simpler scenes. Mocap-UMPM is a mixed dataset of Mocap and UMPM containing synthesized human interaction between three agents [47]. 3DPW is a dataset with 2 agents traversing a real-world environment. We report the test results on each after separate training on respective datasets. + +# 4.2. Metrics + +We use the following widely-used metrics. For a detailed definition, please refer to the supplementary material. + +APE: Aligned mean per joint Position Error is used as a metric to evaluate the forecasted local motion. $L2$ distance of each joint in the hip joint coordinate is averaged over all joints for a given timestep. + +FDE: Final Distance Error evaluates the forecasted global trajectory by calculating the $L2$ distance of a given timestep. + +JPE: Joint Precision Error evaluates both global and local predictions by mean $L2$ distance of all joints for a timestep. + +# 4.3. Implementation details + +We train our model on a single A6000 GPU. 2 layers of pose encoder transformer are stacked, followed by 2 layers of transformer in pose decoder. Embedding dimensions of 96 and 128 are used for trajectory and pose embeddings, respectively. The transformed key, value dimension of 64 is used for all transformer architectures. A learning rate of 0.003 is used with an AdamW optimizer with weight decay. Further details can be found in the supplementary materials. + +Table 3. Short-term prediction results on CMU-Mocap (UMPM) dataset, where 1s of poses are forecasted given 2s of poses. + +
MetricJPEAPEFDE
Time (s)0.20.61.00.20.61.00.20.61.0
MRT64.515221749.811014039.497.9153
JRT31.510417328.785.912517.763.9120
TBIFormer37.410415832.885.811923.363.7104
Ours37.810215833.884.411614.949.192.6
+ +# 4.4. Baselines + +We compare our method against the latest SOTA methods for multi-agent pose forecasting [47, 69, 74]. To compare the multi-modal predictions of these three methods, we extend their prediction modes by spanning embedding $K$ times in the same manner as ours. All baselines are trained and evaluated on CMU-mocap (UMPM) and 3DPW datasets. CMU-Mocap (UMPM) dataset predicts 2 seconds from 1 second of poses, and 3DPW predicts 1.6 seconds from 0.8 seconds of poses, both from 6 modes. For 3DPW dataset, we slightly lengthen the forecast horizon to evaluate long term predictions. For JRDB-GMP dataset, both short (1s/2s) and long term (2s/5s) predictions are evaluated for all models. Lastly, We use HiVT [88] as the baseline For global trajectory prediction of our model. + +# 5. Results + +# 5.1.Quantitative results + +Table 2 compares the quantitative performances on three datasets. Our method exhibits considerable performance gain against all previous SOTA methods, not only on the proposed long-term multi-agent dataset but also on the existing two datasets. Such generalized competence demonstrates the applicability of our trajectory-conditioned pose forecasting method to various real-world scenarios. In detail, our approach achieves over $10\%$ gain of FDE on all datasets. This improvement on forecasting global locomo + +![](images/10bbec035091361641fd3dfb6ac0716dfbc5352256d96f352c218c3ce2aae723.jpg) +t +2.5s +5s + +![](images/4a0d733b417b0691157108f7eda42efe5ca5e599fe5f92e2043110e954c4d6a3.jpg) +Figure 5. Visualization of a long-term forecasting scene from JRDB-GMP (2/5) dataset. Past poses for input are shown on the leftmost column, GT future poses on the next, and forecasts by ours, MRT, TBIFormer, and JRT, respectively. + +![](images/f2a9882d1651207055f03c33321a9385ce57ff85ec9b2489d2431b9c9a4cf832.jpg) + +![](images/b44c99c37110e82209b08963524d194ca4c4042e416e7bbf7e289a55409aec50.jpg) + +![](images/1a6003607be61c44db3936bb5304853d7277c6be956e5bea3528142594565f3f.jpg) + +![](images/6148c95f486c1457ca9e5c06749adb80bd4715c870cbfa4143d5448d43d8daef.jpg) + +![](images/7f23d1a7548a187491fa3133a38a67f1db919411589fe758e0fcd2a8693da24d.jpg) +Figure 6. Visualization of a CMU-Mocap (UMPM) scene. Past poses are shown on the upper row, GT future poses on the next, and forecasts by TBIFormer and ours on the latter two rows. To visualize motion, we stack several frames around the target time stamp. Black/red/blue arrows refer to the direction of the global trajectory, and yellow arrows refer to the direction of foot motion. + +tion could be accredited to the decoupled forecasting of global trajectory and local pose. Previous methods holistically predict both global and local movements, limiting both performances due to superfluous interactions to consider between all joints. Conversely, our approach can extract accurate global intent by decoupling past motion into global and local representations. Moreover, our effective interaction modeling of global and local pose also helps to predict a more accurate global trajectory under multi-agent environment as shown in the latter ablation study. + +For APE metric, our method also surpasses previous SOTA models on all datasets, highlighting the accurate extraction of local pose intent. Such improvement shows that our approach generates plausible local motion due to its proficient sampling from coarse global intents. Our method simplifies the task by learning multi-modality in a coarse-to-fine approach. Its subsequent local motion forecasting is inferred from coarsely modeled multi-modality, a greatly simplified task compared to extracting intent from entangled multi-modality as well as multi-agent interaction. + +These improvements on both global and local scales + +jointly contribute toward lowering the JPE metric, demonstrating proficiency of our method in forecasting overall human motion. Based on such competence, our method which aimed towards improving on forecasting long-term multiagent environments also exhibits similar or better performances on short timeframes as shown in Tab. 3. In addition, our approach also excels even on sole local motion with minimal global displacement, as elaborated in section 7.2 of supplementary materials. + +# 5.2. Qualitative results + +Our method forecasts a more plausible global pose in longer timescales ( $\sim 5\mathrm{s}$ ) as shown in the interacting scene of five agents in Fig. 5. Looking at the input and GT sequences, the leftmost person avoids the traversing couple from right to left. The two people in front are stationary while talking to each other. MRT and TBIFormer forecasts implausible overlapped poses at the final prediction horizon ( $t = 5\mathrm{s}$ ). JRT fails to learn the global locomotion of agents due to the high complexity of its attention mechanism and is stuck in the local minimum of predicting the inactivity of all agents. On the other hand, our model forecasts plausible poses where the closely interacting two agents walk side-by-side. + +Figure 6 illustrates exemplary sequences where more natural local motion has been forecasted by our method. Comparing forecasts on a scene of walking agents, our method generates a much more plausible sequence where the stepping foot remains stationary. On the other hand, the previous SOTA method, TBIFormer, struggles to learn the natural walking mechanism of human legs and a parallel translation of both feet is exhibited. Such discrepancy shows that trajectory-conditioning for inferring local motion from global intent generates more proficient details in human motion than SOTA methods. More visualizations could be found in the supplementary materials. + +# 5.3. Ablation studies + +Different number of modes. The main quantitative results report prediction results with $F$ as 6 to compare the ability to address the multi-modal nature of human motion during pose forecasting. Table. 4 additionally compares forecast + +Table 4. Comparison of performance with different number of modes in CMU-Mocap (UMPM) dataset. + +
F16
Metric @ 2sAPEJPEAPEJPE
MRT163.9366.4164.4280.1
JRT176.7367.4181.6316.9
TBIFormer160.1374.3160.8290.9
Ours154.4366.4151.7262.7
+ +Table 5. Ablation studies on core components of model structures. Experiments are done with JRDB-GMP dataset to evaluate multiagent long-term performance. + +
Exp. #Trajectory encoderPose decoderMetrics
Local pose embeddingAgent interactionTrajectory -conditioningJPE @5sAPE @5sFDE @5s
-471.4101.7457.9
1✓400.595.1370.9
2✓403.394.7374.2
3✓401.293.0372.8
4✓✓395.693.8366.8
5✓✓392.795.2363.4
6✓✓✓391.291.4363.3
+ +results with $F$ as 1. Our method again achieves noticeable improvement in APE over the baseline on single-modal forecasts. With $F = 1$ , although our method barely enjoys improvement in forecasting global motion due to the absence of multi-modality, its superiority in APE shows the validity of our coarse-to-fine forecasting strategy that also effectively captures agent interaction. Our method improves with multi-modal predictions, demonstrating the proficiency of a coarse-to-fine approach in interpreting the stochastic nature of human motion and its intent. Note that our method improves in APE along with an increase in $F$ unlike previous methods, indicating a unique aptitude in addressing the multi-modal nature of not only global locomotion but also local pose intent via trajectory-conditioning. + +Importance of each architecture component. Table. 5 reports the influence of core components of our model. For the trajectory encoder, we evaluate the importance of using local pose embedding and modeling agent interaction. Comparing experiments 1, 5 and 3, 4, both show improvements in JPE and FDE metrics with the use of local pose embedding. Our method has taken advantage of detailed local pose cues to infer an agent's global intention. For interaction modeling, its use is beneficial for both global and local forecasts as compared by experiments 4 and 6. These joint improvements demonstrate the importance of considering local and global motion interactions for their respective forecasts. As for the pose decoder, comparisons of experiments 2,4 and 5,6 both show improvements in APE metric. Such consistent improvement verifies the effectiveness of the trajectory-conditioned local motion forecast approach in generating plausible local motion from global intention. + +Importance of interaction modeling. Accurate model + +Table 6. Ablation studies on agent interaction cutoff distance on JRDB-GMP. + +
JPE @ 5s
TBIFormerOurs
w/o interaction483.8406.0
w/ interaction < 2m-403.5
w/ interaction < 4m-400.5
w/ interaction all481.3390.4
+ +ing of inter-agent interaction becomes more pivotal to forecast in more complex environments. Indeed, its complexity grows in a long-term multi-agent scene. When holistically considering joint-wise interaction for all timesteps, the computation complexity is acquired as $O(T^2 \cdot N^2 \cdot J^2)$ , where $T$ is the number of timesteps, $N$ the number of agents, and $J$ the number of joints. On the contrary, with interaction modeling in global trajectory scale, our method reduces the computation cost by $TJ^2$ into $O(T \cdot N^2)$ . This enables efficient and proficient modeling of intra (pose) and inter (trajectory)-agent interactions as shown by Tab. 6. While body part-wise interaction modeling only improved by $0.52\%$ for TBIFormer, ours improves up to $3.84\%$ with interaction modeling. This demonstrates the proficiency of our efficient interaction modeling-based method in inferring global and local intents from complex interactions. In addition, the gradual improvement of JPE according to a wider interaction range confirms the importance of interaction modeling of more agents, which cannot be learned from the arbitrarily mixed previous datasets. + +# 6. Conclusion + +In this work, we propose a novel interaction-aware trajectory-conditioned approach to handle long-term multi-agent motion forecasting, along with a new dataset suited for such scope. Our proposed model utilizes a coarse-to-fine approach and decouples overall motion prediction into global and local components. Multi-modality of human motion is proficiently modeled via inferring fine local intents from coarse global intents, along with efficient agent-wise interaction modeling. As for the dataset, our JRDB-GMP dataset contains unprecedented long-term $(5s+)$ multi-agent $(6+)$ interactions in a real-world setting. Our method achieves state-of-the-art performance on all previous datasets and JRDB-GMP dataset, offering generalized practical implications in real-world applications. + +Acknowledgements This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601). + +# References + +[1] Vida Adeli, Ehsan Adeli, Ian Reid, Juan Carlos Niebles, and Hamid Rezatofighi. Socially and contextually aware human motion and pose forecasting. IEEE Robotics and Automation Letters, 5(4):6033-6040, 2020. 2 +[2] Vida Adeli, Mahsa Ehsanpour, Ian Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13390-13400, 2021. 2 +[3] Gorkay Aydemir, Adil Kaan Akan, and Fatma Güney. Adapt: Efficient multi-agent trajectory prediction with adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8295-8305, 2023. 2 +[4] German Barquero, Sergio Escalera, and Cristina Palmero. Belfusion: Latent diffusion for behavior-driven human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2317-2327, 2023. 1, 2 +[5] Arij Bouazizi, Adrian Holzbock, Ulrich Kressel, Klaus Dietmayer, and Vasileios Belagiannis. Motionmixer: Mlp-based 3d human body pose forecasting. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 791-798. International Joint Conferences on Artificial Intelligence Organization, 2022. Main Track. 2 +[6] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 387–404. Springer, 2020. 2 +[7] Wenhao Chai, Zhongyu Jiang, Jenq-Neng Hwang, and Gaoang Wang. Global adaptation meets local generalization: Unsupervised domain adaptation for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14655-14665, 2023. 3 +[8] Ling-Hao Chen, JiaWei Zhang, Yewen Li, Yiren Pang, Xiaobo Xia, and Tongliang Liu. Humanmac: Masked motion completion for human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9544-9555, 2023. 2 +[9] Hsu-kuang Chiu, Ehsan Adeli, Borui Wang, De-An Huang, and Juan Carlos Niebles. Action-agnostic human pose forecasting. In 2019 IEEE winter conference on applications of computer vision (WACV), pages 1423–1432. IEEE, 2019. 2 +[10] Sehwan Choi, Jungho Kim, Junyong Yun, and Jun Won Choi. R-pred: Two-stage motion prediction via tube-query attention-based trajectory refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8525-8535, 2023. 2 +[11] Rohan Choudhury, Kris M. Kitani, and László A. Jeni. Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In Proceedings of the IEEE/CVF International + +Conference on Computer Vision (ICCV), pages 14750-14760, 2023. 2 +[12] Andrea Porfiri Dal Cin, Giacomo Boracchi, and Luca Magri. Multi-body depth and camera pose estimation from multiple views. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17804-17814, 2023. 2 +[13] CMU-Graphics-Lab. Cmu graphics lab motion capture database. http://mocap.cs.cmu.edu/, 2003. 2, 6 +[14] Runyang Feng, Yixing Gao, Xueqing Ma, Tze Ho Elden Tse, and Hyung Jin Chang. Mutual information-based temporal difference learning for human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17131-17141, 2023. 2 +[15] Runyang Feng, Yixing Gao, Tze Ho Elden Tse, Xueqing Ma, and Hyung Jin Chang. Diffpose: Spatiotemporal diffusion model for video-based human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14861-14872, 2023. 3 +[16] Xuehao Gao, Shaoyi Du, Yang Wu, and Yang Yang. Decompose more and aggregate better: Two closer looks at frequency representation learning for human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6451-6460, 2023. 2 +[17] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13041-13051, 2023. 3 +[18] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15303-15312, 2021. 2 +[19] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15303-15312, 2021. 2 +[20] Wen Guo, Xiaoyu Bie, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Multi-person extreme motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13053-13064, 2022. 1 +[21] Je-Seok Ham, Dae Hoe Kim, NamKyo Jung, and Jinyoung Moon. Cipf: Crossing intention prediction network based on feature fusion modules for improving pedestrian safety. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3665-3674, 2023. 1 +[22] Karl Holmquist and Bastian Wandt. Diffpose: Multi-hypothesis human pose estimation using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15977-15987, 2023. 2 +[23] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 1 + +[24] Boyuan Jiang, Lei Hu, and Shihong Xia. Probabilistic triangulation for uncalibrated multi-view 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14850-14860, 2023. 2 +[25] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9644-9653, 2023. 2 +[26] Zhehan Kan, Shuoshuo Chen, Ce Zhang, Yushun Tang, and Zhihai He. Self-correctable and adaptable inference for generalizable human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5537-5546, 2023. 2 +[27] Jeongho Kim, Wooksu Shin, Hancheol Park, and Jongwon Baek. Addressing the occlusion problem in multi-camera people tracking with human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5462-5468, 2023. 1 +[28] Mihee Lee, Samuel S Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. Musevae: multi-scale vae for environment-aware long term trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2221-2230, 2022. 2 +[29] Lihuan Li, Maurice Pagnucco, and Yang Song. Graph-based spatial transformer with memory replay for multifuture pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2241, 2022. 2 +[30] Huan Liu, Qiang Chen, Zichang Tan, Jiang-Jiang Liu, Jian Wang, Xiangbo Su, Xiaolong Li, Kun Yao, Junyu Han, Errui Ding, Yao Zhao, and Jingdong Wang. Group pose: A simple baseline for end-to-end multi-person pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15029-15038, 2023. 2 +[31] Tiezheng Ma, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Progressively generating better initial guesses towards next stages for high-quality human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6437-6446, 2022. 1 +[32] Takahiro Maeda and Norimichi Ukita. Motionaug: Augmentation with physical correction for human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6427-6436, 2022. 1 +[33] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15233-15242, 2021. 2 +[34] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9489-9497, 2019. 2 +[35] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Generating smooth pose sequences for diverse human motion pre + +diction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13309-13318, 2021. 1 +[36] Wei Mao, Richard I Hartley, Mathieu Salzmann, et al. Contact-aware human motion forecasting. Advances in Neural Information Processing Systems, 35:7356-7367, 2022. 2 +[37] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Weakly-supervised action transition learning for stochastic human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8151-8160, 2022. 1 +[38] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5517-5526, 2023. 2 +[39] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 2018 International Conference on 3D Vision (3DV), pages 120-130, 2018. 2 +[40] Jiquan Ngiam, Vijay Vasudevan, Benjamin Caine, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, David J Weiss, Benjamin Sapp, Zhifeng Chen, and Jonathon Shlens. Scene transformer: A unified architecture for predicting future trajectories of multiple agents. In International Conference on Learning Representations, 2022. 2 +[41] Daehee Park, Jaewoo Jeong, and Kuk-Jin Yoon. Improving transferability for cross-domain trajectory prediction via neural stochastic differential equation. arXiv preprint arXiv:2312.15906, 2023. 2 +[42] Daehee Park, Hobin Ryu, Yunseo Yang, Jegyeong Cho, Jiwon Kim, and Kuk-Jin Yoon. Leveraging future relationship reasoning for vehicle trajectory prediction. In The Eleventh International Conference on Learning Representations, 2023. 2 +[43] Daehee Park, Jaeseok Jeong, Sung-Hoon Yoon, Jaewoo Jeong, and Kuk-Jin Yoon. T4p: Test-time training of trajectory prediction via masked autoencoder and actor-specific token memory, 2024. 2 +[44] Sungchan Park, Eunyi You, Inhoe Lee, and Joonseok Lee. Towards robust and smooth 3d multi-person pose estimation from monocular videos in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14772-14782, 2023. 3 +[45] Behnam Parsaeifard, Saeed Saadatnejad, Yuejiang Liu, Taylor Mordan, and Alexandre Alahi. Learning decoupled representations for human pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2294-2303, 2021. 2 +[46] Qucheng Peng, Ce Zheng, and Chen Chen. Source-free domain adaptive human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4826-4836, 2023. 3 +[47] Xiaogang Peng, Siyuan Mao, and Zizhao Wu. Trajectory-aware body interaction transformer for multi-person pose forecasting. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 17121- 17130, 2023. 1, 2, 4, 6 +[48] Zhongwei Qiu, Qiansheng Yang, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Chang Xu, Dongmei Fu, and Jingdong Wang. Psvt: End-to-end multi-person 3d pose and shape estimation with progressive video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21254-21263, 2023. 3 +[49] Muhammad Rameez Ur Rahman, Luca Scofano, Edoardo De Matteis, Alessandro Flaborea, Alessio Sampieri, and Fabio Galasso. Best practices for 2-body pose forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3613-3623, 2023. 2 +[50] Dripta S. Raychaudhuri, Calvin-Khang Ta, Arindam Dutta, Rohit Lal, and Amit K. Roy-Chowdhury. Prior-guided source-free domain adaptation for human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14996-15006, 2023. 3 +[51] Luke Rowe, Martin Ethier, Eli-Henry Dykhne, and Krzysztof Czarnecki. Fjmp: Factorized joint multi-agent motion prediction over learned directed acyclic interaction graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13745-13755, 2023. 2 +[52] Saeed Saadatnejad, Ali Rasekh, Mohammadreza Mofayezi, Yasamin Medghalchi, Sara Rajabzadeh, Taylor Mordan, and Alexandre Alahi. A generic diffusion-based approach for 3d human pose prediction in the wild. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8246-8253. IEEE, 2023. 2 +[53] Tim Salzmann, Marco Pavone, and Markus Ryll. Motron: Multimodal probabilistic human motion forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6457-6466, 2022. 1 +[54] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. IEEE Robotics and Automation Letters, 2023. 1, 2 +[55] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14761-14771, 2023. 3 +[56] Xiaolong Shen, Zongxin Yang, Xiaohan Wang, Jianxin Ma, Chang Zhou, and Yi Yang. Global-to-local modeling for video-based 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8887-8896, 2023. 3 +[57] Mingyi Shi, Sebastian Starke, Yuting Ye, Taku Komura, and Jungdam Won. Phasemp: Robust 3d pose estimation via phase-conditioned human motion prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14725-14737, 2023. 2 +[58] Theodoros Sofianos, Alessio Sampieri, Luca Franco, and Fabio Galasso. Space-time-separable graph convolutional + +network for pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11209-11218, 2021. 2 +[59] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13243-13252, 2022. 2, 3, 5 +[60] Yuran Sun, Alan William Dougherty, Zhuoying Zhang, Yi King Choi, and Chuan Wu. Mixsynthformer: A transformer encoder-like structure with mixed synthetic self-attention for efficient human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14884–14893, 2023. 2 +[61] Zhenhua Tang, Zhaofan Qiu, Yanbin Hao, Richang Hong, and Ting Yao. 3d human pose estimation with spatiotemporal criss-cross attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4790-4799, 2023. 2 +[62] Julian Tanke, Chintan Zaveri, and Juergen Gall. Intention-based long-term human motion anticipation. In 2021 International Conference on 3D Vision (3DV), pages 596-605. IEEE, 2021. 2 +[63] Julian Tanke, Linguang Zhang, Amy Zhao, Chengcheng Tang, Yujun Cai, Lezi Wang, Po-Chen Wu, Juergen Gall, and Cem Keskin. Social diffusion: Long-term multiple human motion anticipation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9601-9611, 2023. 2 +[64] N.P. van der Aa, X. Luo, G.J. Giezeman, R.T. Tan, and R.C. Veltkamp. Umpm benchmark: A multi-person dataset with synchronized video and motion capture data for evaluation of articulated human motion and interaction. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 1264-1269, 2011. 2, 6 +[65] Edward Vendrow, Satyajit Kumar, Ehsan Adeli, and Hamid Rezatofighi. Somoformer: Multi-person pose forecasting with transformers. arXiv preprint arXiv:2208.14023, 2022. 1 +[66] Edward Vendrow, Duy Tho Le, Jianfei Cai, and Hamid Rezatofighi. Jrdb-pose: A large-scale dataset for multiperson pose estimation and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4811–4820, 2023. 2, 3, 5 +[67] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2, 6 +[68] Chenxi Wang, Yunfeng Wang, Zixuan Huang, and Zhiwen Chen. Simple baseline for single human motion forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2260-2265, 2021. 2 +[69] Jiashun Wang, Huazhe Xu, Medhini Narasimhan, and Xiaolong Wang. Multi-person 3d motion prediction with multirange transformers. Advances in Neural Information Processing Systems, 34:6036-6049, 2021. 1, 2, 6 + +[70] Mingkun Wang, Xinge Zhu, Changqian Yu, Wei Li, Yuexin Ma, Ruochun Jin, Xiaoguang Ren, Dongchun Ren, Mingxu Wang, and Wenjing Yang. Ganet: Goal area network for motion forecasting. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1609-1615. IEEE, 2023. 2 +[71] Yucheng Xing and Xin Wang. Hdg-ode: A hierarchical continuous-time model for human pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14700–14712, 2023. 1, 2 +[72] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6488-6497, 2022. 2 +[73] Chenxin Xu, Robby T. Tan, Yuhong Tan, Siheng Chen, Xin-chao Wang, and Yanfeng Wang. Auxiliary tasks benefit 3d skeleton-based human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9509–9520, 2023. 2 +[74] Qingyao Xu, Weibo Mao, Jingze Gong, Chenxin Xu, Siheng Chen, Weidi Xie, Ya Zhang, and Yanfeng Wang. Joint-relation transformer for multi-person motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9816-9826, 2023. 1, 6 +[75] Sirui Xu, Yu-Xiong Wang, and Liangyan Gui. Stochastic multi-person 3d motion forecasting. In The Eleventh International Conference on Learning Representations, 2023. 1, 2 +[76] Yi Xu, Armin Bazarjani, Hyung-gun Chi, Chiho Choi, and Yun Fu. Uncovering the missing pattern: Unified framework towards trajectory imputation and prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9632-9643, 2023. 2 +[77] Maosheng Ye, Jiamiao Xu, Xunnong Xu, Tengfei Wang, Tongyi Cao, and Qifeng Chen. Bootstrap motion forecasting with self-consistent constraints. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8504-8514, 2023. 2 +[78] Yingxuan You, Hong Liu, Ti Wang, Wenhao Li, Runwei Ding, and Xia Li. Co-evolution of pose and mesh for 3d human body estimation from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14963-14973, 2023. 2 +[79] Bruce X.B. Yu, Zhi Zhang, Yongxu Liu, Sheng-hua Zhong, Yan Liu, and Chang Wen Chen. Gla-gen: Global-local adaptive graph convolutional network for 3d human pose estimation from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8818-8829, 2023. 2 +[80] Kai Zhai, Qiang Nie, Bo Ouyang, Xiang Li, and Shanlin Yang. Hopfir: Hop-wise graphformer with intragroup joint refinement for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14985-14995, 2023. 2 +[81] Yi Zhang, Pengliang Ji, Angtian Wang, Jieru Mei, Adam Kortylewski, and Alan Yuille. 3d-aware neural body fitting for + +occlusion robust 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9399-9410, 2023. 2 +[82] He Zhao and Richard P. Wildes. Where are you heading? dynamic trajectory prediction with expert goal examples. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7629-7638, 2021. 2 +[83] Qitao Zhao, Ce Zheng, Mengyuan Liu, Pichao Wang, and Chen Chen. Poseformerv2: Exploring frequency domain for efficient and robust 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8877-8886, 2023. 2 +[84] Fang Zheng, Le Wang, Sanping Zhou, Wei Tang, Zhenxing Niu, Nanning Zheng, and Gang Hua. Unlimited neighborhood interaction for heterogeneous trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13168-13177, 2021. 2 +[85] Jingxiao Zheng, Xinwei Shi, Alexander Gorban, Junhua Mao, Yang Song, Charles R Qi, Ting Liu, Visesh Chari, Andre Cornman, Yin Zhou, et al. Multi-modal 3d human pose estimation with 2d weak supervision in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4478-4487, 2022. 1 +[86] Chongyang Zhong, Lei Hu, Zihao Zhang, Yongjing Ye, and Shihong Xia. Spatio-temporal gating-adjacency gen for human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6447–6456, 2022. 1 +[87] Mu Zhou, Lucas Stoffl, Mackenzie Weygandt Mathis, and Alexander Mathis. Rethinking pose estimation in crowds: Overcoming the detection information bottleneck and ambiguity. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14689-14699, 2023. 3 +[88] Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Kejie Lu. Hivt: Hierarchical vector transformer for multi-agent motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8823-8833, 2022. 2, 4, 6 +[89] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 2 +[90] Dekai Zhu, Guangyao Zhai, Yan Di, Fabian Manhardt, Hendrik Berkemeyer, Tuan Tran, Nassir Navab, Federico Tombari, and Benjamin Busam. Ipcc-tp: Utilizing incremental pearson correlation coefficient for joint multi-agent trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5507-5516, 2023. 2 +[91] Tao Zhuo, Zhiyong Cheng, Peng Zhang, Yongkang Wong, and Mohan Kankanhalli. Unsupervised online video object segmentation with motion property understanding. IEEE Transactions on Image Processing, 29:237-249, 2019. 1 \ No newline at end of file diff --git a/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/images.zip b/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ae20c6ebcae6968689951032320ab7f96a04c4c0 --- /dev/null +++ b/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee0de1c3bda130bf94bfb45d24234dc4f70b5f2dc8a42ca57daf306e3a61b0be +size 657599 diff --git a/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/layout.json b/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a8f739705b8bb77cfe2b6f5d07ff24069aae4511 --- /dev/null +++ b/2024/Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning/layout.json @@ -0,0 +1,10302 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "type": "text", + "content": "Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 177, + 161, + 416, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 161, + 416, + 187 + ], + "spans": [ + { + "bbox": [ + 177, + 161, + 416, + 187 + ], + "type": "text", + "content": "Jaewoo Jeong*, Daehee Park*, and Kuk-Jin Yoon KAIST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 197, + 191, + 398, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 191, + 398, + 204 + ], + "spans": [ + { + "bbox": [ + 197, + 191, + 398, + 204 + ], + "type": "text", + "content": "{jeong207,bag2824,kjyoon}@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 256, + 290, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 256, + 290, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 290, + 544 + ], + "type": "text", + "content": "Human pose forecasting garners attention for its diverse applications. However, challenges in modeling the multi-modal nature of human motion and intricate interactions among agents persist, particularly with longer timescales and more agents. In this paper, we propose an interaction-aware trajectory-conditioned long-term multi-agent human pose forecasting model, utilizing a coarse-to-fine prediction approach: multi-modal global trajectories are initially forecasted, followed by respective local pose forecasts conditioned on each mode. In doing so, our Trajectory2Pose model introduces a graph-based agent-wise interaction module for a reciprocal forecast of local motion-conditioned global trajectory and trajectory-conditioned local pose. Our model effectively handles the multi-modality of human motion and the complexity of long-term multi-agent interactions, improving performance in complex environments. Furthermore, we address the lack of long-term " + }, + { + "bbox": [ + 47, + 256, + 290, + 544 + ], + "type": "inline_equation", + "content": "(6s+)" + }, + { + "bbox": [ + 47, + 256, + 290, + 544 + ], + "type": "text", + "content": " multi-agent " + }, + { + "bbox": [ + 47, + 256, + 290, + 544 + ], + "type": "inline_equation", + "content": "(5+)" + }, + { + "bbox": [ + 47, + 256, + 290, + 544 + ], + "type": "text", + "content": " datasets by constructing a new dataset from real-world images and 2D annotations, enabling a comprehensive evaluation of our proposed model. State-of-the-art prediction performance on both complex and simpler datasets confirms the generalized effectiveness of our method. The code is available at https://github.com/Jaewoo97/T2P." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 586, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 695 + ], + "type": "text", + "content": "Human pose forecasting aims to predict future human motion based on observed past motion [20, 31, 32, 35, 37, 53, 86]. Humans instinctively perform such tasks, allowing them to naturally navigate in crowded areas or identify and circumvent potential dangers. For this reason, human pose forecasting plays an important role in various computer vision tasks [21, 23, 27, 54, 85, 91]. Indeed, recent years have seen a proliferation of work on multi-agent motion forecasting which aim towards modeling complex multi-agent inter" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 228, + 545, + 289 + ], + "blocks": [ + { + "bbox": [ + 307, + 228, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 289 + ], + "type": "image", + "image_path": "1a55f1d7b76b2c2e415c064861262d024e6c7ec4a7e55e3fbc6e46e53ca2d746.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 338, + 547, + 405 + ], + "lines": [ + { + "bbox": [ + 305, + 338, + 547, + 405 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 547, + 405 + ], + "type": "text", + "content": "Figure 1. Human motion is goal-directed and influenced by other entities. Therefore, global intention contains hints for local intention, allowing us to infer local pose from global trajectories. Our method first forecasts global trajectories, upon which local poses are conditioned for subsequent forecasts. Pose and trajectory-wise inter-agent interactions are considered for both predictions." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 292, + 416, + 334 + ], + "blocks": [ + { + "bbox": [ + 308, + 292, + 416, + 334 + ], + "lines": [ + { + "bbox": [ + 308, + 292, + 416, + 334 + ], + "spans": [ + { + "bbox": [ + 308, + 292, + 416, + 334 + ], + "type": "table", + "html": "
Global Intent
ABC
Walk side-by-side\nAvoid CKneel
", + "image_path": "3f7c3ec175f3756e35f2530be803e0cdb77f6ecfb0c264decb0f66dd2b118f2f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 436, + 292, + 545, + 334 + ], + "blocks": [ + { + "bbox": [ + 436, + 292, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 436, + 292, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 436, + 292, + 545, + 334 + ], + "type": "table", + "html": "
Local Intent
ABC
Walking motion\nTalking gesturesTie shoe
", + "image_path": "6b18546f0257669459a9bba4f8ef1e76bf67c65ca930d9e3126525d9e431a813.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 423, + 416, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 423, + 416, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 423, + 416, + 436 + ], + "type": "text", + "content": "action [20, 47, 53, 71, 74]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "type": "text", + "content": "Although various methods have been proposed, they share two major limitations. The first is a limitation on long-term predictions, as previous studies predicted up to 3 seconds at most [4, 47, 74, 75]. However, a sufficiently long forecast horizon is essential to fully leverage human pose forecasting for diverse downstream tasks in the scope of identifying potential danger or understanding human behavior. The second is that multi-person interactions are not proficiently learned. Existing methods consider the joints of multiple people all at once as objects of interaction [47, 65, 74], resulting in an excessive complexity with respect to the number of joints. Due to such inefficient modeling, these approaches are found to be incompetent in long-term " + }, + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "type": "inline_equation", + "content": "(3s+)" + }, + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "type": "text", + "content": " multi-agent " + }, + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "type": "inline_equation", + "content": "(6+)" + }, + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "type": "text", + "content": " settings, limiting their practicality on complex real-world environments." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "Moreover, these challenges are also due to the limitations of datasets. Existing pose forecasting datasets have limited sequence length (" + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\sim 3\\mathrm{s}" + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": ") and number of agents (" + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\sim 2" + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "). Therefore, previous works [47, 69, 75] have randomly blended disparate datasets to model multi-agent interaction with up to 10 agents. Yet, such naively merged data lacks authentic interaction as agents from different scenes remain uninfluenced. As such, there was no opportunity to develop and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 153, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 153, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 153, + 712 + ], + "type": "text", + "content": "*Denotes equal contribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1617" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 279, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 279, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 279, + 83 + ], + "type": "text", + "content": "evaluate a model in a long-term multi-agent environment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 86, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 86, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 86, + 287, + 373 + ], + "type": "text", + "content": "To this end, we present a solution from both model and dataset perspectives to tackle long-term multi-agent human pose forecasting. First, from a model perspective, we propose an interaction-aware trajectory-conditioned pose forecasting method. We point out that the limitations of existing methods on long-term multi-agent environments lead to poor performance in handling the multi-modal nature of human motion and correspondingly complex interactions. To improve upon handling multi-modality in these complex settings, we use a coarse-to-fine approach to enjoy effective interaction modeling by propagating agent-wise coarse representations. Agent-wise pose and trajectory embeddings are obtained in their respective local coordinates, followed by a holistic interaction modeling via our proposed Traj-pose module. Interaction-aware forecasts are then made by initial coarse global hip joint trajectory forecast followed by fine local pose forecasts in its hip joint coordinates, conditioned on the global trajectory as shown in Fig. 1. As discovered in previous research [2, 54], learning an agent-wise global intention as coarse trajectories is less challenging than predicting every joint-wise motion. We leverage these hints from global trajectories, which are further conditioned towards forecasting local motion that embodies the interaction-aware spatio-temporal context." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 374, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 601 + ], + "type": "text", + "content": "From a dataset perspective, we parsed a novel real-world dataset for long-term multi-agent human pose forecasting. We utilize JRDB dataset [66] which consists of multi-view video and collected in various environments. Since 3D pose annotations are not provided in the original JRDB, we extracted sequences of 3D human pose from visible agents in omnidirectional images using the latest algorithm for 3D pose extraction from image [59]. We then ensure the reliability of 3D pose information by filtering and adjusting the extracted 3D poses based on 2D pose and 3D bounding box annotations. As a result, we construct a real-world 3D human pose forecasting dataset, JRDB-GlobMultiPose (JRDB-GMP), where up to 24 agents exist for up to 5 seconds. The proposed pose forecasting model is validated on both previous datasets and newly created JRDB-GMP dataset. Our method shows state-of-the-art forecasting performance in both global and local accuracy metrics, not only on JRDB-GMP but also on all previous datasets. Therefore, our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 603, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 603, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 603, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 603, + 287, + 639 + ], + "type": "text", + "content": "- We propose a interaction-aware trajectory-conditioned pose forecasting method " + }, + { + "bbox": [ + 47, + 603, + 287, + 639 + ], + "type": "inline_equation", + "content": "(T2P)" + }, + { + "bbox": [ + 47, + 603, + 287, + 639 + ], + "type": "text", + "content": " for long-term multi-agent 3D human pose forecasting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 640, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 640, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 47, + 640, + 287, + 675 + ], + "type": "text", + "content": "- We propose a long-term, multi-agent real-world 3D human pose forecasting dataset which contains up to 24 persons and forecasts up to 5 seconds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "- We validate our " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "T2P" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " model on both previous datasets and our new JRDB-GMP dataset. Our method achieves state-of-the-art forecasting performance on all datasets." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 394, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 394, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 394, + 83 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 88, + 441, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 88, + 441, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 88, + 441, + 102 + ], + "type": "text", + "content": "2.1. Human pose forecasting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 107, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 107, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 107, + 545, + 370 + ], + "type": "text", + "content": "Human Pose Forecasting involves predicting a future pose sequence with temporal length of a prediction horizon, given a historical pose sequence [5, 8, 11, 16, 36, 45, 49, 52, 68, 71]. In the early stage, methods were developed to forecast single person motion within a short timeframe (" + }, + { + "bbox": [ + 304, + 107, + 545, + 370 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 304, + 107, + 545, + 370 + ], + "type": "text", + "content": " 1s) [9, 34, 58, 73]. However, to improve applicability on diverse downstream computer vision tasks, forecasts are to be made on multi-person poses [1, 2, 63] for longer prediction horizons [6, 62]. Forecasting future inherently involves a stochastic nature, and handling such multi-modality has been attempted by forecasting multiple future poses of a single agent [4]. However, comparatively marginal efforts have been employed in the more complex long-term multi-agent scenes [75]. Such absence is mostly due to the lack of a proper dataset. The commonly used evaluation datasets are CMU-Mocap [13], 3DPW [67], UMPM [64], MuPoTS-3D [39], all of which contain 2 agents at most in a given scene and have short prediction horizons within 3 seconds. Most recent research arbitrarily combines individual scenes to create datasets with more than three individuals [47, 69, 75]. However, such a synthetic approach does not account for authentic agent interactions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 380, + 429, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 429, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 429, + 392 + ], + "type": "text", + "content": "2.2. Trajectory prediction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 398, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 545, + 613 + ], + "type": "text", + "content": "Trajectory prediction involves predicting the future path of an object given its past trajectory [3, 10, 25, 38, 40, 41, 51, 76, 77, 89, 90]. Unlike human pose forecasting which aims to predict every joint position, trajectory prediction regards each agent as a point mass, typically the center of mass or center point of a detected bounding box. Research in trajectory forecasting is interested in not only vehicles but also many types of agents including humans, cyclists, and so on [29, 43, 72, 84]. One substantial direction in research within this field is the Goal-conditioned prediction approach [18, 33, 82]. Goal-conditioned prediction approach first predicts the final destination within the prediction horizon with multiple goal proposals [28, 70]. Then, a thorough future path is conditioned on each mode of the multi-modal proposals. Compared to directly predicting full trajectories, the goal-conditioned approach follows a coarse-to-fine prediction and is effective in learning highly stochastic multi-modality of complex scenes [19, 42, 88]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 623, + 495, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 495, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 495, + 635 + ], + "type": "text", + "content": "2.3. Human pose estimation from image" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Human pose estimation is the task of inferring the pose of a person from an image or a video [22, 24, 26, 30, 57, 60, 78, 80, 81]. Initial deep learning-based methods first utilized convolutional neural networks to estimate 2D and 3D poses from single or multiple images [12, 14, 61, 79, 83]. Recent approaches engage in more challenging tasks such as" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1618" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 547, + 270 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 547, + 270 + ], + "type": "image", + "image_path": "1341b7a715ef1637d68c209a56efc4686d3ca20c761d52245beea7e57a33c25d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 277, + 547, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 547, + 323 + ], + "type": "text", + "content": "Figure 2. Illustration of our T2P framework. We decompose global motion into global trajectory and local pose. Multi-modal global trajectory proposals are predicted from past global trajectory and local pose embeddings. Then, future local poses are conditioned and forecasted on each trajectory proposal to compose the final human pose prediction. Predicted local poses are added to their mode-specific global trajectories in a joint-wise manner, obtaining the global human poses as the final output." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 334, + 289, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 289, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 289, + 490 + ], + "type": "text", + "content": "estimating 3D poses from monocular videos [7, 46, 48, 50, 55, 56] using self-supervised learning and generative methods [15, 17]. Most recent methods estimate multi-person poses in a crowded environment with considerable occlusions [44, 87]. We account for the aforementioned need of a complex dataset by extracting 3D pose from images using these methods. Specifically, we use a monocular 3D pose estimation method " + }, + { + "bbox": [ + 46, + 334, + 289, + 490 + ], + "type": "inline_equation", + "content": "BEV" + }, + { + "bbox": [ + 46, + 334, + 289, + 490 + ], + "type": "text", + "content": " [59] to construct a 3D human motion forecasting dataset with long-term multi-agent characteristics from real-world image sequences. " + }, + { + "bbox": [ + 46, + 334, + 289, + 490 + ], + "type": "inline_equation", + "content": "BEV" + }, + { + "bbox": [ + 46, + 334, + 289, + 490 + ], + "type": "text", + "content": " robustly estimates human pose in a scale-ambiguous and crowded environment, reliably extracting 3D poses from the omnidirectional image sequences of JRDB dataset [66]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 505, + 104, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 104, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 104, + 517 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 526, + 157, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 157, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 157, + 538 + ], + "type": "text", + "content": "3.1. Problem definition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": "Multi-agent human pose forecasting aims to learn a mapping function between the observed 3D pose of " + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "inline_equation", + "content": "N_{A}" + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": " agents composed of " + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": " joints, " + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}:\\left\\{\\mathbf{x}_t^{n,j}\\right\\}_{-T_p:0}^{N_A,J}" + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": ", and future pose " + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}:F\\times \\left\\{\\mathbf{x}_t^{n,j}\\right\\}_{0:T_f}^{N_A,J}" + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": " in global coordinates where " + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": " denotes the number of modes. Here, " + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "inline_equation", + "content": "T_{f}" + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": " are history length and prediction horizon while " + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^{n,j} = (x_t^{n,j},y_t^{n,j},z_t^{n,j})" + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": " is 3D global coordinate of joint " + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": " of agent " + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": ". While the global position of joint is represented in " + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": ", we additionally define local position " + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": ". The local position is defined in local coordinate of each agent, calculated by subtracting the global position of the hip joint of each agent. Therefore, local position of joint is defined as " + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{n,j} = \\mathbf{x}^{n,j} - \\mathbf{x}^{n,\\mathrm{hip}}" + }, + { + "bbox": [ + 46, + 547, + 289, + 714 + ], + "type": "text", + "content": ". We" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "text", + "content": "define the trajectory of global hip joint position as global trajectory, " + }, + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "inline_equation", + "content": "Tr:\\{\\mathbf{x}^{n,\\mathrm{hip}}\\}^{N_A}" + }, + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "text", + "content": ". We also define local pose as local position of all joints, " + }, + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "inline_equation", + "content": "Po:\\{\\mathbf{p}^{n,\\mathrm{j}}\\}^{N_A,J}" + }, + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "text", + "content": ". We denote past and future timesteps of global trajectory and local motion as " + }, + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "inline_equation", + "content": "Tr_{\\mathcal{P}},Tr_{\\mathcal{F}}\\in Tr" + }, + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "inline_equation", + "content": "Po_{\\mathcal{P}},Po_{\\mathcal{F}}\\in Po" + }, + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "text", + "content": " respectively denotes past and future." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 420, + 419, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 420, + 419, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 419, + 431 + ], + "type": "text", + "content": "3.2. Overall framework" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "type": "text", + "content": "We disentangle the overall human motion into global trajectories and local poses, as depicted in top left of Fig. 2. Following a coarse-to-fine strategy, multiple global trajectories are first forecasted to model the coarse modes of global intentions. Based on these forecasts, local pose predictions are conditioned on each mode to jointly constitute a thorough motion. In doing so, our model is widely divided into two portions: Trajectory predictor consists of trajectory encoder and decoder and pose predictor consists of pose encoder and decoder. Both predictors engage in the reciprocal exchange of both trajectory and pose information, facilitating the inference of cues between global and local motion. The detailed methods of each stage are described below:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 603, + 405, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 603, + 405, + 615 + ], + "spans": [ + { + "bbox": [ + 306, + 603, + 405, + 615 + ], + "type": "text", + "content": "3.3. Model structure" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 622, + 394, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 394, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 394, + 632 + ], + "type": "text", + "content": "3.3.1 Pose encoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": "Unlike the holistic approach of previous works that encode and decode all agents' joint motions in global coordinates, our pose encoder encodes the pose dynamics in local coordinates. In addition, our pose encoder only considers intra-agent joint interaction. As a result, the encoded pose embedding represents agent-specific local mo" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1619" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "tion, containing insights on global intent. We follow our baseline [47] and construct the encoder with Multi-Person Body-Part (MPBP) module and transformer networks. As depicted in Fig. 2, body part sequences are constructed in frequency domain, followed by intra-agent attention-based encoding of the body parts to acquire pose embedding " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "Z_{Po}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 163, + 160, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 160, + 175 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 160, + 175 + ], + "type": "text", + "content": "3.3.2 Trajectory module" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "type": "text", + "content": "Trajectory module aims to extract embeddings from the agents' past global trajectory. Using an encoder structure from [88], multi-agent interaction-based trajectory embedding " + }, + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "type": "inline_equation", + "content": "Z_{Tr}" + }, + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "type": "text", + "content": " is extracted which contains rudimentary insight on global intent. Interaction between agent trajectories is represented based on the reference agent " + }, + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "type": "text", + "content": "'s global trajectory segment vector " + }, + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t^i = \\mathbf{x}_t^{i,\\mathrm{hip}} - \\mathbf{x}_{t-1}^{i,\\mathrm{hip}}" + }, + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "type": "text", + "content": ". For rotational invariance, neighbor actor " + }, + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "type": "text", + "content": "'s vector is normalized by the reference vector's orientation at latest timestep " + }, + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "type": "inline_equation", + "content": "t=0" + }, + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "type": "text", + "content": ". Separate MLP layers then compute the reference agent and neighboring agent embeddings " + }, + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "type": "inline_equation", + "content": "z_{Tr_i}^t, z_{Tr_j}^t" + }, + { + "bbox": [ + 46, + 183, + 289, + 317 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 328, + 286, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 328, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 102, + 328, + 286, + 361 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z _ {T r _ {i}} ^ {t} = \\phi_ {r e f} (R _ {i} ^ {T} \\mathbf {v} _ {t} ^ {i}) \\\\ z _ {T r _ {j}} ^ {t} = \\phi_ {n b r} ([ R _ {i} ^ {T} (\\mathbf {v} _ {t} ^ {j}), R _ {i} ^ {T} (\\mathbf {v} _ {t} ^ {i}) ]) \\tag {1} \\\\ \\end{array}", + "image_path": "146fcbe4388c3bf40df0920cb145a775ad65620171582b4c270b54f825fd07bb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\phi_{ref}" + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\phi_{nbr}" + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "text", + "content": " are different MLP blocks, " + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "inline_equation", + "content": "R_i \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "text", + "content": " is the rotation matrix of agent " + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "text", + "content": " against agent " + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "text", + "content": " is concatenation. The resulting agent-specific reference and neighbor embeddings constitute trajectory embedding " + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "inline_equation", + "content": "z_{Tr}^t" + }, + { + "bbox": [ + 47, + 372, + 287, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 439, + 155, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 439, + 155, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 155, + 451 + ], + "type": "text", + "content": "3.3.3 Traj-posed module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 460, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 579 + ], + "type": "text", + "content": "Human maneuver contains various dynamic activities characterized by the agent's multi-modal intents. Auxiliary human motion such as arm gesture, rotational orientation of upper body and head implies the agent's intent in global motion. In that sense, harvesting meaningful insights from past local joint motion helps proficient modeling of coarse multimodality as future trajectory proposals. Therefore, we propose Traj-Pose Module that fuses agent-wise embeddings of both trajectory and pose to fully utilize these information in modeling global intentions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "text", + "content": "First, MLP is used to match the temporal domain of pose embedding " + }, + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "inline_equation", + "content": "Z_{Po}" + }, + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "text", + "content": " to that of " + }, + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "inline_equation", + "content": "Z_{Tr}" + }, + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "text", + "content": ", after which both are concatenated as agent-wise traj-posed embedding " + }, + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 627, + 287, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 627, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 116, + 627, + 287, + 640 + ], + "type": "interline_equation", + "content": "Z = \\left[ Z _ {T r}, \\phi_ {M L P} \\left(Z _ {P o}\\right) \\right] \\tag {2}", + "image_path": "79f92449cad9ffb966203eb5eacaa4dc55e215836104a03003c37d4b250d0d37.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 651, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 287, + 715 + ], + "type": "text", + "content": "The resulting " + }, + { + "bbox": [ + 47, + 651, + 287, + 715 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 47, + 651, + 287, + 715 + ], + "type": "text", + "content": " is comprised of agent and timestep-respective trajectory and pose embeddings: " + }, + { + "bbox": [ + 47, + 651, + 287, + 715 + ], + "type": "inline_equation", + "content": "z_{i}^{t},z_{j}^{t}\\in z^{t}\\in Z" + }, + { + "bbox": [ + 47, + 651, + 287, + 715 + ], + "type": "text", + "content": " Then, " + }, + { + "bbox": [ + 47, + 651, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\widetilde{Z}" + }, + { + "bbox": [ + 47, + 651, + 287, + 715 + ], + "type": "text", + "content": " is acquired from the graph attention with an agent-wise update where each agent embedding " + }, + { + "bbox": [ + 47, + 651, + 287, + 715 + ], + "type": "inline_equation", + "content": "z_{i}^{t}" + }, + { + "bbox": [ + 47, + 651, + 287, + 715 + ], + "type": "text", + "content": " and its neighbor embedding " + }, + { + "bbox": [ + 47, + 651, + 287, + 715 + ], + "type": "inline_equation", + "content": "z_{j}^{t}" + }, + { + "bbox": [ + 47, + 651, + 287, + 715 + ], + "type": "text", + "content": " are used as query and key/value. Similar" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "to trajectory interaction encoder of HiVT [88], the graph attention operation is operated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 357, + 104, + 545, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 104, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 357, + 104, + 545, + 159 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\alpha_ {i} ^ {t} = \\operatorname {s o f t m a x} \\left(\\frac {q _ {i} ^ {t \\top}}{\\sqrt {d _ {k}}} \\cdot \\left[ \\left\\{k _ {j} ^ {t} \\right\\} _ {j \\in N _ {i}} \\right]\\right), \\\\ m _ {i} ^ {t} = \\sum_ {j \\in N _ {i}} \\alpha_ {i} ^ {t} v _ {j} ^ {t}, \\tag {3} \\\\ \\end{array}", + "image_path": "ab59ccfd753f3e85928e95c7c834e09f983a72109ac7e198a253103c33edbc9b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 378, + 162, + 498, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 162, + 498, + 175 + ], + "spans": [ + { + "bbox": [ + 378, + 162, + 498, + 175 + ], + "type": "interline_equation", + "content": "g _ {i} ^ {t} = \\operatorname {s i g m o i d} \\left(W ^ {\\text {g a t e}} \\left[ z _ {i} ^ {t}, m _ {i} ^ {t} \\right]\\right),", + "image_path": "ed3072eb4c6ec71533fb82c4a1df4a9142f05d2c677e7ae491829cc2cad84595.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 177, + 499, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 177, + 499, + 192 + ], + "spans": [ + { + "bbox": [ + 353, + 177, + 499, + 192 + ], + "type": "interline_equation", + "content": "\\widetilde {z} _ {i} ^ {t} = g _ {i} ^ {t} \\odot W ^ {\\text {s e l f}} z _ {i} ^ {t} + \\left(1 - g _ {i} ^ {t}\\right) \\odot m _ {i} ^ {t}", + "image_path": "6743e509133c2824a1589c05d0a0630e1c0f84ca60caaf050a6e8c4445d2da84.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 201, + 545, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 201, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 545, + 226 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 201, + 545, + 226 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 306, + 201, + 545, + 226 + ], + "type": "text", + "content": " is a set of agent " + }, + { + "bbox": [ + 306, + 201, + 545, + 226 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 306, + 201, + 545, + 226 + ], + "type": "text", + "content": "'s neighbors, " + }, + { + "bbox": [ + 306, + 201, + 545, + 226 + ], + "type": "inline_equation", + "content": "W^{gate}" + }, + { + "bbox": [ + 306, + 201, + 545, + 226 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 201, + 545, + 226 + ], + "type": "inline_equation", + "content": "W^{self}" + }, + { + "bbox": [ + 306, + 201, + 545, + 226 + ], + "type": "text", + "content": " are learnable matrices, and " + }, + { + "bbox": [ + 306, + 201, + 545, + 226 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 306, + 201, + 545, + 226 + ], + "type": "text", + "content": " is element-wise product." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 239, + 421, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 421, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 421, + 251 + ], + "type": "text", + "content": "3.3.4 Trajectory decoder" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 258, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 450 + ], + "type": "text", + "content": "Trajectory is forecasted from the output of trajectory encoder which encodes both past global trajectory and local pose information. Since its graph operation is operated by each timestep, a temporal encoder is used as a temporal encoder to integrate " + }, + { + "bbox": [ + 305, + 258, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\widehat{Z}" + }, + { + "bbox": [ + 305, + 258, + 545, + 450 + ], + "type": "text", + "content": " in the temporal dimension. A multihead self-attention temporal encoder is used as the temporal encoder. Aggregator then takes into account variations in local coordinate frames to accurately represent geometric relationships within the global coordinate system via a graph operation. MLP is subsequently applied to span embedding " + }, + { + "bbox": [ + 305, + 258, + 545, + 450 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 258, + 545, + 450 + ], + "type": "text", + "content": " times for multi-modal prediction, which is residually added to the " + }, + { + "bbox": [ + 305, + 258, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\times F" + }, + { + "bbox": [ + 305, + 258, + 545, + 450 + ], + "type": "text", + "content": " repeated embedding before the aggregator. Finally, another MLP is used to extract multi-modal future global trajectory proposals of hip joint " + }, + { + "bbox": [ + 305, + 258, + 545, + 450 + ], + "type": "inline_equation", + "content": "Tr_{\\mathcal{F}} \\in \\mathbb{R}^{F \\times T_f \\times 3}" + }, + { + "bbox": [ + 305, + 258, + 545, + 450 + ], + "type": "text", + "content": ". The multi-modal embedding is also passed onto the Pose Decoder to forecast local poses." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 464, + 394, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 394, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 394, + 475 + ], + "type": "text", + "content": "3.3.5 Pose decoder" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "type": "text", + "content": "Future human pose depends on past human poses and global intention. The pose decoder is designed to consider these factors while generating local poses via mode-specific trajectory conditioning. A transformer (TRM) decoder is used to decode local motions, where pose embedding " + }, + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "type": "inline_equation", + "content": "Z_{Po}" + }, + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "type": "text", + "content": " is used as key/value and concatenation of trajectory and pose query." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 329, + 566, + 545, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 566, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 329, + 566, + 545, + 579 + ], + "type": "interline_equation", + "content": "Q = \\phi_ {M L P} ([ Q _ {P o}, Q _ {T r} ]), K = Z _ {P o}, V = Z _ {P o} \\tag {4}", + "image_path": "8aa607d4ef94a8aa69a08f2d125fee9f04e43078b2a758d23dd559c9114b46d7.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 584, + 545, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 584, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 584, + 545, + 690 + ], + "type": "text", + "content": "Using both pose and trajectory queries, past pose embedding " + }, + { + "bbox": [ + 304, + 584, + 545, + 690 + ], + "type": "inline_equation", + "content": "Z_{Po}" + }, + { + "bbox": [ + 304, + 584, + 545, + 690 + ], + "type": "text", + "content": " is conditioned on both MPBP sequence at " + }, + { + "bbox": [ + 304, + 584, + 545, + 690 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 304, + 584, + 545, + 690 + ], + "type": "text", + "content": " and the multi-modal trajectory proposals which contain global intent. Subsequently, inverse discrete cosine transform (idct) is applied to convert the future pose proposals from frequency domain to local coordinate domain, " + }, + { + "bbox": [ + 304, + 584, + 545, + 690 + ], + "type": "inline_equation", + "content": "P_{o_{\\mathcal{F}}}" + }, + { + "bbox": [ + 304, + 584, + 545, + 690 + ], + "type": "text", + "content": ". The final multi-modal future pose in global coordinates is acquired as Eq. 5 where " + }, + { + "bbox": [ + 304, + 584, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 584, + 545, + 690 + ], + "type": "text", + "content": " is a joint-wise addition operation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 342, + 700, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 342, + 700, + 545, + 714 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} = T r _ {\\mathcal {F}} \\oplus P o _ {\\mathcal {F}}, \\quad \\mathbf {Y} \\in \\mathbb {R} ^ {F \\times N _ {A} \\times T _ {f} \\times 3} \\tag {5}", + "image_path": "62d5faf7bd1637ad2b2da47a4e7b76a269d479019055a270a46d2375a0961365.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1620" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 96, + 287, + 217 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "type": "text", + "content": "Table 1. Comparison of statistics between existing human pose forecasting datasets and newly proposed JRDB-GMP dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 96, + 287, + 217 + ], + "lines": [ + { + "bbox": [ + 50, + 96, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 50, + 96, + 287, + 217 + ], + "type": "table", + "html": "
Dataset
CMU-Mocap (UMPM)MuPoTs -3D3DPWJRDB-GMP 1s/2s2s/5s
Duration (s)400026717001863
Location #-20-27
Sample #1300019243211534593
avg. agent3326.86.8
med. agent #33255
max agent #3322422
avg. vel. (m/s)0.30.260.570.460.38
avg. disp.(m)0.630.551.130.640.79
max. disp.(m)4.622.4510.718.4411.0
", + "image_path": "02c21887687e40ab21d1c50bdc7c2f9e7df8636a629ffde9346a16fdf0453a20.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 224, + 296, + 365 + ], + "blocks": [ + { + "bbox": [ + 51, + 224, + 296, + 365 + ], + "lines": [ + { + "bbox": [ + 51, + 224, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 51, + 224, + 296, + 365 + ], + "type": "image", + "image_path": "fbebee859bddfbb4d3d177925974b79304b1a13b0ffdfc0cf8b14adb1eab21cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 367, + 287, + 388 + ], + "lines": [ + { + "bbox": [ + 46, + 367, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 287, + 388 + ], + "type": "text", + "content": "Figure 3. Example scenes from the JRDB-GMP dataset, illustrating its long-term, multi-agent nature." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 399, + 155, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 155, + 412 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 155, + 412 + ], + "type": "text", + "content": "3.4. Training objective" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 418, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 466 + ], + "type": "text", + "content": "Both objectives of global trajectory and local pose forecasting are trained jointly. For both global trajectory and local pose prediction, " + }, + { + "bbox": [ + 46, + 418, + 287, + 466 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 46, + 418, + 287, + 466 + ], + "type": "text", + "content": " loss is propagated to the mode with minimal " + }, + { + "bbox": [ + 46, + 418, + 287, + 466 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 46, + 418, + 287, + 466 + ], + "type": "text", + "content": " distance with the ground truth." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 474, + 245, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 474, + 245, + 507 + ], + "spans": [ + { + "bbox": [ + 115, + 474, + 245, + 507 + ], + "type": "interline_equation", + "content": "L _ {T r} = \\sum_ {n = 1} ^ {N _ {A}} \\sum_ {t = 1} ^ {T _ {f}} \\| \\widetilde {y} _ {T r, n} ^ {t} - \\hat {y} _ {T r, n} ^ {t} \\|", + "image_path": "d396614f0f9265270cdfd1a8c3ad7f581088635f449c3393e646e28d770d6bd9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 90, + 510, + 287, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 510, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 90, + 510, + 287, + 544 + ], + "type": "interline_equation", + "content": "L _ {P o} = \\sum_ {n = 1} ^ {N _ {A}} \\sum_ {t = 1} ^ {T _ {f}} \\sum_ {j = 1} ^ {J - 1} \\| \\left(\\widetilde {y} _ {P o, n} ^ {t, j} - \\hat {y} _ {P o, n} ^ {t, j}\\right) \\| \\tag {6}", + "image_path": "177d009da6ae6492c921bab6a75903578bd573dde0b935b0b67aeeeee1c12b3f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 177, + 548, + 245, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 548, + 245, + 559 + ], + "spans": [ + { + "bbox": [ + 177, + 548, + 245, + 559 + ], + "type": "interline_equation", + "content": "L = L _ {T r} + L _ {P o}", + "image_path": "073e9f27959350006c1c8e14baf93a87b145bf291482d5a3947ccce305d2171b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 574, + 164, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 164, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 164, + 586 + ], + "type": "text", + "content": "3.5.JRDB-GMP dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Due to the absence of existing long-term " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "(3s+)" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " multi-agent " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "(6+)" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " dataset, we compose a unique 3D human pose forecasting dataset in a real-world environment from JRDB [66]. The original JRDB dataset is constructed by a moving robot that records human activity around a school campus using 5 omnidirectional cameras and LiDAR. Image sequences along with 2D pose annotation and 3D bounding box annotations are provided in the original dataset. However, since 3D human pose annotations are unavailable, we separately parse accurate 3D human pose from provided inputs and" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 68, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 310, + 68, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 310, + 68, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 310, + 68, + 553, + 196 + ], + "type": "image", + "image_path": "507fc764ff5e37ab86f96fc8431046ff42ffa209d9888a29300d9bf19aa962c2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 198, + 553, + 313 + ], + "blocks": [ + { + "bbox": [ + 310, + 198, + 553, + 313 + ], + "lines": [ + { + "bbox": [ + 310, + 198, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 310, + 198, + 553, + 313 + ], + "type": "image", + "image_path": "a7dce04996df690df884af1d9a2c03334fe254bd55eadb8f4382bf933af9807b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 323, + 545, + 345 + ], + "lines": [ + { + "bbox": [ + 305, + 323, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 545, + 345 + ], + "type": "text", + "content": "Figure 4. Various motions from the JRDB-MultiGlobPose dataset, providing rich motion queues for inter-agent interaction inference." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "content": "annotations. First, a SOTA monocular 3D pose extraction method [59] is used to extract raw 3D joint positions from image sequences. Then, 2D pose and 3D bounding box annotations are used to refine the raw joint positions and minimize noise. We use 2D pose annotations to initially filter out the 3D poses with noise. With camera parameters and refined 3D pose, we project it on 2D image plane, then L2 distance between projected 2D pose and GT 2D pose annotation is calculated. If the mean L2 distance per each agent at a time stamp is over a threshold, that instance is filtered out. 2D pose annotations are also projected in 3D space to refine the remaining 3D poses, ensuring the accuracy of the 3D pose information of our JRDB-GMP dataset. Further details are elaborated in the supplementary materials." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": "Figure 3 visualizes some scenes of the constructed dataset. Accurate extraction of 3D poses has been made even with considerable occlusion via the use of 2D poses. The dataset includes agents with both long and short traverse distances and rich inter-agent interactions in both trajectory and local pose aspects. Figure 4 illustrates diverse local poses included in the dataset, which serve as motion cues of inter-agent interaction. Both figures confirm our method's accuracy in extracting 3D multi-human pose, even in crowded environments. Table 1 scrutinizes the statistics compared to previously used datasets. Compared to earlier datasets, the average number of agents is more than twice as high. In addition, comparing JRDB-GMP 1s/2s to CMU-Mocap and MuPoTs datasets, JRDB contains more diverse and longer motion as shown by a similar magnitude of average displacement but longer maximum displacement." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1621" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 93, + 498, + 271 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 92 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of our method to previous methods on CMU-mocap (UMPM), 3DPW, and JRDB-GlobMultiPose datasets with number of prediction modes " + }, + { + "bbox": [ + 47, + 70, + 545, + 92 + ], + "type": "inline_equation", + "content": "(F)" + }, + { + "bbox": [ + 47, + 70, + 545, + 92 + ], + "type": "text", + "content": " as 6. Lower is better for all metrics. The best results are marked in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 93, + 498, + 271 + ], + "lines": [ + { + "bbox": [ + 94, + 93, + 498, + 271 + ], + "spans": [ + { + "bbox": [ + 94, + 93, + 498, + 271 + ], + "type": "table", + "html": "
DatasetCMU-mocap (UMPM)3DPWJRDB-GlobMultiPose
In/out length (s)1/20.8/1.61/22/5
Evaluation time (s)120.81.6122.55
JPEMRT [69]164.7280.1159.1251.2259.3349.3438.4474.0
JRT [74]168.5316.9181.9287.3237.9373.1351.9538.8
TBIFormer [47]170.0290.9153.9265.8257.1339.3443.2481.3
Ours152.4262.7142.6236.2224.0301.4341.6390.4
APEMRT [69]127.0164.4117.9153.272.387.388.5101.9
JRT [74]121.2181.6133.4178.0112.6154.396.7120.2
TBIFormer [47]125.1160.8115.4152.770.683.388.2102.9
Ours114.4151.7114.6150.070.883.382.294.7
FDEMRT [69]99.6204.7102.7185.3235.2325.2418.2454.8
JRT [74]117.7250.8133.7235.4211.4337.4318.5497.2
TBIFormer [47]112.1228.5106.7215.9232.4314.6423.9458.8
Ours88.7188.974.1158.2194.7271.5313.9361.0
", + "image_path": "64422717c5f9b02cc9d8c810df24b468e44699a5df6f66ab64673ae7626420d7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 278, + 123, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 123, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 123, + 293 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 299, + 105, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 105, + 311 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 105, + 311 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 319, + 287, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 438 + ], + "type": "text", + "content": "We test our model on three datasets: CMU-Mocap (UMPM) [13, 64], 3DPW [67], and our JRDB-GMP. Although our model is designed to forecast human poses in a long-term multi-agent environment, we also report experimental results on previous benchmark datasets with simpler scenes. Mocap-UMPM is a mixed dataset of Mocap and UMPM containing synthesized human interaction between three agents [47]. 3DPW is a dataset with 2 agents traversing a real-world environment. We report the test results on each after separate training on respective datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 449, + 106, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 106, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 106, + 460 + ], + "type": "text", + "content": "4.2. Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 468, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 491 + ], + "type": "text", + "content": "We use the following widely-used metrics. For a detailed definition, please refer to the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 492, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 287, + 539 + ], + "type": "text", + "content": "APE: Aligned mean per joint Position Error is used as a metric to evaluate the forecasted local motion. " + }, + { + "bbox": [ + 46, + 492, + 287, + 539 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 46, + 492, + 287, + 539 + ], + "type": "text", + "content": " distance of each joint in the hip joint coordinate is averaged over all joints for a given timestep." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 540, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 563 + ], + "type": "text", + "content": "FDE: Final Distance Error evaluates the forecasted global trajectory by calculating the " + }, + { + "bbox": [ + 46, + 540, + 287, + 563 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 46, + 540, + 287, + 563 + ], + "type": "text", + "content": " distance of a given timestep." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 563, + 286, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 286, + 588 + ], + "type": "text", + "content": "JPE: Joint Precision Error evaluates both global and local predictions by mean " + }, + { + "bbox": [ + 46, + 563, + 286, + 588 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 46, + 563, + 286, + 588 + ], + "type": "text", + "content": " distance of all joints for a timestep." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 598, + 178, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 178, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 178, + 611 + ], + "type": "text", + "content": "4.3. Implementation details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "We train our model on a single A6000 GPU. 2 layers of pose encoder transformer are stacked, followed by 2 layers of transformer in pose decoder. Embedding dimensions of 96 and 128 are used for trajectory and pose embeddings, respectively. The transformed key, value dimension of 64 is used for all transformer architectures. A learning rate of 0.003 is used with an AdamW optimizer with weight decay. Further details can be found in the supplementary materials." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 300, + 544, + 366 + ], + "blocks": [ + { + "bbox": [ + 305, + 277, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 305, + 277, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 545, + 300 + ], + "type": "text", + "content": "Table 3. Short-term prediction results on CMU-Mocap (UMPM) dataset, where 1s of poses are forecasted given 2s of poses." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 300, + 544, + 366 + ], + "lines": [ + { + "bbox": [ + 309, + 300, + 544, + 366 + ], + "spans": [ + { + "bbox": [ + 309, + 300, + 544, + 366 + ], + "type": "table", + "html": "
MetricJPEAPEFDE
Time (s)0.20.61.00.20.61.00.20.61.0
MRT64.515221749.811014039.497.9153
JRT31.510417328.785.912517.763.9120
TBIFormer37.410415832.885.811923.363.7104
Ours37.810215833.884.411614.949.192.6
", + "image_path": "ec862d630b93d64b62f7b919c299579825ae63720035418019c972b381231aa1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 371, + 372, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 372, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 372, + 384 + ], + "type": "text", + "content": "4.4. Baselines" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 390, + 545, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 558 + ], + "type": "text", + "content": "We compare our method against the latest SOTA methods for multi-agent pose forecasting [47, 69, 74]. To compare the multi-modal predictions of these three methods, we extend their prediction modes by spanning embedding " + }, + { + "bbox": [ + 304, + 390, + 545, + 558 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 390, + 545, + 558 + ], + "type": "text", + "content": " times in the same manner as ours. All baselines are trained and evaluated on CMU-mocap (UMPM) and 3DPW datasets. CMU-Mocap (UMPM) dataset predicts 2 seconds from 1 second of poses, and 3DPW predicts 1.6 seconds from 0.8 seconds of poses, both from 6 modes. For 3DPW dataset, we slightly lengthen the forecast horizon to evaluate long term predictions. For JRDB-GMP dataset, both short (1s/2s) and long term (2s/5s) predictions are evaluated for all models. Lastly, We use HiVT [88] as the baseline For global trajectory prediction of our model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 567, + 359, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 359, + 580 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 359, + 580 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 587, + 421, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 421, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 421, + 600 + ], + "type": "text", + "content": "5.1.Quantitative results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Table 2 compares the quantitative performances on three datasets. Our method exhibits considerable performance gain against all previous SOTA methods, not only on the proposed long-term multi-agent dataset but also on the existing two datasets. Such generalized competence demonstrates the applicability of our trajectory-conditioned pose forecasting method to various real-world scenarios. In detail, our approach achieves over " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " gain of FDE on all datasets. This improvement on forecasting global locomo" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1622" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 169, + 180 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 169, + 180 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 169, + 180 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 169, + 180 + ], + "type": "image", + "image_path": "10bbec035091361641fd3dfb6ac0716dfbc5352256d96f352c218c3ce2aae723.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 71, + 188, + 165 + ], + "lines": [ + { + "bbox": [ + 171, + 71, + 188, + 165 + ], + "spans": [ + { + "bbox": [ + 171, + 71, + 188, + 165 + ], + "type": "text", + "content": "t \n2.5s \n5s" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 190, + 71, + 261, + 182 + ], + "blocks": [ + { + "bbox": [ + 190, + 71, + 261, + 182 + ], + "lines": [ + { + "bbox": [ + 190, + 71, + 261, + 182 + ], + "spans": [ + { + "bbox": [ + 190, + 71, + 261, + 182 + ], + "type": "image", + "image_path": "4a0d733b417b0691157108f7eda42efe5ca5e599fe5f92e2043110e954c4d6a3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 185, + 546, + 207 + ], + "lines": [ + { + "bbox": [ + 46, + 185, + 546, + 207 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 546, + 207 + ], + "type": "text", + "content": "Figure 5. Visualization of a long-term forecasting scene from JRDB-GMP (2/5) dataset. Past poses for input are shown on the leftmost column, GT future poses on the next, and forecasts by ours, MRT, TBIFormer, and JRT, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 263, + 72, + 333, + 182 + ], + "blocks": [ + { + "bbox": [ + 263, + 72, + 333, + 182 + ], + "lines": [ + { + "bbox": [ + 263, + 72, + 333, + 182 + ], + "spans": [ + { + "bbox": [ + 263, + 72, + 333, + 182 + ], + "type": "image", + "image_path": "f2a9882d1651207055f03c33321a9385ce57ff85ec9b2489d2431b9c9a4cf832.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 334, + 72, + 393, + 182 + ], + "blocks": [ + { + "bbox": [ + 334, + 72, + 393, + 182 + ], + "lines": [ + { + "bbox": [ + 334, + 72, + 393, + 182 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 393, + 182 + ], + "type": "image", + "image_path": "b44c99c37110e82209b08963524d194ca4c4042e416e7bbf7e289a55409aec50.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 394, + 72, + 469, + 182 + ], + "blocks": [ + { + "bbox": [ + 394, + 72, + 469, + 182 + ], + "lines": [ + { + "bbox": [ + 394, + 72, + 469, + 182 + ], + "spans": [ + { + "bbox": [ + 394, + 72, + 469, + 182 + ], + "type": "image", + "image_path": "1a6003607be61c44db3936bb5304853d7277c6be956e5bea3528142594565f3f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 471, + 72, + 544, + 182 + ], + "blocks": [ + { + "bbox": [ + 471, + 72, + 544, + 182 + ], + "lines": [ + { + "bbox": [ + 471, + 72, + 544, + 182 + ], + "spans": [ + { + "bbox": [ + 471, + 72, + 544, + 182 + ], + "type": "image", + "image_path": "6148c95f486c1457ca9e5c06749adb80bd4715c870cbfa4143d5448d43d8daef.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 211, + 286, + 381 + ], + "blocks": [ + { + "bbox": [ + 48, + 211, + 286, + 381 + ], + "lines": [ + { + "bbox": [ + 48, + 211, + 286, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 211, + 286, + 381 + ], + "type": "image", + "image_path": "7f23d1a7548a187491fa3133a38a67f1db919411589fe758e0fcd2a8693da24d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 385, + 288, + 452 + ], + "lines": [ + { + "bbox": [ + 46, + 385, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 288, + 452 + ], + "type": "text", + "content": "Figure 6. Visualization of a CMU-Mocap (UMPM) scene. Past poses are shown on the upper row, GT future poses on the next, and forecasts by TBIFormer and ours on the latter two rows. To visualize motion, we stack several frames around the target time stamp. Black/red/blue arrows refer to the direction of the global trajectory, and yellow arrows refer to the direction of foot motion." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 460, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 581 + ], + "type": "text", + "content": "tion could be accredited to the decoupled forecasting of global trajectory and local pose. Previous methods holistically predict both global and local movements, limiting both performances due to superfluous interactions to consider between all joints. Conversely, our approach can extract accurate global intent by decoupling past motion into global and local representations. Moreover, our effective interaction modeling of global and local pose also helps to predict a more accurate global trajectory under multi-agent environment as shown in the latter ablation study." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 581, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 700 + ], + "type": "text", + "content": "For APE metric, our method also surpasses previous SOTA models on all datasets, highlighting the accurate extraction of local pose intent. Such improvement shows that our approach generates plausible local motion due to its proficient sampling from coarse global intents. Our method simplifies the task by learning multi-modality in a coarse-to-fine approach. Its subsequent local motion forecasting is inferred from coarsely modeled multi-modality, a greatly simplified task compared to extracting intent from entangled multi-modality as well as multi-agent interaction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "type": "text", + "content": "These improvements on both global and local scales" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 213, + 547, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 321 + ], + "type": "text", + "content": "jointly contribute toward lowering the JPE metric, demonstrating proficiency of our method in forecasting overall human motion. Based on such competence, our method which aimed towards improving on forecasting long-term multiagent environments also exhibits similar or better performances on short timeframes as shown in Tab. 3. In addition, our approach also excels even on sole local motion with minimal global displacement, as elaborated in section 7.2 of supplementary materials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 331, + 414, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 331, + 414, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 414, + 342 + ], + "type": "text", + "content": "5.2. Qualitative results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 350, + 546, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 546, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 546, + 493 + ], + "type": "text", + "content": "Our method forecasts a more plausible global pose in longer timescales (" + }, + { + "bbox": [ + 304, + 350, + 546, + 493 + ], + "type": "inline_equation", + "content": "\\sim 5\\mathrm{s}" + }, + { + "bbox": [ + 304, + 350, + 546, + 493 + ], + "type": "text", + "content": ") as shown in the interacting scene of five agents in Fig. 5. Looking at the input and GT sequences, the leftmost person avoids the traversing couple from right to left. The two people in front are stationary while talking to each other. MRT and TBIFormer forecasts implausible overlapped poses at the final prediction horizon (" + }, + { + "bbox": [ + 304, + 350, + 546, + 493 + ], + "type": "inline_equation", + "content": "t = 5\\mathrm{s}" + }, + { + "bbox": [ + 304, + 350, + 546, + 493 + ], + "type": "text", + "content": "). JRT fails to learn the global locomotion of agents due to the high complexity of its attention mechanism and is stuck in the local minimum of predicting the inactivity of all agents. On the other hand, our model forecasts plausible poses where the closely interacting two agents walk side-by-side." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 494, + 547, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 638 + ], + "type": "text", + "content": "Figure 6 illustrates exemplary sequences where more natural local motion has been forecasted by our method. Comparing forecasts on a scene of walking agents, our method generates a much more plausible sequence where the stepping foot remains stationary. On the other hand, the previous SOTA method, TBIFormer, struggles to learn the natural walking mechanism of human legs and a parallel translation of both feet is exhibited. Such discrepancy shows that trajectory-conditioning for inferring local motion from global intent generates more proficient details in human motion than SOTA methods. More visualizations could be found in the supplementary materials." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 647, + 405, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 405, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 405, + 658 + ], + "type": "text", + "content": "5.3. Ablation studies" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Different number of modes. The main quantitative results report prediction results with " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " as 6 to compare the ability to address the multi-modal nature of human motion during pose forecasting. Table. 4 additionally compares forecast" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1623" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 97, + 261, + 171 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 93 + ], + "type": "text", + "content": "Table 4. Comparison of performance with different number of modes in CMU-Mocap (UMPM) dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 97, + 261, + 171 + ], + "lines": [ + { + "bbox": [ + 72, + 97, + 261, + 171 + ], + "spans": [ + { + "bbox": [ + 72, + 97, + 261, + 171 + ], + "type": "table", + "html": "
F16
Metric @ 2sAPEJPEAPEJPE
MRT163.9366.4164.4280.1
JRT176.7367.4181.6316.9
TBIFormer160.1374.3160.8290.9
Ours154.4366.4151.7262.7
", + "image_path": "64970b331bc7a6b1f21847f375f9cd2dbb57b93508af6d744446afb5b3c212f9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 218, + 287, + 315 + ], + "blocks": [ + { + "bbox": [ + 47, + 181, + 287, + 214 + ], + "lines": [ + { + "bbox": [ + 47, + 181, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 47, + 181, + 287, + 214 + ], + "type": "text", + "content": "Table 5. Ablation studies on core components of model structures. Experiments are done with JRDB-GMP dataset to evaluate multiagent long-term performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 218, + 287, + 315 + ], + "lines": [ + { + "bbox": [ + 50, + 218, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 50, + 218, + 287, + 315 + ], + "type": "table", + "html": "
Exp. #Trajectory encoderPose decoderMetrics
Local pose embeddingAgent interactionTrajectory -conditioningJPE @5sAPE @5sFDE @5s
-471.4101.7457.9
1✓400.595.1370.9
2✓403.394.7374.2
3✓401.293.0372.8
4✓✓395.693.8366.8
5✓✓392.795.2363.4
6✓✓✓391.291.4363.3
", + "image_path": "90cf8cf8052668432edfe65e6e5ab7da69fb50d116ca64af34f49ad32cf5aad9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "text", + "content": "results with " + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "text", + "content": " as 1. Our method again achieves noticeable improvement in APE over the baseline on single-modal forecasts. With " + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "inline_equation", + "content": "F = 1" + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "text", + "content": ", although our method barely enjoys improvement in forecasting global motion due to the absence of multi-modality, its superiority in APE shows the validity of our coarse-to-fine forecasting strategy that also effectively captures agent interaction. Our method improves with multi-modal predictions, demonstrating the proficiency of a coarse-to-fine approach in interpreting the stochastic nature of human motion and its intent. Note that our method improves in APE along with an increase in " + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "text", + "content": " unlike previous methods, indicating a unique aptitude in addressing the multi-modal nature of not only global locomotion but also local pose intent via trajectory-conditioning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 496, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 700 + ], + "type": "text", + "content": "Importance of each architecture component. Table. 5 reports the influence of core components of our model. For the trajectory encoder, we evaluate the importance of using local pose embedding and modeling agent interaction. Comparing experiments 1, 5 and 3, 4, both show improvements in JPE and FDE metrics with the use of local pose embedding. Our method has taken advantage of detailed local pose cues to infer an agent's global intention. For interaction modeling, its use is beneficial for both global and local forecasts as compared by experiments 4 and 6. These joint improvements demonstrate the importance of considering local and global motion interactions for their respective forecasts. As for the pose decoder, comparisons of experiments 2,4 and 5,6 both show improvements in APE metric. Such consistent improvement verifies the effectiveness of the trajectory-conditioned local motion forecast approach in generating plausible local motion from global intention." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": "Importance of interaction modeling. Accurate model" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 339, + 97, + 512, + 171 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 91 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 91 + ], + "type": "text", + "content": "Table 6. Ablation studies on agent interaction cutoff distance on JRDB-GMP." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 97, + 512, + 171 + ], + "lines": [ + { + "bbox": [ + 339, + 97, + 512, + 171 + ], + "spans": [ + { + "bbox": [ + 339, + 97, + 512, + 171 + ], + "type": "table", + "html": "
JPE @ 5s
TBIFormerOurs
w/o interaction483.8406.0
w/ interaction < 2m-403.5
w/ interaction < 4m-400.5
w/ interaction all481.3390.4
", + "image_path": "5381b8d451966346c75439fb30ea8b4350813992a9c09099f1f9e742849f00a4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "text", + "content": "ing of inter-agent interaction becomes more pivotal to forecast in more complex environments. Indeed, its complexity grows in a long-term multi-agent scene. When holistically considering joint-wise interaction for all timesteps, the computation complexity is acquired as " + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "inline_equation", + "content": "O(T^2 \\cdot N^2 \\cdot J^2)" + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "text", + "content": " is the number of timesteps, " + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "text", + "content": " the number of agents, and " + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "text", + "content": " the number of joints. On the contrary, with interaction modeling in global trajectory scale, our method reduces the computation cost by " + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "inline_equation", + "content": "TJ^2" + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "inline_equation", + "content": "O(T \\cdot N^2)" + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "text", + "content": ". This enables efficient and proficient modeling of intra (pose) and inter (trajectory)-agent interactions as shown by Tab. 6. While body part-wise interaction modeling only improved by " + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "inline_equation", + "content": "0.52\\%" + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "text", + "content": " for TBIFormer, ours improves up to " + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "inline_equation", + "content": "3.84\\%" + }, + { + "bbox": [ + 304, + 178, + 545, + 418 + ], + "type": "text", + "content": " with interaction modeling. This demonstrates the proficiency of our efficient interaction modeling-based method in inferring global and local intents from complex interactions. In addition, the gradual improvement of JPE according to a wider interaction range confirms the importance of interaction modeling of more agents, which cannot be learned from the arbitrarily mixed previous datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 430, + 378, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 378, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 378, + 442 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 450, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 617 + ], + "type": "text", + "content": "In this work, we propose a novel interaction-aware trajectory-conditioned approach to handle long-term multi-agent motion forecasting, along with a new dataset suited for such scope. Our proposed model utilizes a coarse-to-fine approach and decouples overall motion prediction into global and local components. Multi-modality of human motion is proficiently modeled via inferring fine local intents from coarse global intents, along with efficient agent-wise interaction modeling. As for the dataset, our JRDB-GMP dataset contains unprecedented long-term " + }, + { + "bbox": [ + 304, + 450, + 545, + 617 + ], + "type": "inline_equation", + "content": "(5s+)" + }, + { + "bbox": [ + 304, + 450, + 545, + 617 + ], + "type": "text", + "content": " multi-agent " + }, + { + "bbox": [ + 304, + 450, + 545, + 617 + ], + "type": "inline_equation", + "content": "(6+)" + }, + { + "bbox": [ + 304, + 450, + 545, + 617 + ], + "type": "text", + "content": " interactions in a real-world setting. Our method achieves state-of-the-art performance on all previous datasets and JRDB-GMP dataset, offering generalized practical implications in real-world applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1624" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Vida Adeli, Ehsan Adeli, Ian Reid, Juan Carlos Niebles, and Hamid Rezatofighi. Socially and contextually aware human motion and pose forecasting. IEEE Robotics and Automation Letters, 5(4):6033-6040, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "type": "text", + "content": "[2] Vida Adeli, Mahsa Ehsanpour, Ian Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13390-13400, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "type": "text", + "content": "[3] Gorkay Aydemir, Adil Kaan Akan, and Fatma Güney. Adapt: Efficient multi-agent trajectory prediction with adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8295-8305, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 251, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 305 + ], + "type": "text", + "content": "[4] German Barquero, Sergio Escalera, and Cristina Palmero. Belfusion: Latent diffusion for behavior-driven human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2317-2327, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 307, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 288, + 384 + ], + "type": "text", + "content": "[5] Arij Bouazizi, Adrian Holzbock, Ulrich Kressel, Klaus Dietmayer, and Vasileios Belagiannis. Motionmixer: Mlp-based 3d human body pose forecasting. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 791-798. International Joint Conferences on Artificial Intelligence Organization, 2022. Main Track. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 386, + 288, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 450 + ], + "type": "text", + "content": "[6] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 387–404. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 453, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 287, + 518 + ], + "type": "text", + "content": "[7] Wenhao Chai, Zhongyu Jiang, Jenq-Neng Hwang, and Gaoang Wang. Global adaptation meets local generalization: Unsupervised domain adaptation for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14655-14665, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 521, + 288, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 288, + 576 + ], + "type": "text", + "content": "[8] Ling-Hao Chen, JiaWei Zhang, Yewen Li, Yiren Pang, Xiaobo Xia, and Tongliang Liu. Humanmac: Masked motion completion for human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9544-9555, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 578, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 578, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 53, + 578, + 288, + 622 + ], + "type": "text", + "content": "[9] Hsu-kuang Chiu, Ehsan Adeli, Borui Wang, De-An Huang, and Juan Carlos Niebles. Action-agnostic human pose forecasting. In 2019 IEEE winter conference on applications of computer vision (WACV), pages 1423–1432. IEEE, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[10] Sehwan Choi, Jungho Kim, Junyong Yun, and Jun Won Choi. R-pred: Two-stage motion prediction via tube-query attention-based trajectory refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8525-8535, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[11] Rohan Choudhury, Kris M. Kitani, and László A. Jeni. Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In Proceedings of the IEEE/CVF International" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision (ICCV), pages 14750-14760, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "text", + "content": "[12] Andrea Porfiri Dal Cin, Giacomo Boracchi, and Luca Magri. Multi-body depth and camera pose estimation from multiple views. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17804-17814, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 151, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 172 + ], + "type": "text", + "content": "[13] CMU-Graphics-Lab. Cmu graphics lab motion capture database. http://mocap.cs.cmu.edu/, 2003. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 174, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 174, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 174, + 545, + 237 + ], + "type": "text", + "content": "[14] Runyang Feng, Yixing Gao, Xueqing Ma, Tze Ho Elden Tse, and Hyung Jin Chang. Mutual information-based temporal difference learning for human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17131-17141, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 239, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 294 + ], + "type": "text", + "content": "[15] Runyang Feng, Yixing Gao, Tze Ho Elden Tse, Xueqing Ma, and Hyung Jin Chang. Diffpose: Spatiotemporal diffusion model for video-based human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14861-14872, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 295, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 295, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 545, + 350 + ], + "type": "text", + "content": "[16] Xuehao Gao, Shaoyi Du, Yang Wu, and Yang Yang. Decompose more and aggregate better: Two closer looks at frequency representation learning for human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6451-6460, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 350, + 545, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 405 + ], + "type": "text", + "content": "[17] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13041-13051, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 405, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 449 + ], + "type": "text", + "content": "[18] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15303-15312, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "text", + "content": "[19] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15303-15312, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "text", + "content": "[20] Wen Guo, Xiaoyu Bie, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Multi-person extreme motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13053-13064, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "type": "text", + "content": "[21] Je-Seok Ham, Dae Hoe Kim, NamKyo Jung, and Jinyoung Moon. Cipf: Crossing intention prediction network based on feature fusion modules for improving pedestrian safety. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3665-3674, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 604, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 647 + ], + "type": "text", + "content": "[22] Karl Holmquist and Bastian Wandt. Diffpose: Multi-hypothesis human pose estimation using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15977-15987, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[23] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1625" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[24] Boyuan Jiang, Lei Hu, and Shihong Xia. Probabilistic triangulation for uncalibrated multi-view 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14850-14860, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "text", + "content": "[25] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9644-9653, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 239 + ], + "type": "text", + "content": "[26] Zhehan Kan, Shuoshuo Chen, Ce Zhang, Yushun Tang, and Zhihai He. Self-correctable and adaptable inference for generalizable human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5537-5546, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "type": "text", + "content": "[27] Jeongho Kim, Wooksu Shin, Hancheol Park, and Jongwon Baek. Addressing the occlusion problem in multi-camera people tracking with human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5462-5468, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 293, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 288, + 358 + ], + "type": "text", + "content": "[28] Mihee Lee, Samuel S Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. Musevae: multi-scale vae for environment-aware long term trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2221-2230, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "type": "text", + "content": "[29] Lihuan Li, Maurice Pagnucco, and Yang Song. Graph-based spatial transformer with memory replay for multifuture pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2241, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 415, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 481 + ], + "type": "text", + "content": "[30] Huan Liu, Qiang Chen, Zichang Tan, Jiang-Jiang Liu, Jian Wang, Xiangbo Su, Xiaolong Li, Kun Yao, Junyu Han, Errui Ding, Yao Zhao, and Jingdong Wang. Group pose: A simple baseline for end-to-end multi-person pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15029-15038, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 481, + 288, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 547 + ], + "type": "text", + "content": "[31] Tiezheng Ma, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Progressively generating better initial guesses towards next stages for high-quality human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6437-6446, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "type": "text", + "content": "[32] Takahiro Maeda and Norimichi Ukita. Motionaug: Augmentation with physical correction for human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6427-6436, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 592, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 646 + ], + "type": "text", + "content": "[33] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15233-15242, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "text", + "content": "[34] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9489-9497, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[35] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Generating smooth pose sequences for diverse human motion pre" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "diction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13309-13318, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[36] Wei Mao, Richard I Hartley, Mathieu Salzmann, et al. Contact-aware human motion forecasting. Advances in Neural Information Processing Systems, 35:7356-7367, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 130, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 185 + ], + "type": "text", + "content": "[37] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Weakly-supervised action transition learning for stochastic human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8151-8160, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "type": "text", + "content": "[38] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5517-5526, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "type": "text", + "content": "[39] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 2018 International Conference on 3D Vision (3DV), pages 120-130, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 298, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 376 + ], + "type": "text", + "content": "[40] Jiquan Ngiam, Vijay Vasudevan, Benjamin Caine, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, David J Weiss, Benjamin Sapp, Zhifeng Chen, and Jonathon Shlens. Scene transformer: A unified architecture for predicting future trajectories of multiple agents. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "type": "text", + "content": "[41] Daehee Park, Jaewoo Jeong, and Kuk-Jin Yoon. Improving transferability for cross-domain trajectory prediction via neural stochastic differential equation. arXiv preprint arXiv:2312.15906, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "text", + "content": "[42] Daehee Park, Hobin Ryu, Yunseo Yang, Jegyeong Cho, Jiwon Kim, and Kuk-Jin Yoon. Leveraging future relationship reasoning for vehicle trajectory prediction. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 477, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 477, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 477, + 545, + 521 + ], + "type": "text", + "content": "[43] Daehee Park, Jaeseok Jeong, Sung-Hoon Yoon, Jaewoo Jeong, and Kuk-Jin Yoon. T4p: Test-time training of trajectory prediction via masked autoencoder and actor-specific token memory, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 522, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 578 + ], + "type": "text", + "content": "[44] Sungchan Park, Eunyi You, Inhoe Lee, and Joonseok Lee. Towards robust and smooth 3d multi-person pose estimation from monocular videos in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14772-14782, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 578, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 634 + ], + "type": "text", + "content": "[45] Behnam Parsaeifard, Saeed Saadatnejad, Yuejiang Liu, Taylor Mordan, and Alexandre Alahi. Learning decoupled representations for human pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2294-2303, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[46] Qucheng Peng, Ce Zheng, and Chen Chen. Source-free domain adaptive human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4826-4836, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 680, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 715 + ], + "type": "text", + "content": "[47] Xiaogang Peng, Siyuan Mao, and Zizhao Wu. Trajectory-aware body interaction transformer for multi-person pose forecasting. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1626" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 17121- 17130, 2023. 1, 2, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 173 + ], + "type": "text", + "content": "[48] Zhongwei Qiu, Qiansheng Yang, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Chang Xu, Dongmei Fu, and Jingdong Wang. Psvt: End-to-end multi-person 3d pose and shape estimation with progressive video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21254-21263, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "text", + "content": "[49] Muhammad Rameez Ur Rahman, Luca Scofano, Edoardo De Matteis, Alessandro Flaborea, Alessio Sampieri, and Fabio Galasso. Best practices for 2-body pose forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3613-3623, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "type": "text", + "content": "[50] Dripta S. Raychaudhuri, Calvin-Khang Ta, Arindam Dutta, Rohit Lal, and Amit K. Roy-Chowdhury. Prior-guided source-free domain adaptation for human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14996-15006, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 287, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 288, + 352 + ], + "type": "text", + "content": "[51] Luke Rowe, Martin Ethier, Eli-Henry Dykhne, and Krzysztof Czarnecki. Fjmp: Factorized joint multi-agent motion prediction over learned directed acyclic interaction graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13745-13755, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 288, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 420 + ], + "type": "text", + "content": "[52] Saeed Saadatnejad, Ali Rasekh, Mohammadreza Mofayezi, Yasamin Medghalchi, Sara Rajabzadeh, Taylor Mordan, and Alexandre Alahi. A generic diffusion-based approach for 3d human pose prediction in the wild. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8246-8253. IEEE, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "type": "text", + "content": "[53] Tim Salzmann, Marco Pavone, and Markus Ryll. Motron: Multimodal probabilistic human motion forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6457-6466, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "type": "text", + "content": "[54] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. IEEE Robotics and Automation Letters, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 512, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 288, + 577 + ], + "type": "text", + "content": "[55] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14761-14771, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 578, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 633 + ], + "type": "text", + "content": "[56] Xiaolong Shen, Zongxin Yang, Xiaohan Wang, Jianxin Ma, Chang Zhou, and Yi Yang. Global-to-local modeling for video-based 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8887-8896, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[57] Mingyi Shi, Sebastian Starke, Yuting Ye, Taku Komura, and Jungdam Won. Phasemp: Robust 3d pose estimation via phase-conditioned human motion prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14725-14737, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[58] Theodoros Sofianos, Alessio Sampieri, Luca Franco, and Fabio Galasso. Space-time-separable graph convolutional" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "network for pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11209-11218, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "type": "text", + "content": "[59] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13243-13252, 2022. 2, 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 164, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 229 + ], + "type": "text", + "content": "[60] Yuran Sun, Alan William Dougherty, Zhuoying Zhang, Yi King Choi, and Chuan Wu. Mixsynthformer: A transformer encoder-like structure with mixed synthetic self-attention for efficient human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14884–14893, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "type": "text", + "content": "[61] Zhenhua Tang, Zhaofan Qiu, Yanbin Hao, Richang Hong, and Ting Yao. 3d human pose estimation with spatiotemporal criss-cross attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4790-4799, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "text", + "content": "[62] Julian Tanke, Chintan Zaveri, and Juergen Gall. Intention-based long-term human motion anticipation. In 2021 International Conference on 3D Vision (3DV), pages 596-605. IEEE, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 332, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 397 + ], + "type": "text", + "content": "[63] Julian Tanke, Linguang Zhang, Amy Zhao, Chengcheng Tang, Yujun Cai, Lezi Wang, Po-Chen Wu, Juergen Gall, and Cem Keskin. Social diffusion: Long-term multiple human motion anticipation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9601-9611, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 399, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 465 + ], + "type": "text", + "content": "[64] N.P. van der Aa, X. Luo, G.J. Giezeman, R.T. Tan, and R.C. Veltkamp. Umpm benchmark: A multi-person dataset with synchronized video and motion capture data for evaluation of articulated human motion and interaction. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 1264-1269, 2011. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "type": "text", + "content": "[65] Edward Vendrow, Satyajit Kumar, Ehsan Adeli, and Hamid Rezatofighi. Somoformer: Multi-person pose forecasting with transformers. arXiv preprint arXiv:2208.14023, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "type": "text", + "content": "[66] Edward Vendrow, Duy Tho Le, Jianfei Cai, and Hamid Rezatofighi. Jrdb-pose: A large-scale dataset for multiperson pose estimation and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4811–4820, 2023. 2, 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "type": "text", + "content": "[67] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "text", + "content": "[68] Chenxi Wang, Yunfeng Wang, Zixuan Huang, and Zhiwen Chen. Simple baseline for single human motion forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2260-2265, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[69] Jiashun Wang, Huazhe Xu, Medhini Narasimhan, and Xiaolong Wang. Multi-person 3d motion prediction with multirange transformers. Advances in Neural Information Processing Systems, 34:6036-6049, 2021. 1, 2, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1627" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[70] Mingkun Wang, Xinge Zhu, Changqian Yu, Wei Li, Yuexin Ma, Ruochun Jin, Xiaoguang Ren, Dongchun Ren, Mingxu Wang, and Wenjing Yang. Ganet: Goal area network for motion forecasting. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1609-1615. IEEE, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 184 + ], + "type": "text", + "content": "[71] Yucheng Xing and Xin Wang. Hdg-ode: A hierarchical continuous-time model for human pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14700–14712, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 239 + ], + "type": "text", + "content": "[72] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6488-6497, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 242, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 288, + 297 + ], + "type": "text", + "content": "[73] Chenxin Xu, Robby T. Tan, Yuhong Tan, Siheng Chen, Xin-chao Wang, and Yanfeng Wang. Auxiliary tasks benefit 3d skeleton-based human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9509–9520, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "type": "text", + "content": "[74] Qingyao Xu, Weibo Mao, Jingze Gong, Chenxin Xu, Siheng Chen, Weidi Xie, Ya Zhang, and Yanfeng Wang. Joint-relation transformer for multi-person motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9816-9826, 2023. 1, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 396 + ], + "type": "text", + "content": "[75] Sirui Xu, Yu-Xiong Wang, and Liangyan Gui. Stochastic multi-person 3d motion forecasting. In The Eleventh International Conference on Learning Representations, 2023. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "type": "text", + "content": "[76] Yi Xu, Armin Bazarjani, Hyung-gun Chi, Chiho Choi, and Yun Fu. Uncovering the missing pattern: Unified framework towards trajectory imputation and prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9632-9643, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "text", + "content": "[77] Maosheng Ye, Jiamiao Xu, Xunnong Xu, Tengfei Wang, Tongyi Cao, and Qifeng Chen. Bootstrap motion forecasting with self-consistent constraints. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8504-8514, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "text", + "content": "[78] Yingxuan You, Hong Liu, Ti Wang, Wenhao Li, Runwei Ding, and Xia Li. Co-evolution of pose and mesh for 3d human body estimation from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14963-14973, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 568, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 632 + ], + "type": "text", + "content": "[79] Bruce X.B. Yu, Zhi Zhang, Yongxu Liu, Sheng-hua Zhong, Yan Liu, and Chang Wen Chen. Gla-gen: Global-local adaptive graph convolutional network for 3d human pose estimation from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8818-8829, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[80] Kai Zhai, Qiang Nie, Bo Ouyang, Xiang Li, and Shanlin Yang. Hopfir: Hop-wise graphformer with intragroup joint refinement for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14985-14995, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[81] Yi Zhang, Pengliang Ji, Angtian Wang, Jieru Mei, Adam Kortylewski, and Alan Yuille. 3d-aware neural body fitting for" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 675 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "occlusion robust 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9399-9410, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[82] He Zhao and Richard P. Wildes. Where are you heading? dynamic trajectory prediction with expert goal examples. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7629-7638, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "type": "text", + "content": "[83] Qitao Zhao, Ce Zheng, Mengyuan Liu, Pichao Wang, and Chen Chen. Poseformerv2: Exploring frequency domain for efficient and robust 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8877-8886, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "text", + "content": "[84] Fang Zheng, Le Wang, Sanping Zhou, Wei Tang, Zhenxing Niu, Nanning Zheng, and Gang Hua. Unlimited neighborhood interaction for heterogeneous trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13168-13177, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 263, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 545, + 329 + ], + "type": "text", + "content": "[85] Jingxiao Zheng, Xinwei Shi, Alexander Gorban, Junhua Mao, Yang Song, Charles R Qi, Ting Liu, Visesh Chari, Andre Cornman, Yin Zhou, et al. Multi-modal 3d human pose estimation with 2d weak supervision in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4478-4487, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 331, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 384 + ], + "type": "text", + "content": "[86] Chongyang Zhong, Lei Hu, Zihao Zhang, Yongjing Ye, and Shihong Xia. Spatio-temporal gating-adjacency gen for human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6447–6456, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 386, + 545, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 386, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 308, + 386, + 545, + 450 + ], + "type": "text", + "content": "[87] Mu Zhou, Lucas Stoffl, Mackenzie Weygandt Mathis, and Alexander Mathis. Rethinking pose estimation in crowds: Overcoming the detection information bottleneck and ambiguity. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14689-14699, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 453, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 453, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 308, + 453, + 545, + 507 + ], + "type": "text", + "content": "[88] Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Kejie Lu. Hivt: Hierarchical vector transformer for multi-agent motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8823-8833, 2022. 2, 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 509, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 509, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 509, + 545, + 552 + ], + "type": "text", + "content": "[89] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 554, + 545, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 545, + 629 + ], + "type": "text", + "content": "[90] Dekai Zhu, Guangyao Zhai, Yan Di, Fabian Manhardt, Hendrik Berkemeyer, Tuan Tran, Nassir Navab, Federico Tombari, and Benjamin Busam. Ipcc-tp: Utilizing incremental pearson correlation coefficient for joint multi-agent trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5507-5516, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 632, + 545, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 632, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 308, + 632, + 545, + 675 + ], + "type": "text", + "content": "[91] Tao Zhuo, Zhiyong Cheng, Peng Zhang, Yongkang Wong, and Mohan Kankanhalli. Unsupervised online video object segmentation with motion property understanding. IEEE Transactions on Image Processing, 29:237-249, 2019. 1" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1628" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/c38e4100-b4f3-4d92-aff9-1fdadd7cfa0a_content_list.json b/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/c38e4100-b4f3-4d92-aff9-1fdadd7cfa0a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..543688e0904f0c20dc4e99115f36892a91beb38f --- /dev/null +++ b/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/c38e4100-b4f3-4d92-aff9-1fdadd7cfa0a_content_list.json @@ -0,0 +1,1724 @@ +[ + { + "type": "text", + "text": "Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers", + "text_level": 1, + "bbox": [ + 186, + 130, + 782, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sanghyeok Lee* Joonmyung Choi* Hyunwoo J. Kim†", + "bbox": [ + 246, + 202, + 722, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science and Engineering, Korea University", + "bbox": [ + 212, + 234, + 756, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cat0626, pizard, hyunwoojkim}@korea.ac.kr", + "bbox": [ + 295, + 255, + 666, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 305, + 313, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Transformer (ViT) has emerged as a prominent backbone for computer vision. For more efficient ViTs, recent works lessen the quadratic cost of the self-attention layer by pruning or fusing the redundant tokens. However, these works faced the speed-accuracy trade-off caused by the loss of information. Here, we argue that token fusion needs to consider diverse relations between tokens to minimize information loss. In this paper, we propose a Multi-criteria Token Fusion (MCTF), that gradually fuses the tokens based on multi-criteria (i.e., similarity, informativeness, and size of fused tokens). Further, we utilize the one-step-ahead attention, which is the improved approach to capture the informativeness of the tokens. By training the model equipped with MCTF using a token reduction consistency, we achieve the best speed-accuracy trade-off in the image classification (ImageNet1K). Experimental results prove that MCTF consistently surpasses the previous reduction methods with and without training. Specifically, DeiT-T and DeiT-S with MCTF reduce FLOPs by about $44\\%$ while improving the performance $(+0.5\\%, \\text{and} +0.3\\%)$ over the base model, respectively. We also demonstrate the applicability of MCTF in various Vision Transformers (e.g., T2T-ViT, LV-ViT), achieving at least $31\\%$ speedup without performance degradation. Code is available at https://github.com/mlvlab/MCTF.", + "bbox": [ + 76, + 338, + 473, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 746, + 209, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Transformer [11] (ViT) has been proposed to tackle the vision tasks with self-attention, originally developed for natural language processing tasks. With the advent of ViT, Transformers are the prevalent architectures for a wide range of vision tasks, e.g., classification [11, 20, 28, 29, 31], object detection [5, 31, 41], segmentation [20, 27, 33], etc. ViTs,", + "bbox": [ + 75, + 771, + 470, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "built only with self-attention and MLP, provide great flexibility and impressive performance compared to conventional approaches, e.g., convolutional neural networks (CNNs). However, despite these advantages, the quadratic computational complexity of self-attention with respect to the number of tokens is the major bottleneck for Transformers. This limitation becomes more substantial with the growing interest in large-scale foundation models such as CLIP [25]. To this end, several works [3, 16, 30, 34] have proposed efficient self-attention mechanisms including local self-attention within predefined windows [1, 9, 20].", + "bbox": [ + 500, + 306, + 893, + 473 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "More recently, there has been increasing interest in token-reduction methods for optimizing ViTs without altering their architecture. Earlier works [12, 23, 24, 26, 37] primarily focused on pruning the uninformative tokens to reduce the number of tokens. Another line of works [4, 17, 18, 21, 22] attempted to fuse the tokens instead of discarding them to minimize the information loss. However, performance degradation is still commonly observed in most token fusion methods. We notice that the token fusion methods usually consider only one criterion, such as the similarity or informativeness of tokens, leading to suboptimal token matching. For instance, similarity-based token fusion is prone to combine the foreground tokens, whereas informativeness-based fusion often merges substantially dissimilar tokens, resulting in collapsed representations. Furthermore, if too many tokens are fused into one token, then information loss is inevitable.", + "bbox": [ + 496, + 473, + 893, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the problems, we introduce Multi-Criteria Token Fusion (MCTF) that optimizes vision transformers by fusing tokens based on multi-criteria. Unlike previous works that consider a single criterion for token fusion, MCTF measures the relationship between the tokens with multi-criteria as follows; (1) similarity to fuse the redundant tokens, (2) informativeness to reduce the uninformative tokens, (3) the size of the tokens to prevent the large-sized tokens that boost the loss of information. Also, to tackle the inconsistency between attention maps of consecutive layers, we adopt one-step-ahead attention, which explicitly estimates the informativeness of", + "bbox": [ + 496, + 729, + 895, + 896 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15741", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/afb4d8e49b0642017b3ab2ced053a4fc3c9392bafd01ad9dff78a810b9d0e898.jpg", + "image_caption": [ + "Figure 1. Comparison of the token reduction methods with DeiT-T (left), and DeiT-S (right). Given a base model marked as blue circle, previous token reduction methods accelerate the speed with the trade-off between accuracy and computational cost. Our MCTF, marked as a star, even brings performance improvements while lessening the complexity of DeiT. Note that after only one finetuning with the specific reduced number of tokens marked as red star, we simply evaluate it with the diverse FLOPs by adjusting the reduced numbers." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 472, + 316 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e3627aae79897e1cb056d4f57193d13d32a28d6e895d8b71f5c6b25c951ec20c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 90, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the tokens in the next layer. Finally, by introducing a token reduction consistency for finetuning the model, we achieve superior performance to the existing works as in Figure 1. Surprisingly, our MCTF even performs better than the 'full' base model (red dotted line) with a reduced computational complexity. Specifically, it brings a $0.5\\%$ , and $0.3\\%$ gain while reducing FLOPs by about $44\\%$ in DeiT-T, and DeiT-S [28], respectively. We have observed a similar speed-up $(31\\%)$ in T2T-ViT [39], and LV-ViT [15] without any performance degradation.", + "bbox": [ + 75, + 401, + 470, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized in fourfold.", + "bbox": [ + 76, + 551, + 383, + 565 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Multi-criteria Token Fusion, a novel token fusion method that considers multi-criteria, e.g., similarity, informativeness, and size, to capture the complex relationship of tokens and minimize information loss.", + "- For measuring the informativeness of the tokens, we utilize one-step-ahead attention to retain the attentive tokens in the following layers.", + "- We propose a new fine-tuning scheme with token reduction consistency to boost the generalization performance of transformers equipped with MCTF.", + "- The extensive experiments demonstrate that MCTF achieves the best speed-accuracy trade-off in diverse ViTs, surpassing all previous token reduction methods." + ], + "bbox": [ + 76, + 568, + 468, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 76, + 782, + 220, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Transformers. Vision Transformer [11] is introduced to tackle the vision tasks. Later, DeiT [28] and CaiT [29] are proposed to handle the data efficiency and scalability of ViT, respectively. Recent works [6, 10, 14, 20, 31] tried to insert the inductive biases of CNNs on ViT, such as the locality or pyramid-architecture. In parallel, there is a", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "line of works that boosts the vanilla ViT by scaling [29, 40] or self-supervised learning [2, 13, 32]. Despite the promising results of these works, the quadratic complexity of ViTs is still the major constraint for scaling the model. For the sake of mitigating the complexity, Reformer [16] lessens the quadratic complexity to $O(N \\log N)$ through the hashing function, and Linformer [30], performer [8], and Nyströmformer [34] achieve the linear cost with the approximated linear attention. Also, several works [1, 9, 10, 20] utilize sparse attention with the reduced key or query. Swin [20] and Twins [9] utilize the local attention within the fixed size of the window to mitigate the complexity.", + "bbox": [ + 496, + 401, + 893, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Token reduction in ViTs. Most of the computational burden in ViTs arises from the self-attention. To reduce the quadratic cost in the number of tokens, recent works [4, 12, 17, 18, 21–24, 26, 37] have an interest in reducing the token itself. These works have the advantage of utilizing the original ViTs architecture without modification. In earlier works [12, 23, 24, 26, 37], the uninformative tokens are simply dropped during the forward process, leading to the information loss. To compensate for this, SPViT [17] and EViT [18] first split the tokens into informative and uninformative token sets based on attention scores, then fuse these uninformative token sets into a single token. In parallel, token pooling [22] and ToMe [4] combine the semantically similar tokens to reduce redundancies. A more recent study BAT [21] first split the tokens based on informativeness then fuse the tokens considering the diversity of the tokens. Despite the advantage of each criterion, successful integration of multi-criteria is still less explored.", + "bbox": [ + 496, + 628, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15742", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/11e3b5e1793a889a3d9c9a678880b8c0c8877ac09c70b47122b1d2da106c5011.jpg", + "image_caption": [ + "(a) Origin" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 261, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fa08f1eaca5e8102ae9ab000f75eb90d923a88562bae277388ff7344d72d338b.jpg", + "image_caption": [ + "(b) $\\mathbf{W}^{\\mathrm{sim}}$" + ], + "image_footnote": [], + "bbox": [ + 287, + 88, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b627314e338ea533b512dc8a9dc206fd81ef73b25967bd2250fdc197147f4cd2.jpg", + "image_caption": [ + "(c) $\\mathbf{W}^{\\mathrm{sim}}\\& \\mathbf{W}^{\\mathrm{info}}$", + "Figure 2. Visualization of the fused tokens. Given (a) the leftmost image, (b) fusing the tokens with a single criterion $\\mathbf{W}^{\\mathrm{sim}}$ often results in the excessive fusion of the foreground object. (c) Then considering both similarity and informativeness ( $\\mathbf{W}^{\\mathrm{sim}}\\& \\mathbf{W}^{\\mathrm{info}}$ ), tokens in the foreground objects are less fused while the tokens in the background are largely fused. (d) Finally, MCTF helps retain the information of each component in the image by preventing the large-size token with the multi-criteria ( $\\mathbf{W}^{\\mathrm{sim}}\\& \\mathbf{W}^{\\mathrm{info}}\\& \\mathbf{W}^{\\mathrm{size}}$ )." + ], + "image_footnote": [], + "bbox": [ + 498, + 88, + 681, + 228 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f94704b97ecdeb4baccac06247d87493beb88ccbf29c6ee2cfb9b01417a5efb4.jpg", + "image_caption": [ + "(d) $\\mathbf{W}^{\\mathrm{sim}}\\& \\mathbf{W}^{\\mathrm{info}}\\& \\mathbf{W}^{\\mathrm{size}}$" + ], + "image_footnote": [], + "bbox": [ + 709, + 88, + 890, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 335, + 166, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first review the self-attention and token reduction approaches (Section 3.1). Then, we present our multi-criteria token fusion (Section 3.2) that leverages one-step-ahead attention (Section 3.3). Lastly, we introduce a training strategy with token reduction consistency in Section 3.4.", + "bbox": [ + 75, + 362, + 470, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 449, + 217, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Transformers, tokens $\\mathbf{X} \\in \\mathbb{R}^{N \\times C}$ are processed by self-attention defined as", + "bbox": [ + 75, + 470, + 472, + 502 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S A} (\\mathbf {X}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {C}}\\right) \\mathbf {V}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 511, + 468, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{Q},\\mathbf{K},\\mathbf{V} = \\mathbf{X}\\mathbf{W}_{\\mathbf{Q}},\\mathbf{X}\\mathbf{W}_{\\mathbf{K}},\\mathbf{X}\\mathbf{W}_{\\mathbf{V}}$ and $\\mathbf{W}_{\\mathbf{Q}},\\mathbf{W}_{\\mathbf{K}}$ $\\mathbf{W}_{\\mathbf{V}}\\in \\mathbb{R}^{C\\times C}$ are learnable weight matrices. Despite its outstanding expressive power, the self-attention does not scale well with the number of tokens $N$ due to its quadratic time complexity $O(N^{2}C + NC^{2})$ . To address this problem, a line of works [12, 23, 24, 26, 37] reduces the number of tokens simply by pruning uninformative tokens. These approaches often cause significant performance degradation due to the loss of information. Thus, another line of works [4, 17, 18, 21, 22] fuses the uninformative or redundant tokens $\\hat{\\mathbf{X}}\\subset \\mathbf{X}$ into a new token $\\hat{\\mathbf{x}} = \\delta (\\hat{\\mathbf{X}})$ , where $\\mathbf{X}$ is the set of original tokens, and $\\delta$ denotes a merging function, e.g., max-pooling or averaging. In this work, we also adopt 'token fusion' rather than 'token pruning' with multiple criteria to minimize the loss of information by token reduction.", + "bbox": [ + 75, + 563, + 472, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-criteria token fusion", + "text_level": 1, + "bbox": [ + 76, + 801, + 316, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of input tokens $\\mathbf{X} \\in \\mathbb{R}^{N \\times C}$ , the goal of MCTF is to fuse the tokens into output tokens $\\hat{\\mathbf{X}} \\in \\mathbb{R}^{(N - r) \\times C}$ , where $r$ is the number of fused tokens. To minimize the information loss, we first evaluate the relations between the tokens based on multi-criteria, then group and merge the", + "bbox": [ + 75, + 823, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tokens through bidirectional bipartite soft matching.", + "bbox": [ + 498, + 338, + 841, + 353 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-criteria attraction function. We first define an attraction function $\\mathbf{W}$ based on multiple criteria as", + "bbox": [ + 498, + 353, + 893, + 382 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\Pi_ {k = 1} ^ {M} \\left(\\mathbf {W} ^ {k} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right)\\right) ^ {\\tau_ {k}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 392, + 890, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{W}^k: \\mathbb{R}^C \\times \\mathbb{R}^C \\to \\mathbb{R}_+$ is an attraction function computed by $k$ -th criterion, and $\\tau^k \\in \\mathbb{R}_+$ is the temperature parameter to adjust the influence of $k$ -th criterion. The higher attraction score between two tokens indicates a higher chance of being fused. In this work, we consider the following three criteria: similarity, informativeness, and size.", + "bbox": [ + 496, + 421, + 890, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarity. The first criterion is the similarity of tokens to reduce redundant information. Akin to the previous works [4, 22] requiring the proximity of tokens, we leverage the cosine similarity between the set of tokens for", + "bbox": [ + 498, + 513, + 893, + 574 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} ^ {\\text {s i m}} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\frac {1}{2} \\left(\\frac {\\mathbf {x} _ {i} \\cdot \\mathbf {x} _ {j}}{\\| \\mathbf {x} _ {i} \\| \\| \\mathbf {x} _ {j} \\|} + 1\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 584, + 890, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Token fusion with similarity effectively eliminates the redundant tokens, yet it often excessively combines the informative tokens as in Figure 2b, causing the loss of information. Informativeness. To minimize the information loss, we introduce informativeness to avoid the fusion of informative tokens. To quantify the informativeness, we measure the averaged attention scores $\\mathbf{a}\\in [0,1]^N$ in the self-attention layer, which indicates the impact of each token on others: $\\mathbf{a}_j = \\frac{1}{N}\\sum_i^N\\mathbf{A}_{ij}$ , where $\\mathbf{A}_{ij} = \\mathrm{softmax}\\left(\\frac{\\mathbf{Q}_i\\mathbf{K}_j^T}{\\sqrt{C}}\\right)$ . When $\\mathbf{a}_i\\rightarrow 0$ , there's no influence from $\\mathbf{x}_i$ to other tokens. With the informativeness scores, we define an informativeness-based attraction function as", + "bbox": [ + 496, + 626, + 893, + 820 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} ^ {\\text {i n f o}} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\frac {1}{\\mathbf {a} _ {i} \\mathbf {a} _ {j}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 829, + 890, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{a}_i, \\mathbf{a}_j$ are the informative scores of $\\mathbf{x}_i, \\mathbf{x}_j$ , respectively. When both tokens are uninformative ( $\\mathbf{a}_i, \\mathbf{a}_j \\to 0$ ),", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15743", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ac4bf9f15015a684ae0c54cb8542749ffb0d741b46ac00e74d63c75d8d9ce802.jpg", + "image_caption": [ + "Figure 3. Bidirectional bipartite soft matching. The set of tokens $\\mathbf{X}$ is split into two groups $\\mathbf{X}^{\\alpha},\\mathbf{X}^{\\beta}$ , and bidirectional bipartite soft matching are conducted through Step 1-4. The intensity of the lines indicates the multi-criteria weights $\\mathbf{W}^t$ ." + ], + "image_footnote": [], + "bbox": [ + 116, + 89, + 851, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the weight gets higher $(\\mathbf{W}^{\\mathrm{info}}(\\mathbf{x}_i,\\mathbf{x}_j)\\to \\infty)$ , making two tokens prone to be fused. In Figure 2c, with the weights combined with the similarity and informativeness, the tokens in the foreground object are less fused.", + "bbox": [ + 75, + 291, + 468, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Size. The last criterion is the size of the tokens, which indicates the number of fused tokens. Although tokens are not dropped but merged via a merging function, e.g., averaging pooling or max pooling, it is difficult to preserve all the information as the number of constituent tokens increases. So, the fusion between smaller tokens is preferred. To this end, we initially set the size $\\mathbf{s} \\in \\mathbb{N}^N$ of tokens $\\mathbf{X}$ as 1 and track the number of constituent (fused) tokens of each token, and define a size-based attraction function as", + "bbox": [ + 75, + 352, + 468, + 487 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} ^ {\\text {s i z e}} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\frac {1}{\\mathbf {s} _ {i} \\mathbf {s} _ {j}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 500, + 468, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Figure 2d, tokens are merged based on the multi-criteria: similarity, informativeness, and size. We observed that the fusion happens between similar tokens and the fusion of foreground tokens or large tokens is properly suppressed.", + "bbox": [ + 75, + 544, + 468, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bidirectional bipartite soft matching. Given the multicriteria-based attraction function $\\mathbf{W}$ , our MCTF performs a relaxed bidirectional bipartite matching called bipartite soft matching [4]. One advantage of bipartite matching is that it alleviates the quadratic cost of similarity computation between tokens, i.e., $O(N^2) \\to O(N'^2)$ , where $N' = \\left\\lfloor \\frac{N}{2} \\right\\rfloor$ . In addition, by relaxing the one-to-one correspondence constraints, the solution can be obtained by an efficient algorithm. In this relaxed matching problem, the set of tokens $\\mathbf{X}$ is first split into the source and target $\\mathbf{X}^\\alpha$ , $\\mathbf{X}^\\beta \\in \\mathbb{R}^{N' \\times C}$ as in Step 1 of Figure 3. Given a set of binary decision variables, i.e., the edge matrix $\\mathbf{E} \\in \\{0,1\\}^{N' \\times N'}$ between $\\mathbf{X}^\\alpha$ , and $\\mathbf{X}^\\beta$ , bipartite soft matching is formulated as", + "bbox": [ + 75, + 632, + 468, + 828 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} ^ {*} = \\arg \\max _ {\\mathbf {E}} \\sum_ {i j} \\mathbf {w} _ {i j} ^ {\\prime} \\mathbf {e} _ {i j} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 843, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s u b j e c t} \\sum_ {i j} \\mathbf {e} _ {i j} = r, \\sum_ {j} \\mathbf {e} _ {i j} \\leq 1 \\forall i, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 873, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 500, + 292, + 544, + 304 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {i j} ^ {\\prime} = \\left\\{ \\begin{array}{l l} \\mathbf {w} _ {i j} & \\text {i f} j \\neq \\arg \\max _ {j ^ {\\prime}} \\mathbf {w} _ {i j ^ {\\prime}} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 314, + 890, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{e}_{ij}$ indicates the presence of the edge between $i, j$ -th token of $\\mathbf{X}^{\\alpha}, \\mathbf{X}^{\\beta}$ , and, $\\mathbf{w}_{ij} = \\mathbf{W}(\\mathbf{x}_i^\\alpha, \\mathbf{x}_j^\\beta)$ . This optimization problem can be solved by two simple steps: 1) find the best edge that maximizes $\\mathbf{w}_{ij}$ for each $i$ , and 2) choose the top- $r$ edges with the largest attraction scores. Then, based on the soft matching result $\\mathbf{E}^*$ , we group the tokens as", + "bbox": [ + 496, + 363, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {j} ^ {\\alpha \\rightarrow \\beta} = \\left\\{\\mathbf {x} _ {i} ^ {\\alpha} \\in \\mathbf {X} ^ {\\alpha} \\mid \\mathbf {e} _ {i j} = 1 \\right\\} \\cup \\left\\{\\mathbf {x} _ {j} ^ {\\beta} \\right\\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 464, + 890, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_i^{\\alpha \\rightarrow \\beta}$ indicates the set of tokens matched with $\\mathbf{x}_i^\\beta$ . Finally, the results of the fusion $\\tilde{\\mathbf{X}}$ are obtained as", + "bbox": [ + 496, + 494, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {X}} = \\tilde {\\mathbf {X}} ^ {\\alpha} \\cup \\tilde {\\mathbf {X}} ^ {\\beta}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 537, + 890, + 555 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\tilde {\\mathbf {X}} ^ {\\alpha} = \\mathbf {X} ^ {\\alpha} - \\bigcup_ {i} ^ {N ^ {\\prime}} \\mathbf {X} _ {i} ^ {\\alpha \\rightarrow \\beta}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 559, + 890, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {X}} ^ {\\beta} = \\bigcup_ {i} ^ {N ^ {\\prime}} \\left\\{\\delta \\left(\\mathbf {X} _ {i} ^ {\\alpha \\rightarrow \\beta}\\right)\\right\\}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 588, + 890, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\delta (\\mathbf{X}) = \\delta (\\{\\mathbf{x}_i\\} _i) = \\sum_i\\frac{\\mathbf{a}_i\\mathbf{s}_i\\mathbf{x}_i}{\\sum_{i'}\\mathbf{a}_{i'}\\mathbf{s}_{i'}}$ is the pooling operation considering the attention scores a and the size s of the tokens. Still, as shown in Step2 of Figure 3, the number of target tokens $\\mathbf{X}^{\\beta}$ cannot be reduced. To handle this issue, MCTF performs bidirectional bipartite soft matching by conducting the matching in the opposite direction with the updated token sets $\\hat{\\mathbf{X}}^{\\alpha}$ , and $\\hat{\\mathbf{X}}^{\\beta}$ as in Step 3, 4 of Figure 3. The final output tokens $\\hat{\\mathbf{X}} = \\hat{\\mathbf{X}}^{\\alpha}\\cup \\hat{\\mathbf{X}}^{\\beta}$ are defined with the following.", + "bbox": [ + 496, + 625, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {X}} ^ {\\alpha} = \\bigcup_ {i} ^ {N ^ {\\prime} - r} \\left\\{\\delta \\left(\\tilde {\\mathbf {X}} _ {i} ^ {\\beta \\rightarrow \\alpha}\\right)\\right\\}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 758, + 890, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {X}} ^ {\\beta} = \\tilde {\\mathbf {X}} ^ {\\beta} - \\bigcup_ {i} ^ {N ^ {\\prime} - r} \\tilde {\\mathbf {X}} _ {i} ^ {\\beta \\rightarrow \\alpha}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 787, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that calculating the pairwise weights with updated two sets of tokens $\\tilde{\\mathbf{w}}_{ij} = \\mathbf{W}(\\tilde{\\mathbf{x}}_i^\\beta ,\\tilde{\\mathbf{x}}_j^\\alpha)$ introduces the additional computational costs of $O(N^{\\prime}(N^{\\prime} - r))$ . To avoid this overhead, we approximate the attraction function by the attraction scores before fusion. In short, we just reuse the", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15744", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1efd23ccb0b313862831b3fdb49f340803b9b2724c0de026249128d145b79577.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 89, + 173, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/59ea783f68764fae6e21c35b758f3ae1726e2bf6a52391ab04ef35f15b40c197.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 90, + 271, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/99df14cd404d71f06b1341db84ff09835c40033236f720e376c46c068fa4eece.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 90, + 370, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5206213c74a72aa3df8b6a2ac0ea81103293bdbcf10101d41fd45eca7b50fba3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 90, + 467, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1b0f0a9ad6831157ff29dfd8c8e311b5d41e5186b2abd6c31db8d01892b96c68.jpg", + "image_caption": [ + "Figure 4. Visualization of attentiveness in consecutive layers.", + "Figure 5. Illustration of attention map in the consecutive layers and approximated attention. (Left) The attention score $\\hat{\\mathbf{A}}^l$ is the past influence of the tokens to generate $\\mathbf{X}^l$ . If we fuse the tokens $\\mathbf{X}^l$ based on $\\mathbf{A}^l$ , $\\mathbf{x}_1$ is prone to be fused despite the highest informativeness score in the following attention. So, we instead leverage the informativeness based on the one-step-ahead attention $\\mathbf{A}^{l + 1}$ . (Right) After the fusion, we also aggregate the $\\mathbf{A}^{l + 1}$ to approximate the attention map $\\hat{\\mathbf{A}}^{l + 1}$ for updating fused tokens $\\hat{\\mathbf{X}}^l$ ." + ], + "image_footnote": [], + "bbox": [ + 76, + 200, + 470, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pre-calculated weights since $\\tilde{\\mathbf{X}}^{\\alpha}$ is the subset of $\\mathbf{X}^{\\alpha}$ . This allows MCTF to efficiently reduce tokens considering bidirectional relations between two subsets with negligible extra costs compared to uni-directional bipartite soft matching.", + "bbox": [ + 75, + 444, + 468, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. One-step-ahead attention for informativeness", + "text_level": 1, + "bbox": [ + 76, + 513, + 460, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In assessing informativeness, prior works [17, 18, 21] have leveraged the attention scores from the previous self-attention layer. As illustrated in Figure 5, previous approaches use the attention $\\mathbf{A}^l$ from the previous layer to fuse tokens $\\mathbf{X}^l$ . This technique allows efficient assessment under the assumption that the attention maps in consecutive layers are similar. However, we observed that the attention maps often substantially differ, as shown in Figure 4, and the attention from a previous layer may lead to suboptimal token fusion. Thus, we proposed one-step-ahead attention, which measures the informativeness of tokens based on the attention map in the next layer, i.e., $\\mathbf{A}^{l + 1}$ . Then, the informativeness scores $\\mathbf{a}$ in Equation (4) is calculated with $\\mathbf{A}^{l + 1}\\in \\mathbb{R}^{N\\times N}$ . This simple remedy provides a considerable improvement; see Figure 7b in Section 4.2. After token fusion, we efficiently compute the attention map $\\hat{\\mathbf{A}}^{l + 1}\\in \\mathbb{R}^{(N - r)\\times (N - r)}$ of fused tokens $\\hat{\\mathbf{X}}^l\\in \\mathbb{R}^{(N - r)\\times C}$ by simply aggregating $\\mathbf{A}^{l + 1}\\in \\mathbb{R}^{N\\times N}$ without recomputing the dot-product self-attention. To be specific, when the tokens are fused as $\\delta (\\{\\mathbf{x}_i\\}_i)$ during Equations (10) to (14), their corresponding one-step-ahead attention scores are also fused as $\\delta (\\{\\mathbf{A}_i^{l + 1}\\}_i)$ in both query and key direction. Note that when fusing attention scores for queries we use simple sum for $\\delta ,i.e.$ $\\forall_{i}\\sum_{j}\\hat{\\mathbf{A}}_{ij}^{l + 1} = 1$ . For fusing attention", + "bbox": [ + 75, + 537, + 472, + 904 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/68a1062d07f2665f5fa55a6beae484cf46cc2ef602536b54cc1a028426f376e9.jpg", + "image_caption": [ + "Figure 6. Illustration of training with token reduction consistency. During training, we forward the input $x$ as $f(x; r)$ , and $f(x; r')$ , respectively. To obtain the augmented representation, $r'$ is randomly selected in every step, and the model is updated with supervisory signals $\\mathcal{L}_{\\mathrm{CE}}$ , and consistency loss $\\mathcal{L}_{\\mathrm{MSE}}$ ." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 890, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "scores for queries, we use simple sum for $\\delta$ to guarantee $\\forall_{i}\\sum_{j}\\hat{\\mathbf{A}}_{ij}^{l + 1} = 1.$", + "bbox": [ + 498, + 362, + 890, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Token reduction consistency", + "text_level": 1, + "bbox": [ + 500, + 410, + 751, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We here propose a new fine-tuning scheme to further improve the performance of vision Transformer $f_{\\theta}(\\cdot ;r)$ with MCTF. We observe that a different number of reduced tokens per layer, denoted as $r$ , may lead to different representations of samples. By training Transformers with different $r$ and encouraging the consistency between them, namely, token reduction consistency, we achieve the additional performance gain. The objective function of our method is given as", + "bbox": [ + 496, + 435, + 893, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} \\left(f _ {\\theta} (x; r), y\\right) + \\mathcal {L} _ {\\mathrm {C E}} \\left(f _ {\\theta} \\left(x; r ^ {\\prime}\\right), y\\right) \\\\ + \\lambda \\mathcal {L} _ {\\mathrm {M S E}} \\left(\\mathbf {x} _ {r} ^ {\\mathrm {c l s}}, \\mathbf {x} _ {r ^ {\\prime}} ^ {\\mathrm {c l s}}\\right), \\tag {15} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 573, + 890, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(x,y)$ is a supervised sample, $r,r^{\\prime}$ is the fixed and dynamic reduced token numbers, $\\lambda$ is the coefficient for consistency loss, and $\\mathbf{x}_r^{\\mathrm{cls}},\\mathbf{x}_{r'}^{\\mathrm{cls}}$ are the class tokens in the last layer of models $f_{\\theta}(x;r),f_{\\theta}(x;r')$ . In this objective, we first calculate the cross-entropy loss $\\mathcal{L}_{\\mathrm{CE}}(f_{\\theta}(x;r),y)$ with fixed $r$ , which is the target reduction number that will be used in the evaluation. At the same time, we generate another representation of the input $x$ with smaller but randomly drawn $r^\\prime \\sim$ uniform $(0,r)$ , and calculate the loss $\\mathcal{L}_{\\mathrm{CE}}(f_{\\theta}(x;r'),y)$ Then, we impose the token consistency loss $\\mathcal{L}_{\\mathrm{MSE}}(\\mathbf{x}_r^{\\mathrm{cls}},\\mathbf{x}_{r'}^{\\mathrm{cls}})$ on the class tokens, to retain the consistent representation across the diverse reduced token numbers $r^\\prime$ . The proposed method can be viewed as a new type of token-level data augmentation [7, 19] and consistency regularization. Our token reduction consistency encourages the representation $\\mathbf{x}_r^{\\mathrm{cls}}$ obtained by the target reduction number $r$ to mimic the slightly augmented representation $\\mathbf{x}_{r'}^{\\mathrm{cls}}$ , which is more similar to ones with no token reduction since $r^{\\prime} < r$", + "bbox": [ + 496, + 628, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15745", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/de693853223cf1de64d9c40cc1486f52c134ddff46a9582307a30eb947ec2600.jpg", + "table_caption": [ + "Table 1. Image classification results" + ], + "table_footnote": [], + "table_body": "
MethodFLOPs (G)Params (M)Top-1 Acc (%)
DeiT-T [28]1.2572.2 (-)
+EvoViT[AAAI '22] [36]0.8572.0 (-0.2)
+A-ViT[CVPR '22] [37]0.8571.0 (-1.2)
+SPViT[ECCV '22] [17]0.9572.1 (-0.1)
+ToMe[ICLR '23] [4]0.7571.3 (-0.9)
+BAT[CVPR '23] [21]0.8572.3 (+0.1)
+MCTFr=160.7572.7 (+0.5)
DeiT-S [28]4.62279.8 (-)
+IA-RED2[NeurIPS '21] [24]3.22279.1 (-0.7)
+DynamicViT[NeurIPS '21] [26]2.92379.3 (-0.5)
+EvoViT[AAAI '22] [36]3.02279.4 (-0.4)
+EViT[ICLR '22] [18]3.02279.5 (-0.3)
+A-ViT[CVPR '22] [37]3.62278.6 (-1.2)
+ATS[ECCV '22] [12]2.92279.7 (-0.1)
+SPViT[ECCV '22] [17]2.62279.3 (-0.5)
+ToMe[ICLR '23] [4]2.72279.4 (-0.4)
+BAT[CVPR '23] [21]3.02279.6 (-0.2)
+MCTFr=162.62280.1 (+0.3)
", + "bbox": [ + 78, + 108, + 470, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 458, + 209, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. To validate the effectiveness of the proposed methods, we compare MCTF with the previous token reduction methods. For comparison, we opt the token pruning methods (A-ViT [37], IA-RED² [24], DynamicViT [26], EvoViT [36], ATS [12]) and token fusion methods (SPViT [17], EViT [18], ToMe [4], BAT [21]) in DeiT [28], and report the efficiency (FLOPs (G)) and the performance (Top-1 Acc (\\%)) of each method. Further, to validate MCTF on other Vision Transformers (T2T-ViT [39], LV-ViT [15]), we report the results of MCTF and compare it with the official number of existing works. We denote the number of reduced tokens per layer $r$ with the subscript in Tables 1 and 2. The gray color in the table indicates the base model, and the green and red color indicates the improvements and degradations of the performance compared to the base model, respectively.", + "bbox": [ + 75, + 484, + 470, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 739, + 279, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison of the token reduction methods. The comparison with existing token reduction methods is summarized in Table 1. We demonstrate that our MCTF achieves the best performance with the lowest FLOPs in DeiT [28] surpassing all previous works. Further, it is worth noting that MCTF is the only work that avoids performance degradation with the lowest FLOPs in both DeiT-T and DeiT-S. Through Finetuning DeiT-T for 30 epochs, MCTF brings a significant gain of $+0.5\\%$ in accuracy over the base model with nearly half", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b3fda42c58ff08306cbd9a258474eda445e0262d530b4be488603b9a690650df.jpg", + "table_caption": [ + "Table 2. Comparison with other Vision Transformers" + ], + "table_footnote": [], + "table_body": "
ModelsFLOPs (G)Params (M)Acc (%)
PVT-Small[31]3.824.579.8
PVT-Medium [31]6.744.281.2
CoaT Mini [35]6.810.080.8
CoaT-Lite Small [35]4.020.081.9
Swin-T [20]4.529.081.3
Swin-S [20]8.750.083.0
PoolFormer-S36 [38]5.031.081.4
PoolFormer-M48 [38]11.673.082.5
T2T-ViTt-14 [39]6.121.581.7
+MCTFr=134.221.581.8 (↑)
T2T-ViTt-19 [39]9.839.282.4
+MCTFr=96.439.282.4 (-)
LV-ViT-S [15]6.626.283.3
+EVIT(ICLR '22] [18]4.726.283.0 (↓)
+BAT[CVPR '23] [21]4.726.283.1 (↓)
+DynamicViT[NeurIPS '21] [26]4.626.983.0 (↓)
+SPViT[ECCV '22] [17]4.326.283.1 (↓)
+MCTFr=124.226.283.4 (↑)
", + "bbox": [ + 501, + 108, + 890, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FLOPs. Similarly, we observe a gain of $+0.3\\%$ with DeiT-S while boosting the FLOPs by $-2.0$ (G). We believe that multi-criteria with one-step-ahead attention helps the model to minimize the loss of information; further consistency loss on the class token through the token reduction improves the generalizability of the model.", + "bbox": [ + 496, + 457, + 893, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MCTF with other Vision Transformers. To validate the applicability of MCTF in various ViTs, we demonstrate MCTF with other transformer architectures in Table 2. Following previous works [17, 18, 21, 26], we apply MCTF with LVViT. Also, we present the results of MCTF with T2T-ViT. As presented in the table, our experimental results are promising. MCTF in these architectures gets at least $31\\%$ speedup without performance degradation. Further, MCTF combined with LV-ViT outperforms all other Transformers and token reduction methods regarding FLOPs, and accuracy. Especially, it is worth noting that all token reduction methods except for MCTF bring the performance degradation in LVViT. These results reveal that MCTF is the efficient token reduction method for the diverse Vision Transformers.", + "bbox": [ + 496, + 550, + 895, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Token reduction without training. Similar to ToMe [4], MCTF is applicable with pre-trained ViTs without any additional training since MCTF does not require any learnable parameters. We here apply the two reduction methods to the pre-trained DeiT without finetuning and provide the results in Table 3. Regardless of the reduced number of tokens $r$ in each layer, MCTF consistently surpasses ToMe. Especially, in the most sparse setting $r = 20$ , the performance gap is significant ( $+7.0\\%$ in DeiT-T, $+3.8\\%$ in DeiT-S). Note", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15746", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d0626cb853e7b60ab729c17db8b9c741ba4b3d0a046296ae10bed2ad55a86723.jpg", + "image_caption": [ + "Figure 7. Ablations on (a) multi-criteria, (b) one-step-ahead-attention, and token reduction consistency. Each marker indicates the model with $r \\in [1,20]$ , and we highlight $r \\in \\{5,10,15,20\\}$ as bordered circle. We also denote the model as star when $r = 16$ , which is used for finetuning the model." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 441, + 314 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14a979a63359a92a0d6246733b9e8852dc305f08f05e303a29162dd01f72d07d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 95, + 883, + 314 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0c9375cae11861ed99e8b3641386a4db2f38e22eb12b8b755bf6d0de19f9e0b9.jpg", + "table_caption": [ + "Table 3. Image classification results without training" + ], + "table_footnote": [], + "table_body": "
Methodr
Base1248121620
DeiT-T
ToMe [4]72.272.172.072.071.670.868.761.5
MCTF72.272.272.172.172.071.771.068.5
DeiT-S
ToMe [4]79.879.879.779.779.479.077.974.2
MCTF79.879.879.879.879.879.679.278.0
", + "bbox": [ + 78, + 386, + 467, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that without any additional training, our $\\mathrm{MCTF}_{r=16}$ with pre-trained DeiT-S still shows a competitive performance of $79.2\\%$ compared to reduction methods requiring training (e.g., $78.6\\%$ of A-ViT, $79.1\\%$ of IA-RED $^2$ , and $79.3\\%$ of DynamicViT, and SPViT in Table 1).", + "bbox": [ + 75, + 546, + 470, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation studies on MCTF", + "text_level": 1, + "bbox": [ + 76, + 633, + 318, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide ablation studies to validate each component of MCTF. Unless otherwise stated, we conduct whole experiments with DeiT-S finetuned with MCTF ( $r = 16$ ). We provide the FLOPs-Accuracy graph by adjusting the reduced number of tokens per layer $r \\in [1,20]$ .", + "bbox": [ + 75, + 657, + 468, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-criteria. We explore the effectiveness of multi-criteria in Figure 7a. First, regarding the multi-criteria, we utilize three criteria for MCTF, i.e., similarity (sim.), informativeness (info.), and size. Each single criterion of similarity and informativeness shows a relatively inferior performance compared to dual (sim. & info.) and multi-criteria (sim. & info. & size). Specifically, when $r = 16$ , the performance of a single criterion is $79.7\\%$ , and $79.4\\%$ with similarity and informativeness, respectively. Then, adopting dual criteria (sim. & info.), MCTF achieves $79.8\\%$ . Finally, we get an accuracy of $80.1\\%$ with a gain of $+0.3\\%$ by respecting all", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "three criteria (sim. & info. & size). These performance gaps get larger as $r$ increases, which proves the importance of the multi-criteria for token fusion.", + "bbox": [ + 496, + 369, + 890, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "One-step-ahead attention and token reduction consistency. To show the validity of one-step-ahead attention and token reduction consistency, we also provide the results of MCTF with and without each component in Figure 7b. When eliminating either one-step-ahead attention or token reduction consistency, the accuracies are dropped in every FLOP. This significant drop indicates that both approaches matter for MCTF. In short, by adopting one-step-ahead attention and token reduction consistency, MCTF effectively mitigates the performance degradation in a wide range of FLOPs.", + "bbox": [ + 496, + 416, + 893, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of design choices. The ablations on design choices are presented in Table 4. First, our bidirectional bipartite matching, which enables capturing the bidirectional relation in two sets, enhances the accuracy compared to one-way bipartite matching. Next, for pooling operation $\\delta$ , the weighted sum considering the size $s$ and attentiveness $a$ is a better choice than others like max-pool or average. Lastly, we compare the results with the precise and approximated attention for $\\hat{\\mathbf{A}}^l$ . For precise attention, we just conduct the similarity calculation for one-step-ahead attention and the attention in the self-attention layer after fusion, separately. Otherwise, we approximate it with one-step-ahead attention as described in Section 3.3. As presented in the table, our approximated attention maintains the performance with the substantial improvement in efficiency (-0.4 (G) FLOPs).", + "bbox": [ + 496, + 566, + 893, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analyse of MCTF", + "text_level": 1, + "bbox": [ + 500, + 801, + 671, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. For a better understanding of MCTF, we provide the qualitative results of MCTF in Figure 8. We visualize the fused tokens at the last block of DeiT-S on ImageNet-1K and denote the fused tokens by the same border color. As shown in the figure, since the tokens are merged", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15747", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4afb32ad419c3fe1b70f201ed7fd7bb8d8f34b37a36121c9d3f0cc1473a41798.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 171, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a7af67b7dff5a9a184c724a23658843b70f2c386b9abe9aad747cde52cc0bf70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 88, + 274, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/16e8435975dad32dea31b872d46f7f5bcac044f0edceb4ca319eb878667bf030.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 88, + 379, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6efb986832f72af6d0b7121735852b9ed8049f6f2b69570d354f47487483cdb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 88, + 480, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/49851728a0fed6734a5653551e1148a59857223fff9303ea43adb2d4db91361a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 88, + 583, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f60e0ff937b8bf353c7c02c965d198f6c9426570f202c1384ebaaa28a411ed7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 89, + 684, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/074f72928dd8ff2f3faae726dfdc8a6d58953ccc68ab5bebfa8a52f864fb930b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 88, + 787, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4ed23b403ec7dd12409d09838c3f3faddf2138b4f05e96ce10c16036149a95e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 88, + 890, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c70c77a25b8cdc52e1440a7bf0e80025cdcd67bb8b0b6d8cbfa1d423c0d69e1.jpg", + "image_caption": [ + "Figure 8. Visualization of the fused tokens with MCTF. Given the input images of ImageNet-1K (Top), the qualitative results of MCTF with DeiT-S are provided at the bottom. The same border color of the patches indicates the fused tokens." + ], + "image_footnote": [], + "bbox": [ + 80, + 172, + 171, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b4778632b1db8ab333fb7c8584d5dd0ffd43d9b20816691c2420b71db2d3be3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 172, + 274, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e7645866783e0d1187c2758d549d0fe9fc6cf2a3819135065ab030e51d3a5510.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 172, + 379, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c90f29dbe07f3c44d77a75c67fea121eb4bfa54f825644dd11c11e011a983732.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 172, + 480, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f6ef5e81f64d1587a7ed67f95bfffedcf8ed205fbfa5e05f000a7ef2fa000bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 172, + 583, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d19e33361ea236cea5de533f7f20db979196958c0194cb91a05a2727bca37ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 172, + 684, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb064ac90519e417a61cd0e87b153a55c8f261e3fc1c9d0fe6f32b2b28ba8db4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 172, + 787, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d99bb75710b808bd952169ccfe6cf094e5115c09973b59c11439c1b3654a574.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 172, + 890, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f48d4114f8cb2f3d917835f0e3727e983dffd619fe05ecdf81782a3b1f76fff7.jpg", + "table_caption": [ + "Table 4. Ablations of the design choices." + ], + "table_footnote": [], + "table_body": "
MethodFLOPs ↓ (G)Acc ↑ (%)
DeiT-S4.679.8
bipartite soft matching
One-way2.680.0
Bidirectional2.680.1
pooling function δ
average2.680.0
max2.679.8
weighted average2.680.1
approximation of attention map
precise attention3.080.1
approximated attention2.680.1
", + "bbox": [ + 96, + 314, + 446, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with multi-criteria (e.g., similarity, informativeness, size), we maintain the more diverse tokens in the informative foreground object. For instance, in the third image of the hamster, while the background patches including the hand are fused into one token, the foreground tokens are less fused while maintaining the details like the eye, ear, and face of the hamster. In short, compared to the background, the foreground tokens are less fused with the moderate size retaining the information of the main content.", + "bbox": [ + 75, + 559, + 468, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Soundness of size criterion. Figure 9 presents the histogram of sizes of tokens after token reduction with and without size criterion. Specifically, we measure the size of the largest token at the last block and provide the histogram. With our size criterion, the merged tokens tend to have smaller sizes s showing the average size of 39.3/49.2 with and without the Size criterion, respectively. As intended, MCTF successfully suppresses the large-sized tokens, which are a source of information loss, leading to performance improvement.", + "bbox": [ + 75, + 695, + 470, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced the Multi-Criteria Token Fusion (MCTF), a novel strategy aimed at reducing the complex", + "bbox": [ + 75, + 869, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/213dabcea98cb5373ee53e8066e805dda7e83b670889a3b52385efb5f4cd4343.jpg", + "image_caption": [ + "Figure 9. Histogram of the size of tokens after reduction." + ], + "image_footnote": [], + "bbox": [ + 522, + 295, + 870, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ity inherent in ViTs while mitigating performance degradation. MCTF effectively discerns the relation of tokens based on multiple criteria, including similarity, informativeness, and the size of the tokens. Our comprehensive ablation studies and detailed analyses demonstrate the efficacy of MCTF particularly with our innovative one-step-ahead attention and token reduction consistency. Remarkably, DeiT-T and DeiT-S with MCTF achieve considerable improvements, with $+0.5\\%$ , and $+0.3\\%$ increase in Top-1 Accuracy over the vanilla models, accompanied by about $44\\%$ fewer FLOPs, respectively. We also observe that our MCTF outperforms all of the previous token reduction methods in diverse vision Transformers with and without training.", + "bbox": [ + 496, + 487, + 893, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 717, + 658, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by ICT Creative Consilience Program through the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT)(IITP-2024-2020-0-01819), the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT)(NRF-2023R1A2C2005373), and a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI) funded by the Ministry of Health & Welfare Republic of Korea (HR20C0021).", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15748", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Moab Arar, Ariel Shamir, and Amit H Bermano. Learned queries for efficient local attention. In CVPR, 2022. 1, 2", + "[2] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In ECCV, 2022. 2", + "[3] Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv:2004.05150, 2020. 1", + "[4] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. ICLR, 2022. 1, 2, 3, 4, 6, 7", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1", + "[6] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, 2021. 2", + "[7] Hyeong Kyu Choi, Joonmyung Choi, and Hyunwoo J. Kim. Tokenmixup: Efficient attention-guided token-level data augmentation for transformers. In NeurIPS, 2022. 5", + "[8] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. *ICLR*, 2021. 2", + "[9] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. NeurIPS, 2021. 1, 2", + "[10] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, 2022. 2", + "[11] 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. ICLR, 2020. 1, 2", + "[12] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In ECCV, 2022. 1, 2, 3, 6", + "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2", + "[14] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, 2021. 2", + "[15] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. NeurIPS, 2021. 2, 6", + "[16] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. In ICLR, 2020. 1, 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In ECCV, 2022. 1, 2, 3, 5, 6", + "[18] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. ICLR, 2022. 1, 2, 3, 5, 6", + "[19] Jihao Liu, Boxiao Liu, Hang Zhou, Hongsheng Li, and Yu Liu. Tokenmix: Rethinking image mixing for data augmentation in vision transformers. In ECCV, 2023. 5", + "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 6", + "[21] Sifan Long, Zhen Zhao, Jimin Pi, Shengsheng Wang, and Jingdong Wang. Beyond attentive tokens: Incorporating token importance and diversity for efficient vision transformers. In CVPR, 2023. 1, 2, 3, 5, 6", + "[22] Dmitrii Marin, Jen-Hao Rick Chang, Anurag Ranjan, Anish Prabhu, Mohammad Rastegari, and Oncel Tuzel. Token pooling in vision transformers for image classification. In WACV, 2023. 1, 2, 3", + "[23] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In CVPR, 2022. 1, 2, 3", + "[24] Bowen Pan, Rameswar Panda, Yifan Jiang, Zhangyang Wang, Rogerio Feris, and Aude Oliva. IA-RED2: Interpretability-aware redundancy reduction for vision transformers. NeurIPS, 2021. 1, 2, 3, 6", + "[25] 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 ICML, 2021. 1", + "[26] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. NeurIPS, 2021. 1, 2, 3, 6", + "[27] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, 2021. 1", + "[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 1, 2, 6", + "[29] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, 2021. 1, 2", + "[30] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv:2006.04768, 2020. 1, 2", + "[31] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense predic" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15749", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 1, 2, 6", + "[32] QuanLin Wu, Hang Ye, Yuntian Gu, Huishuai Zhang, Liwei Wang, and Di He. Denoising masked autoencoders help robust classification. In ICLR, 2023. 2", + "[33] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS, 2021. 1", + "[34] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. In AAAI, 2021. 1, 2", + "[35] Weijian Xu, Yifan Xu, Tyler Chang, and Zhuowen Tu. Coscale conv-attentional image transformers. In ICCV, 2021. 6", + "[36] Yifan Xu, Zhijie Zhang, Mengdan Zhang, Kekai Sheng, Ke Li, Weiming Dong, Liqing Zhang, Changsheng Xu, and Xing Sun. Evo-vit: Slow-fast token evolution for dynamic vision transformer. In AAAI, 2022. 6", + "[37] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In CVPR, 2022. 1, 2, 3, 6", + "[38] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 6", + "[39] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In ICCV, 2021. 2, 6", + "[40] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In CVPR, 2022. 2", + "[41] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. *ICLR*, 2021. 1" + ], + "bbox": [ + 78, + 90, + 470, + 588 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15750", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/c38e4100-b4f3-4d92-aff9-1fdadd7cfa0a_model.json b/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/c38e4100-b4f3-4d92-aff9-1fdadd7cfa0a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..989a6cfd3da3c4fa7ddde9344da9edf0639c5b7c --- /dev/null +++ b/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/c38e4100-b4f3-4d92-aff9-1fdadd7cfa0a_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.784, + 0.175 + ], + "angle": 0, + "content": "Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.203, + 0.723, + 0.223 + ], + "angle": 0, + "content": "Sanghyeok Lee* Joonmyung Choi* Hyunwoo J. Kim†" + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.236, + 0.757, + 0.254 + ], + "angle": 0, + "content": "Department of Computer Science and Engineering, Korea University" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.256, + 0.668, + 0.272 + ], + "angle": 0, + "content": "{cat0626, pizard, hyunwoojkim}@korea.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.306, + 0.314, + 0.323 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.339, + 0.474, + 0.716 + ], + "angle": 0, + "content": "Vision Transformer (ViT) has emerged as a prominent backbone for computer vision. For more efficient ViTs, recent works lessen the quadratic cost of the self-attention layer by pruning or fusing the redundant tokens. However, these works faced the speed-accuracy trade-off caused by the loss of information. Here, we argue that token fusion needs to consider diverse relations between tokens to minimize information loss. In this paper, we propose a Multi-criteria Token Fusion (MCTF), that gradually fuses the tokens based on multi-criteria (i.e., similarity, informativeness, and size of fused tokens). Further, we utilize the one-step-ahead attention, which is the improved approach to capture the informativeness of the tokens. By training the model equipped with MCTF using a token reduction consistency, we achieve the best speed-accuracy trade-off in the image classification (ImageNet1K). Experimental results prove that MCTF consistently surpasses the previous reduction methods with and without training. Specifically, DeiT-T and DeiT-S with MCTF reduce FLOPs by about \\(44\\%\\) while improving the performance \\((+0.5\\%, \\text{and} +0.3\\%)\\) over the base model, respectively. We also demonstrate the applicability of MCTF in various Vision Transformers (e.g., T2T-ViT, LV-ViT), achieving at least \\(31\\%\\) speedup without performance degradation. Code is available at https://github.com/mlvlab/MCTF." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.747, + 0.21, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.471, + 0.864 + ], + "angle": 0, + "content": "Vision Transformer [11] (ViT) has been proposed to tackle the vision tasks with self-attention, originally developed for natural language processing tasks. With the advent of ViT, Transformers are the prevalent architectures for a wide range of vision tasks, e.g., classification [11, 20, 28, 29, 31], object detection [5, 31, 41], segmentation [20, 27, 33], etc. ViTs," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.308, + 0.895, + 0.474 + ], + "angle": 0, + "content": "built only with self-attention and MLP, provide great flexibility and impressive performance compared to conventional approaches, e.g., convolutional neural networks (CNNs). However, despite these advantages, the quadratic computational complexity of self-attention with respect to the number of tokens is the major bottleneck for Transformers. This limitation becomes more substantial with the growing interest in large-scale foundation models such as CLIP [25]. To this end, several works [3, 16, 30, 34] have proposed efficient self-attention mechanisms including local self-attention within predefined windows [1, 9, 20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.895, + 0.729 + ], + "angle": 0, + "content": "More recently, there has been increasing interest in token-reduction methods for optimizing ViTs without altering their architecture. Earlier works [12, 23, 24, 26, 37] primarily focused on pruning the uninformative tokens to reduce the number of tokens. Another line of works [4, 17, 18, 21, 22] attempted to fuse the tokens instead of discarding them to minimize the information loss. However, performance degradation is still commonly observed in most token fusion methods. We notice that the token fusion methods usually consider only one criterion, such as the similarity or informativeness of tokens, leading to suboptimal token matching. For instance, similarity-based token fusion is prone to combine the foreground tokens, whereas informativeness-based fusion often merges substantially dissimilar tokens, resulting in collapsed representations. Furthermore, if too many tokens are fused into one token, then information loss is inevitable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.897, + 0.897 + ], + "angle": 0, + "content": "To address the problems, we introduce Multi-Criteria Token Fusion (MCTF) that optimizes vision transformers by fusing tokens based on multi-criteria. Unlike previous works that consider a single criterion for token fusion, MCTF measures the relationship between the tokens with multi-criteria as follows; (1) similarity to fuse the redundant tokens, (2) informativeness to reduce the uninformative tokens, (3) the size of the tokens to prevent the large-sized tokens that boost the loss of information. Also, to tackle the inconsistency between attention maps of consecutive layers, we adopt one-step-ahead attention, which explicitly estimates the informativeness of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15741" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.092, + 0.891, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.897, + 0.39 + ], + "angle": 0, + "content": "Figure 1. Comparison of the token reduction methods with DeiT-T (left), and DeiT-S (right). Given a base model marked as blue circle, previous token reduction methods accelerate the speed with the trade-off between accuracy and computational cost. Our MCTF, marked as a star, even brings performance improvements while lessening the complexity of DeiT. Note that after only one finetuning with the specific reduced number of tokens marked as red star, we simply evaluate it with the diverse FLOPs by adjusting the reduced numbers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.472, + 0.553 + ], + "angle": 0, + "content": "the tokens in the next layer. Finally, by introducing a token reduction consistency for finetuning the model, we achieve superior performance to the existing works as in Figure 1. Surprisingly, our MCTF even performs better than the 'full' base model (red dotted line) with a reduced computational complexity. Specifically, it brings a \\(0.5\\%\\), and \\(0.3\\%\\) gain while reducing FLOPs by about \\(44\\%\\) in DeiT-T, and DeiT-S [28], respectively. We have observed a similar speed-up \\((31\\%)\\) in T2T-ViT [39], and LV-ViT [15] without any performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.553, + 0.384, + 0.566 + ], + "angle": 0, + "content": "Our contributions are summarized in fourfold." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.569, + 0.47, + 0.628 + ], + "angle": 0, + "content": "- We propose Multi-criteria Token Fusion, a novel token fusion method that considers multi-criteria, e.g., similarity, informativeness, and size, to capture the complex relationship of tokens and minimize information loss." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.63, + 0.469, + 0.675 + ], + "angle": 0, + "content": "- For measuring the informativeness of the tokens, we utilize one-step-ahead attention to retain the attentive tokens in the following layers." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.675, + 0.47, + 0.719 + ], + "angle": 0, + "content": "- We propose a new fine-tuning scheme with token reduction consistency to boost the generalization performance of transformers equipped with MCTF." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.72, + 0.47, + 0.765 + ], + "angle": 0, + "content": "- The extensive experiments demonstrate that MCTF achieves the best speed-accuracy trade-off in diverse ViTs, surpassing all previous token reduction methods." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.569, + 0.47, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.784, + 0.222, + 0.799 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Vision Transformers. Vision Transformer [11] is introduced to tackle the vision tasks. Later, DeiT [28] and CaiT [29] are proposed to handle the data efficiency and scalability of ViT, respectively. Recent works [6, 10, 14, 20, 31] tried to insert the inductive biases of CNNs on ViT, such as the locality or pyramid-architecture. In parallel, there is a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.895, + 0.584 + ], + "angle": 0, + "content": "line of works that boosts the vanilla ViT by scaling [29, 40] or self-supervised learning [2, 13, 32]. Despite the promising results of these works, the quadratic complexity of ViTs is still the major constraint for scaling the model. For the sake of mitigating the complexity, Reformer [16] lessens the quadratic complexity to \\( O(N \\log N) \\) through the hashing function, and Linformer [30], performer [8], and Nyströmformer [34] achieve the linear cost with the approximated linear attention. Also, several works [1, 9, 10, 20] utilize sparse attention with the reduced key or query. Swin [20] and Twins [9] utilize the local attention within the fixed size of the window to mitigate the complexity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Token reduction in ViTs. Most of the computational burden in ViTs arises from the self-attention. To reduce the quadratic cost in the number of tokens, recent works [4, 12, 17, 18, 21–24, 26, 37] have an interest in reducing the token itself. These works have the advantage of utilizing the original ViTs architecture without modification. In earlier works [12, 23, 24, 26, 37], the uninformative tokens are simply dropped during the forward process, leading to the information loss. To compensate for this, SPViT [17] and EViT [18] first split the tokens into informative and uninformative token sets based on attention scores, then fuse these uninformative token sets into a single token. In parallel, token pooling [22] and ToMe [4] combine the semantically similar tokens to reduce redundancies. A more recent study BAT [21] first split the tokens based on informativeness then fuse the tokens considering the diversity of the tokens. Despite the advantage of each criterion, successful integration of multi-criteria is still less explored." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15742" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.262, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.231, + 0.198, + 0.244 + ], + "angle": 0, + "content": "(a) Origin" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.089, + 0.473, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.231, + 0.406, + 0.242 + ], + "angle": 0, + "content": "(b) \\(\\mathbf{W}^{\\mathrm{sim}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.089, + 0.682, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.23, + 0.638, + 0.242 + ], + "angle": 0, + "content": "(c) \\(\\mathbf{W}^{\\mathrm{sim}}\\& \\mathbf{W}^{\\mathrm{info}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.089, + 0.892, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.23, + 0.871, + 0.242 + ], + "angle": 0, + "content": "(d) \\(\\mathbf{W}^{\\mathrm{sim}}\\& \\mathbf{W}^{\\mathrm{info}}\\& \\mathbf{W}^{\\mathrm{size}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.256, + 0.894, + 0.312 + ], + "angle": 0, + "content": "Figure 2. Visualization of the fused tokens. Given (a) the leftmost image, (b) fusing the tokens with a single criterion \\(\\mathbf{W}^{\\mathrm{sim}}\\) often results in the excessive fusion of the foreground object. (c) Then considering both similarity and informativeness (\\(\\mathbf{W}^{\\mathrm{sim}}\\& \\mathbf{W}^{\\mathrm{info}}\\)), tokens in the foreground objects are less fused while the tokens in the background are largely fused. (d) Finally, MCTF helps retain the information of each component in the image by preventing the large-size token with the multi-criteria (\\(\\mathbf{W}^{\\mathrm{sim}}\\& \\mathbf{W}^{\\mathrm{info}}\\& \\mathbf{W}^{\\mathrm{size}}\\))." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.337, + 0.168, + 0.352 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.471, + 0.439 + ], + "angle": 0, + "content": "We first review the self-attention and token reduction approaches (Section 3.1). Then, we present our multi-criteria token fusion (Section 3.2) that leverages one-step-ahead attention (Section 3.3). Lastly, we introduce a training strategy with token reduction consistency in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.45, + 0.218, + 0.463 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.473, + 0.503 + ], + "angle": 0, + "content": "In Transformers, tokens \\(\\mathbf{X} \\in \\mathbb{R}^{N \\times C}\\) are processed by self-attention defined as" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.512, + 0.47, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S A} (\\mathbf {X}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {C}}\\right) \\mathbf {V}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.473, + 0.791 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q},\\mathbf{K},\\mathbf{V} = \\mathbf{X}\\mathbf{W}_{\\mathbf{Q}},\\mathbf{X}\\mathbf{W}_{\\mathbf{K}},\\mathbf{X}\\mathbf{W}_{\\mathbf{V}}\\) and \\(\\mathbf{W}_{\\mathbf{Q}},\\mathbf{W}_{\\mathbf{K}}\\) \\(\\mathbf{W}_{\\mathbf{V}}\\in \\mathbb{R}^{C\\times C}\\) are learnable weight matrices. Despite its outstanding expressive power, the self-attention does not scale well with the number of tokens \\(N\\) due to its quadratic time complexity \\(O(N^{2}C + NC^{2})\\). To address this problem, a line of works [12, 23, 24, 26, 37] reduces the number of tokens simply by pruning uninformative tokens. These approaches often cause significant performance degradation due to the loss of information. Thus, another line of works [4, 17, 18, 21, 22] fuses the uninformative or redundant tokens \\(\\hat{\\mathbf{X}}\\subset \\mathbf{X}\\) into a new token \\(\\hat{\\mathbf{x}} = \\delta (\\hat{\\mathbf{X}})\\), where \\(\\mathbf{X}\\) is the set of original tokens, and \\(\\delta\\) denotes a merging function, e.g., max-pooling or averaging. In this work, we also adopt 'token fusion' rather than 'token pruning' with multiple criteria to minimize the loss of information by token reduction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.318, + 0.815 + ], + "angle": 0, + "content": "3.2. Multi-criteria token fusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.824, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Given a set of input tokens \\(\\mathbf{X} \\in \\mathbb{R}^{N \\times C}\\), the goal of MCTF is to fuse the tokens into output tokens \\(\\hat{\\mathbf{X}} \\in \\mathbb{R}^{(N - r) \\times C}\\), where \\(r\\) is the number of fused tokens. To minimize the information loss, we first evaluate the relations between the tokens based on multi-criteria, then group and merge the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.339, + 0.843, + 0.354 + ], + "angle": 0, + "content": "tokens through bidirectional bipartite soft matching." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.354, + 0.894, + 0.383 + ], + "angle": 0, + "content": "Multi-criteria attraction function. We first define an attraction function \\(\\mathbf{W}\\) based on multiple criteria as" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.393, + 0.892, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\Pi_ {k = 1} ^ {M} \\left(\\mathbf {W} ^ {k} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right)\\right) ^ {\\tau_ {k}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.892, + 0.514 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}^k: \\mathbb{R}^C \\times \\mathbb{R}^C \\to \\mathbb{R}_+\\) is an attraction function computed by \\(k\\)-th criterion, and \\(\\tau^k \\in \\mathbb{R}_+\\) is the temperature parameter to adjust the influence of \\(k\\)-th criterion. The higher attraction score between two tokens indicates a higher chance of being fused. In this work, we consider the following three criteria: similarity, informativeness, and size." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.515, + 0.894, + 0.575 + ], + "angle": 0, + "content": "Similarity. The first criterion is the similarity of tokens to reduce redundant information. Akin to the previous works [4, 22] requiring the proximity of tokens, we leverage the cosine similarity between the set of tokens for" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.585, + 0.892, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} ^ {\\text {s i m}} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\frac {1}{2} \\left(\\frac {\\mathbf {x} _ {i} \\cdot \\mathbf {x} _ {j}}{\\| \\mathbf {x} _ {i} \\| \\| \\mathbf {x} _ {j} \\|} + 1\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.895, + 0.821 + ], + "angle": 0, + "content": "Token fusion with similarity effectively eliminates the redundant tokens, yet it often excessively combines the informative tokens as in Figure 2b, causing the loss of information. Informativeness. To minimize the information loss, we introduce informativeness to avoid the fusion of informative tokens. To quantify the informativeness, we measure the averaged attention scores \\(\\mathbf{a}\\in [0,1]^N\\) in the self-attention layer, which indicates the impact of each token on others: \\(\\mathbf{a}_j = \\frac{1}{N}\\sum_i^N\\mathbf{A}_{ij}\\), where \\(\\mathbf{A}_{ij} = \\mathrm{softmax}\\left(\\frac{\\mathbf{Q}_i\\mathbf{K}_j^T}{\\sqrt{C}}\\right)\\). When \\(\\mathbf{a}_i\\rightarrow 0\\), there's no influence from \\(\\mathbf{x}_i\\) to other tokens. With the informativeness scores, we define an informativeness-based attraction function as" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.83, + 0.892, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} ^ {\\text {i n f o}} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\frac {1}{\\mathbf {a} _ {i} \\mathbf {a} _ {j}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{a}_i, \\mathbf{a}_j\\) are the informative scores of \\(\\mathbf{x}_i, \\mathbf{x}_j\\), respectively. When both tokens are uninformative (\\(\\mathbf{a}_i, \\mathbf{a}_j \\to 0\\))," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15743" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.09, + 0.852, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.89, + 0.278 + ], + "angle": 0, + "content": "Figure 3. Bidirectional bipartite soft matching. The set of tokens \\(\\mathbf{X}\\) is split into two groups \\(\\mathbf{X}^{\\alpha},\\mathbf{X}^{\\beta}\\), and bidirectional bipartite soft matching are conducted through Step 1-4. The intensity of the lines indicates the multi-criteria weights \\(\\mathbf{W}^t\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.469, + 0.352 + ], + "angle": 0, + "content": "the weight gets higher \\((\\mathbf{W}^{\\mathrm{info}}(\\mathbf{x}_i,\\mathbf{x}_j)\\to \\infty)\\), making two tokens prone to be fused. In Figure 2c, with the weights combined with the similarity and informativeness, the tokens in the foreground object are less fused." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.469, + 0.488 + ], + "angle": 0, + "content": "Size. The last criterion is the size of the tokens, which indicates the number of fused tokens. Although tokens are not dropped but merged via a merging function, e.g., averaging pooling or max pooling, it is difficult to preserve all the information as the number of constituent tokens increases. So, the fusion between smaller tokens is preferred. To this end, we initially set the size \\( \\mathbf{s} \\in \\mathbb{N}^N \\) of tokens \\( \\mathbf{X} \\) as 1 and track the number of constituent (fused) tokens of each token, and define a size-based attraction function as" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.5, + 0.47, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} ^ {\\text {s i z e}} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\frac {1}{\\mathbf {s} _ {i} \\mathbf {s} _ {j}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.47, + 0.606 + ], + "angle": 0, + "content": "In Figure 2d, tokens are merged based on the multi-criteria: similarity, informativeness, and size. We observed that the fusion happens between similar tokens and the fusion of foreground tokens or large tokens is properly suppressed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.47, + 0.829 + ], + "angle": 0, + "content": "Bidirectional bipartite soft matching. Given the multicriteria-based attraction function \\(\\mathbf{W}\\), our MCTF performs a relaxed bidirectional bipartite matching called bipartite soft matching [4]. One advantage of bipartite matching is that it alleviates the quadratic cost of similarity computation between tokens, i.e., \\(O(N^2) \\to O(N'^2)\\), where \\(N' = \\left\\lfloor \\frac{N}{2} \\right\\rfloor\\). In addition, by relaxing the one-to-one correspondence constraints, the solution can be obtained by an efficient algorithm. In this relaxed matching problem, the set of tokens \\(\\mathbf{X}\\) is first split into the source and target \\(\\mathbf{X}^\\alpha\\), \\(\\mathbf{X}^\\beta \\in \\mathbb{R}^{N' \\times C}\\) as in Step 1 of Figure 3. Given a set of binary decision variables, i.e., the edge matrix \\(\\mathbf{E} \\in \\{0,1\\}^{N' \\times N'}\\) between \\(\\mathbf{X}^\\alpha\\), and \\(\\mathbf{X}^\\beta\\), bipartite soft matching is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.844, + 0.47, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} ^ {*} = \\arg \\max _ {\\mathbf {E}} \\sum_ {i j} \\mathbf {w} _ {i j} ^ {\\prime} \\mathbf {e} _ {i j} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.874, + 0.47, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\text {s u b j e c t} \\sum_ {i j} \\mathbf {e} _ {i j} = r, \\sum_ {j} \\mathbf {e} _ {i j} \\leq 1 \\forall i, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.294, + 0.545, + 0.305 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.315, + 0.892, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {i j} ^ {\\prime} = \\left\\{ \\begin{array}{l l} \\mathbf {w} _ {i j} & \\text {i f} j \\neq \\arg \\max _ {j ^ {\\prime}} \\mathbf {w} _ {i j ^ {\\prime}} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.457 + ], + "angle": 0, + "content": "\\(\\mathbf{e}_{ij}\\) indicates the presence of the edge between \\(i, j\\)-th token of \\(\\mathbf{X}^{\\alpha}, \\mathbf{X}^{\\beta}\\), and, \\(\\mathbf{w}_{ij} = \\mathbf{W}(\\mathbf{x}_i^\\alpha, \\mathbf{x}_j^\\beta)\\). This optimization problem can be solved by two simple steps: 1) find the best edge that maximizes \\(\\mathbf{w}_{ij}\\) for each \\(i\\), and 2) choose the top-\\(r\\) edges with the largest attraction scores. Then, based on the soft matching result \\(\\mathbf{E}^*\\), we group the tokens as" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.465, + 0.892, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {j} ^ {\\alpha \\rightarrow \\beta} = \\left\\{\\mathbf {x} _ {i} ^ {\\alpha} \\in \\mathbf {X} ^ {\\alpha} \\mid \\mathbf {e} _ {i j} = 1 \\right\\} \\cup \\left\\{\\mathbf {x} _ {j} ^ {\\beta} \\right\\}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.892, + 0.529 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_i^{\\alpha \\rightarrow \\beta}\\) indicates the set of tokens matched with \\(\\mathbf{x}_i^\\beta\\). Finally, the results of the fusion \\(\\tilde{\\mathbf{X}}\\) are obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.538, + 0.892, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {X}} = \\tilde {\\mathbf {X}} ^ {\\alpha} \\cup \\tilde {\\mathbf {X}} ^ {\\beta}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.56, + 0.892, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\tilde {\\mathbf {X}} ^ {\\alpha} = \\mathbf {X} ^ {\\alpha} - \\bigcup_ {i} ^ {N ^ {\\prime}} \\mathbf {X} _ {i} ^ {\\alpha \\rightarrow \\beta}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.589, + 0.892, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {X}} ^ {\\beta} = \\bigcup_ {i} ^ {N ^ {\\prime}} \\left\\{\\delta \\left(\\mathbf {X} _ {i} ^ {\\alpha \\rightarrow \\beta}\\right)\\right\\}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.75 + ], + "angle": 0, + "content": "\\(\\delta (\\mathbf{X}) = \\delta (\\{\\mathbf{x}_i\\} _i) = \\sum_i\\frac{\\mathbf{a}_i\\mathbf{s}_i\\mathbf{x}_i}{\\sum_{i'}\\mathbf{a}_{i'}\\mathbf{s}_{i'}}\\) is the pooling operation considering the attention scores a and the size s of the tokens. Still, as shown in Step2 of Figure 3, the number of target tokens \\(\\mathbf{X}^{\\beta}\\) cannot be reduced. To handle this issue, MCTF performs bidirectional bipartite soft matching by conducting the matching in the opposite direction with the updated token sets \\(\\hat{\\mathbf{X}}^{\\alpha}\\), and \\(\\hat{\\mathbf{X}}^{\\beta}\\) as in Step 3, 4 of Figure 3. The final output tokens \\(\\hat{\\mathbf{X}} = \\hat{\\mathbf{X}}^{\\alpha}\\cup \\hat{\\mathbf{X}}^{\\beta}\\) are defined with the following." + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.759, + 0.892, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {X}} ^ {\\alpha} = \\bigcup_ {i} ^ {N ^ {\\prime} - r} \\left\\{\\delta \\left(\\tilde {\\mathbf {X}} _ {i} ^ {\\beta \\rightarrow \\alpha}\\right)\\right\\}, \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.789, + 0.892, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {X}} ^ {\\beta} = \\tilde {\\mathbf {X}} ^ {\\beta} - \\bigcup_ {i} ^ {N ^ {\\prime} - r} \\tilde {\\mathbf {X}} _ {i} ^ {\\beta \\rightarrow \\alpha}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Note that calculating the pairwise weights with updated two sets of tokens \\(\\tilde{\\mathbf{w}}_{ij} = \\mathbf{W}(\\tilde{\\mathbf{x}}_i^\\beta ,\\tilde{\\mathbf{x}}_j^\\alpha)\\) introduces the additional computational costs of \\(O(N^{\\prime}(N^{\\prime} - r))\\) . To avoid this overhead, we approximate the attraction function by the attraction scores before fusion. In short, we just reuse the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15744" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.174, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.091, + 0.272, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.091, + 0.371, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.091, + 0.468, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.185, + 0.465, + 0.199 + ], + "angle": 0, + "content": "Figure 4. Visualization of attentiveness in consecutive layers." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.201, + 0.472, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.32, + 0.47, + 0.432 + ], + "angle": 0, + "content": "Figure 5. Illustration of attention map in the consecutive layers and approximated attention. (Left) The attention score \\(\\hat{\\mathbf{A}}^l\\) is the past influence of the tokens to generate \\(\\mathbf{X}^l\\). If we fuse the tokens \\(\\mathbf{X}^l\\) based on \\(\\mathbf{A}^l\\), \\(\\mathbf{x}_1\\) is prone to be fused despite the highest informativeness score in the following attention. So, we instead leverage the informativeness based on the one-step-ahead attention \\(\\mathbf{A}^{l + 1}\\). (Right) After the fusion, we also aggregate the \\(\\mathbf{A}^{l + 1}\\) to approximate the attention map \\(\\hat{\\mathbf{A}}^{l + 1}\\) for updating fused tokens \\(\\hat{\\mathbf{X}}^l\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.506 + ], + "angle": 0, + "content": "pre-calculated weights since \\(\\tilde{\\mathbf{X}}^{\\alpha}\\) is the subset of \\(\\mathbf{X}^{\\alpha}\\). This allows MCTF to efficiently reduce tokens considering bidirectional relations between two subsets with negligible extra costs compared to uni-directional bipartite soft matching." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.462, + 0.531 + ], + "angle": 0, + "content": "3.3. One-step-ahead attention for informativeness" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.473, + 0.905 + ], + "angle": 0, + "content": "In assessing informativeness, prior works [17, 18, 21] have leveraged the attention scores from the previous self-attention layer. As illustrated in Figure 5, previous approaches use the attention \\(\\mathbf{A}^l\\) from the previous layer to fuse tokens \\(\\mathbf{X}^l\\). This technique allows efficient assessment under the assumption that the attention maps in consecutive layers are similar. However, we observed that the attention maps often substantially differ, as shown in Figure 4, and the attention from a previous layer may lead to suboptimal token fusion. Thus, we proposed one-step-ahead attention, which measures the informativeness of tokens based on the attention map in the next layer, i.e., \\(\\mathbf{A}^{l + 1}\\). Then, the informativeness scores \\(\\mathbf{a}\\) in Equation (4) is calculated with \\(\\mathbf{A}^{l + 1}\\in \\mathbb{R}^{N\\times N}\\). This simple remedy provides a considerable improvement; see Figure 7b in Section 4.2. After token fusion, we efficiently compute the attention map \\(\\hat{\\mathbf{A}}^{l + 1}\\in \\mathbb{R}^{(N - r)\\times (N - r)}\\) of fused tokens \\(\\hat{\\mathbf{X}}^l\\in \\mathbb{R}^{(N - r)\\times C}\\) by simply aggregating \\(\\mathbf{A}^{l + 1}\\in \\mathbb{R}^{N\\times N}\\) without recomputing the dot-product self-attention. To be specific, when the tokens are fused as \\(\\delta (\\{\\mathbf{x}_i\\}_i)\\) during Equations (10) to (14), their corresponding one-step-ahead attention scores are also fused as \\(\\delta (\\{\\mathbf{A}_i^{l + 1}\\}_i)\\) in both query and key direction. Note that when fusing attention scores for queries we use simple sum for \\(\\delta ,i.e.\\) \\(\\forall_{i}\\sum_{j}\\hat{\\mathbf{A}}_{ij}^{l + 1} = 1\\). For fusing attention" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.891, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.274, + 0.895, + 0.344 + ], + "angle": 0, + "content": "Figure 6. Illustration of training with token reduction consistency. During training, we forward the input \\( x \\) as \\( f(x; r) \\), and \\( f(x; r') \\), respectively. To obtain the augmented representation, \\( r' \\) is randomly selected in every step, and the model is updated with supervisory signals \\( \\mathcal{L}_{\\mathrm{CE}} \\), and consistency loss \\( \\mathcal{L}_{\\mathrm{MSE}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.363, + 0.892, + 0.397 + ], + "angle": 0, + "content": "scores for queries, we use simple sum for \\(\\delta\\) to guarantee \\(\\forall_{i}\\sum_{j}\\hat{\\mathbf{A}}_{ij}^{l + 1} = 1.\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.411, + 0.753, + 0.427 + ], + "angle": 0, + "content": "3.4. Token reduction consistency" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.895, + 0.558 + ], + "angle": 0, + "content": "We here propose a new fine-tuning scheme to further improve the performance of vision Transformer \\( f_{\\theta}(\\cdot ;r) \\) with MCTF. We observe that a different number of reduced tokens per layer, denoted as \\( r \\), may lead to different representations of samples. By training Transformers with different \\( r \\) and encouraging the consistency between them, namely, token reduction consistency, we achieve the additional performance gain. The objective function of our method is given as" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.574, + 0.892, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} \\left(f _ {\\theta} (x; r), y\\right) + \\mathcal {L} _ {\\mathrm {C E}} \\left(f _ {\\theta} \\left(x; r ^ {\\prime}\\right), y\\right) \\\\ + \\lambda \\mathcal {L} _ {\\mathrm {M S E}} \\left(\\mathbf {x} _ {r} ^ {\\mathrm {c l s}}, \\mathbf {x} _ {r ^ {\\prime}} ^ {\\mathrm {c l s}}\\right), \\tag {15} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.901 + ], + "angle": 0, + "content": "where \\((x,y)\\) is a supervised sample, \\(r,r^{\\prime}\\) is the fixed and dynamic reduced token numbers, \\(\\lambda\\) is the coefficient for consistency loss, and \\(\\mathbf{x}_r^{\\mathrm{cls}},\\mathbf{x}_{r'}^{\\mathrm{cls}}\\) are the class tokens in the last layer of models \\(f_{\\theta}(x;r),f_{\\theta}(x;r')\\) . In this objective, we first calculate the cross-entropy loss \\(\\mathcal{L}_{\\mathrm{CE}}(f_{\\theta}(x;r),y)\\) with fixed \\(r\\) , which is the target reduction number that will be used in the evaluation. At the same time, we generate another representation of the input \\(x\\) with smaller but randomly drawn \\(r^\\prime \\sim\\) uniform \\((0,r)\\) , and calculate the loss \\(\\mathcal{L}_{\\mathrm{CE}}(f_{\\theta}(x;r'),y)\\) Then, we impose the token consistency loss \\(\\mathcal{L}_{\\mathrm{MSE}}(\\mathbf{x}_r^{\\mathrm{cls}},\\mathbf{x}_{r'}^{\\mathrm{cls}})\\) on the class tokens, to retain the consistent representation across the diverse reduced token numbers \\(r^\\prime\\) . The proposed method can be viewed as a new type of token-level data augmentation [7, 19] and consistency regularization. Our token reduction consistency encourages the representation \\(\\mathbf{x}_r^{\\mathrm{cls}}\\) obtained by the target reduction number \\(r\\) to mimic the slightly augmented representation \\(\\mathbf{x}_{r'}^{\\mathrm{cls}}\\) , which is more similar to ones with no token reduction since \\(r^{\\prime} < r\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15745" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.16, + 0.09, + 0.388, + 0.104 + ], + "angle": 0, + "content": "Table 1. Image classification results" + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.109, + 0.472, + 0.447 + ], + "angle": 0, + "content": "
MethodFLOPs (G)Params (M)Top-1 Acc (%)
DeiT-T [28]1.2572.2 (-)
+EvoViT[AAAI '22] [36]0.8572.0 (-0.2)
+A-ViT[CVPR '22] [37]0.8571.0 (-1.2)
+SPViT[ECCV '22] [17]0.9572.1 (-0.1)
+ToMe[ICLR '23] [4]0.7571.3 (-0.9)
+BAT[CVPR '23] [21]0.8572.3 (+0.1)
+MCTFr=160.7572.7 (+0.5)
DeiT-S [28]4.62279.8 (-)
+IA-RED2[NeurIPS '21] [24]3.22279.1 (-0.7)
+DynamicViT[NeurIPS '21] [26]2.92379.3 (-0.5)
+EvoViT[AAAI '22] [36]3.02279.4 (-0.4)
+EViT[ICLR '22] [18]3.02279.5 (-0.3)
+A-ViT[CVPR '22] [37]3.62278.6 (-1.2)
+ATS[ECCV '22] [12]2.92279.7 (-0.1)
+SPViT[ECCV '22] [17]2.62279.3 (-0.5)
+ToMe[ICLR '23] [4]2.72279.4 (-0.4)
+BAT[CVPR '23] [21]3.02279.6 (-0.2)
+MCTFr=162.62280.1 (+0.3)
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.459, + 0.21, + 0.476 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.472, + 0.729 + ], + "angle": 0, + "content": "Baselines. To validate the effectiveness of the proposed methods, we compare MCTF with the previous token reduction methods. For comparison, we opt the token pruning methods (A-ViT [37], IA-RED² [24], DynamicViT [26], EvoViT [36], ATS [12]) and token fusion methods (SPViT [17], EViT [18], ToMe [4], BAT [21]) in DeiT [28], and report the efficiency (FLOPs (G)) and the performance (Top-1 Acc (\\%)) of each method. Further, to validate MCTF on other Vision Transformers (T2T-ViT [39], LV-ViT [15]), we report the results of MCTF and compare it with the official number of existing works. We denote the number of reduced tokens per layer \\( r \\) with the subscript in Tables 1 and 2. The gray color in the table indicates the base model, and the green and red color indicates the improvements and degradations of the performance compared to the base model, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.281, + 0.758 + ], + "angle": 0, + "content": "4.1. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Comparison of the token reduction methods. The comparison with existing token reduction methods is summarized in Table 1. We demonstrate that our MCTF achieves the best performance with the lowest FLOPs in DeiT [28] surpassing all previous works. Further, it is worth noting that MCTF is the only work that avoids performance degradation with the lowest FLOPs in both DeiT-T and DeiT-S. Through Finetuning DeiT-T for 30 epochs, MCTF brings a significant gain of \\(+0.5\\%\\) in accuracy over the base model with nearly half" + }, + { + "type": "table_caption", + "bbox": [ + 0.529, + 0.09, + 0.864, + 0.104 + ], + "angle": 0, + "content": "Table 2. Comparison with other Vision Transformers" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.109, + 0.892, + 0.44 + ], + "angle": 0, + "content": "
ModelsFLOPs (G)Params (M)Acc (%)
PVT-Small[31]3.824.579.8
PVT-Medium [31]6.744.281.2
CoaT Mini [35]6.810.080.8
CoaT-Lite Small [35]4.020.081.9
Swin-T [20]4.529.081.3
Swin-S [20]8.750.083.0
PoolFormer-S36 [38]5.031.081.4
PoolFormer-M48 [38]11.673.082.5
T2T-ViTt-14 [39]6.121.581.7
+MCTFr=134.221.581.8 (↑)
T2T-ViTt-19 [39]9.839.282.4
+MCTFr=96.439.282.4 (-)
LV-ViT-S [15]6.626.283.3
+EVIT(ICLR '22] [18]4.726.283.0 (↓)
+BAT[CVPR '23] [21]4.726.283.1 (↓)
+DynamicViT[NeurIPS '21] [26]4.626.983.0 (↓)
+SPViT[ECCV '22] [17]4.326.283.1 (↓)
+MCTFr=124.226.283.4 (↑)
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.895, + 0.549 + ], + "angle": 0, + "content": "FLOPs. Similarly, we observe a gain of \\(+0.3\\%\\) with DeiT-S while boosting the FLOPs by \\(-2.0\\) (G). We believe that multi-criteria with one-step-ahead attention helps the model to minimize the loss of information; further consistency loss on the class token through the token reduction improves the generalizability of the model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.897, + 0.763 + ], + "angle": 0, + "content": "MCTF with other Vision Transformers. To validate the applicability of MCTF in various ViTs, we demonstrate MCTF with other transformer architectures in Table 2. Following previous works [17, 18, 21, 26], we apply MCTF with LVViT. Also, we present the results of MCTF with T2T-ViT. As presented in the table, our experimental results are promising. MCTF in these architectures gets at least \\(31\\%\\) speedup without performance degradation. Further, MCTF combined with LV-ViT outperforms all other Transformers and token reduction methods regarding FLOPs, and accuracy. Especially, it is worth noting that all token reduction methods except for MCTF bring the performance degradation in LVViT. These results reveal that MCTF is the efficient token reduction method for the diverse Vision Transformers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Token reduction without training. Similar to ToMe [4], MCTF is applicable with pre-trained ViTs without any additional training since MCTF does not require any learnable parameters. We here apply the two reduction methods to the pre-trained DeiT without finetuning and provide the results in Table 3. Regardless of the reduced number of tokens \\( r \\) in each layer, MCTF consistently surpasses ToMe. Especially, in the most sparse setting \\( r = 20 \\), the performance gap is significant (\\(+7.0\\%\\) in DeiT-T, \\(+3.8\\%\\) in DeiT-S). Note" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "15746" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.442, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.096, + 0.885, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.316, + 0.895, + 0.358 + ], + "angle": 0, + "content": "Figure 7. Ablations on (a) multi-criteria, (b) one-step-ahead-attention, and token reduction consistency. Each marker indicates the model with \\( r \\in [1,20] \\), and we highlight \\( r \\in \\{5,10,15,20\\} \\) as bordered circle. We also denote the model as star when \\( r = 16 \\), which is used for finetuning the model." + }, + { + "type": "table_caption", + "bbox": [ + 0.108, + 0.37, + 0.439, + 0.384 + ], + "angle": 0, + "content": "Table 3. Image classification results without training" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.387, + 0.468, + 0.535 + ], + "angle": 0, + "content": "
Methodr
Base1248121620
DeiT-T
ToMe [4]72.272.172.072.071.670.868.761.5
MCTF72.272.272.172.172.071.771.068.5
DeiT-S
ToMe [4]79.879.879.779.779.479.077.974.2
MCTF79.879.879.879.879.879.679.278.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.471, + 0.623 + ], + "angle": 0, + "content": "that without any additional training, our \\(\\mathrm{MCTF}_{r=16}\\) with pre-trained DeiT-S still shows a competitive performance of \\(79.2\\%\\) compared to reduction methods requiring training (e.g., \\(78.6\\%\\) of A-ViT, \\(79.1\\%\\) of IA-RED\\(^2\\), and \\(79.3\\%\\) of DynamicViT, and SPViT in Table 1)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.32, + 0.649 + ], + "angle": 0, + "content": "4.2. Ablation studies on MCTF" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.735 + ], + "angle": 0, + "content": "We provide ablation studies to validate each component of MCTF. Unless otherwise stated, we conduct whole experiments with DeiT-S finetuned with MCTF (\\(r = 16\\)). We provide the FLOPs-Accuracy graph by adjusting the reduced number of tokens per layer \\(r \\in [1,20]\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Multi-criteria. We explore the effectiveness of multi-criteria in Figure 7a. First, regarding the multi-criteria, we utilize three criteria for MCTF, i.e., similarity (sim.), informativeness (info.), and size. Each single criterion of similarity and informativeness shows a relatively inferior performance compared to dual (sim. & info.) and multi-criteria (sim. & info. & size). Specifically, when \\( r = 16 \\), the performance of a single criterion is \\( 79.7\\% \\), and \\( 79.4\\% \\) with similarity and informativeness, respectively. Then, adopting dual criteria (sim. & info.), MCTF achieves \\( 79.8\\% \\). Finally, we get an accuracy of \\( 80.1\\% \\) with a gain of \\( +0.3\\% \\) by respecting all" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.415 + ], + "angle": 0, + "content": "three criteria (sim. & info. & size). These performance gaps get larger as \\( r \\) increases, which proves the importance of the multi-criteria for token fusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.895, + 0.568 + ], + "angle": 0, + "content": "One-step-ahead attention and token reduction consistency. To show the validity of one-step-ahead attention and token reduction consistency, we also provide the results of MCTF with and without each component in Figure 7b. When eliminating either one-step-ahead attention or token reduction consistency, the accuracies are dropped in every FLOP. This significant drop indicates that both approaches matter for MCTF. In short, by adopting one-step-ahead attention and token reduction consistency, MCTF effectively mitigates the performance degradation in a wide range of FLOPs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.895, + 0.794 + ], + "angle": 0, + "content": "Comparison of design choices. The ablations on design choices are presented in Table 4. First, our bidirectional bipartite matching, which enables capturing the bidirectional relation in two sets, enhances the accuracy compared to one-way bipartite matching. Next, for pooling operation \\(\\delta\\), the weighted sum considering the size \\(s\\) and attentiveness \\(a\\) is a better choice than others like max-pool or average. Lastly, we compare the results with the precise and approximated attention for \\(\\hat{\\mathbf{A}}^l\\). For precise attention, we just conduct the similarity calculation for one-step-ahead attention and the attention in the self-attention layer after fusion, separately. Otherwise, we approximate it with one-step-ahead attention as described in Section 3.3. As presented in the table, our approximated attention maintains the performance with the substantial improvement in efficiency (-0.4 (G) FLOPs)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.673, + 0.817 + ], + "angle": 0, + "content": "4.3. Analyse of MCTF" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Qualitative results. For a better understanding of MCTF, we provide the qualitative results of MCTF in Figure 8. We visualize the fused tokens at the last block of DeiT-S on ImageNet-1K and denote the fused tokens by the same border color. As shown in the figure, since the tokens are merged" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15747" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.172, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.089, + 0.275, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.38, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.089, + 0.481, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.089, + 0.584, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.09, + 0.686, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.089, + 0.789, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.089, + 0.892, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.173, + 0.172, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.173, + 0.275, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.173, + 0.38, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.173, + 0.481, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.173, + 0.584, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.173, + 0.686, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.173, + 0.789, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.173, + 0.892, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.894, + 0.283 + ], + "angle": 0, + "content": "Figure 8. Visualization of the fused tokens with MCTF. Given the input images of ImageNet-1K (Top), the qualitative results of MCTF with DeiT-S are provided at the bottom. The same border color of the patches indicates the fused tokens." + }, + { + "type": "table_caption", + "bbox": [ + 0.147, + 0.295, + 0.4, + 0.31 + ], + "angle": 0, + "content": "Table 4. Ablations of the design choices." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.315, + 0.447, + 0.546 + ], + "angle": 0, + "content": "
MethodFLOPs ↓ (G)Acc ↑ (%)
DeiT-S4.679.8
bipartite soft matching
One-way2.680.0
Bidirectional2.680.1
pooling function δ
average2.680.0
max2.679.8
weighted average2.680.1
approximation of attention map
precise attention3.080.1
approximated attention2.680.1
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.695 + ], + "angle": 0, + "content": "with multi-criteria (e.g., similarity, informativeness, size), we maintain the more diverse tokens in the informative foreground object. For instance, in the third image of the hamster, while the background patches including the hand are fused into one token, the foreground tokens are less fused while maintaining the details like the eye, ear, and face of the hamster. In short, compared to the background, the foreground tokens are less fused with the moderate size retaining the information of the main content." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.471, + 0.833 + ], + "angle": 0, + "content": "Soundness of size criterion. Figure 9 presents the histogram of sizes of tokens after token reduction with and without size criterion. Specifically, we measure the size of the largest token at the last block and provide the histogram. With our size criterion, the merged tokens tend to have smaller sizes s showing the average size of 39.3/49.2 with and without the Size criterion, respectively. As intended, MCTF successfully suppresses the large-sized tokens, which are a source of information loss, leading to performance improvement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.196, + 0.861 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "In this work, we introduced the Multi-Criteria Token Fusion (MCTF), a novel strategy aimed at reducing the complex" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.296, + 0.871, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.445, + 0.874, + 0.46 + ], + "angle": 0, + "content": "Figure 9. Histogram of the size of tokens after reduction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.895, + 0.685 + ], + "angle": 0, + "content": "ity inherent in ViTs while mitigating performance degradation. MCTF effectively discerns the relation of tokens based on multiple criteria, including similarity, informativeness, and the size of the tokens. Our comprehensive ablation studies and detailed analyses demonstrate the efficacy of MCTF particularly with our innovative one-step-ahead attention and token reduction consistency. Remarkably, DeiT-T and DeiT-S with MCTF achieve considerable improvements, with \\(+0.5\\%\\), and \\(+0.3\\%\\) increase in Top-1 Accuracy over the vanilla models, accompanied by about \\(44\\%\\) fewer FLOPs, respectively. We also observe that our MCTF outperforms all of the previous token reduction methods in diverse vision Transformers with and without training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.718, + 0.66, + 0.736 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "This work was supported by ICT Creative Consilience Program through the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT)(IITP-2024-2020-0-01819), the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT)(NRF-2023R1A2C2005373), and a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI) funded by the Ministry of Health & Welfare Republic of Korea (HR20C0021)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15748" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.143 + ], + "angle": 0, + "content": "[1] Moab Arar, Ariel Shamir, and Amit H Bermano. Learned queries for efficient local attention. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.471, + 0.185 + ], + "angle": 0, + "content": "[2] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[3] Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv:2004.05150, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[4] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. ICLR, 2022. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.471, + 0.327 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.471, + 0.373 + ], + "angle": 0, + "content": "[6] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.471, + 0.415 + ], + "angle": 0, + "content": "[7] Hyeong Kyu Choi, Joonmyung Choi, and Hyunwoo J. Kim. Tokenmixup: Efficient attention-guided token-level data augmentation for transformers. In NeurIPS, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.471, + 0.472 + ], + "angle": 0, + "content": "[8] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. *ICLR*, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.471, + 0.528 + ], + "angle": 0, + "content": "[9] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. NeurIPS, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.471, + 0.586 + ], + "angle": 0, + "content": "[10] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[11] 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. ICLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.471, + 0.728 + ], + "angle": 0, + "content": "[12] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In ECCV, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.471, + 0.771 + ], + "angle": 0, + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.471, + 0.814 + ], + "angle": 0, + "content": "[14] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[15] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. NeurIPS, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[16] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. In ICLR, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.16 + ], + "angle": 0, + "content": "[17] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In ECCV, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[18] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. ICLR, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[19] Jihao Liu, Boxiao Liu, Hang Zhou, Hongsheng Li, and Yu Liu. Tokenmix: Rethinking image mixing for data augmentation in vision transformers. In ECCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.894, + 0.317 + ], + "angle": 0, + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.894, + 0.374 + ], + "angle": 0, + "content": "[21] Sifan Long, Zhen Zhao, Jimin Pi, Shengsheng Wang, and Jingdong Wang. Beyond attentive tokens: Incorporating token importance and diversity for efficient vision transformers. In CVPR, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.894, + 0.431 + ], + "angle": 0, + "content": "[22] Dmitrii Marin, Jen-Hao Rick Chang, Anurag Ranjan, Anish Prabhu, Mohammad Rastegari, and Oncel Tuzel. Token pooling in vision transformers for image classification. In WACV, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.894, + 0.488 + ], + "angle": 0, + "content": "[23] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In CVPR, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.894, + 0.544 + ], + "angle": 0, + "content": "[24] Bowen Pan, Rameswar Panda, Yifan Jiang, Zhangyang Wang, Rogerio Feris, and Aude Oliva. IA-RED2: Interpretability-aware redundancy reduction for vision transformers. NeurIPS, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.894, + 0.614 + ], + "angle": 0, + "content": "[25] 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 ICML, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.894, + 0.671 + ], + "angle": 0, + "content": "[26] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. NeurIPS, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.894, + 0.714 + ], + "angle": 0, + "content": "[27] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.894, + 0.771 + ], + "angle": 0, + "content": "[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.894, + 0.814 + ], + "angle": 0, + "content": "[29] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[30] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv:2006.04768, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[31] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense predic" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15749" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "tion without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[32] QuanLin Wu, Hang Ye, Yuntian Gu, Huishuai Zhang, Liwei Wang, and Di He. Denoising masked autoencoders help robust classification. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[33] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.472, + 0.276 + ], + "angle": 0, + "content": "[34] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. In AAAI, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.472, + 0.318 + ], + "angle": 0, + "content": "[35] Weijian Xu, Yifan Xu, Tyler Chang, and Zhuowen Tu. Coscale conv-attentional image transformers. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[36] Yifan Xu, Zhijie Zhang, Mengdan Zhang, Kekai Sheng, Ke Li, Weiming Dong, Liqing Zhang, Changsheng Xu, and Xing Sun. Evo-vit: Slow-fast token evolution for dynamic vision transformer. In AAAI, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.47, + 0.418 + ], + "angle": 0, + "content": "[37] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In CVPR, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.47, + 0.461 + ], + "angle": 0, + "content": "[38] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.47, + 0.517 + ], + "angle": 0, + "content": "[39] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In ICCV, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.47, + 0.547 + ], + "angle": 0, + "content": "[40] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.549, + 0.47, + 0.589 + ], + "angle": 0, + "content": "[41] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. *ICLR*, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15750" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/c38e4100-b4f3-4d92-aff9-1fdadd7cfa0a_origin.pdf b/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/c38e4100-b4f3-4d92-aff9-1fdadd7cfa0a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..748010209e232bd55e55bb8bf9d7dc860012b2ec --- /dev/null +++ b/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/c38e4100-b4f3-4d92-aff9-1fdadd7cfa0a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f60941508852c201924340256702edc62db4c43e40594f9fed905b5eec844888 +size 6309065 diff --git a/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/full.md b/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5723d243c2cb00fb0bd768fa5eed99d51522fc9d --- /dev/null +++ b/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/full.md @@ -0,0 +1,342 @@ +# Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers + +Sanghyeok Lee* Joonmyung Choi* Hyunwoo J. Kim† + +Department of Computer Science and Engineering, Korea University + +{cat0626, pizard, hyunwoojkim}@korea.ac.kr + +# Abstract + +Vision Transformer (ViT) has emerged as a prominent backbone for computer vision. For more efficient ViTs, recent works lessen the quadratic cost of the self-attention layer by pruning or fusing the redundant tokens. However, these works faced the speed-accuracy trade-off caused by the loss of information. Here, we argue that token fusion needs to consider diverse relations between tokens to minimize information loss. In this paper, we propose a Multi-criteria Token Fusion (MCTF), that gradually fuses the tokens based on multi-criteria (i.e., similarity, informativeness, and size of fused tokens). Further, we utilize the one-step-ahead attention, which is the improved approach to capture the informativeness of the tokens. By training the model equipped with MCTF using a token reduction consistency, we achieve the best speed-accuracy trade-off in the image classification (ImageNet1K). Experimental results prove that MCTF consistently surpasses the previous reduction methods with and without training. Specifically, DeiT-T and DeiT-S with MCTF reduce FLOPs by about $44\%$ while improving the performance $(+0.5\%, \text{and} +0.3\%)$ over the base model, respectively. We also demonstrate the applicability of MCTF in various Vision Transformers (e.g., T2T-ViT, LV-ViT), achieving at least $31\%$ speedup without performance degradation. Code is available at https://github.com/mlvlab/MCTF. + +# 1. Introduction + +Vision Transformer [11] (ViT) has been proposed to tackle the vision tasks with self-attention, originally developed for natural language processing tasks. With the advent of ViT, Transformers are the prevalent architectures for a wide range of vision tasks, e.g., classification [11, 20, 28, 29, 31], object detection [5, 31, 41], segmentation [20, 27, 33], etc. ViTs, + +built only with self-attention and MLP, provide great flexibility and impressive performance compared to conventional approaches, e.g., convolutional neural networks (CNNs). However, despite these advantages, the quadratic computational complexity of self-attention with respect to the number of tokens is the major bottleneck for Transformers. This limitation becomes more substantial with the growing interest in large-scale foundation models such as CLIP [25]. To this end, several works [3, 16, 30, 34] have proposed efficient self-attention mechanisms including local self-attention within predefined windows [1, 9, 20]. + +More recently, there has been increasing interest in token-reduction methods for optimizing ViTs without altering their architecture. Earlier works [12, 23, 24, 26, 37] primarily focused on pruning the uninformative tokens to reduce the number of tokens. Another line of works [4, 17, 18, 21, 22] attempted to fuse the tokens instead of discarding them to minimize the information loss. However, performance degradation is still commonly observed in most token fusion methods. We notice that the token fusion methods usually consider only one criterion, such as the similarity or informativeness of tokens, leading to suboptimal token matching. For instance, similarity-based token fusion is prone to combine the foreground tokens, whereas informativeness-based fusion often merges substantially dissimilar tokens, resulting in collapsed representations. Furthermore, if too many tokens are fused into one token, then information loss is inevitable. + +To address the problems, we introduce Multi-Criteria Token Fusion (MCTF) that optimizes vision transformers by fusing tokens based on multi-criteria. Unlike previous works that consider a single criterion for token fusion, MCTF measures the relationship between the tokens with multi-criteria as follows; (1) similarity to fuse the redundant tokens, (2) informativeness to reduce the uninformative tokens, (3) the size of the tokens to prevent the large-sized tokens that boost the loss of information. Also, to tackle the inconsistency between attention maps of consecutive layers, we adopt one-step-ahead attention, which explicitly estimates the informativeness of + +![](images/afb4d8e49b0642017b3ab2ced053a4fc3c9392bafd01ad9dff78a810b9d0e898.jpg) +Figure 1. Comparison of the token reduction methods with DeiT-T (left), and DeiT-S (right). Given a base model marked as blue circle, previous token reduction methods accelerate the speed with the trade-off between accuracy and computational cost. Our MCTF, marked as a star, even brings performance improvements while lessening the complexity of DeiT. Note that after only one finetuning with the specific reduced number of tokens marked as red star, we simply evaluate it with the diverse FLOPs by adjusting the reduced numbers. + +![](images/e3627aae79897e1cb056d4f57193d13d32a28d6e895d8b71f5c6b25c951ec20c.jpg) + +the tokens in the next layer. Finally, by introducing a token reduction consistency for finetuning the model, we achieve superior performance to the existing works as in Figure 1. Surprisingly, our MCTF even performs better than the 'full' base model (red dotted line) with a reduced computational complexity. Specifically, it brings a $0.5\%$ , and $0.3\%$ gain while reducing FLOPs by about $44\%$ in DeiT-T, and DeiT-S [28], respectively. We have observed a similar speed-up $(31\%)$ in T2T-ViT [39], and LV-ViT [15] without any performance degradation. + +Our contributions are summarized in fourfold. + +- We propose Multi-criteria Token Fusion, a novel token fusion method that considers multi-criteria, e.g., similarity, informativeness, and size, to capture the complex relationship of tokens and minimize information loss. +- For measuring the informativeness of the tokens, we utilize one-step-ahead attention to retain the attentive tokens in the following layers. +- We propose a new fine-tuning scheme with token reduction consistency to boost the generalization performance of transformers equipped with MCTF. +- The extensive experiments demonstrate that MCTF achieves the best speed-accuracy trade-off in diverse ViTs, surpassing all previous token reduction methods. + +# 2. Related works + +Vision Transformers. Vision Transformer [11] is introduced to tackle the vision tasks. Later, DeiT [28] and CaiT [29] are proposed to handle the data efficiency and scalability of ViT, respectively. Recent works [6, 10, 14, 20, 31] tried to insert the inductive biases of CNNs on ViT, such as the locality or pyramid-architecture. In parallel, there is a + +line of works that boosts the vanilla ViT by scaling [29, 40] or self-supervised learning [2, 13, 32]. Despite the promising results of these works, the quadratic complexity of ViTs is still the major constraint for scaling the model. For the sake of mitigating the complexity, Reformer [16] lessens the quadratic complexity to $O(N \log N)$ through the hashing function, and Linformer [30], performer [8], and Nyströmformer [34] achieve the linear cost with the approximated linear attention. Also, several works [1, 9, 10, 20] utilize sparse attention with the reduced key or query. Swin [20] and Twins [9] utilize the local attention within the fixed size of the window to mitigate the complexity. + +Token reduction in ViTs. Most of the computational burden in ViTs arises from the self-attention. To reduce the quadratic cost in the number of tokens, recent works [4, 12, 17, 18, 21–24, 26, 37] have an interest in reducing the token itself. These works have the advantage of utilizing the original ViTs architecture without modification. In earlier works [12, 23, 24, 26, 37], the uninformative tokens are simply dropped during the forward process, leading to the information loss. To compensate for this, SPViT [17] and EViT [18] first split the tokens into informative and uninformative token sets based on attention scores, then fuse these uninformative token sets into a single token. In parallel, token pooling [22] and ToMe [4] combine the semantically similar tokens to reduce redundancies. A more recent study BAT [21] first split the tokens based on informativeness then fuse the tokens considering the diversity of the tokens. Despite the advantage of each criterion, successful integration of multi-criteria is still less explored. + +![](images/11e3b5e1793a889a3d9c9a678880b8c0c8877ac09c70b47122b1d2da106c5011.jpg) +(a) Origin + +![](images/fa08f1eaca5e8102ae9ab000f75eb90d923a88562bae277388ff7344d72d338b.jpg) +(b) $\mathbf{W}^{\mathrm{sim}}$ + +![](images/b627314e338ea533b512dc8a9dc206fd81ef73b25967bd2250fdc197147f4cd2.jpg) +(c) $\mathbf{W}^{\mathrm{sim}}\& \mathbf{W}^{\mathrm{info}}$ +Figure 2. Visualization of the fused tokens. Given (a) the leftmost image, (b) fusing the tokens with a single criterion $\mathbf{W}^{\mathrm{sim}}$ often results in the excessive fusion of the foreground object. (c) Then considering both similarity and informativeness ( $\mathbf{W}^{\mathrm{sim}}\& \mathbf{W}^{\mathrm{info}}$ ), tokens in the foreground objects are less fused while the tokens in the background are largely fused. (d) Finally, MCTF helps retain the information of each component in the image by preventing the large-size token with the multi-criteria ( $\mathbf{W}^{\mathrm{sim}}\& \mathbf{W}^{\mathrm{info}}\& \mathbf{W}^{\mathrm{size}}$ ). + +![](images/f94704b97ecdeb4baccac06247d87493beb88ccbf29c6ee2cfb9b01417a5efb4.jpg) +(d) $\mathbf{W}^{\mathrm{sim}}\& \mathbf{W}^{\mathrm{info}}\& \mathbf{W}^{\mathrm{size}}$ + +# 3. Method + +We first review the self-attention and token reduction approaches (Section 3.1). Then, we present our multi-criteria token fusion (Section 3.2) that leverages one-step-ahead attention (Section 3.3). Lastly, we introduce a training strategy with token reduction consistency in Section 3.4. + +# 3.1. Preliminaries + +In Transformers, tokens $\mathbf{X} \in \mathbb{R}^{N \times C}$ are processed by self-attention defined as + +$$ +\operatorname {S A} (\mathbf {X}) = \operatorname {s o f t m a x} \left(\frac {\mathbf {Q K} ^ {\top}}{\sqrt {C}}\right) \mathbf {V}, \tag {1} +$$ + +where $\mathbf{Q},\mathbf{K},\mathbf{V} = \mathbf{X}\mathbf{W}_{\mathbf{Q}},\mathbf{X}\mathbf{W}_{\mathbf{K}},\mathbf{X}\mathbf{W}_{\mathbf{V}}$ and $\mathbf{W}_{\mathbf{Q}},\mathbf{W}_{\mathbf{K}}$ $\mathbf{W}_{\mathbf{V}}\in \mathbb{R}^{C\times C}$ are learnable weight matrices. Despite its outstanding expressive power, the self-attention does not scale well with the number of tokens $N$ due to its quadratic time complexity $O(N^{2}C + NC^{2})$ . To address this problem, a line of works [12, 23, 24, 26, 37] reduces the number of tokens simply by pruning uninformative tokens. These approaches often cause significant performance degradation due to the loss of information. Thus, another line of works [4, 17, 18, 21, 22] fuses the uninformative or redundant tokens $\hat{\mathbf{X}}\subset \mathbf{X}$ into a new token $\hat{\mathbf{x}} = \delta (\hat{\mathbf{X}})$ , where $\mathbf{X}$ is the set of original tokens, and $\delta$ denotes a merging function, e.g., max-pooling or averaging. In this work, we also adopt 'token fusion' rather than 'token pruning' with multiple criteria to minimize the loss of information by token reduction. + +# 3.2. Multi-criteria token fusion + +Given a set of input tokens $\mathbf{X} \in \mathbb{R}^{N \times C}$ , the goal of MCTF is to fuse the tokens into output tokens $\hat{\mathbf{X}} \in \mathbb{R}^{(N - r) \times C}$ , where $r$ is the number of fused tokens. To minimize the information loss, we first evaluate the relations between the tokens based on multi-criteria, then group and merge the + +tokens through bidirectional bipartite soft matching. + +Multi-criteria attraction function. We first define an attraction function $\mathbf{W}$ based on multiple criteria as + +$$ +\mathbf {W} \left(\mathbf {x} _ {i}, \mathbf {x} _ {j}\right) = \Pi_ {k = 1} ^ {M} \left(\mathbf {W} ^ {k} \left(\mathbf {x} _ {i}, \mathbf {x} _ {j}\right)\right) ^ {\tau_ {k}}, \tag {2} +$$ + +where $\mathbf{W}^k: \mathbb{R}^C \times \mathbb{R}^C \to \mathbb{R}_+$ is an attraction function computed by $k$ -th criterion, and $\tau^k \in \mathbb{R}_+$ is the temperature parameter to adjust the influence of $k$ -th criterion. The higher attraction score between two tokens indicates a higher chance of being fused. In this work, we consider the following three criteria: similarity, informativeness, and size. + +Similarity. The first criterion is the similarity of tokens to reduce redundant information. Akin to the previous works [4, 22] requiring the proximity of tokens, we leverage the cosine similarity between the set of tokens for + +$$ +\mathbf {W} ^ {\text {s i m}} \left(\mathbf {x} _ {i}, \mathbf {x} _ {j}\right) = \frac {1}{2} \left(\frac {\mathbf {x} _ {i} \cdot \mathbf {x} _ {j}}{\| \mathbf {x} _ {i} \| \| \mathbf {x} _ {j} \|} + 1\right). \tag {3} +$$ + +Token fusion with similarity effectively eliminates the redundant tokens, yet it often excessively combines the informative tokens as in Figure 2b, causing the loss of information. Informativeness. To minimize the information loss, we introduce informativeness to avoid the fusion of informative tokens. To quantify the informativeness, we measure the averaged attention scores $\mathbf{a}\in [0,1]^N$ in the self-attention layer, which indicates the impact of each token on others: $\mathbf{a}_j = \frac{1}{N}\sum_i^N\mathbf{A}_{ij}$ , where $\mathbf{A}_{ij} = \mathrm{softmax}\left(\frac{\mathbf{Q}_i\mathbf{K}_j^T}{\sqrt{C}}\right)$ . When $\mathbf{a}_i\rightarrow 0$ , there's no influence from $\mathbf{x}_i$ to other tokens. With the informativeness scores, we define an informativeness-based attraction function as + +$$ +\mathbf {W} ^ {\text {i n f o}} \left(\mathbf {x} _ {i}, \mathbf {x} _ {j}\right) = \frac {1}{\mathbf {a} _ {i} \mathbf {a} _ {j}}, \tag {4} +$$ + +where $\mathbf{a}_i, \mathbf{a}_j$ are the informative scores of $\mathbf{x}_i, \mathbf{x}_j$ , respectively. When both tokens are uninformative ( $\mathbf{a}_i, \mathbf{a}_j \to 0$ ), + +![](images/ac4bf9f15015a684ae0c54cb8542749ffb0d741b46ac00e74d63c75d8d9ce802.jpg) +Figure 3. Bidirectional bipartite soft matching. The set of tokens $\mathbf{X}$ is split into two groups $\mathbf{X}^{\alpha},\mathbf{X}^{\beta}$ , and bidirectional bipartite soft matching are conducted through Step 1-4. The intensity of the lines indicates the multi-criteria weights $\mathbf{W}^t$ . + +the weight gets higher $(\mathbf{W}^{\mathrm{info}}(\mathbf{x}_i,\mathbf{x}_j)\to \infty)$ , making two tokens prone to be fused. In Figure 2c, with the weights combined with the similarity and informativeness, the tokens in the foreground object are less fused. + +Size. The last criterion is the size of the tokens, which indicates the number of fused tokens. Although tokens are not dropped but merged via a merging function, e.g., averaging pooling or max pooling, it is difficult to preserve all the information as the number of constituent tokens increases. So, the fusion between smaller tokens is preferred. To this end, we initially set the size $\mathbf{s} \in \mathbb{N}^N$ of tokens $\mathbf{X}$ as 1 and track the number of constituent (fused) tokens of each token, and define a size-based attraction function as + +$$ +\mathbf {W} ^ {\text {s i z e}} \left(\mathbf {x} _ {i}, \mathbf {x} _ {j}\right) = \frac {1}{\mathbf {s} _ {i} \mathbf {s} _ {j}}. \tag {5} +$$ + +In Figure 2d, tokens are merged based on the multi-criteria: similarity, informativeness, and size. We observed that the fusion happens between similar tokens and the fusion of foreground tokens or large tokens is properly suppressed. + +Bidirectional bipartite soft matching. Given the multicriteria-based attraction function $\mathbf{W}$ , our MCTF performs a relaxed bidirectional bipartite matching called bipartite soft matching [4]. One advantage of bipartite matching is that it alleviates the quadratic cost of similarity computation between tokens, i.e., $O(N^2) \to O(N'^2)$ , where $N' = \left\lfloor \frac{N}{2} \right\rfloor$ . In addition, by relaxing the one-to-one correspondence constraints, the solution can be obtained by an efficient algorithm. In this relaxed matching problem, the set of tokens $\mathbf{X}$ is first split into the source and target $\mathbf{X}^\alpha$ , $\mathbf{X}^\beta \in \mathbb{R}^{N' \times C}$ as in Step 1 of Figure 3. Given a set of binary decision variables, i.e., the edge matrix $\mathbf{E} \in \{0,1\}^{N' \times N'}$ between $\mathbf{X}^\alpha$ , and $\mathbf{X}^\beta$ , bipartite soft matching is formulated as + +$$ +\mathbf {E} ^ {*} = \arg \max _ {\mathbf {E}} \sum_ {i j} \mathbf {w} _ {i j} ^ {\prime} \mathbf {e} _ {i j} \tag {6} +$$ + +$$ +\text {s u b j e c t} \sum_ {i j} \mathbf {e} _ {i j} = r, \sum_ {j} \mathbf {e} _ {i j} \leq 1 \forall i, \tag {7} +$$ + +where + +$$ +\mathbf {w} _ {i j} ^ {\prime} = \left\{ \begin{array}{l l} \mathbf {w} _ {i j} & \text {i f} j \neq \arg \max _ {j ^ {\prime}} \mathbf {w} _ {i j ^ {\prime}} \\ 0 & \text {o t h e r w i s e} \end{array} , \right. \tag {8} +$$ + +$\mathbf{e}_{ij}$ indicates the presence of the edge between $i, j$ -th token of $\mathbf{X}^{\alpha}, \mathbf{X}^{\beta}$ , and, $\mathbf{w}_{ij} = \mathbf{W}(\mathbf{x}_i^\alpha, \mathbf{x}_j^\beta)$ . This optimization problem can be solved by two simple steps: 1) find the best edge that maximizes $\mathbf{w}_{ij}$ for each $i$ , and 2) choose the top- $r$ edges with the largest attraction scores. Then, based on the soft matching result $\mathbf{E}^*$ , we group the tokens as + +$$ +\mathbf {X} _ {j} ^ {\alpha \rightarrow \beta} = \left\{\mathbf {x} _ {i} ^ {\alpha} \in \mathbf {X} ^ {\alpha} \mid \mathbf {e} _ {i j} = 1 \right\} \cup \left\{\mathbf {x} _ {j} ^ {\beta} \right\}, \tag {9} +$$ + +where $\mathbf{X}_i^{\alpha \rightarrow \beta}$ indicates the set of tokens matched with $\mathbf{x}_i^\beta$ . Finally, the results of the fusion $\tilde{\mathbf{X}}$ are obtained as + +$$ +\tilde {\mathbf {X}} = \tilde {\mathbf {X}} ^ {\alpha} \cup \tilde {\mathbf {X}} ^ {\beta}, \tag {10} +$$ + +$$ +\text {w h e r e} \tilde {\mathbf {X}} ^ {\alpha} = \mathbf {X} ^ {\alpha} - \bigcup_ {i} ^ {N ^ {\prime}} \mathbf {X} _ {i} ^ {\alpha \rightarrow \beta}, \tag {11} +$$ + +$$ +\tilde {\mathbf {X}} ^ {\beta} = \bigcup_ {i} ^ {N ^ {\prime}} \left\{\delta \left(\mathbf {X} _ {i} ^ {\alpha \rightarrow \beta}\right)\right\}, \tag {12} +$$ + +$\delta (\mathbf{X}) = \delta (\{\mathbf{x}_i\} _i) = \sum_i\frac{\mathbf{a}_i\mathbf{s}_i\mathbf{x}_i}{\sum_{i'}\mathbf{a}_{i'}\mathbf{s}_{i'}}$ is the pooling operation considering the attention scores a and the size s of the tokens. Still, as shown in Step2 of Figure 3, the number of target tokens $\mathbf{X}^{\beta}$ cannot be reduced. To handle this issue, MCTF performs bidirectional bipartite soft matching by conducting the matching in the opposite direction with the updated token sets $\hat{\mathbf{X}}^{\alpha}$ , and $\hat{\mathbf{X}}^{\beta}$ as in Step 3, 4 of Figure 3. The final output tokens $\hat{\mathbf{X}} = \hat{\mathbf{X}}^{\alpha}\cup \hat{\mathbf{X}}^{\beta}$ are defined with the following. + +$$ +\hat {\mathbf {X}} ^ {\alpha} = \bigcup_ {i} ^ {N ^ {\prime} - r} \left\{\delta \left(\tilde {\mathbf {X}} _ {i} ^ {\beta \rightarrow \alpha}\right)\right\}, \tag {13} +$$ + +$$ +\hat {\mathbf {X}} ^ {\beta} = \tilde {\mathbf {X}} ^ {\beta} - \bigcup_ {i} ^ {N ^ {\prime} - r} \tilde {\mathbf {X}} _ {i} ^ {\beta \rightarrow \alpha}. \tag {14} +$$ + +Note that calculating the pairwise weights with updated two sets of tokens $\tilde{\mathbf{w}}_{ij} = \mathbf{W}(\tilde{\mathbf{x}}_i^\beta ,\tilde{\mathbf{x}}_j^\alpha)$ introduces the additional computational costs of $O(N^{\prime}(N^{\prime} - r))$ . To avoid this overhead, we approximate the attraction function by the attraction scores before fusion. In short, we just reuse the + +![](images/1efd23ccb0b313862831b3fdb49f340803b9b2724c0de026249128d145b79577.jpg) + +![](images/59ea783f68764fae6e21c35b758f3ae1726e2bf6a52391ab04ef35f15b40c197.jpg) + +![](images/99df14cd404d71f06b1341db84ff09835c40033236f720e376c46c068fa4eece.jpg) + +![](images/5206213c74a72aa3df8b6a2ac0ea81103293bdbcf10101d41fd45eca7b50fba3.jpg) + +![](images/1b0f0a9ad6831157ff29dfd8c8e311b5d41e5186b2abd6c31db8d01892b96c68.jpg) +Figure 4. Visualization of attentiveness in consecutive layers. +Figure 5. Illustration of attention map in the consecutive layers and approximated attention. (Left) The attention score $\hat{\mathbf{A}}^l$ is the past influence of the tokens to generate $\mathbf{X}^l$ . If we fuse the tokens $\mathbf{X}^l$ based on $\mathbf{A}^l$ , $\mathbf{x}_1$ is prone to be fused despite the highest informativeness score in the following attention. So, we instead leverage the informativeness based on the one-step-ahead attention $\mathbf{A}^{l + 1}$ . (Right) After the fusion, we also aggregate the $\mathbf{A}^{l + 1}$ to approximate the attention map $\hat{\mathbf{A}}^{l + 1}$ for updating fused tokens $\hat{\mathbf{X}}^l$ . + +pre-calculated weights since $\tilde{\mathbf{X}}^{\alpha}$ is the subset of $\mathbf{X}^{\alpha}$ . This allows MCTF to efficiently reduce tokens considering bidirectional relations between two subsets with negligible extra costs compared to uni-directional bipartite soft matching. + +# 3.3. One-step-ahead attention for informativeness + +In assessing informativeness, prior works [17, 18, 21] have leveraged the attention scores from the previous self-attention layer. As illustrated in Figure 5, previous approaches use the attention $\mathbf{A}^l$ from the previous layer to fuse tokens $\mathbf{X}^l$ . This technique allows efficient assessment under the assumption that the attention maps in consecutive layers are similar. However, we observed that the attention maps often substantially differ, as shown in Figure 4, and the attention from a previous layer may lead to suboptimal token fusion. Thus, we proposed one-step-ahead attention, which measures the informativeness of tokens based on the attention map in the next layer, i.e., $\mathbf{A}^{l + 1}$ . Then, the informativeness scores $\mathbf{a}$ in Equation (4) is calculated with $\mathbf{A}^{l + 1}\in \mathbb{R}^{N\times N}$ . This simple remedy provides a considerable improvement; see Figure 7b in Section 4.2. After token fusion, we efficiently compute the attention map $\hat{\mathbf{A}}^{l + 1}\in \mathbb{R}^{(N - r)\times (N - r)}$ of fused tokens $\hat{\mathbf{X}}^l\in \mathbb{R}^{(N - r)\times C}$ by simply aggregating $\mathbf{A}^{l + 1}\in \mathbb{R}^{N\times N}$ without recomputing the dot-product self-attention. To be specific, when the tokens are fused as $\delta (\{\mathbf{x}_i\}_i)$ during Equations (10) to (14), their corresponding one-step-ahead attention scores are also fused as $\delta (\{\mathbf{A}_i^{l + 1}\}_i)$ in both query and key direction. Note that when fusing attention scores for queries we use simple sum for $\delta ,i.e.$ $\forall_{i}\sum_{j}\hat{\mathbf{A}}_{ij}^{l + 1} = 1$ . For fusing attention + +![](images/68a1062d07f2665f5fa55a6beae484cf46cc2ef602536b54cc1a028426f376e9.jpg) +Figure 6. Illustration of training with token reduction consistency. During training, we forward the input $x$ as $f(x; r)$ , and $f(x; r')$ , respectively. To obtain the augmented representation, $r'$ is randomly selected in every step, and the model is updated with supervisory signals $\mathcal{L}_{\mathrm{CE}}$ , and consistency loss $\mathcal{L}_{\mathrm{MSE}}$ . + +scores for queries, we use simple sum for $\delta$ to guarantee $\forall_{i}\sum_{j}\hat{\mathbf{A}}_{ij}^{l + 1} = 1.$ + +# 3.4. Token reduction consistency + +We here propose a new fine-tuning scheme to further improve the performance of vision Transformer $f_{\theta}(\cdot ;r)$ with MCTF. We observe that a different number of reduced tokens per layer, denoted as $r$ , may lead to different representations of samples. By training Transformers with different $r$ and encouraging the consistency between them, namely, token reduction consistency, we achieve the additional performance gain. The objective function of our method is given as + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {L} _ {\mathrm {C E}} \left(f _ {\theta} (x; r), y\right) + \mathcal {L} _ {\mathrm {C E}} \left(f _ {\theta} \left(x; r ^ {\prime}\right), y\right) \\ + \lambda \mathcal {L} _ {\mathrm {M S E}} \left(\mathbf {x} _ {r} ^ {\mathrm {c l s}}, \mathbf {x} _ {r ^ {\prime}} ^ {\mathrm {c l s}}\right), \tag {15} \\ \end{array} +$$ + +where $(x,y)$ is a supervised sample, $r,r^{\prime}$ is the fixed and dynamic reduced token numbers, $\lambda$ is the coefficient for consistency loss, and $\mathbf{x}_r^{\mathrm{cls}},\mathbf{x}_{r'}^{\mathrm{cls}}$ are the class tokens in the last layer of models $f_{\theta}(x;r),f_{\theta}(x;r')$ . In this objective, we first calculate the cross-entropy loss $\mathcal{L}_{\mathrm{CE}}(f_{\theta}(x;r),y)$ with fixed $r$ , which is the target reduction number that will be used in the evaluation. At the same time, we generate another representation of the input $x$ with smaller but randomly drawn $r^\prime \sim$ uniform $(0,r)$ , and calculate the loss $\mathcal{L}_{\mathrm{CE}}(f_{\theta}(x;r'),y)$ Then, we impose the token consistency loss $\mathcal{L}_{\mathrm{MSE}}(\mathbf{x}_r^{\mathrm{cls}},\mathbf{x}_{r'}^{\mathrm{cls}})$ on the class tokens, to retain the consistent representation across the diverse reduced token numbers $r^\prime$ . The proposed method can be viewed as a new type of token-level data augmentation [7, 19] and consistency regularization. Our token reduction consistency encourages the representation $\mathbf{x}_r^{\mathrm{cls}}$ obtained by the target reduction number $r$ to mimic the slightly augmented representation $\mathbf{x}_{r'}^{\mathrm{cls}}$ , which is more similar to ones with no token reduction since $r^{\prime} < r$ + +Table 1. Image classification results + +
MethodFLOPs (G)Params (M)Top-1 Acc (%)
DeiT-T [28]1.2572.2 (-)
+EvoViT[AAAI '22] [36]0.8572.0 (-0.2)
+A-ViT[CVPR '22] [37]0.8571.0 (-1.2)
+SPViT[ECCV '22] [17]0.9572.1 (-0.1)
+ToMe[ICLR '23] [4]0.7571.3 (-0.9)
+BAT[CVPR '23] [21]0.8572.3 (+0.1)
+MCTFr=160.7572.7 (+0.5)
DeiT-S [28]4.62279.8 (-)
+IA-RED2[NeurIPS '21] [24]3.22279.1 (-0.7)
+DynamicViT[NeurIPS '21] [26]2.92379.3 (-0.5)
+EvoViT[AAAI '22] [36]3.02279.4 (-0.4)
+EViT[ICLR '22] [18]3.02279.5 (-0.3)
+A-ViT[CVPR '22] [37]3.62278.6 (-1.2)
+ATS[ECCV '22] [12]2.92279.7 (-0.1)
+SPViT[ECCV '22] [17]2.62279.3 (-0.5)
+ToMe[ICLR '23] [4]2.72279.4 (-0.4)
+BAT[CVPR '23] [21]3.02279.6 (-0.2)
+MCTFr=162.62280.1 (+0.3)
+ +# 4. Experiments + +Baselines. To validate the effectiveness of the proposed methods, we compare MCTF with the previous token reduction methods. For comparison, we opt the token pruning methods (A-ViT [37], IA-RED² [24], DynamicViT [26], EvoViT [36], ATS [12]) and token fusion methods (SPViT [17], EViT [18], ToMe [4], BAT [21]) in DeiT [28], and report the efficiency (FLOPs (G)) and the performance (Top-1 Acc (\%)) of each method. Further, to validate MCTF on other Vision Transformers (T2T-ViT [39], LV-ViT [15]), we report the results of MCTF and compare it with the official number of existing works. We denote the number of reduced tokens per layer $r$ with the subscript in Tables 1 and 2. The gray color in the table indicates the base model, and the green and red color indicates the improvements and degradations of the performance compared to the base model, respectively. + +# 4.1. Experimental Results + +Comparison of the token reduction methods. The comparison with existing token reduction methods is summarized in Table 1. We demonstrate that our MCTF achieves the best performance with the lowest FLOPs in DeiT [28] surpassing all previous works. Further, it is worth noting that MCTF is the only work that avoids performance degradation with the lowest FLOPs in both DeiT-T and DeiT-S. Through Finetuning DeiT-T for 30 epochs, MCTF brings a significant gain of $+0.5\%$ in accuracy over the base model with nearly half + +Table 2. Comparison with other Vision Transformers + +
ModelsFLOPs (G)Params (M)Acc (%)
PVT-Small[31]3.824.579.8
PVT-Medium [31]6.744.281.2
CoaT Mini [35]6.810.080.8
CoaT-Lite Small [35]4.020.081.9
Swin-T [20]4.529.081.3
Swin-S [20]8.750.083.0
PoolFormer-S36 [38]5.031.081.4
PoolFormer-M48 [38]11.673.082.5
T2T-ViTt-14 [39]6.121.581.7
+MCTFr=134.221.581.8 (↑)
T2T-ViTt-19 [39]9.839.282.4
+MCTFr=96.439.282.4 (-)
LV-ViT-S [15]6.626.283.3
+EVIT(ICLR '22] [18]4.726.283.0 (↓)
+BAT[CVPR '23] [21]4.726.283.1 (↓)
+DynamicViT[NeurIPS '21] [26]4.626.983.0 (↓)
+SPViT[ECCV '22] [17]4.326.283.1 (↓)
+MCTFr=124.226.283.4 (↑)
+ +FLOPs. Similarly, we observe a gain of $+0.3\%$ with DeiT-S while boosting the FLOPs by $-2.0$ (G). We believe that multi-criteria with one-step-ahead attention helps the model to minimize the loss of information; further consistency loss on the class token through the token reduction improves the generalizability of the model. + +MCTF with other Vision Transformers. To validate the applicability of MCTF in various ViTs, we demonstrate MCTF with other transformer architectures in Table 2. Following previous works [17, 18, 21, 26], we apply MCTF with LVViT. Also, we present the results of MCTF with T2T-ViT. As presented in the table, our experimental results are promising. MCTF in these architectures gets at least $31\%$ speedup without performance degradation. Further, MCTF combined with LV-ViT outperforms all other Transformers and token reduction methods regarding FLOPs, and accuracy. Especially, it is worth noting that all token reduction methods except for MCTF bring the performance degradation in LVViT. These results reveal that MCTF is the efficient token reduction method for the diverse Vision Transformers. + +Token reduction without training. Similar to ToMe [4], MCTF is applicable with pre-trained ViTs without any additional training since MCTF does not require any learnable parameters. We here apply the two reduction methods to the pre-trained DeiT without finetuning and provide the results in Table 3. Regardless of the reduced number of tokens $r$ in each layer, MCTF consistently surpasses ToMe. Especially, in the most sparse setting $r = 20$ , the performance gap is significant ( $+7.0\%$ in DeiT-T, $+3.8\%$ in DeiT-S). Note + +![](images/d0626cb853e7b60ab729c17db8b9c741ba4b3d0a046296ae10bed2ad55a86723.jpg) +Figure 7. Ablations on (a) multi-criteria, (b) one-step-ahead-attention, and token reduction consistency. Each marker indicates the model with $r \in [1,20]$ , and we highlight $r \in \{5,10,15,20\}$ as bordered circle. We also denote the model as star when $r = 16$ , which is used for finetuning the model. + +![](images/14a979a63359a92a0d6246733b9e8852dc305f08f05e303a29162dd01f72d07d.jpg) + +Table 3. Image classification results without training + +
Methodr
Base1248121620
DeiT-T
ToMe [4]72.272.172.072.071.670.868.761.5
MCTF72.272.272.172.172.071.771.068.5
DeiT-S
ToMe [4]79.879.879.779.779.479.077.974.2
MCTF79.879.879.879.879.879.679.278.0
+ +that without any additional training, our $\mathrm{MCTF}_{r=16}$ with pre-trained DeiT-S still shows a competitive performance of $79.2\%$ compared to reduction methods requiring training (e.g., $78.6\%$ of A-ViT, $79.1\%$ of IA-RED $^2$ , and $79.3\%$ of DynamicViT, and SPViT in Table 1). + +# 4.2. Ablation studies on MCTF + +We provide ablation studies to validate each component of MCTF. Unless otherwise stated, we conduct whole experiments with DeiT-S finetuned with MCTF ( $r = 16$ ). We provide the FLOPs-Accuracy graph by adjusting the reduced number of tokens per layer $r \in [1,20]$ . + +Multi-criteria. We explore the effectiveness of multi-criteria in Figure 7a. First, regarding the multi-criteria, we utilize three criteria for MCTF, i.e., similarity (sim.), informativeness (info.), and size. Each single criterion of similarity and informativeness shows a relatively inferior performance compared to dual (sim. & info.) and multi-criteria (sim. & info. & size). Specifically, when $r = 16$ , the performance of a single criterion is $79.7\%$ , and $79.4\%$ with similarity and informativeness, respectively. Then, adopting dual criteria (sim. & info.), MCTF achieves $79.8\%$ . Finally, we get an accuracy of $80.1\%$ with a gain of $+0.3\%$ by respecting all + +three criteria (sim. & info. & size). These performance gaps get larger as $r$ increases, which proves the importance of the multi-criteria for token fusion. + +One-step-ahead attention and token reduction consistency. To show the validity of one-step-ahead attention and token reduction consistency, we also provide the results of MCTF with and without each component in Figure 7b. When eliminating either one-step-ahead attention or token reduction consistency, the accuracies are dropped in every FLOP. This significant drop indicates that both approaches matter for MCTF. In short, by adopting one-step-ahead attention and token reduction consistency, MCTF effectively mitigates the performance degradation in a wide range of FLOPs. + +Comparison of design choices. The ablations on design choices are presented in Table 4. First, our bidirectional bipartite matching, which enables capturing the bidirectional relation in two sets, enhances the accuracy compared to one-way bipartite matching. Next, for pooling operation $\delta$ , the weighted sum considering the size $s$ and attentiveness $a$ is a better choice than others like max-pool or average. Lastly, we compare the results with the precise and approximated attention for $\hat{\mathbf{A}}^l$ . For precise attention, we just conduct the similarity calculation for one-step-ahead attention and the attention in the self-attention layer after fusion, separately. Otherwise, we approximate it with one-step-ahead attention as described in Section 3.3. As presented in the table, our approximated attention maintains the performance with the substantial improvement in efficiency (-0.4 (G) FLOPs). + +# 4.3. Analyse of MCTF + +Qualitative results. For a better understanding of MCTF, we provide the qualitative results of MCTF in Figure 8. We visualize the fused tokens at the last block of DeiT-S on ImageNet-1K and denote the fused tokens by the same border color. As shown in the figure, since the tokens are merged + +![](images/4afb32ad419c3fe1b70f201ed7fd7bb8d8f34b37a36121c9d3f0cc1473a41798.jpg) + +![](images/a7af67b7dff5a9a184c724a23658843b70f2c386b9abe9aad747cde52cc0bf70.jpg) + +![](images/16e8435975dad32dea31b872d46f7f5bcac044f0edceb4ca319eb878667bf030.jpg) + +![](images/e6efb986832f72af6d0b7121735852b9ed8049f6f2b69570d354f47487483cdb.jpg) + +![](images/49851728a0fed6734a5653551e1148a59857223fff9303ea43adb2d4db91361a.jpg) + +![](images/0f60e0ff937b8bf353c7c02c965d198f6c9426570f202c1384ebaaa28a411ed7.jpg) + +![](images/074f72928dd8ff2f3faae726dfdc8a6d58953ccc68ab5bebfa8a52f864fb930b.jpg) + +![](images/c4ed23b403ec7dd12409d09838c3f3faddf2138b4f05e96ce10c16036149a95e.jpg) + +![](images/3c70c77a25b8cdc52e1440a7bf0e80025cdcd67bb8b0b6d8cbfa1d423c0d69e1.jpg) +Figure 8. Visualization of the fused tokens with MCTF. Given the input images of ImageNet-1K (Top), the qualitative results of MCTF with DeiT-S are provided at the bottom. The same border color of the patches indicates the fused tokens. + +![](images/0b4778632b1db8ab333fb7c8584d5dd0ffd43d9b20816691c2420b71db2d3be3.jpg) + +![](images/e7645866783e0d1187c2758d549d0fe9fc6cf2a3819135065ab030e51d3a5510.jpg) + +![](images/c90f29dbe07f3c44d77a75c67fea121eb4bfa54f825644dd11c11e011a983732.jpg) + +![](images/2f6ef5e81f64d1587a7ed67f95bfffedcf8ed205fbfa5e05f000a7ef2fa000bd.jpg) + +![](images/8d19e33361ea236cea5de533f7f20db979196958c0194cb91a05a2727bca37ce.jpg) + +![](images/eb064ac90519e417a61cd0e87b153a55c8f261e3fc1c9d0fe6f32b2b28ba8db4.jpg) + +![](images/8d99bb75710b808bd952169ccfe6cf094e5115c09973b59c11439c1b3654a574.jpg) + +Table 4. Ablations of the design choices. + +
MethodFLOPs ↓ (G)Acc ↑ (%)
DeiT-S4.679.8
bipartite soft matching
One-way2.680.0
Bidirectional2.680.1
pooling function δ
average2.680.0
max2.679.8
weighted average2.680.1
approximation of attention map
precise attention3.080.1
approximated attention2.680.1
+ +with multi-criteria (e.g., similarity, informativeness, size), we maintain the more diverse tokens in the informative foreground object. For instance, in the third image of the hamster, while the background patches including the hand are fused into one token, the foreground tokens are less fused while maintaining the details like the eye, ear, and face of the hamster. In short, compared to the background, the foreground tokens are less fused with the moderate size retaining the information of the main content. + +Soundness of size criterion. Figure 9 presents the histogram of sizes of tokens after token reduction with and without size criterion. Specifically, we measure the size of the largest token at the last block and provide the histogram. With our size criterion, the merged tokens tend to have smaller sizes s showing the average size of 39.3/49.2 with and without the Size criterion, respectively. As intended, MCTF successfully suppresses the large-sized tokens, which are a source of information loss, leading to performance improvement. + +# 5. Conclusion + +In this work, we introduced the Multi-Criteria Token Fusion (MCTF), a novel strategy aimed at reducing the complex + +![](images/213dabcea98cb5373ee53e8066e805dda7e83b670889a3b52385efb5f4cd4343.jpg) +Figure 9. Histogram of the size of tokens after reduction. + +ity inherent in ViTs while mitigating performance degradation. MCTF effectively discerns the relation of tokens based on multiple criteria, including similarity, informativeness, and the size of the tokens. Our comprehensive ablation studies and detailed analyses demonstrate the efficacy of MCTF particularly with our innovative one-step-ahead attention and token reduction consistency. Remarkably, DeiT-T and DeiT-S with MCTF achieve considerable improvements, with $+0.5\%$ , and $+0.3\%$ increase in Top-1 Accuracy over the vanilla models, accompanied by about $44\%$ fewer FLOPs, respectively. We also observe that our MCTF outperforms all of the previous token reduction methods in diverse vision Transformers with and without training. + +# Acknowledgments + +This work was supported by ICT Creative Consilience Program through the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT)(IITP-2024-2020-0-01819), the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT)(NRF-2023R1A2C2005373), and a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI) funded by the Ministry of Health & Welfare Republic of Korea (HR20C0021). + +# References + +[1] Moab Arar, Ariel Shamir, and Amit H Bermano. Learned queries for efficient local attention. In CVPR, 2022. 1, 2 +[2] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In ECCV, 2022. 2 +[3] Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv:2004.05150, 2020. 1 +[4] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. ICLR, 2022. 1, 2, 3, 4, 6, 7 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1 +[6] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, 2021. 2 +[7] Hyeong Kyu Choi, Joonmyung Choi, and Hyunwoo J. Kim. Tokenmixup: Efficient attention-guided token-level data augmentation for transformers. In NeurIPS, 2022. 5 +[8] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. *ICLR*, 2021. 2 +[9] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. NeurIPS, 2021. 1, 2 +[10] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, 2022. 2 +[11] 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. ICLR, 2020. 1, 2 +[12] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In ECCV, 2022. 1, 2, 3, 6 +[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2 +[14] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, 2021. 2 +[15] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. NeurIPS, 2021. 2, 6 +[16] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. In ICLR, 2020. 1, 2 + +[17] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In ECCV, 2022. 1, 2, 3, 5, 6 +[18] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. ICLR, 2022. 1, 2, 3, 5, 6 +[19] Jihao Liu, Boxiao Liu, Hang Zhou, Hongsheng Li, and Yu Liu. Tokenmix: Rethinking image mixing for data augmentation in vision transformers. In ECCV, 2023. 5 +[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 6 +[21] Sifan Long, Zhen Zhao, Jimin Pi, Shengsheng Wang, and Jingdong Wang. Beyond attentive tokens: Incorporating token importance and diversity for efficient vision transformers. In CVPR, 2023. 1, 2, 3, 5, 6 +[22] Dmitrii Marin, Jen-Hao Rick Chang, Anurag Ranjan, Anish Prabhu, Mohammad Rastegari, and Oncel Tuzel. Token pooling in vision transformers for image classification. In WACV, 2023. 1, 2, 3 +[23] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In CVPR, 2022. 1, 2, 3 +[24] Bowen Pan, Rameswar Panda, Yifan Jiang, Zhangyang Wang, Rogerio Feris, and Aude Oliva. IA-RED2: Interpretability-aware redundancy reduction for vision transformers. NeurIPS, 2021. 1, 2, 3, 6 +[25] 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 ICML, 2021. 1 +[26] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. NeurIPS, 2021. 1, 2, 3, 6 +[27] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, 2021. 1 +[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 1, 2, 6 +[29] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, 2021. 1, 2 +[30] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv:2006.04768, 2020. 1, 2 +[31] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense predic + +tion without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 1, 2, 6 +[32] QuanLin Wu, Hang Ye, Yuntian Gu, Huishuai Zhang, Liwei Wang, and Di He. Denoising masked autoencoders help robust classification. In ICLR, 2023. 2 +[33] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS, 2021. 1 +[34] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. In AAAI, 2021. 1, 2 +[35] Weijian Xu, Yifan Xu, Tyler Chang, and Zhuowen Tu. Coscale conv-attentional image transformers. In ICCV, 2021. 6 +[36] Yifan Xu, Zhijie Zhang, Mengdan Zhang, Kekai Sheng, Ke Li, Weiming Dong, Liqing Zhang, Changsheng Xu, and Xing Sun. Evo-vit: Slow-fast token evolution for dynamic vision transformer. In AAAI, 2022. 6 +[37] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In CVPR, 2022. 1, 2, 3, 6 +[38] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 6 +[39] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In ICCV, 2021. 2, 6 +[40] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In CVPR, 2022. 2 +[41] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. *ICLR*, 2021. 1 \ No newline at end of file diff --git a/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/images.zip b/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5af87879f2a7595b02125c749c7b5aea712b5dd1 --- /dev/null +++ b/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92e7a4a762677f08777d1419c765ae085ab3c5b149de8e174ba92bc09db3b227 +size 758027 diff --git a/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/layout.json b/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c69e55c36c154f25df604f67903ec0a7bbbe9c74 --- /dev/null +++ b/2024/Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers/layout.json @@ -0,0 +1,10227 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 479, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 479, + 138 + ], + "type": "text", + "content": "Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 160, + 442, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 160, + 442, + 176 + ], + "spans": [ + { + "bbox": [ + 151, + 160, + 442, + 176 + ], + "type": "text", + "content": "Sanghyeok Lee* Joonmyung Choi* Hyunwoo J. Kim†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 463, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 463, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 463, + 201 + ], + "type": "text", + "content": "Department of Computer Science and Engineering, Korea University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 181, + 202, + 408, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 202, + 408, + 215 + ], + "spans": [ + { + "bbox": [ + 181, + 202, + 408, + 215 + ], + "type": "text", + "content": "{cat0626, pizard, hyunwoojkim}@korea.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 242, + 192, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 242, + 192, + 255 + ], + "spans": [ + { + "bbox": [ + 143, + 242, + 192, + 255 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 268, + 290, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 268, + 290, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 268, + 290, + 567 + ], + "type": "text", + "content": "Vision Transformer (ViT) has emerged as a prominent backbone for computer vision. For more efficient ViTs, recent works lessen the quadratic cost of the self-attention layer by pruning or fusing the redundant tokens. However, these works faced the speed-accuracy trade-off caused by the loss of information. Here, we argue that token fusion needs to consider diverse relations between tokens to minimize information loss. In this paper, we propose a Multi-criteria Token Fusion (MCTF), that gradually fuses the tokens based on multi-criteria (i.e., similarity, informativeness, and size of fused tokens). Further, we utilize the one-step-ahead attention, which is the improved approach to capture the informativeness of the tokens. By training the model equipped with MCTF using a token reduction consistency, we achieve the best speed-accuracy trade-off in the image classification (ImageNet1K). Experimental results prove that MCTF consistently surpasses the previous reduction methods with and without training. Specifically, DeiT-T and DeiT-S with MCTF reduce FLOPs by about " + }, + { + "bbox": [ + 47, + 268, + 290, + 567 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 47, + 268, + 290, + 567 + ], + "type": "text", + "content": " while improving the performance " + }, + { + "bbox": [ + 47, + 268, + 290, + 567 + ], + "type": "inline_equation", + "content": "(+0.5\\%, \\text{and} +0.3\\%)" + }, + { + "bbox": [ + 47, + 268, + 290, + 567 + ], + "type": "text", + "content": " over the base model, respectively. We also demonstrate the applicability of MCTF in various Vision Transformers (e.g., T2T-ViT, LV-ViT), achieving at least " + }, + { + "bbox": [ + 47, + 268, + 290, + 567 + ], + "type": "inline_equation", + "content": "31\\%" + }, + { + "bbox": [ + 47, + 268, + 290, + 567 + ], + "type": "text", + "content": " speedup without performance degradation. Code is available at https://github.com/mlvlab/MCTF." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 611, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 288, + 684 + ], + "type": "text", + "content": "Vision Transformer [11] (ViT) has been proposed to tackle the vision tasks with self-attention, originally developed for natural language processing tasks. With the advent of ViT, Transformers are the prevalent architectures for a wide range of vision tasks, e.g., classification [11, 20, 28, 29, 31], object detection [5, 31, 41], segmentation [20, 27, 33], etc. ViTs," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 243, + 547, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 547, + 375 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 547, + 375 + ], + "type": "text", + "content": "built only with self-attention and MLP, provide great flexibility and impressive performance compared to conventional approaches, e.g., convolutional neural networks (CNNs). However, despite these advantages, the quadratic computational complexity of self-attention with respect to the number of tokens is the major bottleneck for Transformers. This limitation becomes more substantial with the growing interest in large-scale foundation models such as CLIP [25]. To this end, several works [3, 16, 30, 34] have proposed efficient self-attention mechanisms including local self-attention within predefined windows [1, 9, 20]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 375, + 547, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 547, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 547, + 577 + ], + "type": "text", + "content": "More recently, there has been increasing interest in token-reduction methods for optimizing ViTs without altering their architecture. Earlier works [12, 23, 24, 26, 37] primarily focused on pruning the uninformative tokens to reduce the number of tokens. Another line of works [4, 17, 18, 21, 22] attempted to fuse the tokens instead of discarding them to minimize the information loss. However, performance degradation is still commonly observed in most token fusion methods. We notice that the token fusion methods usually consider only one criterion, such as the similarity or informativeness of tokens, leading to suboptimal token matching. For instance, similarity-based token fusion is prone to combine the foreground tokens, whereas informativeness-based fusion often merges substantially dissimilar tokens, resulting in collapsed representations. Furthermore, if too many tokens are fused into one token, then information loss is inevitable." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 578, + 548, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 548, + 710 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 548, + 710 + ], + "type": "text", + "content": "To address the problems, we introduce Multi-Criteria Token Fusion (MCTF) that optimizes vision transformers by fusing tokens based on multi-criteria. Unlike previous works that consider a single criterion for token fusion, MCTF measures the relationship between the tokens with multi-criteria as follows; (1) similarity to fuse the redundant tokens, (2) informativeness to reduce the uninformative tokens, (3) the size of the tokens to prevent the large-sized tokens that boost the loss of information. Also, to tackle the inconsistency between attention maps of consecutive layers, we adopt one-step-ahead attention, which explicitly estimates the informativeness of" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15741" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 289, + 251 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 289, + 251 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 289, + 251 + ], + "type": "image", + "image_path": "afb4d8e49b0642017b3ab2ced053a4fc3c9392bafd01ad9dff78a810b9d0e898.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 548, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 548, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 548, + 308 + ], + "type": "text", + "content": "Figure 1. Comparison of the token reduction methods with DeiT-T (left), and DeiT-S (right). Given a base model marked as blue circle, previous token reduction methods accelerate the speed with the trade-off between accuracy and computational cost. Our MCTF, marked as a star, even brings performance improvements while lessening the complexity of DeiT. Note that after only one finetuning with the specific reduced number of tokens marked as red star, we simply evaluate it with the diverse FLOPs by adjusting the reduced numbers." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 545, + 251 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 251 + ], + "type": "image", + "image_path": "e3627aae79897e1cb056d4f57193d13d32a28d6e895d8b71f5c6b25c951ec20c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 318, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 288, + 437 + ], + "type": "text", + "content": "the tokens in the next layer. Finally, by introducing a token reduction consistency for finetuning the model, we achieve superior performance to the existing works as in Figure 1. Surprisingly, our MCTF even performs better than the 'full' base model (red dotted line) with a reduced computational complexity. Specifically, it brings a " + }, + { + "bbox": [ + 46, + 318, + 288, + 437 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 318, + 288, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 318, + 288, + 437 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 46, + 318, + 288, + 437 + ], + "type": "text", + "content": " gain while reducing FLOPs by about " + }, + { + "bbox": [ + 46, + 318, + 288, + 437 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 46, + 318, + 288, + 437 + ], + "type": "text", + "content": " in DeiT-T, and DeiT-S [28], respectively. We have observed a similar speed-up " + }, + { + "bbox": [ + 46, + 318, + 288, + 437 + ], + "type": "inline_equation", + "content": "(31\\%)" + }, + { + "bbox": [ + 46, + 318, + 288, + 437 + ], + "type": "text", + "content": " in T2T-ViT [39], and LV-ViT [15] without any performance degradation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 437, + 235, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 235, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 235, + 448 + ], + "type": "text", + "content": "Our contributions are summarized in fourfold." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 450, + 287, + 605 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 47, + 450, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 287, + 497 + ], + "type": "text", + "content": "- We propose Multi-criteria Token Fusion, a novel token fusion method that considers multi-criteria, e.g., similarity, informativeness, and size, to capture the complex relationship of tokens and minimize information loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 498, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 287, + 534 + ], + "type": "text", + "content": "- For measuring the informativeness of the tokens, we utilize one-step-ahead attention to retain the attentive tokens in the following layers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 534, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 287, + 569 + ], + "type": "text", + "content": "- We propose a new fine-tuning scheme with token reduction consistency to boost the generalization performance of transformers equipped with MCTF." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 570, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 287, + 605 + ], + "type": "text", + "content": "- The extensive experiments demonstrate that MCTF achieves the best speed-accuracy trade-off in diverse ViTs, surpassing all previous token reduction methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 620, + 135, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 135, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 135, + 632 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Vision Transformers. Vision Transformer [11] is introduced to tackle the vision tasks. Later, DeiT [28] and CaiT [29] are proposed to handle the data efficiency and scalability of ViT, respectively. Recent works [6, 10, 14, 20, 31] tried to insert the inductive biases of CNNs on ViT, such as the locality or pyramid-architecture. In parallel, there is a" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 318, + 547, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 547, + 462 + ], + "type": "text", + "content": "line of works that boosts the vanilla ViT by scaling [29, 40] or self-supervised learning [2, 13, 32]. Despite the promising results of these works, the quadratic complexity of ViTs is still the major constraint for scaling the model. For the sake of mitigating the complexity, Reformer [16] lessens the quadratic complexity to " + }, + { + "bbox": [ + 304, + 318, + 547, + 462 + ], + "type": "inline_equation", + "content": "O(N \\log N)" + }, + { + "bbox": [ + 304, + 318, + 547, + 462 + ], + "type": "text", + "content": " through the hashing function, and Linformer [30], performer [8], and Nyströmformer [34] achieve the linear cost with the approximated linear attention. Also, several works [1, 9, 10, 20] utilize sparse attention with the reduced key or query. Swin [20] and Twins [9] utilize the local attention within the fixed size of the window to mitigate the complexity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": "Token reduction in ViTs. Most of the computational burden in ViTs arises from the self-attention. To reduce the quadratic cost in the number of tokens, recent works [4, 12, 17, 18, 21–24, 26, 37] have an interest in reducing the token itself. These works have the advantage of utilizing the original ViTs architecture without modification. In earlier works [12, 23, 24, 26, 37], the uninformative tokens are simply dropped during the forward process, leading to the information loss. To compensate for this, SPViT [17] and EViT [18] first split the tokens into informative and uninformative token sets based on attention scores, then fuse these uninformative token sets into a single token. In parallel, token pooling [22] and ToMe [4] combine the semantically similar tokens to reduce redundancies. A more recent study BAT [21] first split the tokens based on informativeness then fuse the tokens considering the diversity of the tokens. Despite the advantage of each criterion, successful integration of multi-criteria is still less explored." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15742" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 160, + 178 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 160, + 178 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 160, + 178 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 160, + 178 + ], + "type": "image", + "image_path": "11e3b5e1793a889a3d9c9a678880b8c0c8877ac09c70b47122b1d2da106c5011.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 182, + 121, + 193 + ], + "lines": [ + { + "bbox": [ + 87, + 182, + 121, + 193 + ], + "spans": [ + { + "bbox": [ + 87, + 182, + 121, + 193 + ], + "type": "text", + "content": "(a) Origin" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 70, + 289, + 180 + ], + "blocks": [ + { + "bbox": [ + 176, + 70, + 289, + 180 + ], + "lines": [ + { + "bbox": [ + 176, + 70, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 176, + 70, + 289, + 180 + ], + "type": "image", + "image_path": "fa08f1eaca5e8102ae9ab000f75eb90d923a88562bae277388ff7344d72d338b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 182, + 248, + 191 + ], + "lines": [ + { + "bbox": [ + 216, + 182, + 248, + 191 + ], + "spans": [ + { + "bbox": [ + 216, + 182, + 248, + 191 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 216, + 182, + 248, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\mathrm{sim}}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 305, + 70, + 417, + 181 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 417, + 181 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 417, + 181 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 417, + 181 + ], + "type": "image", + "image_path": "b627314e338ea533b512dc8a9dc206fd81ef73b25967bd2250fdc197147f4cd2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 182, + 390, + 191 + ], + "lines": [ + { + "bbox": [ + 332, + 182, + 390, + 191 + ], + "spans": [ + { + "bbox": [ + 332, + 182, + 390, + 191 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 332, + 182, + 390, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\mathrm{sim}}\\& \\mathbf{W}^{\\mathrm{info}}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "type": "text", + "content": "Figure 2. Visualization of the fused tokens. Given (a) the leftmost image, (b) fusing the tokens with a single criterion " + }, + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\mathrm{sim}}" + }, + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "type": "text", + "content": " often results in the excessive fusion of the foreground object. (c) Then considering both similarity and informativeness (" + }, + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\mathrm{sim}}\\& \\mathbf{W}^{\\mathrm{info}}" + }, + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "type": "text", + "content": "), tokens in the foreground objects are less fused while the tokens in the background are largely fused. (d) Finally, MCTF helps retain the information of each component in the image by preventing the large-size token with the multi-criteria (" + }, + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\mathrm{sim}}\\& \\mathbf{W}^{\\mathrm{info}}\\& \\mathbf{W}^{\\mathrm{size}}" + }, + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 434, + 70, + 545, + 181 + ], + "blocks": [ + { + "bbox": [ + 434, + 70, + 545, + 181 + ], + "lines": [ + { + "bbox": [ + 434, + 70, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 434, + 70, + 545, + 181 + ], + "type": "image", + "image_path": "f94704b97ecdeb4baccac06247d87493beb88ccbf29c6ee2cfb9b01417a5efb4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 182, + 533, + 191 + ], + "lines": [ + { + "bbox": [ + 447, + 182, + 533, + 191 + ], + "spans": [ + { + "bbox": [ + 447, + 182, + 533, + 191 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 447, + 182, + 533, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\mathrm{sim}}\\& \\mathbf{W}^{\\mathrm{info}}\\& \\mathbf{W}^{\\mathrm{size}}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 266, + 102, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 102, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 102, + 278 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 287, + 288, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 288, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 288, + 347 + ], + "type": "text", + "content": "We first review the self-attention and token reduction approaches (Section 3.1). Then, we present our multi-criteria token fusion (Section 3.2) that leverages one-step-ahead attention (Section 3.3). Lastly, we introduce a training strategy with token reduction consistency in Section 3.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 356, + 133, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 133, + 366 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 133, + 366 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 373, + 289, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 289, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 289, + 398 + ], + "type": "text", + "content": "In Transformers, tokens " + }, + { + "bbox": [ + 46, + 373, + 289, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 46, + 373, + 289, + 398 + ], + "type": "text", + "content": " are processed by self-attention defined as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 99, + 405, + 287, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 405, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 99, + 405, + 287, + 439 + ], + "type": "interline_equation", + "content": "\\operatorname {S A} (\\mathbf {X}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {C}}\\right) \\mathbf {V}, \\tag {1}", + "image_path": "1a265276c90a69b94d5fd9ce8e2e4759fd68026da1d7f64e9ab107d206bbf8cc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{Q},\\mathbf{K},\\mathbf{V} = \\mathbf{X}\\mathbf{W}_{\\mathbf{Q}},\\mathbf{X}\\mathbf{W}_{\\mathbf{K}},\\mathbf{X}\\mathbf{W}_{\\mathbf{V}}" + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{Q}},\\mathbf{W}_{\\mathbf{K}}" + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{V}}\\in \\mathbb{R}^{C\\times C}" + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "text", + "content": " are learnable weight matrices. Despite its outstanding expressive power, the self-attention does not scale well with the number of tokens " + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "text", + "content": " due to its quadratic time complexity " + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "inline_equation", + "content": "O(N^{2}C + NC^{2})" + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "text", + "content": ". To address this problem, a line of works [12, 23, 24, 26, 37] reduces the number of tokens simply by pruning uninformative tokens. These approaches often cause significant performance degradation due to the loss of information. Thus, another line of works [4, 17, 18, 21, 22] fuses the uninformative or redundant tokens " + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}\\subset \\mathbf{X}" + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "text", + "content": " into a new token " + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} = \\delta (\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "text", + "content": " is the set of original tokens, and " + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 446, + 289, + 626 + ], + "type": "text", + "content": " denotes a merging function, e.g., max-pooling or averaging. In this work, we also adopt 'token fusion' rather than 'token pruning' with multiple criteria to minimize the loss of information by token reduction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 635, + 194, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 194, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 194, + 645 + ], + "type": "text", + "content": "3.2. Multi-criteria token fusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 652, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 652, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 652, + 288, + 713 + ], + "type": "text", + "content": "Given a set of input tokens " + }, + { + "bbox": [ + 46, + 652, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 46, + 652, + 288, + 713 + ], + "type": "text", + "content": ", the goal of MCTF is to fuse the tokens into output tokens " + }, + { + "bbox": [ + 46, + 652, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}} \\in \\mathbb{R}^{(N - r) \\times C}" + }, + { + "bbox": [ + 46, + 652, + 288, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 652, + 288, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 652, + 288, + 713 + ], + "type": "text", + "content": " is the number of fused tokens. To minimize the information loss, we first evaluate the relations between the tokens based on multi-criteria, then group and merge the" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 268, + 515, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 268, + 515, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 268, + 515, + 280 + ], + "type": "text", + "content": "tokens through bidirectional bipartite soft matching." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 280, + 547, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 280, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 305, + 280, + 547, + 303 + ], + "type": "text", + "content": "Multi-criteria attraction function. We first define an attraction function " + }, + { + "bbox": [ + 305, + 280, + 547, + 303 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 305, + 280, + 547, + 303 + ], + "type": "text", + "content": " based on multiple criteria as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 350, + 311, + 545, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 311, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 350, + 311, + 545, + 326 + ], + "type": "interline_equation", + "content": "\\mathbf {W} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\Pi_ {k = 1} ^ {M} \\left(\\mathbf {W} ^ {k} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right)\\right) ^ {\\tau_ {k}}, \\tag {2}", + "image_path": "d7fcf0cc2079eb4ee1dabf7b972173e0b0fe8558d29f262ccd4b5c426fb8cde4.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 334, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 545, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 334, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^k: \\mathbb{R}^C \\times \\mathbb{R}^C \\to \\mathbb{R}_+" + }, + { + "bbox": [ + 304, + 334, + 545, + 407 + ], + "type": "text", + "content": " is an attraction function computed by " + }, + { + "bbox": [ + 304, + 334, + 545, + 407 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 334, + 545, + 407 + ], + "type": "text", + "content": "-th criterion, and " + }, + { + "bbox": [ + 304, + 334, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\tau^k \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 304, + 334, + 545, + 407 + ], + "type": "text", + "content": " is the temperature parameter to adjust the influence of " + }, + { + "bbox": [ + 304, + 334, + 545, + 407 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 334, + 545, + 407 + ], + "type": "text", + "content": "-th criterion. The higher attraction score between two tokens indicates a higher chance of being fused. In this work, we consider the following three criteria: similarity, informativeness, and size." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 407, + 547, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 547, + 455 + ], + "type": "text", + "content": "Similarity. The first criterion is the similarity of tokens to reduce redundant information. Akin to the previous works [4, 22] requiring the proximity of tokens, we leverage the cosine similarity between the set of tokens for" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 346, + 463, + 545, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 463, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 346, + 463, + 545, + 490 + ], + "type": "interline_equation", + "content": "\\mathbf {W} ^ {\\text {s i m}} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\frac {1}{2} \\left(\\frac {\\mathbf {x} _ {i} \\cdot \\mathbf {x} _ {j}}{\\| \\mathbf {x} _ {i} \\| \\| \\mathbf {x} _ {j} \\|} + 1\\right). \\tag {3}", + "image_path": "21c4f2c8f1f00553b4be21adf471a6434568ea0d3bf59d495643059e84da0efc.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 496, + 547, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 650 + ], + "type": "text", + "content": "Token fusion with similarity effectively eliminates the redundant tokens, yet it often excessively combines the informative tokens as in Figure 2b, causing the loss of information. Informativeness. To minimize the information loss, we introduce informativeness to avoid the fusion of informative tokens. To quantify the informativeness, we measure the averaged attention scores " + }, + { + "bbox": [ + 304, + 496, + 547, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{a}\\in [0,1]^N" + }, + { + "bbox": [ + 304, + 496, + 547, + 650 + ], + "type": "text", + "content": " in the self-attention layer, which indicates the impact of each token on others: " + }, + { + "bbox": [ + 304, + 496, + 547, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_j = \\frac{1}{N}\\sum_i^N\\mathbf{A}_{ij}" + }, + { + "bbox": [ + 304, + 496, + 547, + 650 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 496, + 547, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{ij} = \\mathrm{softmax}\\left(\\frac{\\mathbf{Q}_i\\mathbf{K}_j^T}{\\sqrt{C}}\\right)" + }, + { + "bbox": [ + 304, + 496, + 547, + 650 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 304, + 496, + 547, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i\\rightarrow 0" + }, + { + "bbox": [ + 304, + 496, + 547, + 650 + ], + "type": "text", + "content": ", there's no influence from " + }, + { + "bbox": [ + 304, + 496, + 547, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 496, + 547, + 650 + ], + "type": "text", + "content": " to other tokens. With the informativeness scores, we define an informativeness-based attraction function as" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 378, + 657, + 545, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 657, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 378, + 657, + 545, + 682 + ], + "type": "interline_equation", + "content": "\\mathbf {W} ^ {\\text {i n f o}} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\frac {1}{\\mathbf {a} _ {i} \\mathbf {a} _ {j}}, \\tag {4}", + "image_path": "e13bd4c289360e392a7758367e2bae7a3a4456afb00a26d516caa0ec7bb64535.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i, \\mathbf{a}_j" + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": " are the informative scores of " + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i, \\mathbf{x}_j" + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": ", respectively. When both tokens are uninformative (" + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i, \\mathbf{a}_j \\to 0" + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": ")," + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15743" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 71, + 521, + 190 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 521, + 190 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 521, + 190 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 521, + 190 + ], + "type": "image", + "image_path": "ac4bf9f15015a684ae0c54cb8542749ffb0d741b46ac00e74d63c75d8d9ce802.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 544, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 544, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 544, + 220 + ], + "type": "text", + "content": "Figure 3. Bidirectional bipartite soft matching. The set of tokens " + }, + { + "bbox": [ + 46, + 198, + 544, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 198, + 544, + 220 + ], + "type": "text", + "content": " is split into two groups " + }, + { + "bbox": [ + 46, + 198, + 544, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\alpha},\\mathbf{X}^{\\beta}" + }, + { + "bbox": [ + 46, + 198, + 544, + 220 + ], + "type": "text", + "content": ", and bidirectional bipartite soft matching are conducted through Step 1-4. The intensity of the lines indicates the multi-criteria weights " + }, + { + "bbox": [ + 46, + 198, + 544, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^t" + }, + { + "bbox": [ + 46, + 198, + 544, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 231, + 287, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 287, + 278 + ], + "type": "text", + "content": "the weight gets higher " + }, + { + "bbox": [ + 46, + 231, + 287, + 278 + ], + "type": "inline_equation", + "content": "(\\mathbf{W}^{\\mathrm{info}}(\\mathbf{x}_i,\\mathbf{x}_j)\\to \\infty)" + }, + { + "bbox": [ + 46, + 231, + 287, + 278 + ], + "type": "text", + "content": ", making two tokens prone to be fused. In Figure 2c, with the weights combined with the similarity and informativeness, the tokens in the foreground object are less fused." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 279, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 386 + ], + "type": "text", + "content": "Size. The last criterion is the size of the tokens, which indicates the number of fused tokens. Although tokens are not dropped but merged via a merging function, e.g., averaging pooling or max pooling, it is difficult to preserve all the information as the number of constituent tokens increases. So, the fusion between smaller tokens is preferred. To this end, we initially set the size " + }, + { + "bbox": [ + 46, + 279, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{N}^N" + }, + { + "bbox": [ + 46, + 279, + 287, + 386 + ], + "type": "text", + "content": " of tokens " + }, + { + "bbox": [ + 46, + 279, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 279, + 287, + 386 + ], + "type": "text", + "content": " as 1 and track the number of constituent (fused) tokens of each token, and define a size-based attraction function as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 396, + 287, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 396, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 120, + 396, + 287, + 422 + ], + "type": "interline_equation", + "content": "\\mathbf {W} ^ {\\text {s i z e}} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\frac {1}{\\mathbf {s} _ {i} \\mathbf {s} _ {j}}. \\tag {5}", + "image_path": "4eb022e40361a907100f8c75227969462a7a65423e7b8562f1d2e347ff06a1be.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 431, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 479 + ], + "type": "text", + "content": "In Figure 2d, tokens are merged based on the multi-criteria: similarity, informativeness, and size. We observed that the fusion happens between similar tokens and the fusion of foreground tokens or large tokens is properly suppressed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "text", + "content": "Bidirectional bipartite soft matching. Given the multicriteria-based attraction function " + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "text", + "content": ", our MCTF performs a relaxed bidirectional bipartite matching called bipartite soft matching [4]. One advantage of bipartite matching is that it alleviates the quadratic cost of similarity computation between tokens, i.e., " + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "inline_equation", + "content": "O(N^2) \\to O(N'^2)" + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "inline_equation", + "content": "N' = \\left\\lfloor \\frac{N}{2} \\right\\rfloor" + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "text", + "content": ". In addition, by relaxing the one-to-one correspondence constraints, the solution can be obtained by an efficient algorithm. In this relaxed matching problem, the set of tokens " + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "text", + "content": " is first split into the source and target " + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^\\alpha" + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^\\beta \\in \\mathbb{R}^{N' \\times C}" + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "text", + "content": " as in Step 1 of Figure 3. Given a set of binary decision variables, i.e., the edge matrix " + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\{0,1\\}^{N' \\times N'}" + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^\\alpha" + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^\\beta" + }, + { + "bbox": [ + 46, + 501, + 287, + 656 + ], + "type": "text", + "content": ", bipartite soft matching is formulated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 668, + 287, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 668, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 73, + 668, + 287, + 689 + ], + "type": "interline_equation", + "content": "\\mathbf {E} ^ {*} = \\arg \\max _ {\\mathbf {E}} \\sum_ {i j} \\mathbf {w} _ {i j} ^ {\\prime} \\mathbf {e} _ {i j} \\tag {6}", + "image_path": "f6fb2f3f6d482dace9fa83de06d697368ecbd7925041652e03bbbf1687d839d9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 96, + 692, + 287, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 692, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 96, + 692, + 287, + 712 + ], + "type": "interline_equation", + "content": "\\text {s u b j e c t} \\sum_ {i j} \\mathbf {e} _ {i j} = r, \\sum_ {j} \\mathbf {e} _ {i j} \\leq 1 \\forall i, \\tag {7}", + "image_path": "cd012784d1a7355638bcf63a7230d34a1febc3b9ca08a6aa73ea873134e4e9c6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 232, + 333, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 333, + 241 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 333, + 241 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 346, + 249, + 545, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 249, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 346, + 249, + 545, + 281 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {i j} ^ {\\prime} = \\left\\{ \\begin{array}{l l} \\mathbf {w} _ {i j} & \\text {i f} j \\neq \\arg \\max _ {j ^ {\\prime}} \\mathbf {w} _ {i j ^ {\\prime}} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {8}", + "image_path": "74e333c7702cdc5d29be786441adcfc6456f9ba89b168ee26e9c88167b05f269.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{ij}" + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "text", + "content": " indicates the presence of the edge between " + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "inline_equation", + "content": "i, j" + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "text", + "content": "-th token of " + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\alpha}, \\mathbf{X}^{\\beta}" + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "text", + "content": ", and, " + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{ij} = \\mathbf{W}(\\mathbf{x}_i^\\alpha, \\mathbf{x}_j^\\beta)" + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "text", + "content": ". This optimization problem can be solved by two simple steps: 1) find the best edge that maximizes " + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{ij}" + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "text", + "content": ", and 2) choose the top-" + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "text", + "content": " edges with the largest attraction scores. Then, based on the soft matching result " + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^*" + }, + { + "bbox": [ + 304, + 288, + 545, + 361 + ], + "type": "text", + "content": ", we group the tokens as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 342, + 368, + 545, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 368, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 342, + 368, + 545, + 384 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {j} ^ {\\alpha \\rightarrow \\beta} = \\left\\{\\mathbf {x} _ {i} ^ {\\alpha} \\in \\mathbf {X} ^ {\\alpha} \\mid \\mathbf {e} _ {i j} = 1 \\right\\} \\cup \\left\\{\\mathbf {x} _ {j} ^ {\\beta} \\right\\}, \\tag {9}", + "image_path": "ea8e48f00bd6f167a358e2f1eb9a73075ba2dd8da4eae2680bc3db25ca420e1a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 392, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 392, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i^{\\alpha \\rightarrow \\beta}" + }, + { + "bbox": [ + 304, + 392, + 545, + 418 + ], + "type": "text", + "content": " indicates the set of tokens matched with " + }, + { + "bbox": [ + 304, + 392, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^\\beta" + }, + { + "bbox": [ + 304, + 392, + 545, + 418 + ], + "type": "text", + "content": ". Finally, the results of the fusion " + }, + { + "bbox": [ + 304, + 392, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}}" + }, + { + "bbox": [ + 304, + 392, + 545, + 418 + ], + "type": "text", + "content": " are obtained as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 362, + 426, + 545, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 426, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 362, + 426, + 545, + 440 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {X}} = \\tilde {\\mathbf {X}} ^ {\\alpha} \\cup \\tilde {\\mathbf {X}} ^ {\\beta}, \\tag {10}", + "image_path": "4c3e4546e0d4906edb98c87d34b1a67137d4e6de3440a0a5c1c61e790340f46c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 358, + 443, + 545, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 443, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 358, + 443, + 545, + 464 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\tilde {\\mathbf {X}} ^ {\\alpha} = \\mathbf {X} ^ {\\alpha} - \\bigcup_ {i} ^ {N ^ {\\prime}} \\mathbf {X} _ {i} ^ {\\alpha \\rightarrow \\beta}, \\tag {11}", + "image_path": "d88ecc9d8ffb1b3d3ea8608758b1af261b67fc1d31968e50f8ba3bfbd770882c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 385, + 466, + 545, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 466, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 385, + 466, + 545, + 488 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {X}} ^ {\\beta} = \\bigcup_ {i} ^ {N ^ {\\prime}} \\left\\{\\delta \\left(\\mathbf {X} _ {i} ^ {\\alpha \\rightarrow \\beta}\\right)\\right\\}, \\tag {12}", + "image_path": "be35807c1b0b36c5eb9093d40a6d31156a0bd2a24d04ab6ffeb218aaff5792c2.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 495, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\delta (\\mathbf{X}) = \\delta (\\{\\mathbf{x}_i\\} _i) = \\sum_i\\frac{\\mathbf{a}_i\\mathbf{s}_i\\mathbf{x}_i}{\\sum_{i'}\\mathbf{a}_{i'}\\mathbf{s}_{i'}}" + }, + { + "bbox": [ + 304, + 495, + 545, + 594 + ], + "type": "text", + "content": " is the pooling operation considering the attention scores a and the size s of the tokens. Still, as shown in Step2 of Figure 3, the number of target tokens " + }, + { + "bbox": [ + 304, + 495, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\beta}" + }, + { + "bbox": [ + 304, + 495, + 545, + 594 + ], + "type": "text", + "content": " cannot be reduced. To handle this issue, MCTF performs bidirectional bipartite soft matching by conducting the matching in the opposite direction with the updated token sets " + }, + { + "bbox": [ + 304, + 495, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{\\alpha}" + }, + { + "bbox": [ + 304, + 495, + 545, + 594 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 495, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{\\beta}" + }, + { + "bbox": [ + 304, + 495, + 545, + 594 + ], + "type": "text", + "content": " as in Step 3, 4 of Figure 3. The final output tokens " + }, + { + "bbox": [ + 304, + 495, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}} = \\hat{\\mathbf{X}}^{\\alpha}\\cup \\hat{\\mathbf{X}}^{\\beta}" + }, + { + "bbox": [ + 304, + 495, + 545, + 594 + ], + "type": "text", + "content": " are defined with the following." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 365, + 601, + 545, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 601, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 365, + 601, + 545, + 622 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {X}} ^ {\\alpha} = \\bigcup_ {i} ^ {N ^ {\\prime} - r} \\left\\{\\delta \\left(\\tilde {\\mathbf {X}} _ {i} ^ {\\beta \\rightarrow \\alpha}\\right)\\right\\}, \\tag {13}", + "image_path": "40aec1378140fb4cf3cb80865a72aebb0f208b760f9d7dbc99d1f985d2226cf7.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 366, + 624, + 545, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 624, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 366, + 624, + 545, + 647 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {X}} ^ {\\beta} = \\tilde {\\mathbf {X}} ^ {\\beta} - \\bigcup_ {i} ^ {N ^ {\\prime} - r} \\tilde {\\mathbf {X}} _ {i} ^ {\\beta \\rightarrow \\alpha}. \\tag {14}", + "image_path": "6c3bc11155c6b3300aace8d4de902dc0eb9891f5124a2f976cd50fe4da9862f2.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "Note that calculating the pairwise weights with updated two sets of tokens " + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{w}}_{ij} = \\mathbf{W}(\\tilde{\\mathbf{x}}_i^\\beta ,\\tilde{\\mathbf{x}}_j^\\alpha)" + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": " introduces the additional computational costs of " + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "O(N^{\\prime}(N^{\\prime} - r))" + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": " . To avoid this overhead, we approximate the attraction function by the attraction scores before fusion. In short, we just reuse the" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "15744" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 106, + 137 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 106, + 137 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 106, + 137 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 106, + 137 + ], + "type": "image", + "image_path": "1efd23ccb0b313862831b3fdb49f340803b9b2724c0de026249128d145b79577.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 110, + 72, + 166, + 137 + ], + "blocks": [ + { + "bbox": [ + 110, + 72, + 166, + 137 + ], + "lines": [ + { + "bbox": [ + 110, + 72, + 166, + 137 + ], + "spans": [ + { + "bbox": [ + 110, + 72, + 166, + 137 + ], + "type": "image", + "image_path": "59ea783f68764fae6e21c35b758f3ae1726e2bf6a52391ab04ef35f15b40c197.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 170, + 72, + 227, + 137 + ], + "blocks": [ + { + "bbox": [ + 170, + 72, + 227, + 137 + ], + "lines": [ + { + "bbox": [ + 170, + 72, + 227, + 137 + ], + "spans": [ + { + "bbox": [ + 170, + 72, + 227, + 137 + ], + "type": "image", + "image_path": "99df14cd404d71f06b1341db84ff09835c40033236f720e376c46c068fa4eece.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 230, + 72, + 286, + 137 + ], + "blocks": [ + { + "bbox": [ + 230, + 72, + 286, + 137 + ], + "lines": [ + { + "bbox": [ + 230, + 72, + 286, + 137 + ], + "spans": [ + { + "bbox": [ + 230, + 72, + 286, + 137 + ], + "type": "image", + "image_path": "5206213c74a72aa3df8b6a2ac0ea81103293bdbcf10101d41fd45eca7b50fba3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 47, + 159, + 288, + 244 + ], + "blocks": [ + { + "bbox": [ + 50, + 146, + 284, + 157 + ], + "lines": [ + { + "bbox": [ + 50, + 146, + 284, + 157 + ], + "spans": [ + { + "bbox": [ + 50, + 146, + 284, + 157 + ], + "type": "text", + "content": "Figure 4. Visualization of attentiveness in consecutive layers." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 159, + 288, + 244 + ], + "lines": [ + { + "bbox": [ + 47, + 159, + 288, + 244 + ], + "spans": [ + { + "bbox": [ + 47, + 159, + 288, + 244 + ], + "type": "image", + "image_path": "1b0f0a9ad6831157ff29dfd8c8e311b5d41e5186b2abd6c31db8d01892b96c68.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "text", + "content": "Figure 5. Illustration of attention map in the consecutive layers and approximated attention. (Left) The attention score " + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}^l" + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "text", + "content": " is the past influence of the tokens to generate " + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l" + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "text", + "content": ". If we fuse the tokens " + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l" + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^l" + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "text", + "content": " is prone to be fused despite the highest informativeness score in the following attention. So, we instead leverage the informativeness based on the one-step-ahead attention " + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{l + 1}" + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "text", + "content": ". (Right) After the fusion, we also aggregate the " + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{l + 1}" + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "text", + "content": " to approximate the attention map " + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}^{l + 1}" + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "text", + "content": " for updating fused tokens " + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^l" + }, + { + "bbox": [ + 46, + 253, + 287, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 352, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 400 + ], + "type": "text", + "content": "pre-calculated weights since " + }, + { + "bbox": [ + 46, + 352, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}}^{\\alpha}" + }, + { + "bbox": [ + 46, + 352, + 287, + 400 + ], + "type": "text", + "content": " is the subset of " + }, + { + "bbox": [ + 46, + 352, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\alpha}" + }, + { + "bbox": [ + 46, + 352, + 287, + 400 + ], + "type": "text", + "content": ". This allows MCTF to efficiently reduce tokens considering bidirectional relations between two subsets with negligible extra costs compared to uni-directional bipartite soft matching." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 407, + 282, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 282, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 282, + 420 + ], + "type": "text", + "content": "3.3. One-step-ahead attention for informativeness" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "text", + "content": "In assessing informativeness, prior works [17, 18, 21] have leveraged the attention scores from the previous self-attention layer. As illustrated in Figure 5, previous approaches use the attention " + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^l" + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "text", + "content": " from the previous layer to fuse tokens " + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l" + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "text", + "content": ". This technique allows efficient assessment under the assumption that the attention maps in consecutive layers are similar. However, we observed that the attention maps often substantially differ, as shown in Figure 4, and the attention from a previous layer may lead to suboptimal token fusion. Thus, we proposed one-step-ahead attention, which measures the informativeness of tokens based on the attention map in the next layer, i.e., " + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{l + 1}" + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "text", + "content": ". Then, the informativeness scores " + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "text", + "content": " in Equation (4) is calculated with " + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{l + 1}\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "text", + "content": ". This simple remedy provides a considerable improvement; see Figure 7b in Section 4.2. After token fusion, we efficiently compute the attention map " + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}^{l + 1}\\in \\mathbb{R}^{(N - r)\\times (N - r)}" + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "text", + "content": " of fused tokens " + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^l\\in \\mathbb{R}^{(N - r)\\times C}" + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "text", + "content": " by simply aggregating " + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{l + 1}\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "text", + "content": " without recomputing the dot-product self-attention. To be specific, when the tokens are fused as " + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\delta (\\{\\mathbf{x}_i\\}_i)" + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "text", + "content": " during Equations (10) to (14), their corresponding one-step-ahead attention scores are also fused as " + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\delta (\\{\\mathbf{A}_i^{l + 1}\\}_i)" + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "text", + "content": " in both query and key direction. Note that when fusing attention scores for queries we use simple sum for " + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\delta ,i.e." + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\forall_{i}\\sum_{j}\\hat{\\mathbf{A}}_{ij}^{l + 1} = 1" + }, + { + "bbox": [ + 46, + 426, + 289, + 716 + ], + "type": "text", + "content": ". For fusing attention" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 545, + 209 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 545, + 209 + ], + "type": "image", + "image_path": "68a1062d07f2665f5fa55a6beae484cf46cc2ef602536b54cc1a028426f376e9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "lines": [ + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "type": "text", + "content": "Figure 6. Illustration of training with token reduction consistency. During training, we forward the input " + }, + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "type": "inline_equation", + "content": "f(x; r)" + }, + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "type": "inline_equation", + "content": "f(x; r')" + }, + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "type": "text", + "content": ", respectively. To obtain the augmented representation, " + }, + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "type": "inline_equation", + "content": "r'" + }, + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "type": "text", + "content": " is randomly selected in every step, and the model is updated with supervisory signals " + }, + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "type": "text", + "content": ", and consistency loss " + }, + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 305, + 217, + 547, + 272 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 287, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 287, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 305, + 287, + 545, + 314 + ], + "type": "text", + "content": "scores for queries, we use simple sum for " + }, + { + "bbox": [ + 305, + 287, + 545, + 314 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 305, + 287, + 545, + 314 + ], + "type": "text", + "content": " to guarantee " + }, + { + "bbox": [ + 305, + 287, + 545, + 314 + ], + "type": "inline_equation", + "content": "\\forall_{i}\\sum_{j}\\hat{\\mathbf{A}}_{ij}^{l + 1} = 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 325, + 460, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 325, + 460, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 325, + 460, + 338 + ], + "type": "text", + "content": "3.4. Token reduction consistency" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 345, + 547, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 547, + 441 + ], + "type": "text", + "content": "We here propose a new fine-tuning scheme to further improve the performance of vision Transformer " + }, + { + "bbox": [ + 304, + 345, + 547, + 441 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot ;r)" + }, + { + "bbox": [ + 304, + 345, + 547, + 441 + ], + "type": "text", + "content": " with MCTF. We observe that a different number of reduced tokens per layer, denoted as " + }, + { + "bbox": [ + 304, + 345, + 547, + 441 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 345, + 547, + 441 + ], + "type": "text", + "content": ", may lead to different representations of samples. By training Transformers with different " + }, + { + "bbox": [ + 304, + 345, + 547, + 441 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 345, + 547, + 441 + ], + "type": "text", + "content": " and encouraging the consistency between them, namely, token reduction consistency, we achieve the additional performance gain. The objective function of our method is given as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 339, + 454, + 545, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 454, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 339, + 454, + 545, + 485 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} \\left(f _ {\\theta} (x; r), y\\right) + \\mathcal {L} _ {\\mathrm {C E}} \\left(f _ {\\theta} \\left(x; r ^ {\\prime}\\right), y\\right) \\\\ + \\lambda \\mathcal {L} _ {\\mathrm {M S E}} \\left(\\mathbf {x} _ {r} ^ {\\mathrm {c l s}}, \\mathbf {x} _ {r ^ {\\prime}} ^ {\\mathrm {c l s}}\\right), \\tag {15} \\\\ \\end{array}", + "image_path": "89ec3e41f7525a8c5edc1532ab915e60d558894efc78cb9e335b9bc04cef106e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " is a supervised sample, " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "r,r^{\\prime}" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " is the fixed and dynamic reduced token numbers, " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " is the coefficient for consistency loss, and " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_r^{\\mathrm{cls}},\\mathbf{x}_{r'}^{\\mathrm{cls}}" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " are the class tokens in the last layer of models " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x;r),f_{\\theta}(x;r')" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " . In this objective, we first calculate the cross-entropy loss " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}(f_{\\theta}(x;r),y)" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " with fixed " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " , which is the target reduction number that will be used in the evaluation. At the same time, we generate another representation of the input " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " with smaller but randomly drawn " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "r^\\prime \\sim" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " uniform " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "(0,r)" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " , and calculate the loss " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}(f_{\\theta}(x;r'),y)" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " Then, we impose the token consistency loss " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}(\\mathbf{x}_r^{\\mathrm{cls}},\\mathbf{x}_{r'}^{\\mathrm{cls}})" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " on the class tokens, to retain the consistent representation across the diverse reduced token numbers " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "r^\\prime" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " . The proposed method can be viewed as a new type of token-level data augmentation [7, 19] and consistency regularization. Our token reduction consistency encourages the representation " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_r^{\\mathrm{cls}}" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " obtained by the target reduction number " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " to mimic the slightly augmented representation " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{r'}^{\\mathrm{cls}}" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " , which is more similar to ones with no token reduction since " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "r^{\\prime} < r" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15745" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 86, + 288, + 354 + ], + "blocks": [ + { + "bbox": [ + 97, + 71, + 237, + 82 + ], + "lines": [ + { + "bbox": [ + 97, + 71, + 237, + 82 + ], + "spans": [ + { + "bbox": [ + 97, + 71, + 237, + 82 + ], + "type": "text", + "content": "Table 1. Image classification results" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 86, + 288, + 354 + ], + "lines": [ + { + "bbox": [ + 48, + 86, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 86, + 288, + 354 + ], + "type": "table", + "html": "
MethodFLOPs (G)Params (M)Top-1 Acc (%)
DeiT-T [28]1.2572.2 (-)
+EvoViT[AAAI '22] [36]0.8572.0 (-0.2)
+A-ViT[CVPR '22] [37]0.8571.0 (-1.2)
+SPViT[ECCV '22] [17]0.9572.1 (-0.1)
+ToMe[ICLR '23] [4]0.7571.3 (-0.9)
+BAT[CVPR '23] [21]0.8572.3 (+0.1)
+MCTFr=160.7572.7 (+0.5)
DeiT-S [28]4.62279.8 (-)
+IA-RED2[NeurIPS '21] [24]3.22279.1 (-0.7)
+DynamicViT[NeurIPS '21] [26]2.92379.3 (-0.5)
+EvoViT[AAAI '22] [36]3.02279.4 (-0.4)
+EViT[ICLR '22] [18]3.02279.5 (-0.3)
+A-ViT[CVPR '22] [37]3.62278.6 (-1.2)
+ATS[ECCV '22] [12]2.92279.7 (-0.1)
+SPViT[ECCV '22] [17]2.62279.3 (-0.5)
+ToMe[ICLR '23] [4]2.72279.4 (-0.4)
+BAT[CVPR '23] [21]3.02279.6 (-0.2)
+MCTFr=162.62280.1 (+0.3)
", + "image_path": "de693853223cf1de64d9c40cc1486f52c134ddff46a9582307a30eb947ec2600.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 363, + 128, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 128, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 128, + 376 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 288, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 288, + 577 + ], + "type": "text", + "content": "Baselines. To validate the effectiveness of the proposed methods, we compare MCTF with the previous token reduction methods. For comparison, we opt the token pruning methods (A-ViT [37], IA-RED² [24], DynamicViT [26], EvoViT [36], ATS [12]) and token fusion methods (SPViT [17], EViT [18], ToMe [4], BAT [21]) in DeiT [28], and report the efficiency (FLOPs (G)) and the performance (Top-1 Acc (\\%)) of each method. Further, to validate MCTF on other Vision Transformers (T2T-ViT [39], LV-ViT [15]), we report the results of MCTF and compare it with the official number of existing works. We denote the number of reduced tokens per layer " + }, + { + "bbox": [ + 46, + 384, + 288, + 577 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 384, + 288, + 577 + ], + "type": "text", + "content": " with the subscript in Tables 1 and 2. The gray color in the table indicates the base model, and the green and red color indicates the improvements and degradations of the performance compared to the base model, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 586, + 171, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 171, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 171, + 600 + ], + "type": "text", + "content": "4.1. Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Comparison of the token reduction methods. The comparison with existing token reduction methods is summarized in Table 1. We demonstrate that our MCTF achieves the best performance with the lowest FLOPs in DeiT [28] surpassing all previous works. Further, it is worth noting that MCTF is the only work that avoids performance degradation with the lowest FLOPs in both DeiT-T and DeiT-S. Through Finetuning DeiT-T for 30 epochs, MCTF brings a significant gain of " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "+0.5\\%" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " in accuracy over the base model with nearly half" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 86, + 545, + 348 + ], + "blocks": [ + { + "bbox": [ + 323, + 71, + 528, + 82 + ], + "lines": [ + { + "bbox": [ + 323, + 71, + 528, + 82 + ], + "spans": [ + { + "bbox": [ + 323, + 71, + 528, + 82 + ], + "type": "text", + "content": "Table 2. Comparison with other Vision Transformers" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 86, + 545, + 348 + ], + "lines": [ + { + "bbox": [ + 307, + 86, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 86, + 545, + 348 + ], + "type": "table", + "html": "
ModelsFLOPs (G)Params (M)Acc (%)
PVT-Small[31]3.824.579.8
PVT-Medium [31]6.744.281.2
CoaT Mini [35]6.810.080.8
CoaT-Lite Small [35]4.020.081.9
Swin-T [20]4.529.081.3
Swin-S [20]8.750.083.0
PoolFormer-S36 [38]5.031.081.4
PoolFormer-M48 [38]11.673.082.5
T2T-ViTt-14 [39]6.121.581.7
+MCTFr=134.221.581.8 (↑)
T2T-ViTt-19 [39]9.839.282.4
+MCTFr=96.439.282.4 (-)
LV-ViT-S [15]6.626.283.3
+EVIT(ICLR '22] [18]4.726.283.0 (↓)
+BAT[CVPR '23] [21]4.726.283.1 (↓)
+DynamicViT[NeurIPS '21] [26]4.626.983.0 (↓)
+SPViT[ECCV '22] [17]4.326.283.1 (↓)
+MCTFr=124.226.283.4 (↑)
", + "image_path": "b3fda42c58ff08306cbd9a258474eda445e0262d530b4be488603b9a690650df.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 362, + 547, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 547, + 434 + ], + "type": "text", + "content": "FLOPs. Similarly, we observe a gain of " + }, + { + "bbox": [ + 304, + 362, + 547, + 434 + ], + "type": "inline_equation", + "content": "+0.3\\%" + }, + { + "bbox": [ + 304, + 362, + 547, + 434 + ], + "type": "text", + "content": " with DeiT-S while boosting the FLOPs by " + }, + { + "bbox": [ + 304, + 362, + 547, + 434 + ], + "type": "inline_equation", + "content": "-2.0" + }, + { + "bbox": [ + 304, + 362, + 547, + 434 + ], + "type": "text", + "content": " (G). We believe that multi-criteria with one-step-ahead attention helps the model to minimize the loss of information; further consistency loss on the class token through the token reduction improves the generalizability of the model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 436, + 548, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 548, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 548, + 604 + ], + "type": "text", + "content": "MCTF with other Vision Transformers. To validate the applicability of MCTF in various ViTs, we demonstrate MCTF with other transformer architectures in Table 2. Following previous works [17, 18, 21, 26], we apply MCTF with LVViT. Also, we present the results of MCTF with T2T-ViT. As presented in the table, our experimental results are promising. MCTF in these architectures gets at least " + }, + { + "bbox": [ + 304, + 436, + 548, + 604 + ], + "type": "inline_equation", + "content": "31\\%" + }, + { + "bbox": [ + 304, + 436, + 548, + 604 + ], + "type": "text", + "content": " speedup without performance degradation. Further, MCTF combined with LV-ViT outperforms all other Transformers and token reduction methods regarding FLOPs, and accuracy. Especially, it is worth noting that all token reduction methods except for MCTF bring the performance degradation in LVViT. These results reveal that MCTF is the efficient token reduction method for the diverse Vision Transformers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "Token reduction without training. Similar to ToMe [4], MCTF is applicable with pre-trained ViTs without any additional training since MCTF does not require any learnable parameters. We here apply the two reduction methods to the pre-trained DeiT without finetuning and provide the results in Table 3. Regardless of the reduced number of tokens " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": " in each layer, MCTF consistently surpasses ToMe. Especially, in the most sparse setting " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "r = 20" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": ", the performance gap is significant (" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "+7.0\\%" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": " in DeiT-T, " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "+3.8\\%" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": " in DeiT-S). Note" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15746" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 270, + 249 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 270, + 249 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 270, + 249 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 270, + 249 + ], + "type": "image", + "image_path": "d0626cb853e7b60ab729c17db8b9c741ba4b3d0a046296ae10bed2ad55a86723.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 250, + 547, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 250, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 547, + 283 + ], + "type": "text", + "content": "Figure 7. Ablations on (a) multi-criteria, (b) one-step-ahead-attention, and token reduction consistency. Each marker indicates the model with " + }, + { + "bbox": [ + 46, + 250, + 547, + 283 + ], + "type": "inline_equation", + "content": "r \\in [1,20]" + }, + { + "bbox": [ + 46, + 250, + 547, + 283 + ], + "type": "text", + "content": ", and we highlight " + }, + { + "bbox": [ + 46, + 250, + 547, + 283 + ], + "type": "inline_equation", + "content": "r \\in \\{5,10,15,20\\}" + }, + { + "bbox": [ + 46, + 250, + 547, + 283 + ], + "type": "text", + "content": " as bordered circle. We also denote the model as star when " + }, + { + "bbox": [ + 46, + 250, + 547, + 283 + ], + "type": "inline_equation", + "content": "r = 16" + }, + { + "bbox": [ + 46, + 250, + 547, + 283 + ], + "type": "text", + "content": ", which is used for finetuning the model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 323, + 76, + 541, + 249 + ], + "blocks": [ + { + "bbox": [ + 323, + 76, + 541, + 249 + ], + "lines": [ + { + "bbox": [ + 323, + 76, + 541, + 249 + ], + "spans": [ + { + "bbox": [ + 323, + 76, + 541, + 249 + ], + "type": "image", + "image_path": "14a979a63359a92a0d6246733b9e8852dc305f08f05e303a29162dd01f72d07d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 306, + 286, + 423 + ], + "blocks": [ + { + "bbox": [ + 66, + 293, + 268, + 304 + ], + "lines": [ + { + "bbox": [ + 66, + 293, + 268, + 304 + ], + "spans": [ + { + "bbox": [ + 66, + 293, + 268, + 304 + ], + "type": "text", + "content": "Table 3. Image classification results without training" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 306, + 286, + 423 + ], + "lines": [ + { + "bbox": [ + 48, + 306, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 48, + 306, + 286, + 423 + ], + "type": "table", + "html": "
Methodr
Base1248121620
DeiT-T
ToMe [4]72.272.172.072.071.670.868.761.5
MCTF72.272.272.172.172.071.771.068.5
DeiT-S
ToMe [4]79.879.879.779.779.479.077.974.2
MCTF79.879.879.879.879.879.679.278.0
", + "image_path": "0c9375cae11861ed99e8b3641386a4db2f38e22eb12b8b755bf6d0de19f9e0b9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "type": "text", + "content": "that without any additional training, our " + }, + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "type": "inline_equation", + "content": "\\mathrm{MCTF}_{r=16}" + }, + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "type": "text", + "content": " with pre-trained DeiT-S still shows a competitive performance of " + }, + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "type": "inline_equation", + "content": "79.2\\%" + }, + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "type": "text", + "content": " compared to reduction methods requiring training (e.g., " + }, + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "type": "inline_equation", + "content": "78.6\\%" + }, + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "type": "text", + "content": " of A-ViT, " + }, + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "type": "inline_equation", + "content": "79.1\\%" + }, + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "type": "text", + "content": " of IA-RED" + }, + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "type": "inline_equation", + "content": "79.3\\%" + }, + { + "bbox": [ + 46, + 433, + 288, + 493 + ], + "type": "text", + "content": " of DynamicViT, and SPViT in Table 1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 502, + 195, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 195, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 195, + 514 + ], + "type": "text", + "content": "4.2. Ablation studies on MCTF" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "text", + "content": "We provide ablation studies to validate each component of MCTF. Unless otherwise stated, we conduct whole experiments with DeiT-S finetuned with MCTF (" + }, + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "inline_equation", + "content": "r = 16" + }, + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "text", + "content": "). We provide the FLOPs-Accuracy graph by adjusting the reduced number of tokens per layer " + }, + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "inline_equation", + "content": "r \\in [1,20]" + }, + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Multi-criteria. We explore the effectiveness of multi-criteria in Figure 7a. First, regarding the multi-criteria, we utilize three criteria for MCTF, i.e., similarity (sim.), informativeness (info.), and size. Each single criterion of similarity and informativeness shows a relatively inferior performance compared to dual (sim. & info.) and multi-criteria (sim. & info. & size). Specifically, when " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "r = 16" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ", the performance of a single criterion is " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "79.7\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "79.4\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " with similarity and informativeness, respectively. Then, adopting dual criteria (sim. & info.), MCTF achieves " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "79.8\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ". Finally, we get an accuracy of " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "80.1\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " with a gain of " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "+0.3\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " by respecting all" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "type": "text", + "content": "three criteria (sim. & info. & size). These performance gaps get larger as " + }, + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "type": "text", + "content": " increases, which proves the importance of the multi-criteria for token fusion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 330, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 547, + 449 + ], + "type": "text", + "content": "One-step-ahead attention and token reduction consistency. To show the validity of one-step-ahead attention and token reduction consistency, we also provide the results of MCTF with and without each component in Figure 7b. When eliminating either one-step-ahead attention or token reduction consistency, the accuracies are dropped in every FLOP. This significant drop indicates that both approaches matter for MCTF. In short, by adopting one-step-ahead attention and token reduction consistency, MCTF effectively mitigates the performance degradation in a wide range of FLOPs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 628 + ], + "type": "text", + "content": "Comparison of design choices. The ablations on design choices are presented in Table 4. First, our bidirectional bipartite matching, which enables capturing the bidirectional relation in two sets, enhances the accuracy compared to one-way bipartite matching. Next, for pooling operation " + }, + { + "bbox": [ + 304, + 449, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 449, + 547, + 628 + ], + "type": "text", + "content": ", the weighted sum considering the size " + }, + { + "bbox": [ + 304, + 449, + 547, + 628 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 449, + 547, + 628 + ], + "type": "text", + "content": " and attentiveness " + }, + { + "bbox": [ + 304, + 449, + 547, + 628 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 449, + 547, + 628 + ], + "type": "text", + "content": " is a better choice than others like max-pool or average. Lastly, we compare the results with the precise and approximated attention for " + }, + { + "bbox": [ + 304, + 449, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}^l" + }, + { + "bbox": [ + 304, + 449, + 547, + 628 + ], + "type": "text", + "content": ". For precise attention, we just conduct the similarity calculation for one-step-ahead attention and the attention in the self-attention layer after fusion, separately. Otherwise, we approximate it with one-step-ahead attention as described in Section 3.3. As presented in the table, our approximated attention maintains the performance with the substantial improvement in efficiency (-0.4 (G) FLOPs)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 635, + 411, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 411, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 411, + 647 + ], + "type": "text", + "content": "4.3. Analyse of MCTF" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Qualitative results. For a better understanding of MCTF, we provide the qualitative results of MCTF in Figure 8. We visualize the fused tokens at the last block of DeiT-S on ImageNet-1K and denote the fused tokens by the same border color. As shown in the figure, since the tokens are merged" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15747" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 105, + 128 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 105, + 128 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 105, + 128 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 105, + 128 + ], + "type": "image", + "image_path": "4afb32ad419c3fe1b70f201ed7fd7bb8d8f34b37a36121c9d3f0cc1473a41798.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 111, + 70, + 168, + 128 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 168, + 128 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 168, + 128 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 168, + 128 + ], + "type": "image", + "image_path": "a7af67b7dff5a9a184c724a23658843b70f2c386b9abe9aad747cde52cc0bf70.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 232, + 128 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 232, + 128 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 232, + 128 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 232, + 128 + ], + "type": "image", + "image_path": "16e8435975dad32dea31b872d46f7f5bcac044f0edceb4ca319eb878667bf030.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 70, + 294, + 128 + ], + "blocks": [ + { + "bbox": [ + 238, + 70, + 294, + 128 + ], + "lines": [ + { + "bbox": [ + 238, + 70, + 294, + 128 + ], + "spans": [ + { + "bbox": [ + 238, + 70, + 294, + 128 + ], + "type": "image", + "image_path": "e6efb986832f72af6d0b7121735852b9ed8049f6f2b69570d354f47487483cdb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 299, + 70, + 357, + 128 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 357, + 128 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 357, + 128 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 357, + 128 + ], + "type": "image", + "image_path": "49851728a0fed6734a5653551e1148a59857223fff9303ea43adb2d4db91361a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 362, + 71, + 419, + 128 + ], + "blocks": [ + { + "bbox": [ + 362, + 71, + 419, + 128 + ], + "lines": [ + { + "bbox": [ + 362, + 71, + 419, + 128 + ], + "spans": [ + { + "bbox": [ + 362, + 71, + 419, + 128 + ], + "type": "image", + "image_path": "0f60e0ff937b8bf353c7c02c965d198f6c9426570f202c1384ebaaa28a411ed7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 426, + 70, + 482, + 128 + ], + "blocks": [ + { + "bbox": [ + 426, + 70, + 482, + 128 + ], + "lines": [ + { + "bbox": [ + 426, + 70, + 482, + 128 + ], + "spans": [ + { + "bbox": [ + 426, + 70, + 482, + 128 + ], + "type": "image", + "image_path": "074f72928dd8ff2f3faae726dfdc8a6d58953ccc68ab5bebfa8a52f864fb930b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 488, + 70, + 545, + 128 + ], + "blocks": [ + { + "bbox": [ + 488, + 70, + 545, + 128 + ], + "lines": [ + { + "bbox": [ + 488, + 70, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 488, + 70, + 545, + 128 + ], + "type": "image", + "image_path": "c4ed23b403ec7dd12409d09838c3f3faddf2138b4f05e96ce10c16036149a95e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 49, + 137, + 105, + 194 + ], + "blocks": [ + { + "bbox": [ + 49, + 137, + 105, + 194 + ], + "lines": [ + { + "bbox": [ + 49, + 137, + 105, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 137, + 105, + 194 + ], + "type": "image", + "image_path": "3c70c77a25b8cdc52e1440a7bf0e80025cdcd67bb8b0b6d8cbfa1d423c0d69e1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 224 + ], + "type": "text", + "content": "Figure 8. Visualization of the fused tokens with MCTF. Given the input images of ImageNet-1K (Top), the qualitative results of MCTF with DeiT-S are provided at the bottom. The same border color of the patches indicates the fused tokens." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 111, + 137, + 168, + 194 + ], + "blocks": [ + { + "bbox": [ + 111, + 137, + 168, + 194 + ], + "lines": [ + { + "bbox": [ + 111, + 137, + 168, + 194 + ], + "spans": [ + { + "bbox": [ + 111, + 137, + 168, + 194 + ], + "type": "image", + "image_path": "0b4778632b1db8ab333fb7c8584d5dd0ffd43d9b20816691c2420b71db2d3be3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 174, + 137, + 232, + 194 + ], + "blocks": [ + { + "bbox": [ + 174, + 137, + 232, + 194 + ], + "lines": [ + { + "bbox": [ + 174, + 137, + 232, + 194 + ], + "spans": [ + { + "bbox": [ + 174, + 137, + 232, + 194 + ], + "type": "image", + "image_path": "e7645866783e0d1187c2758d549d0fe9fc6cf2a3819135065ab030e51d3a5510.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 238, + 137, + 294, + 194 + ], + "blocks": [ + { + "bbox": [ + 238, + 137, + 294, + 194 + ], + "lines": [ + { + "bbox": [ + 238, + 137, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 238, + 137, + 294, + 194 + ], + "type": "image", + "image_path": "c90f29dbe07f3c44d77a75c67fea121eb4bfa54f825644dd11c11e011a983732.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 300, + 137, + 357, + 194 + ], + "blocks": [ + { + "bbox": [ + 300, + 137, + 357, + 194 + ], + "lines": [ + { + "bbox": [ + 300, + 137, + 357, + 194 + ], + "spans": [ + { + "bbox": [ + 300, + 137, + 357, + 194 + ], + "type": "image", + "image_path": "2f6ef5e81f64d1587a7ed67f95bfffedcf8ed205fbfa5e05f000a7ef2fa000bd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 362, + 137, + 419, + 194 + ], + "blocks": [ + { + "bbox": [ + 362, + 137, + 419, + 194 + ], + "lines": [ + { + "bbox": [ + 362, + 137, + 419, + 194 + ], + "spans": [ + { + "bbox": [ + 362, + 137, + 419, + 194 + ], + "type": "image", + "image_path": "8d19e33361ea236cea5de533f7f20db979196958c0194cb91a05a2727bca37ce.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 426, + 137, + 482, + 194 + ], + "blocks": [ + { + "bbox": [ + 426, + 137, + 482, + 194 + ], + "lines": [ + { + "bbox": [ + 426, + 137, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 426, + 137, + 482, + 194 + ], + "type": "image", + "image_path": "eb064ac90519e417a61cd0e87b153a55c8f261e3fc1c9d0fe6f32b2b28ba8db4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 488, + 137, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 488, + 137, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 488, + 137, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 488, + 137, + 545, + 194 + ], + "type": "image", + "image_path": "8d99bb75710b808bd952169ccfe6cf094e5115c09973b59c11439c1b3654a574.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 59, + 249, + 273, + 432 + ], + "blocks": [ + { + "bbox": [ + 89, + 233, + 244, + 245 + ], + "lines": [ + { + "bbox": [ + 89, + 233, + 244, + 245 + ], + "spans": [ + { + "bbox": [ + 89, + 233, + 244, + 245 + ], + "type": "text", + "content": "Table 4. Ablations of the design choices." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 249, + 273, + 432 + ], + "lines": [ + { + "bbox": [ + 59, + 249, + 273, + 432 + ], + "spans": [ + { + "bbox": [ + 59, + 249, + 273, + 432 + ], + "type": "table", + "html": "
MethodFLOPs ↓ (G)Acc ↑ (%)
DeiT-S4.679.8
bipartite soft matching
One-way2.680.0
Bidirectional2.680.1
pooling function δ
average2.680.0
max2.679.8
weighted average2.680.1
approximation of attention map
precise attention3.080.1
approximated attention2.680.1
", + "image_path": "f48d4114f8cb2f3d917835f0e3727e983dffd619fe05ecdf81782a3b1f76fff7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 443, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 550 + ], + "type": "text", + "content": "with multi-criteria (e.g., similarity, informativeness, size), we maintain the more diverse tokens in the informative foreground object. For instance, in the third image of the hamster, while the background patches including the hand are fused into one token, the foreground tokens are less fused while maintaining the details like the eye, ear, and face of the hamster. In short, compared to the background, the foreground tokens are less fused with the moderate size retaining the information of the main content." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 551, + 288, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 288, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 288, + 659 + ], + "type": "text", + "content": "Soundness of size criterion. Figure 9 presents the histogram of sizes of tokens after token reduction with and without size criterion. Specifically, we measure the size of the largest token at the last block and provide the histogram. With our size criterion, the merged tokens tend to have smaller sizes s showing the average size of 39.3/49.2 with and without the Size criterion, respectively. As intended, MCTF successfully suppresses the large-sized tokens, which are a source of information loss, leading to performance improvement." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": "In this work, we introduced the Multi-Criteria Token Fusion (MCTF), a novel strategy aimed at reducing the complex" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 320, + 234, + 533, + 342 + ], + "blocks": [ + { + "bbox": [ + 320, + 234, + 533, + 342 + ], + "lines": [ + { + "bbox": [ + 320, + 234, + 533, + 342 + ], + "spans": [ + { + "bbox": [ + 320, + 234, + 533, + 342 + ], + "type": "image", + "image_path": "213dabcea98cb5373ee53e8066e805dda7e83b670889a3b52385efb5f4cd4343.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 352, + 534, + 364 + ], + "lines": [ + { + "bbox": [ + 316, + 352, + 534, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 352, + 534, + 364 + ], + "type": "text", + "content": "Figure 9. Histogram of the size of tokens after reduction." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 386, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 547, + 542 + ], + "type": "text", + "content": "ity inherent in ViTs while mitigating performance degradation. MCTF effectively discerns the relation of tokens based on multiple criteria, including similarity, informativeness, and the size of the tokens. Our comprehensive ablation studies and detailed analyses demonstrate the efficacy of MCTF particularly with our innovative one-step-ahead attention and token reduction consistency. Remarkably, DeiT-T and DeiT-S with MCTF achieve considerable improvements, with " + }, + { + "bbox": [ + 304, + 386, + 547, + 542 + ], + "type": "inline_equation", + "content": "+0.5\\%" + }, + { + "bbox": [ + 304, + 386, + 547, + 542 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 386, + 547, + 542 + ], + "type": "inline_equation", + "content": "+0.3\\%" + }, + { + "bbox": [ + 304, + 386, + 547, + 542 + ], + "type": "text", + "content": " increase in Top-1 Accuracy over the vanilla models, accompanied by about " + }, + { + "bbox": [ + 304, + 386, + 547, + 542 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 304, + 386, + 547, + 542 + ], + "type": "text", + "content": " fewer FLOPs, respectively. We also observe that our MCTF outperforms all of the previous token reduction methods in diverse vision Transformers with and without training." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 568, + 403, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 568, + 403, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 568, + 403, + 582 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "This work was supported by ICT Creative Consilience Program through the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT)(IITP-2024-2020-0-01819), the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT)(NRF-2023R1A2C2005373), and a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI) funded by the Ministry of Health & Welfare Republic of Korea (HR20C0021)." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15748" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "type": "text", + "content": "[1] Moab Arar, Ariel Shamir, and Amit H Bermano. Learned queries for efficient local attention. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 288, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 288, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 288, + 146 + ], + "type": "text", + "content": "[2] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In ECCV, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "type": "text", + "content": "[3] Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv:2004.05150, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "type": "text", + "content": "[4] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. ICLR, 2022. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 217, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 258 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 262, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 288, + 295 + ], + "type": "text", + "content": "[6] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 297, + 288, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 288, + 328 + ], + "type": "text", + "content": "[7] Hyeong Kyu Choi, Joonmyung Choi, and Hyunwoo J. Kim. Tokenmixup: Efficient attention-guided token-level data augmentation for transformers. In NeurIPS, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 331, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 288, + 373 + ], + "type": "text", + "content": "[8] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. *ICLR*, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "type": "text", + "content": "[9] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. NeurIPS, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 420, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 288, + 464 + ], + "type": "text", + "content": "[10] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 466, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 288, + 520 + ], + "type": "text", + "content": "[11] 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. ICLR, 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 522, + 288, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 576 + ], + "type": "text", + "content": "[12] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In ECCV, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 578, + 288, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 610 + ], + "type": "text", + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 612, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 288, + 644 + ], + "type": "text", + "content": "[14] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[15] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. NeurIPS, 2021. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "text", + "content": "[16] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. In ICLR, 2020. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "type": "text", + "content": "[17] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In ECCV, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "text", + "content": "[18] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. ICLR, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 175, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 206 + ], + "type": "text", + "content": "[19] Jihao Liu, Boxiao Liu, Hang Zhou, Hongsheng Li, and Yu Liu. Tokenmix: Rethinking image mixing for data augmentation in vision transformers. In ECCV, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 208, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 547, + 251 + ], + "type": "text", + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 253, + 547, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 547, + 296 + ], + "type": "text", + "content": "[21] Sifan Long, Zhen Zhao, Jimin Pi, Shengsheng Wang, and Jingdong Wang. Beyond attentive tokens: Incorporating token importance and diversity for efficient vision transformers. In CVPR, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 298, + 547, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 547, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 547, + 341 + ], + "type": "text", + "content": "[22] Dmitrii Marin, Jen-Hao Rick Chang, Anurag Ranjan, Anish Prabhu, Mohammad Rastegari, and Oncel Tuzel. Token pooling in vision transformers for image classification. In WACV, 2023. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 343, + 547, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 547, + 386 + ], + "type": "text", + "content": "[23] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In CVPR, 2022. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 547, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 547, + 430 + ], + "type": "text", + "content": "[24] Bowen Pan, Rameswar Panda, Yifan Jiang, Zhangyang Wang, Rogerio Feris, and Aude Oliva. IA-RED2: Interpretability-aware redundancy reduction for vision transformers. NeurIPS, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 433, + 547, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 547, + 486 + ], + "type": "text", + "content": "[25] 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 ICML, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 488, + 547, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 547, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 547, + 531 + ], + "type": "text", + "content": "[26] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. NeurIPS, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 533, + 547, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 547, + 565 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 547, + 565 + ], + "type": "text", + "content": "[27] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, 2021. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 567, + 547, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 547, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 547, + 610 + ], + "type": "text", + "content": "[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 1, 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 613, + 547, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 547, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 547, + 644 + ], + "type": "text", + "content": "[29] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 647, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 678 + ], + "type": "text", + "content": "[30] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv:2006.04768, 2020. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "type": "text", + "content": "[31] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense predic" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15749" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 466 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "tion without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "text", + "content": "[32] QuanLin Wu, Hang Ye, Yuntian Gu, Huishuai Zhang, Liwei Wang, and Di He. Denoising masked autoencoders help robust classification. In ICLR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 288, + 173 + ], + "type": "text", + "content": "[33] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "type": "text", + "content": "[34] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. In AAAI, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 251 + ], + "type": "text", + "content": "[35] Weijian Xu, Yifan Xu, Tyler Chang, and Zhuowen Tu. Coscale conv-attentional image transformers. In ICCV, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "text", + "content": "[36] Yifan Xu, Zhijie Zhang, Mengdan Zhang, Kekai Sheng, Ke Li, Weiming Dong, Liqing Zhang, Changsheng Xu, and Xing Sun. Evo-vit: Slow-fast token evolution for dynamic vision transformer. In AAAI, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 299, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 331 + ], + "type": "text", + "content": "[37] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In CVPR, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 332, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 365 + ], + "type": "text", + "content": "[38] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "type": "text", + "content": "[39] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In ICCV, 2021. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 411, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 433 + ], + "type": "text", + "content": "[40] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In CVPR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "type": "text", + "content": "[41] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. *ICLR*, 2021. 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15750" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/bf855ad8-e560-4eb0-9886-16b2ff2266f1_content_list.json b/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/bf855ad8-e560-4eb0-9886-16b2ff2266f1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c0d5edaf81405f73762e0e826493cdf7e99131e3 --- /dev/null +++ b/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/bf855ad8-e560-4eb0-9886-16b2ff2266f1_content_list.json @@ -0,0 +1,1812 @@ +[ + { + "type": "text", + "text": "Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer", + "text_level": 1, + "bbox": [ + 78, + 130, + 893, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhen Zhao $^{1,2,*}$ , Jingqun Tang $^{2,\\dagger}$ , Chunhui Lin $^{2}$ , Binghong Wu $^{2}$ , Can Huang $^{2}$ , Hao Liu $^{2}$ , Xin Tan $^{1}$ , Zhizhong Zhang $^{1}$ , Yuan Xie $^{1,\\dagger}$ $^{1}$ East China Normal University $^{2}$ ByteDance", + "bbox": [ + 179, + 179, + 787, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{51255901056}@stu.ecnu.edu.cn, {zzzhang,xtan,yxie}@cs.ecnu.edu.cn", + "bbox": [ + 191, + 236, + 772, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{tangjingqun,linchunhui,wubinghong,liuhao,can.huang}@bytedance.com", + "bbox": [ + 191, + 253, + 772, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene text recognition (STR) in the wild frequently encounters challenges when coping with domain variations, font diversity, shape deformations, etc. A straightforward solution is performing model fine-tuning tailored to a specific scenario, but it is computationally intensive and requires multiple model copies for various scenarios. Recent studies indicate that large language models (LLMs) can learn from a few demonstration examples in a training-free manner, termed \"In-Context Learning\" (ICL). Nevertheless, applying LLMs as a text recognizer is unacceptably resource-consuming. Moreover, our pilot experiments on LLMs show that ICL fails in STR, mainly attributed to the insufficient incorporation of contextual information from diverse samples in the training stage. To this end, we introduce $E^2$ STR, a STR model trained with context-rich scene text sequences, where the sequences are generated via our proposed in-context training strategy. $E^2$ STR demonstrates that a regular-sized model is sufficient to achieve effective ICL capabilities in STR. Extensive experiments show that $E^2$ STR exhibits remarkable training-free adaptation in various scenarios and outperforms even the fine-tuned state-of-the-art approaches on public benchmarks. The code is released at https://github.com/bytedance/E2STR.", + "bbox": [ + 73, + 335, + 473, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 714, + 209, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene Text Recognition (STR) is a fundamental task in computer vision, with extensive applications in several domains such as autonomous driving [45], augmented reality [30, 33], industrial print recognition [28] and visual understanding [26].", + "bbox": [ + 75, + 739, + 468, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current progress in STR [3, 17, 20, 35] has demonstrated remarkable performance in numerous scenarios.", + "bbox": [ + 76, + 816, + 468, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, as shown in Figure 1 (a), STR models are sup", + "bbox": [ + 96, + 847, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2bb54acc35d7fb57ff0359e0d3752e9e0be937b0180d2a18ac947d2f68e0fd1f.jpg", + "image_caption": [ + "(a) Diversified scenarios of scene text" + ], + "image_footnote": [], + "bbox": [ + 506, + 306, + 888, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0332d3b9d72166f36a2f7dac608dcdec76fdcc86fa7080a9ce1debda48c46650.jpg", + "image_caption": [ + "(b) Scenario-specific fine-tuning pipeline of current methods" + ], + "image_footnote": [], + "bbox": [ + 508, + 405, + 888, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/933747a14fd1c940284ad7451e6b7d89a3af0d25cd0710451878dda39a6e3859.jpg", + "image_caption": [ + "(c) In-context learning pipeline of $\\mathbf{E}^2\\mathbf{STR}$", + "Figure 1. Demonstration of real-world scene text scenarios and the adaptation pipeline. (a) Diversified scenarios of scene text in the real world. (b) The adaptation pipeline of current methods. They typically have to fine-tune upon a trained STR model with the training set, under a specific scenario. (c) The adaptation pipeline of our proposed $\\mathrm{E}^2\\mathrm{STR}$ . Our method automatically selects in-context prompts and performs training-free adaptation when faced with novel scenarios. Blue characters denote ambiguous scene text that is easily recognized." + ], + "image_footnote": [], + "bbox": [ + 508, + 507, + 888, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "posed to perform robustly over diversified scenarios in the real world, where the scene text is hard to recognize because of domain variation, font diversity, shape deformation, etc. As shown in Figure 1 (b), a straightforward solution involves collecting the corresponding data and then fine-tuning the model for the specific scenario [3, 17, 20]. This process is computationally intensive and requires multiple model copies for diverse scenarios.", + "bbox": [ + 496, + 763, + 890, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The development of a comprehensive and reliable STR", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR 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": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 99, + 875, + 236, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work is done when Zhen Zhao is an intern at ByteDance.", + "bbox": [ + 99, + 887, + 434, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15567", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3dba2f422435df1238e2ef183a714acb1715d1b4ae84fd8b39037bde1982db78.jpg", + "image_caption": [ + "(b) Context-Rich scene text sequence" + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 267, + 159 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cf0ef15b006408fe6b66e44f03c601ec822a7c646f5ed00a51c575f73ea054f7.jpg", + "image_caption": [ + "(c) ICL results on cross-domain benchmark", + "Figure 2. Our pilot experiments. (a) The randomly concatenated scene text sequence. (b) Our proposed context-rich scene text sequence. (c) By training an STR model based on the randomly concatenated scene text sequence, we evaluate the model on three cross-domain datasets." + ], + "image_footnote": [], + "bbox": [ + 282, + 88, + 468, + 159 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model that can effectively handle many real-world scenarios remains a significant challenge.", + "bbox": [ + 75, + 279, + 468, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fortunately, plenty of studies [1, 6, 21, 38] have shown that Large Language Models (LLMs) can easily adapt without additional training. This adaptation is achieved by leveraging only a handful of input-label pairs as context (prompting information), a phenomenon known as \"In-Context Learning\" (ICL). The advantages of ICL inspire our interest in implementing it in STR, such that by fetching a few in-context prompts, a single model can be rapidly adapted to various scenarios without fine-tuning.", + "bbox": [ + 75, + 310, + 468, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the equipment of ICL in STR still poses challenges under the existing circumstances. Firstly, it is deemed excessively costly to apply Multi-Modal Large Language Models (M-LLMs) with billions of parameters as a scene text recognizer. And the ICL capabilities in regular-sized models have been barely explored currently.", + "bbox": [ + 75, + 446, + 468, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Secondly, it is hard to acquire ICL capabilities for a STR model with current training strategies. Previous studies have observed that sending image-text sequences for training would naturally endow ICL for M-LLMs [1, 21, 38], while such a phenomenon is hard to achieve in STR. As shown in Figure 2 (a), we generate sequential training data by randomly concatenating scene text samples. This practice fails as the trained model does not exhibit any performance improvement even when provided with in-domain prompts (Figure 2 (c)). The major cause of this failure is the lack of context in the generated scene text sequences during the training phase. The arbitrary concatenation of scene text does not provide any contextual information (i.e., sample connections) between different samples (Figure 2 (a)). Consequently, the model lacks the ability to effectively use information derived from in-context prompts (Figure 2 (c)), which implies that in-context training is essentially important for the effective implementation of ICL in STR.", + "bbox": [ + 75, + 537, + 468, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above analysis, we propose $\\mathrm{E}^2\\mathrm{STR}$ (Ego-Evolving STR), a paradigm that facilitates adaptation across diverse scenarios in a training-free manner. Specifically, we propose an in-context training strategy, which enables the model to exploit contextual information from the generated context-rich scene text sequences (Figure 2", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(b)). The context-rich scene text sequences are formed using our ST-strategy, which involves random Splitting and Transformation of scene text, hence generating a set of \"sub-samples\". The sub-samples are inner-connected in terms of both visual and linguistic aspects. In the inference stage, $\\mathrm{E}^2\\mathrm{STR}$ fetches in-context prompts based on visual similarities, and utilizes the prompts to assist the recognition, shown in Figure 1 (c). In practice, it is found that with proper training and inference strategies, ICL capabilities can also be observed in regular-sized STR models (hundreds of millions of parameters).", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, the proposed $\\mathrm{E}^2\\mathrm{STR}$ effectively captures contextual information from the in-context prompts and performs rapid adaptation in various novel scenarios in a training-free manner (Please refer to Section 4.2). On common benchmarks, $\\mathrm{E}^2\\mathrm{STR}$ achieves SOTA results, with an average improvement of $0.8\\%$ over previous methods and $1.1\\%$ over itself without ICL. Most importantly, when evaluated on unseen domains, $\\mathrm{E}^2\\mathrm{STR}$ achieves impressive performance with only a few prompts, even outperforming the fine-tuning results of SOTA methods by $1.2\\%$ . Our contributions are summarized below:", + "bbox": [ + 496, + 257, + 890, + 422 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We propose $\\mathrm{E}^2\\mathrm{STR}$ , a robust STR paradigm that can perform rapid adaptation over diverse scenarios in a training-free manner.", + "(2) We provide an in-context training strategy for equipping STR models with ICL capabilities, as well as an in-context inference strategy for STR models to leverage contextual information from in-context prompts.", + "(3) We demonstrate that ICL capabilities can be effectively incorporated into regular-sized STR models via appropriate training and inference strategies.", + "(4) Extensive experiments show that $\\mathrm{E}^2\\mathrm{STR}$ exceeds state-of-the-art performance across diverse benchmarks, even surpassing the fine-tuned approaches in unseen domains." + ], + "bbox": [ + 496, + 424, + 890, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 654, + 640, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Scene Text Recognition", + "text_level": 1, + "bbox": [ + 500, + 679, + 714, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent years have witnessed extensive studies in STR, which can be generally divided into Language-free methods and Language-aware methods.", + "bbox": [ + 496, + 703, + 890, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language-free STR. Language-free models directly utilize visual features for prediction, without considering the relationship between the characters. In this branch CTC-based [11] methods [5, 24] play the most prominent part. They typically consist of a CNN for feature extraction and an RNN for sequential feature processing, which are trained end-to-end with the CTC loss [11]. Other methods like [23, 40] focus on treating STR as a character-level segmentation task. The lack of linguistic information limits the application of language-free methods in scenarios with", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15568", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "occluded or incomplete characters.", + "bbox": [ + 76, + 90, + 308, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Language-aware STR. Language-aware models leverage linguistic information to assist the recognition, typically utilizing an external language model (LM) [10, 44] or training internal LMs [2, 7, 36]. SRN [44] and ABINet [10] feed visual predictions to an external LM for linguistic refinement. The direct application of an external LM without considering visual features leads to possible erroneous correction. On the other hand, methods like PARSeq [3] and MAERec [17] implicitly train an internal LM in an auto-regressive manner, which have achieved decent performance. In this paper we base our model on the language-aware design, training a transformer-based language decoder inner-connected with the vision encoder.", + "bbox": [ + 75, + 106, + 468, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Multi-Modal In-Context Learning", + "text_level": 1, + "bbox": [ + 76, + 313, + 375, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent large language models (LLMs) [6, 46] have demonstrated their excellent few-shot adaptation capabilities. By concatenating a few examples with the input as the prompt at reference time, LLMs quickly adapt to novel tasks without parameter updating. This phenomenon introduces a novel learning paradigm termed \"In-Context Learning\". Meanwhile, unlike LLMs, vision-language models (VLMs) struggle to understand complex multi-modal prompts [47]. A large set of approaches [13-15, 34] have been proposed to empower VLMs with multi-modal in-context learning (M-ICL) capabilities, but they typically utilize vision models (like image caption models) to translate images to text [15, 34, 43], or view the LLM as a scheduler learning to call vision experts based on a few examples [13]. These approaches do not truly establish a VLM with M-ICL capabilities. Recently, several work [1, 21, 38] proposes to train VLMs with sequential multi-modal data, and have achieved great success in prompting VLMs with multi-modal examples. In this paper, we aim to train a scene text recognizer equipped with M-ICL capabilities based on this sequential training paradigm. We demonstrate that the arbitrary concatenation of scene text fails as stated above, which motivates us to generate context-rich scene text sequences.", + "bbox": [ + 75, + 337, + 472, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 699, + 210, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary of Multi-Modal In-Context Learning", + "text_level": 1, + "bbox": [ + 76, + 724, + 468, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modal in-context Learning enables M-LLMs to perform quick adaptation in downstream tasks in a training-free manner, hence eliminating the redundant computation and time expenses of fine-tuning. In this subsection, we introduce how to formulate multi-modal in-context learning for addressing the STR task.", + "bbox": [ + 75, + 763, + 468, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a scene text tuple $(\\mathbf{x},\\mathbf{y})$ where $\\pmb{x}$ is the scene image and $\\pmb{y}$ is the ground-truth text, the STR task involves generating the label $\\pmb{y}$ by maximizing the conditional probabil", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ity under the classic auto-regressive paradigm as follows: $p(\\boldsymbol{y}|\\boldsymbol{x}) = \\prod_{l=1}^{L} p(\\boldsymbol{y}_l|\\boldsymbol{x}, \\boldsymbol{y}_{$ . The training in this phase makes use of the next-token prediction loss:", + "bbox": [ + 75, + 752, + 470, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y}) \\sim \\mathcal {D}} \\left[ - \\sum_ {l = 1} ^ {L} \\log p (\\boldsymbol {y} _ {l} | \\boldsymbol {y} _ {< l}, \\boldsymbol {x}) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 820, + 468, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ is the training set.", + "bbox": [ + 76, + 885, + 259, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 In-Context Training", + "text_level": 1, + "bbox": [ + 500, + 523, + 694, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The objective of the in-context training phase is to equip $\\mathrm{E}^2\\mathrm{STR}$ with the capability of In-Context Learning. As depicted in the top of Figure 3, the model is trained with context-rich scene text sequences as stated before. In these sequences, we interleave a placeholder $$ in the text for each image. This serves to make the language decoder distinguish between different samples following [1]. In this stage, we propose two strategies to generate context-rich scene text sequences: the Split Strategy and the Transform Strategy (the ST strategy).", + "bbox": [ + 496, + 550, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Split Strategy. As shown in Figure 4 (a), when presented with a training tuple $(\\pmb{x},\\pmb{y})$ , we split the sample and hence generating a set of \"sub-samples\". It is evident that the sub-samples exhibit a strong connection to the original training sample. Furthermore, the sub-samples themselves demonstrate interconnectivity as they overlap with one another. Next, we proceed to concatenate the sub-samples with $(\\pmb{x},\\pmb{y})$ and additional randomly selected samples to form a context-rich sample sequence.", + "bbox": [ + 496, + 702, + 890, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We randomly shuffle the whole sequence before generating the actual input text (i.e., interleaving the $\\langle \\mathbf{\\Pi} / \\mathbf{i}\\rangle$ token to the text sequence).", + "bbox": [ + 496, + 839, + 890, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, to accurately split the training samples, we", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15570", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "synthesize 600k scene text images based on [4] and record the accurate bounding boxes of every single character. Our subsequent experiments show that the synthesized data does not change $\\mathrm{E}^2\\mathrm{STR}$ 's non-context text recognition ability, but the Split Strategy based on them equips $\\mathrm{E}^2\\mathrm{STR}$ with a strong capability of in-context learning.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Transform Strategy. As shown in Figure 4 (b), given a training tuple $(\\boldsymbol{x}, \\boldsymbol{y})$ (whether with character-wise bounding boxes or not), we perform data augmentation (a set of image transformations, e.g., color/direction transformations) on $\\boldsymbol{x}$ . In this way, we also generate a set of sub-samples with the same label but different image patterns from the original sample.", + "bbox": [ + 75, + 181, + 470, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practice, as depicted in Figure 4 (c), we hybrid the above strategies. The training set is formed by concatenating the synthesized data and the original training data used in the first training phase. For the synthesized data with character-wise bounding boxes, both the Split Strategy and the Transform Strategy are utilized. For the original training data, only the Transform Strategy is implemented.", + "bbox": [ + 75, + 287, + 468, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, after generating the sample sequence $(X,Y)$ , where $X$ is the image sequence and $Y$ is the text sequence, $X$ is fed into the vision encoder, while $Y$ is processed by the language decoder under the auto-regressive framework. The loss function is formulated as:", + "bbox": [ + 76, + 393, + 470, + 470 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\left(\\boldsymbol {X}, \\boldsymbol {Y}\\right)} = - \\sum_ {l = 1} ^ {L} \\log p \\left(\\boldsymbol {Y} _ {l} \\mid \\boldsymbol {Y} _ {< l}, \\boldsymbol {X} _ {\\leq l}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 479, + 468, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_{\\leq l}$ is the set of image tokens preceding token $\\mathbf{Y}_l$ in the input sequence.", + "bbox": [ + 75, + 532, + 468, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. In-Context Inference", + "text_level": 1, + "bbox": [ + 76, + 575, + 274, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The In-Context Learning ability is acquired by our $\\mathrm{E}^2\\mathrm{STR}$ model through the above two-stage training approach. AS shown in the bottom of Figure 3, when presented with a test image $\\pmb{x}$ , the framework selects $N$ samples $\\{(x_i^c,y_i^c)\\}_{i = 1}^N$ from a in-context pool $\\mathcal{D}^c$ . The selected samples have the highest visual similarities to $\\pmb{x}$ in the latent space. Specifically, we calculate the image embedding $\\pmb{I}$ of $\\pmb{x}$ by averaging the visual token sequence Encoder $(\\pmb{x})$ . The in-context prompts are then formed by choosing $N$ samples from $\\mathcal{D}^c$ , where the image embeddings of these samples have the top-N highest cosine similarity with $\\pmb{I}$ , i.e.,", + "bbox": [ + 75, + 598, + 468, + 765 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} = \\underset {i \\in 1, 2, \\dots , | \\mathcal {D} ^ {c} |} {\\arg \\operatorname {T o p N}} \\frac {\\boldsymbol {I} ^ {\\mathrm {T}} \\boldsymbol {I} _ {i} ^ {c}}{\\| \\boldsymbol {I} \\| _ {2} \\| \\boldsymbol {I} _ {i} ^ {c} \\| _ {2}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 777, + 468, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{I}$ is the index set of the top-N similar samples in $\\mathcal{D}^c$ and $I_{i}^{c}$ is the image embedding of the $i$ -th sample in $\\mathcal{D}^c$ . The in-context prompts are then defined as:", + "bbox": [ + 75, + 825, + 468, + 871 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {E} = \\left\\{\\left(\\boldsymbol {x} _ {i} ^ {c}, \\boldsymbol {y} _ {i} ^ {c}\\right) | i \\in \\mathcal {I} \\right\\}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 883, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in the bottom of Figure 3, $\\pmb{E}$ is concatenated with the test sample $\\pmb{x}$ and our in-context prediction is formulated as $p(\\pmb{y}|\\pmb{E},\\pmb{x})$ . In practice, the in-context pool $\\mathcal{D}^c$ retains solely the output tokens generated by the vision encoder, resulting in a highly efficient selection process. Furthermore, because the in-context pool is tiny and we do straight inference without training, the extra consumption is kept to a minimum (Please refer to Section 4.3).", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 498, + 224, + 624, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 498, + 248, + 689, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Following MAERec [17], we choose Vision Transformer [9] pre-trained under the MAE [16] framework as the vision encoder. The default language decoder is set as OPT-125M [46]. We use the cosine learning rate scheduler without warm-up and the AdamW optimizer with a weight decay of 0.01. We train our model for 10 epochs with an initial learning rate of 1e-4 during the first training stage, and 5 epochs with an initial learning rate of 5e-6 during the second in-context training stage. The training batch size is 64 for the first stage and 8 for the second stage. During inference for $\\mathrm{E}^2\\mathrm{STR}$ -ICL, we select two in-context prompts based on the kNN selection strategy.", + "bbox": [ + 496, + 271, + 890, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Metrics. We use the real-world training dataset Union14M-L[17] for the two-stage training. The same training dataset (including the synthesized data) is adopted for all compared methods. $\\mathrm{E}^2\\mathrm{STR}$ is evaluated under various publicly available benchmarks, including Regular Benchmarks IIIT5k [29], SVT [37], IC13 [18], Irregular Benchmarks IC15 [19], SVTP [31], CUTE80 (CT80) [32], COCO Text (COCO) [39], CTW [25], Total Text (TT) [8], Occluded Benchmarks OST (HOST and WOST) [41] and artistic benchmark WordArt [42]. In cross domain scenarios the evaluated datasets including the metal-surface benchmark MPSC [12] and the handwriting benchmark IAM [27], as well as a more difficult real-world industrial text recognition dataset EIST (Enhanced Industrial Scene Text) collected by us. EIST is collected from the real-world industrial scenario, which contains 200 training samples and 8000 test samples. We use Word Accuracy[17] as the evaluation metric for all compared methods.", + "bbox": [ + 496, + 454, + 890, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 498, + 732, + 638, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 Results on Common Benchmarks", + "text_level": 1, + "bbox": [ + 498, + 756, + 785, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 presents the performance of $\\mathrm{E}^2\\mathrm{STR}$ on common benchmarks. We evaluate $\\mathrm{E}^2\\mathrm{STR}$ on 12 commonly used STR benchmarks and compare with SOTA methods. $\\mathrm{E}^2\\mathrm{STR}$ -base refers to non-context prediction without prompts. For $\\mathrm{E}^2\\mathrm{STR}$ -ICL, a tiny in-context pool is maintained by randomly sampling 1000 images from the training data (less than $0.025\\%$ of the number of training samples). As we can see, $\\mathrm{E}^2\\mathrm{STR}$ -base achieves $90.25\\%$ average word", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15571", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2ebaca2d40bfde807c2fb3e072e2e1d5fa23294e2669a32335bbdbba04300329.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVenueRegularIrregularOccludedOthersAVG
IIIT3000SVT647IC131015IC152077SVTP645CT80288COCO9896CTW1572TT2201HOST2416WOST2416WordArt1511
ASTER [36]PAMI'1895.0389.4993.7985.4882.0290.2862.2576.5378.6943.3464.6565.5977.26
NRTR [35]ICDAR'1997.4393.8296.0685.1584.0391.3265.9481.7481.8350.8371.5264.0680.31
SAR [22]AAAI'1997.7094.1396.3587.4787.6093.0667.4183.9186.2346.3670.3272.4081.91
SATRN [20]AAAI'2097.8395.8397.4489.4690.8596.1873.0684.6187.9156.7175.6275.7185.10
ABINet [10]CVPR'2197.9095.9896.1691.6690.2393.7571.4683.7286.0156.5475.7575.2584.53
PARSeq* [3]ECCV'2299.1097.8498.1389.2296.9098.61-------
MAERec [17]ICCV'2398.9397.9998.6293.0494.5798.9678.8488.8793.9173.9785.7282.5990.50
\\( E^2STR \\)-base99.1098.1598.0392.9996.4398.9677.2988.3693.4673.3085.5181.4790.25
\\( E^2STR-ICL \\)99.2398.6198.7293.8296.7499.3178.3888.9994.6874.7586.5986.1791.33
", + "bbox": [ + 81, + 87, + 890, + 258 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/44cbb7d7ba427f50a90a8d97ec55d8075b034a4776dcab633ed4ee85b8e86bae.jpg", + "table_caption": [ + "Table 1. Results on common benchmarks. All methods are trained on the same dataset except for PARSeq.*: PARSeq is trained on its self-collected real-world dataset and we directly quote the results from its original paper. Red and blue values denote the best and the secondary performance. ${\\mathrm{E}}^{2}$ STR-base refers to non-context inference." + ], + "table_footnote": [], + "table_body": "
MethodIndustrialHandwritingAVG
MPSC 2941EIST 8000IAM 3000
ASTER [36]63.4848.7652.5054.91
NRTR [35]73.2461.7759.5364.85
SAR [22]73.8558.2656.6362.91
ABINet [10]75.3562.8561.5766.59
SATRN [20]76.1065.4259.4767.00
MAERec [17]81.8170.3370.2774.14
E²STR-base81.2669.6669.5173.48
E²STR-ICL83.6476.7774.1078.17
", + "bbox": [ + 78, + 333, + 473, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Results on cross domain scenarios. Three datasets under two unseen domains are evaluated. All approaches are evaluated in a training-free manner.", + "bbox": [ + 75, + 513, + 470, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "accuracy over 12 datasets, $0.25\\%$ lower than MAERec [17]. However, by fetching in-context prompts and exploiting incontext information, $\\mathrm{E}^2$ STR-ICL achieves an average word accuracy of $91.33\\%$ , which is $1.08\\%$ higher than $\\mathrm{E}^2$ STR-base and $0.83\\%$ higher than MAERec. Please note that this improvement is automatic and training-free.", + "bbox": [ + 75, + 584, + 468, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, on the six traditional STR benchmarks (i.e., IIIT, SVT, IC13, IC15, SVTP, and CT80) which have nearly reached saturation in recent years[17], $\\mathrm{E}^2\\mathrm{STR}$ still push the performance limit from $97.02\\%$ to $97.74\\%$ , leading to a $24\\%$ error rate decrease. On the 6 larger and harder STR benchmarks (i.e., COCO Text, CTW, TT, HOST, and WOST), $\\mathrm{E}^2\\mathrm{STR - ICL}$ outperforms MAERec by $0.94\\%$ .", + "bbox": [ + 75, + 675, + 468, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Results on Cross Domain Scenarios", + "text_level": 1, + "bbox": [ + 76, + 800, + 377, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare with SOTA methods on cross domain benchmarks. Two novel scenarios are introduced: the industrial scenario (MPSC and EIST) and the handwriting scenario (IAM). In each dataset, only 100 training samples are provided. For $\\mathrm{E}^2$ STR-ICL we simply use the training samples", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/01bd4147d21069695247f3d54bd9690b436b9a72d6043d12c65248ee53eaa125.jpg", + "image_caption": [ + "Figure 5. Comparison with the fine-tuned models. We report the average performance on three cross-domain datasets. Please note that ABINet [10], SATRN [20] and MAERec [17] are fine-tuned with the in-domain data, while our $\\mathrm{E}^2$ STR-ICL is training-free." + ], + "image_footnote": [], + "bbox": [ + 501, + 333, + 893, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as the in-context pool. We compare the training-free results in Table 2 and the fine-tuning results in Figure 5.", + "bbox": [ + 498, + 592, + 890, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As we can see, on both industrial and handwriting scenarios our $\\mathrm{E}^2$ STR-ICL reaches SOTA performance. As shown in Table 2, under the training-free constraint $\\mathrm{E}^2$ STR-ICL reaches an average performance of $78.17\\%$ , which is $4.69\\%$ higher than $\\mathrm{E}^2$ STR-base and $4.03\\%$ higher than the SOTA method MAERec. Specifically, on EIST and IAM the application of ICL brings a huge improvement of $7.11\\%$ and $4.59\\%$ , which demonstrates the extraordinary adaptation ability of $\\mathrm{E}^2$ STR-ICL.", + "bbox": [ + 496, + 625, + 890, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further compare the fine-tuned methods and our $\\mathrm{E}^2$ STR-ICL. We fine-tune ABINet [10], SATRN [20] and MAERec [17] with the same data preserved in the in-context pool. As shown in Figure 5, $\\mathrm{E}^2$ STR-ICL outperforms MAERec by $1.16\\%$ even if the latter is fine-tuned with in-domain data, which is an exciting result given that $\\mathrm{E}^2$ STR-ICL requires no parameter updating. In a word, our $\\mathrm{E}^2$ STR can be rapidly implemented in a training-free manner in various novel scenarios and even achieves better per", + "bbox": [ + 496, + 763, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15572", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/80c22ec14a1a2a3b520e29614c808b82ca4de267495f3cb6157fa7c7853f25f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
COCOHOSTWordArt
annotation rate10%20%10%20%10%20%
MAERec [17]000000
w/ fine-tuning0.821.671.031.721.342.23
E²STR-base000000
E²STR-ICL10.1212.9212.4313.7626.2232.02
", + "bbox": [ + 81, + 88, + 467, + 167 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1d8d6cf8e04cf8ffb53138266ee6dc97b70c6da5d3b4fa301e0d11cd6dc7bba9.jpg", + "table_caption": [ + "Table 3. Results on hard case rectification. \"Hard Cases\" are test samples misrecognized by both MAERec [17] and our $\\mathrm{E}^2$ STR-base. By providing annotation of a small part of the hard cases, we compare the performance increase in the rest test samples between the fine-tuned MAERec and our $\\mathrm{E}^2$ STR-ICL." + ], + "table_footnote": [], + "table_body": "
Training TaskWord Accuracy
VTTSSSNon-ContextIn-Context
✓69.6926.82
✓✓69.8075.60
✓✓69.6673.09
✓✓✓69.6676.77
", + "bbox": [ + 106, + 260, + 439, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation of our proposed training strategies, where VT, TS, and SS refer to vanilla STR training, the Transform Strategy, and the Split Strategy. The experiment is performed on EIST.", + "bbox": [ + 75, + 366, + 467, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "formance than the fine-tuned SOTA methods.", + "bbox": [ + 76, + 434, + 375, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.3 Results on Hard Case Rectification", + "text_level": 1, + "bbox": [ + 76, + 467, + 372, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate the rectification ability of $\\mathrm{E}^2\\mathrm{STR}$ , which can handle hard cases in STR conveniently and effectively, in a training-free manner. Specifically, we define \"hard cases\" as the scene text samples that are wrongly recognized by both $\\mathrm{E}^2\\mathrm{STR}$ -base and the SOTA method MAERec. A small number of hard cases are then annotated, and we study how the model can benefit from the annotated hard cases and decrease the error rate of the rest hard cases. Shown in Table 3, we perform experiments on COCO Text, HOST, and WordArt. As we can see, by annotating $10\\%$ to $20\\%$ of the hard cases, $\\mathrm{E}^2\\mathrm{STR}$ -ICL decreases the error rate of the rest hard cases by up to $32\\%$ . This improvement is achieved by simply putting the annotated hard cases into the in-context pool, without any hassle of re-training the model. By comparison, by fine-tuning on the annotated hard cases, MAERec only decreases the error rate by up to $2.23\\%$ , completely incomparable to our $\\mathrm{E}^2\\mathrm{STR}$ -ICL.", + "bbox": [ + 75, + 489, + 467, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 756, + 240, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Split-and-Transform Training Strategies. We perform an experiment to show the effectiveness of our proposed Split Strategy and Transform Strategy. Shown in Table 4, the vanilla STR training brings a word accuracy of $69.69\\%$ , but the model cannot understand context information, and the performance even severely decreases to $26.82\\%$ when provided with in-context prompts. The application of the Transform Strategy and the Split Strategy in", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6781f97025bd314b694923578d0f14fad12bdc7a6a8c7bae8cfcc1a141fead86.jpg", + "image_caption": [ + "Figure 6. Comparison between different in-context prompt selection strategies. \"Random Selection\" refers to randomly selecting two samples as in-context prompts from the in-context pool. X-axis is the evaluated benchmarks." + ], + "image_footnote": [], + "bbox": [ + 509, + 89, + 880, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the second training stage does not increase the non-context performance (concerning that the synthesized data is typically weaker than the real-world data used in the vanilla training stage), but the model learns to profit from context, and the performance is improved to $75.60\\%$ and $73.09\\%$ respectively when provided with in-context prompts. Finally, the hybrid of the above two strategies further enhances the ICL ability, and the performance reaches $76.77\\%$ .", + "bbox": [ + 496, + 324, + 890, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Nearest Neighbor In-Context Prompt Selection. We compare the kNN in-context prompt selection strategy to Random Selection, i.e., randomly selecting in-context prompts from the in-context pool. Shown in Figure 6, on all three evaluated datasets, random selection can improve the performance of non-context prediction by a small margin, but is far from comparing with kNN selection. Specifically, on EIST random selection improves the performance of non-context from $69.66\\%$ to $70.65\\%$ , while kNN selection reaches $76.77\\%$ word accuracy under the same condition.", + "bbox": [ + 496, + 446, + 890, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of In-Context Pool Size. We next study the impact of varying the size of the in-context pool. Shown in Figure 7, we perform experiments on IAM, EIST, and MPSC, by varying the number of samples maintained in the in-context pool. As we can see, in general, the larger in-context pool brings about better performance, and a small number of samples is adequate to bring about huge performance improvement when deploying $\\mathrm{E}^2$ STR-ICL.", + "bbox": [ + 496, + 612, + 890, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of the Number of In-Context prompts. We analyze the influence of the number of in-context prompts. Shown in Table 5, the experiment is performed on HOST, ToTal Text, and WordArt. Similar to the in-context pool size, the increase in the number of in-context prompts also generally boosts the performance of $\\mathrm{E}^2$ STR-ICL. Moreover, one to two in-context prompts are adequate for improving the performance by a large margin.", + "bbox": [ + 496, + 734, + 890, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational Complexity. We experimentally compare the inference speed of $\\mathrm{E}^2\\mathrm{STR}$ and MAERec [17]. Shown in Table 6, the inference speed of $\\mathrm{E}^2\\mathrm{STR - ICL}$ is on par", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15573", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/93d6f6e5a20b02efc97be0831c144ec55ee061278170774389b547806be59f2d.jpg", + "image_caption": [ + "Figure 7. The performance change brought by different sizes of the in-context pool. The X-axis is the size of the in-context pool and the Y-axis is the word accuracy results." + ], + "image_footnote": [], + "bbox": [ + 88, + 90, + 465, + 247 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/84b2162522ef1958e2ad0999b2207c5741dd93d60f6a0bfad38a6e7624bc5a27.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ICL prompts0124816
HOST73.3074.3474.7574.8374.8374.92
TT93.4694.6894.6894.7894.8894.91
WordArt81.4786.0486.1786.1786.2786.35
", + "bbox": [ + 93, + 311, + 455, + 369 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ecde32a1d263add66585695f10a0b1172437394485061c7fef34a8a655bbb6fd.jpg", + "table_caption": [ + "Table 5. The performance change brought by the different number of in-context prompts." + ], + "table_footnote": [], + "table_body": "
MAERec [17]E2STR-baseE2STR-ICL
Inference Time (s)0.0920.0710.094
", + "bbox": [ + 91, + 419, + 455, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Comparison of the mean inference time of each test sample. All results are reported under the same hardware environment.", + "bbox": [ + 76, + 460, + 468, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with MAERec. Compared to $\\mathrm{E}^2\\mathrm{STR}$ -base, the in-context prompts of $\\mathrm{E}^2\\mathrm{STR}$ -ICL bring extra consumption, but this leads to a limited inference time increase (i.e., from 0.071 to 0.094). It makes sense since we only maintain the visual tokens in the in-context pool and directly feed the visual tokens of the selected prompts to the language model.", + "bbox": [ + 75, + 513, + 468, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization and Further Analysis. We further study how the test sample learns from context. Shown in Figure 8, we select one context prompt for the test sample, and study the model pays attention to which region of the context image. This is achieved by collecting the attention maps between the language tokens and the image features. As we can see, when the language tokens pay close attention to the misrecognized image region, they also focus on the context image region which has similar patterns. For example, on the last row of Figure 8, $\\mathrm{E}^2\\mathrm{STR}$ misrecognized the test image as \"simplest\" without context. By providing a context prompt \"Display\", one language token focuses on the \"la\" region of both images, which have similar image patterns. Finally, $\\mathrm{E}^2\\mathrm{STR}$ rectified the misrecognized \"e\" to \"a\" with the help of context ground-truth \"la\" of the focused region.", + "bbox": [ + 75, + 604, + 468, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 76, + 844, + 200, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "There are two limitations in our study. Firstly, there is a very slim chance that $\\mathrm{E}^2\\mathrm{STR}$ -ICL erroneously rectifies pre", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb73f529238dbad11fb19e80a7ad099dda95c56a0f91d84952d8e3cfa36f5efd.jpg", + "image_caption": [ + "Vitoc" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 625, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d9d82612dc2e15073ed5c84b0a21b97e6a17710f7940946e18c37129fb604d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 141, + 627, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/647e6b809f51ed9fdeaa153f5ab50dd8d614ee65af2418719ce8c386618b4c5a.jpg", + "image_caption": [ + "PORTED" + ], + "image_footnote": [], + "bbox": [ + 504, + 191, + 625, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d87c88179af81fa73d266bcbb6bbe3b83d39ed9edf7c45d127984831206842ea.jpg", + "image_caption": [ + "simplest", + "Non-Context Prediction", + "Figure 8. Cross attention visualization between the language tokens and the vision tokens. Left: Non-context prediction of $\\mathrm{E}^2\\mathrm{STR}$ . Error characters are marked in red. Right: In-context prediction of $\\mathrm{E}^2\\mathrm{STR - ICL}$ , where only one in-context prompt is selected. We visualize how the language tokens attend to the prompt image and the test image." + ], + "image_footnote": [], + "bbox": [ + 504, + 241, + 627, + 276 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21c36fb8edbf9243aa4fc66bd3801cbd202c7520288d34ddb67df5bef3b224bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 90, + 764, + 125 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f192a34e78674eda264ef87513a81c6c5eec554e717784f3096c073d12b57fe4.jpg", + "image_caption": [ + "pal" + ], + "image_footnote": [], + "bbox": [ + 645, + 140, + 764, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5aa92e626be07ba7ea68daaa0da7fa03ab4d7b97ac6793aaf4bca22656e40fbb.jpg", + "image_caption": [ + "integral", + "POWER" + ], + "image_footnote": [], + "bbox": [ + 645, + 188, + 764, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/05e2c5bdd61e9db6780bbb6ae809a78fb703830a582f7813493c3c816fbcad7e.jpg", + "image_caption": [ + "Display", + "In-Context Prompt" + ], + "image_footnote": [], + "bbox": [ + 645, + 239, + 764, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d2cee8497d2796ae8bf38b286327a2d9027d1cacc034f23d77ff1d545691feb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 90, + 887, + 125 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d7eec6f5e561f97bc02e9350d0abd870baed38019a7f39d5560abd91aba403d4.jpg", + "image_caption": [ + "Vital" + ], + "image_footnote": [], + "bbox": [ + 767, + 140, + 888, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b71a7df6b4ff1935ea8ad117e3848d53a00ea68b7ef0eec185362c04798717bb.jpg", + "image_caption": [ + "accepting", + "PORTER" + ], + "image_footnote": [], + "bbox": [ + 767, + 189, + 888, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/848adbcc35db8b1f350e1345d21d0b69074e306452934c3907f3608ffbd04d42.jpg", + "image_caption": [ + "simplest", + "In-Context Prediction" + ], + "image_footnote": [], + "bbox": [ + 767, + 239, + 888, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dictions due to misleading prompts (please refer to supplementary materials). Additionally, our model still could not recognize characters that are not included in the lexicon.", + "bbox": [ + 496, + 433, + 892, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 494, + 617, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose $\\mathrm{E}^2\\mathrm{STR}$ , an ego-evolving scene text recognizer equipped with in-context learning capabilities. Through our proposed in-context training strategy incorporating context-rich scene text sequences, $\\mathrm{E}^2\\mathrm{STR}$ performs rapid adaptation across diverse scenarios without additional fine-tuning. Extensive experiments demonstrate that $\\mathrm{E}^2\\mathrm{STR}$ not only achieves SOTA performance on common STR benchmarks but also outperforms even the approaches that have been fine-tuned specifically for cross-domain scenarios. The model's ability to easily and effectively handle difficult text cases further underscores its potential as a unified text recognizer. Overall, this research represents a significant step toward efficient and highly adaptive text recognition models well-suited for diverse real-world applications.", + "bbox": [ + 496, + 521, + 890, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work is supported by National Natural Science Foundation of China No.62222602, No.62176092, No.62106075, No.U23A20343. Science and Technology Commission of Shanghai No.21511100700, Shanghai Sailing Program (23YF1410500), Natural Science Foundation of Shanghai (23ZR1420400), Natural Science Foundation of Chongqing, China (CSTB2023NSCQ-JQX0007), (CSTB2023NSCQ-MSX0137), Development Project of Ministry of Industry and Information Technology (ZTZB.23-990-016).", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15574", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2, 3, 4", + "[2] Jeonghun Baek, Geewook Kim, Junyeop Lee, Sungrae Park, Dongyoon Han, Sangdoo Yun, Seong Joon Oh, and Hwalsuk Lee. What is wrong with scene text recognition model comparisons? dataset and model analysis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4715-4723, 2019. 3", + "[3] Darwin Bautista and Rowel Atienza. Scene text recognition with permuted autoregressive sequence models. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 1, 3, 6", + "[4] Belval. Generator. https://github.com/Belval/TextRecognitionDataGenerator.5", + "[5] Fedor Borisyuk, Albert Gordo, and Viswanath Sivakumar. Rosetta: Large scale system for text detection and recognition in images. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 71-79, 2018. 2", + "[6] 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, 3", + "[7] Zhanzhan Cheng, Fan Bai, Yunlu Xu, Gang Zheng, Shiliang Pu, and Shuigeng Zhou. Focusing attention: Towards accurate text recognition in natural images. In Proceedings of the IEEE international conference on computer vision, pages 5076-5084, 2017. 3", + "[8] Chee Kheng Ch'ng and Chee Seng Chan. Total-text: A comprehensive dataset for scene text detection and recognition. In 2017 14th IAPR international conference on document analysis and recognition (ICDAR), pages 935-942. IEEE, 2017. 5", + "[9] 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. 5", + "[10] Shancheng Fang, Hongtao Xie, Yuxin Wang, Zhendong Mao, and Yongdong Zhang. Read like humans: Autonomous, bidirectional and iterative language modeling for scene text recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7098-7107, 2021. 3, 6", + "[11] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, pages 369-376, 2006. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Tongkun Guan, Chaochen Gu, Changsheng Lu, Jingzheng Tu, Qi Feng, Kaijie Wu, and Xinping Guan. Industrial scene text detection with refined feature-attentive network. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):6073–6085, 2022. 5", + "[13] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 3", + "[14] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. Icl-d3ie: In-context learning with diverse demonstrations updating for document information extraction. arXiv preprint arXiv:2303.05063, 2023.", + "[15] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. Icl-d3ie: In-context learning with diverse demonstrations updating for document information extraction. arXiv preprint arXiv:2303.05063, 2023. 3", + "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5", + "[17] Qing Jiang, Jiapeng Wang, Dezhi Peng, Chongyu Liu, and Lianwen Jin. Revisiting scene text recognition: A data perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20543-20554, 2023. 1, 3, 5, 6, 7, 8", + "[18] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazan Almazan, and Lluis Pere De Las Heras. Icedar 2013 robust reading competition. In 2013 12th international conference on document analysis and recognition, pages 1484-1493. IEEE, 2013. 5", + "[19] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In 2015 13th international conference on document analysis and recognition (ICDAR), pages 1156-1160. IEEE, 2015. 5", + "[20] Junyeop Lee, Sungrae Park, Jeonghun Baek, Seong Joon Oh, Seonghyeon Kim, and Hwalsuk Lee. On recognizing texts of arbitrary shapes with 2d self-attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 546–547, 2020. 1, 3, 6", + "[21] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2, 3", + "[22] Hui Li, Peng Wang, Chunhua Shen, and Guyu Zhang. Show, attend and read: A simple and strong baseline for irregular text recognition. In Proceedings of the AAAI conference on artificial intelligence, pages 8610-8617, 2019. 6", + "[23] Minghui Liao, Jian Zhang, Zhaoyi Wan, Fengming Xie, Jia-jun Liang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Scene text recognition from two-dimensional perspective. In Pro" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15575", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the AAAI conference on artificial intelligence, pages 8714-8721, 2019. 2", + "[24] Wei Liu, Chaofeng Chen, Kwan-Yee K Wong, Zhizhong Su, and Junyu Han. Star-net: a spatial attention residue network for scene text recognition. In BMVC, page 7, 2016. 2", + "[25] Yuliang Liu, Lianwen Jin, Shuai tao Zhang, Canjie Luo, and Sheng Zhang. Curved scene text detection via transverse and longitudinal sequence connection. Pattern Recognition, 90: 337-345, 2019. 5", + "[26] Mengkai Ma, Qiu-Feng Wang, Shan Huang, Shen Huang, Yannis Goulermas, and Kaizhu Huang. Residual attention-based multi-scale script identification in scene text images. Neurocomputing, 421:222-233, 2021. 1", + "[27] U-V Marti and Horst Bunke. The iam-database: an english sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition, 5: 39-46, 2002. 5", + "[28] Qiang Mei, Qinyou Hu, Chun Yang, Hailin Zheng, and Zhisheng Hu. Port recommendation system for alternative container port destinations using a novel neural language-based algorithm. IEEE Access, 8:199970-199979, 2020. 1", + "[29] Anand Mishra, Karteek Alahari, and CV Jawahar. Top-down and bottom-up cues for scene text recognition. In 2012 IEEE conference on computer vision and pattern recognition, pages 2687-2694. IEEE, 2012. 5", + "[30] Imene OUALI, Mohamed BEN HALIMA, and WALI Ali. Augmented reality for scene text recognition, visualization and reading to assist visually impaired people. Procedia Computer Science, 207:158-167, 2022. 1", + "[31] Trung Quy Phan, Palaiahnakote Shivakumara, Shangxuan Tian, and Chew Lim Tan. Recognizing text with perspective distortion in natural scenes. In Proceedings of the IEEE international conference on computer vision, pages 569-576, 2013. 5", + "[32] Anhar Risnumawan, Palaiahankote Shivakumara, Chee Seng Chan, and Chew Lim Tan. A robust arbitrary text detection system for natural scene images. Expert Systems with Applications, 41(18):8027-8048, 2014. 5", + "[33] Abdul Khader Jilani Saudagar and HabeebVulla Mohammad. Augmented reality mobile application for arabic text extraction, recognition and translation. Journal of Statistics and Management Systems, 21(4):617-629, 2018. 1", + "[34] Zhenwei Shao, Zhou Yu, Meng Wang, and Jun Yu. Prompting large language models with answer heuristics for knowledge-based visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14974-14983, 2023. 3", + "[35] Fenfen Sheng, Zhineng Chen, and Bo Xu. Ntrr: A no-recurrence sequence-to-sequence model for scene text recognition. In 2019 International conference on document analysis and recognition (ICDAR), pages 781-786. IEEE, 2019. 1, 6", + "[36] Baoguang Shi, Mingkun Yang, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Aster: An attentional scene text recognizer with flexible rectification. IEEE transactions on pattern analysis and machine intelligence, 41(9):2035-2048, 2018. 3, 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Cunzhao Shi, Chunheng Wang, Baihua Xiao, Song Gao, and Jinlong Hu. End-to-end scene text recognition using tree-structured models. Pattern Recognition, 47(9):2853-2866, 2014. 5", + "[38] 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. 2, 3", + "[39] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie. Coco-text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016. 5", + "[40] Zhaoyi Wan, Minghang He, Haoran Chen, Xiang Bai, and Cong Yao. Textscanner: Reading characters in order for robust scene text recognition. In Proceedings of the AAAI conference on artificial intelligence, pages 12120-12127, 2020. 2", + "[41] Yuxin Wang, Hongtao Xie, Shancheng Fang, Jing Wang, Shenggao Zhu, and Yongdong Zhang. From two to one: A new scene text recognizer with visual language modeling network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14194–14203, 2021. 5", + "[42] Xudong Xie, Ling Fu, Zhifei Zhang, Zhaowen Wang, and Xiang Bai. Toward understanding wordart: Corner-guided transformer for scene text recognition. In European Conference on Computer Vision, pages 303-321. Springer, 2022. 5", + "[43] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. An empirical study of gpt-3 for few-shot knowledge-based vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3081-3089, 2022. 3", + "[44] Deli Yu, Xuan Li, Chengquan Zhang, Tao Liu, Junyu Han, Jingtuo Liu, and Errui Ding. Towards accurate scene text recognition with semantic reasoning networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12113-12122, 2020. 3", + "[45] Chongsheng Zhang, Yuefeng Tao, Kai Du, Weiping Ding, Bin Wang, Ji Liu, and Wei Wang. Character-level street view text spotting based on deep multisegmentation network for smarter autonomous driving. IEEE Transactions on Artificial Intelligence, 3(2):297-308, 2021. 1", + "[46] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 3, 5", + "[47] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 835 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15576", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/bf855ad8-e560-4eb0-9886-16b2ff2266f1_model.json b/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/bf855ad8-e560-4eb0-9886-16b2ff2266f1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b3b2a774d781c26a8b70f482a654d6d821c884bd --- /dev/null +++ b/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/bf855ad8-e560-4eb0-9886-16b2ff2266f1_model.json @@ -0,0 +1,2541 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR 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." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.895, + 0.155 + ], + "angle": 0, + "content": "Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.18, + 0.788, + 0.235 + ], + "angle": 0, + "content": "Zhen Zhao\\(^{1,2,*}\\), Jingqun Tang\\(^{2,\\dagger}\\), Chunhui Lin\\(^{2}\\), Binghong Wu\\(^{2}\\), Can Huang\\(^{2}\\), Hao Liu\\(^{2}\\), Xin Tan\\(^{1}\\), Zhizhong Zhang\\(^{1}\\), Yuan Xie\\(^{1,\\dagger}\\) \n\\(^{1}\\) East China Normal University \\(^{2}\\)ByteDance" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.237, + 0.774, + 0.253 + ], + "angle": 0, + "content": "{51255901056}@stu.ecnu.edu.cn, {zzzhang,xtan,yxie}@cs.ecnu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.255, + 0.774, + 0.27 + ], + "angle": 0, + "content": "{tangjingqun,linchunhui,wubinghong,liuhao,can.huang}@bytedance.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.337, + 0.474, + 0.685 + ], + "angle": 0, + "content": "Scene text recognition (STR) in the wild frequently encounters challenges when coping with domain variations, font diversity, shape deformations, etc. A straightforward solution is performing model fine-tuning tailored to a specific scenario, but it is computationally intensive and requires multiple model copies for various scenarios. Recent studies indicate that large language models (LLMs) can learn from a few demonstration examples in a training-free manner, termed \"In-Context Learning\" (ICL). Nevertheless, applying LLMs as a text recognizer is unacceptably resource-consuming. Moreover, our pilot experiments on LLMs show that ICL fails in STR, mainly attributed to the insufficient incorporation of contextual information from diverse samples in the training stage. To this end, we introduce \\( E^2 \\)STR, a STR model trained with context-rich scene text sequences, where the sequences are generated via our proposed in-context training strategy. \\( E^2 \\)STR demonstrates that a regular-sized model is sufficient to achieve effective ICL capabilities in STR. Extensive experiments show that \\( E^2 \\)STR exhibits remarkable training-free adaptation in various scenarios and outperforms even the fine-tuned state-of-the-art approaches on public benchmarks. The code is released at https://github.com/bytedance/E2STR." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.715, + 0.21, + 0.73 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.741, + 0.47, + 0.816 + ], + "angle": 0, + "content": "Scene Text Recognition (STR) is a fundamental task in computer vision, with extensive applications in several domains such as autonomous driving [45], augmented reality [30, 33], industrial print recognition [28] and visual understanding [26]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.817, + 0.47, + 0.847 + ], + "angle": 0, + "content": "Current progress in STR [3, 17, 20, 35] has demonstrated remarkable performance in numerous scenarios." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.848, + 0.469, + 0.865 + ], + "angle": 0, + "content": "However, as shown in Figure 1 (a), STR models are sup" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.308, + 0.889, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.389, + 0.817, + 0.402 + ], + "angle": 0, + "content": "(a) Diversified scenarios of scene text" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.406, + 0.89, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.49, + 0.893, + 0.504 + ], + "angle": 0, + "content": "(b) Scenario-specific fine-tuning pipeline of current methods" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.508, + 0.889, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.584, + 0.829, + 0.599 + ], + "angle": 0, + "content": "(c) In-context learning pipeline of \\(\\mathbf{E}^2\\mathbf{STR}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.61, + 0.894, + 0.736 + ], + "angle": 0, + "content": "Figure 1. Demonstration of real-world scene text scenarios and the adaptation pipeline. (a) Diversified scenarios of scene text in the real world. (b) The adaptation pipeline of current methods. They typically have to fine-tune upon a trained STR model with the training set, under a specific scenario. (c) The adaptation pipeline of our proposed \\(\\mathrm{E}^2\\mathrm{STR}\\). Our method automatically selects in-context prompts and performs training-free adaptation when faced with novel scenarios. Blue characters denote ambiguous scene text that is easily recognized." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.884 + ], + "angle": 0, + "content": "posed to perform robustly over diversified scenarios in the real world, where the scene text is hard to recognize because of domain variation, font diversity, shape deformation, etc. As shown in Figure 1 (b), a straightforward solution involves collecting the corresponding data and then fine-tuning the model for the specific scenario [3, 17, 20]. This process is computationally intensive and requires multiple model copies for diverse scenarios." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The development of a comprehensive and reliable STR" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.875, + 0.237, + 0.888 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.888, + 0.436, + 0.9 + ], + "angle": 0, + "content": "*This work is done when Zhen Zhao is an intern at ByteDance." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.875, + 0.436, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15567" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.091, + 0.268, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.162, + 0.243, + 0.17 + ], + "angle": 0, + "content": "(b) Context-Rich scene text sequence" + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.089, + 0.47, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.162, + 0.468, + 0.17 + ], + "angle": 0, + "content": "(c) ICL results on cross-domain benchmark" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.183, + 0.47, + 0.252 + ], + "angle": 0, + "content": "Figure 2. Our pilot experiments. (a) The randomly concatenated scene text sequence. (b) Our proposed context-rich scene text sequence. (c) By training an STR model based on the randomly concatenated scene text sequence, we evaluate the model on three cross-domain datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.28, + 0.469, + 0.31 + ], + "angle": 0, + "content": "model that can effectively handle many real-world scenarios remains a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.469, + 0.446 + ], + "angle": 0, + "content": "Fortunately, plenty of studies [1, 6, 21, 38] have shown that Large Language Models (LLMs) can easily adapt without additional training. This adaptation is achieved by leveraging only a handful of input-label pairs as context (prompting information), a phenomenon known as \"In-Context Learning\" (ICL). The advantages of ICL inspire our interest in implementing it in STR, such that by fetching a few in-context prompts, a single model can be rapidly adapted to various scenarios without fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.469, + 0.537 + ], + "angle": 0, + "content": "However, the equipment of ICL in STR still poses challenges under the existing circumstances. Firstly, it is deemed excessively costly to apply Multi-Modal Large Language Models (M-LLMs) with billions of parameters as a scene text recognizer. And the ICL capabilities in regular-sized models have been barely explored currently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.469, + 0.809 + ], + "angle": 0, + "content": "Secondly, it is hard to acquire ICL capabilities for a STR model with current training strategies. Previous studies have observed that sending image-text sequences for training would naturally endow ICL for M-LLMs [1, 21, 38], while such a phenomenon is hard to achieve in STR. As shown in Figure 2 (a), we generate sequential training data by randomly concatenating scene text samples. This practice fails as the trained model does not exhibit any performance improvement even when provided with in-domain prompts (Figure 2 (c)). The major cause of this failure is the lack of context in the generated scene text sequences during the training phase. The arbitrary concatenation of scene text does not provide any contextual information (i.e., sample connections) between different samples (Figure 2 (a)). Consequently, the model lacks the ability to effectively use information derived from in-context prompts (Figure 2 (c)), which implies that in-context training is essentially important for the effective implementation of ICL in STR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Based on the above analysis, we propose \\(\\mathrm{E}^2\\mathrm{STR}\\) (Ego-Evolving STR), a paradigm that facilitates adaptation across diverse scenarios in a training-free manner. Specifically, we propose an in-context training strategy, which enables the model to exploit contextual information from the generated context-rich scene text sequences (Figure 2" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "(b)). The context-rich scene text sequences are formed using our ST-strategy, which involves random Splitting and Transformation of scene text, hence generating a set of \"sub-samples\". The sub-samples are inner-connected in terms of both visual and linguistic aspects. In the inference stage, \\(\\mathrm{E}^2\\mathrm{STR}\\) fetches in-context prompts based on visual similarities, and utilizes the prompts to assist the recognition, shown in Figure 1 (c). In practice, it is found that with proper training and inference strategies, ICL capabilities can also be observed in regular-sized STR models (hundreds of millions of parameters)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Finally, the proposed \\(\\mathrm{E}^2\\mathrm{STR}\\) effectively captures contextual information from the in-context prompts and performs rapid adaptation in various novel scenarios in a training-free manner (Please refer to Section 4.2). On common benchmarks, \\(\\mathrm{E}^2\\mathrm{STR}\\) achieves SOTA results, with an average improvement of \\(0.8\\%\\) over previous methods and \\(1.1\\%\\) over itself without ICL. Most importantly, when evaluated on unseen domains, \\(\\mathrm{E}^2\\mathrm{STR}\\) achieves impressive performance with only a few prompts, even outperforming the fine-tuning results of SOTA methods by \\(1.2\\%\\). Our contributions are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.47 + ], + "angle": 0, + "content": "(1) We propose \\(\\mathrm{E}^2\\mathrm{STR}\\), a robust STR paradigm that can perform rapid adaptation over diverse scenarios in a training-free manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.892, + 0.532 + ], + "angle": 0, + "content": "(2) We provide an in-context training strategy for equipping STR models with ICL capabilities, as well as an in-context inference strategy for STR models to leverage contextual information from in-context prompts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.578 + ], + "angle": 0, + "content": "(3) We demonstrate that ICL capabilities can be effectively incorporated into regular-sized STR models via appropriate training and inference strategies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.638 + ], + "angle": 0, + "content": "(4) Extensive experiments show that \\(\\mathrm{E}^2\\mathrm{STR}\\) exceeds state-of-the-art performance across diverse benchmarks, even surpassing the fine-tuned approaches in unseen domains." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.655, + 0.642, + 0.67 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.715, + 0.696 + ], + "angle": 0, + "content": "2.1. Scene Text Recognition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.891, + 0.749 + ], + "angle": 0, + "content": "Recent years have witnessed extensive studies in STR, which can be generally divided into Language-free methods and Language-aware methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Language-free STR. Language-free models directly utilize visual features for prediction, without considering the relationship between the characters. In this branch CTC-based [11] methods [5, 24] play the most prominent part. They typically consist of a CNN for feature extraction and an RNN for sequential feature processing, which are trained end-to-end with the CTC loss [11]. Other methods like [23, 40] focus on treating STR as a character-level segmentation task. The lack of linguistic information limits the application of language-free methods in scenarios with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15568" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.31, + 0.106 + ], + "angle": 0, + "content": "occluded or incomplete characters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.47, + 0.304 + ], + "angle": 0, + "content": "Language-aware STR. Language-aware models leverage linguistic information to assist the recognition, typically utilizing an external language model (LM) [10, 44] or training internal LMs [2, 7, 36]. SRN [44] and ABINet [10] feed visual predictions to an external LM for linguistic refinement. The direct application of an external LM without considering visual features leads to possible erroneous correction. On the other hand, methods like PARSeq [3] and MAERec [17] implicitly train an internal LM in an auto-regressive manner, which have achieved decent performance. In this paper we base our model on the language-aware design, training a transformer-based language decoder inner-connected with the vision encoder." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.314, + 0.377, + 0.331 + ], + "angle": 0, + "content": "2.2. Multi-Modal In-Context Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.473, + 0.686 + ], + "angle": 0, + "content": "Recent large language models (LLMs) [6, 46] have demonstrated their excellent few-shot adaptation capabilities. By concatenating a few examples with the input as the prompt at reference time, LLMs quickly adapt to novel tasks without parameter updating. This phenomenon introduces a novel learning paradigm termed \"In-Context Learning\". Meanwhile, unlike LLMs, vision-language models (VLMs) struggle to understand complex multi-modal prompts [47]. A large set of approaches [13-15, 34] have been proposed to empower VLMs with multi-modal in-context learning (M-ICL) capabilities, but they typically utilize vision models (like image caption models) to translate images to text [15, 34, 43], or view the LLM as a scheduler learning to call vision experts based on a few examples [13]. These approaches do not truly establish a VLM with M-ICL capabilities. Recently, several work [1, 21, 38] proposes to train VLMs with sequential multi-modal data, and have achieved great success in prompting VLMs with multi-modal examples. In this paper, we aim to train a scene text recognizer equipped with M-ICL capabilities based on this sequential training paradigm. We demonstrate that the arbitrary concatenation of scene text fails as stated above, which motivates us to generate context-rich scene text sequences." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.7, + 0.212, + 0.718 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.47, + 0.757 + ], + "angle": 0, + "content": "3.1. Preliminary of Multi-Modal In-Context Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.469, + 0.854 + ], + "angle": 0, + "content": "Multi-modal in-context Learning enables M-LLMs to perform quick adaptation in downstream tasks in a training-free manner, hence eliminating the redundant computation and time expenses of fine-tuning. In this subsection, we introduce how to formulate multi-modal in-context learning for addressing the STR task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "For a scene text tuple \\((\\mathbf{x},\\mathbf{y})\\) where \\(\\pmb{x}\\) is the scene image and \\(\\pmb{y}\\) is the ground-truth text, the STR task involves generating the label \\(\\pmb{y}\\) by maximizing the conditional probabil" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "ity under the classic auto-regressive paradigm as follows: \\(p(\\boldsymbol{y}|\\boldsymbol{x}) = \\prod_{l=1}^{L} p(\\boldsymbol{y}_l|\\boldsymbol{x}, \\boldsymbol{y}_{ \\). The training in this phase makes use of the next-token prediction loss:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.821, + 0.47, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y}) \\sim \\mathcal {D}} \\left[ - \\sum_ {l = 1} ^ {L} \\log p (\\boldsymbol {y} _ {l} | \\boldsymbol {y} _ {< l}, \\boldsymbol {x}) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.26, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\) is the training set." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.525, + 0.696, + 0.541 + ], + "angle": 0, + "content": "3.3.2 In-Context Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.702 + ], + "angle": 0, + "content": "The objective of the in-context training phase is to equip \\(\\mathrm{E}^2\\mathrm{STR}\\) with the capability of In-Context Learning. As depicted in the top of Figure 3, the model is trained with context-rich scene text sequences as stated before. In these sequences, we interleave a placeholder \\(\\) in the text for each image. This serves to make the language decoder distinguish between different samples following [1]. In this stage, we propose two strategies to generate context-rich scene text sequences: the Split Strategy and the Transform Strategy (the ST strategy)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.839 + ], + "angle": 0, + "content": "The Split Strategy. As shown in Figure 4 (a), when presented with a training tuple \\((\\pmb{x},\\pmb{y})\\), we split the sample and hence generating a set of \"sub-samples\". It is evident that the sub-samples exhibit a strong connection to the original training sample. Furthermore, the sub-samples themselves demonstrate interconnectivity as they overlap with one another. Next, we proceed to concatenate the sub-samples with \\((\\pmb{x},\\pmb{y})\\) and additional randomly selected samples to form a context-rich sample sequence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.885 + ], + "angle": 0, + "content": "We randomly shuffle the whole sequence before generating the actual input text (i.e., interleaving the \\( \\langle \\mathbf{\\Pi} / \\mathbf{i}\\rangle \\) token to the text sequence)." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In practice, to accurately split the training samples, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15570" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "synthesize 600k scene text images based on [4] and record the accurate bounding boxes of every single character. Our subsequent experiments show that the synthesized data does not change \\(\\mathrm{E}^2\\mathrm{STR}\\)'s non-context text recognition ability, but the Split Strategy based on them equips \\(\\mathrm{E}^2\\mathrm{STR}\\) with a strong capability of in-context learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.471, + 0.288 + ], + "angle": 0, + "content": "The Transform Strategy. As shown in Figure 4 (b), given a training tuple \\((\\boldsymbol{x}, \\boldsymbol{y})\\) (whether with character-wise bounding boxes or not), we perform data augmentation (a set of image transformations, e.g., color/direction transformations) on \\(\\boldsymbol{x}\\). In this way, we also generate a set of sub-samples with the same label but different image patterns from the original sample." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.47, + 0.395 + ], + "angle": 0, + "content": "In practice, as depicted in Figure 4 (c), we hybrid the above strategies. The training set is formed by concatenating the synthesized data and the original training data used in the first training phase. For the synthesized data with character-wise bounding boxes, both the Split Strategy and the Transform Strategy are utilized. For the original training data, only the Transform Strategy is implemented." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.395, + 0.471, + 0.471 + ], + "angle": 0, + "content": "Finally, after generating the sample sequence \\((X,Y)\\), where \\(X\\) is the image sequence and \\(Y\\) is the text sequence, \\(X\\) is fed into the vision encoder, while \\(Y\\) is processed by the language decoder under the auto-regressive framework. The loss function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.481, + 0.47, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\left(\\boldsymbol {X}, \\boldsymbol {Y}\\right)} = - \\sum_ {l = 1} ^ {L} \\log p \\left(\\boldsymbol {Y} _ {l} \\mid \\boldsymbol {Y} _ {< l}, \\boldsymbol {X} _ {\\leq l}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.566 + ], + "angle": 0, + "content": "where \\( \\mathbf{X}_{\\leq l} \\) is the set of image tokens preceding token \\( \\mathbf{Y}_l \\) in the input sequence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.576, + 0.276, + 0.59 + ], + "angle": 0, + "content": "3.4. In-Context Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.766 + ], + "angle": 0, + "content": "The In-Context Learning ability is acquired by our \\(\\mathrm{E}^2\\mathrm{STR}\\) model through the above two-stage training approach. AS shown in the bottom of Figure 3, when presented with a test image \\(\\pmb{x}\\), the framework selects \\(N\\) samples \\(\\{(x_i^c,y_i^c)\\}_{i = 1}^N\\) from a in-context pool \\(\\mathcal{D}^c\\). The selected samples have the highest visual similarities to \\(\\pmb{x}\\) in the latent space. Specifically, we calculate the image embedding \\(\\pmb{I}\\) of \\(\\pmb{x}\\) by averaging the visual token sequence Encoder \\((\\pmb{x})\\). The in-context prompts are then formed by choosing \\(N\\) samples from \\(\\mathcal{D}^c\\), where the image embeddings of these samples have the top-N highest cosine similarity with \\(\\pmb{I}\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.778, + 0.47, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} = \\underset {i \\in 1, 2, \\dots , | \\mathcal {D} ^ {c} |} {\\arg \\operatorname {T o p N}} \\frac {\\boldsymbol {I} ^ {\\mathrm {T}} \\boldsymbol {I} _ {i} ^ {c}}{\\| \\boldsymbol {I} \\| _ {2} \\| \\boldsymbol {I} _ {i} ^ {c} \\| _ {2}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.827, + 0.469, + 0.872 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}\\) is the index set of the top-N similar samples in \\(\\mathcal{D}^c\\) and \\(I_{i}^{c}\\) is the image embedding of the \\(i\\)-th sample in \\(\\mathcal{D}^c\\). The in-context prompts are then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.885, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {E} = \\left\\{\\left(\\boldsymbol {x} _ {i} ^ {c}, \\boldsymbol {y} _ {i} ^ {c}\\right) | i \\in \\mathcal {I} \\right\\}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "As shown in the bottom of Figure 3, \\( \\pmb{E} \\) is concatenated with the test sample \\( \\pmb{x} \\) and our in-context prediction is formulated as \\( p(\\pmb{y}|\\pmb{E},\\pmb{x}) \\). In practice, the in-context pool \\( \\mathcal{D}^c \\) retains solely the output tokens generated by the vision encoder, resulting in a highly efficient selection process. Furthermore, because the in-context pool is tiny and we do straight inference without training, the extra consumption is kept to a minimum (Please refer to Section 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.225, + 0.625, + 0.241 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.249, + 0.691, + 0.266 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.272, + 0.892, + 0.454 + ], + "angle": 0, + "content": "Implementation Details. Following MAERec [17], we choose Vision Transformer [9] pre-trained under the MAE [16] framework as the vision encoder. The default language decoder is set as OPT-125M [46]. We use the cosine learning rate scheduler without warm-up and the AdamW optimizer with a weight decay of 0.01. We train our model for 10 epochs with an initial learning rate of 1e-4 during the first training stage, and 5 epochs with an initial learning rate of 5e-6 during the second in-context training stage. The training batch size is 64 for the first stage and 8 for the second stage. During inference for \\(\\mathrm{E}^2\\mathrm{STR}\\)-ICL, we select two in-context prompts based on the kNN selection strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.726 + ], + "angle": 0, + "content": "Datasets and Metrics. We use the real-world training dataset Union14M-L[17] for the two-stage training. The same training dataset (including the synthesized data) is adopted for all compared methods. \\(\\mathrm{E}^2\\mathrm{STR}\\) is evaluated under various publicly available benchmarks, including Regular Benchmarks IIIT5k [29], SVT [37], IC13 [18], Irregular Benchmarks IC15 [19], SVTP [31], CUTE80 (CT80) [32], COCO Text (COCO) [39], CTW [25], Total Text (TT) [8], Occluded Benchmarks OST (HOST and WOST) [41] and artistic benchmark WordArt [42]. In cross domain scenarios the evaluated datasets including the metal-surface benchmark MPSC [12] and the handwriting benchmark IAM [27], as well as a more difficult real-world industrial text recognition dataset EIST (Enhanced Industrial Scene Text) collected by us. EIST is collected from the real-world industrial scenario, which contains 200 training samples and 8000 test samples. We use Word Accuracy[17] as the evaluation metric for all compared methods." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.733, + 0.64, + 0.748 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.787, + 0.771 + ], + "angle": 0, + "content": "4.2.1 Results on Common Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Table 1 presents the performance of \\(\\mathrm{E}^2\\mathrm{STR}\\) on common benchmarks. We evaluate \\(\\mathrm{E}^2\\mathrm{STR}\\) on 12 commonly used STR benchmarks and compare with SOTA methods. \\(\\mathrm{E}^2\\mathrm{STR}\\)-base refers to non-context prediction without prompts. For \\(\\mathrm{E}^2\\mathrm{STR}\\)-ICL, a tiny in-context pool is maintained by randomly sampling 1000 images from the training data (less than \\(0.025\\%\\) of the number of training samples). As we can see, \\(\\mathrm{E}^2\\mathrm{STR}\\)-base achieves \\(90.25\\%\\) average word" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15571" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.088, + 0.891, + 0.259 + ], + "angle": 0, + "content": "
MethodVenueRegularIrregularOccludedOthersAVG
IIIT3000SVT647IC131015IC152077SVTP645CT80288COCO9896CTW1572TT2201HOST2416WOST2416WordArt1511
ASTER [36]PAMI'1895.0389.4993.7985.4882.0290.2862.2576.5378.6943.3464.6565.5977.26
NRTR [35]ICDAR'1997.4393.8296.0685.1584.0391.3265.9481.7481.8350.8371.5264.0680.31
SAR [22]AAAI'1997.7094.1396.3587.4787.6093.0667.4183.9186.2346.3670.3272.4081.91
SATRN [20]AAAI'2097.8395.8397.4489.4690.8596.1873.0684.6187.9156.7175.6275.7185.10
ABINet [10]CVPR'2197.9095.9896.1691.6690.2393.7571.4683.7286.0156.5475.7575.2584.53
PARSeq* [3]ECCV'2299.1097.8498.1389.2296.9098.61-------
MAERec [17]ICCV'2398.9397.9998.6293.0494.5798.9678.8488.8793.9173.9785.7282.5990.50
\\( E^2STR \\)-base99.1098.1598.0392.9996.4398.9677.2988.3693.4673.3085.5181.4790.25
\\( E^2STR-ICL \\)99.2398.6198.7293.8296.7499.3178.3888.9994.6874.7586.5986.1791.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.268, + 0.893, + 0.311 + ], + "angle": 0, + "content": "Table 1. Results on common benchmarks. All methods are trained on the same dataset except for PARSeq.*: PARSeq is trained on its self-collected real-world dataset and we directly quote the results from its original paper. Red and blue values denote the best and the secondary performance. \\( {\\mathrm{E}}^{2} \\) STR-base refers to non-context inference." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.334, + 0.475, + 0.505 + ], + "angle": 0, + "content": "
MethodIndustrialHandwritingAVG
MPSC 2941EIST 8000IAM 3000
ASTER [36]63.4848.7652.5054.91
NRTR [35]73.2461.7759.5364.85
SAR [22]73.8558.2656.6362.91
ABINet [10]75.3562.8561.5766.59
SATRN [20]76.1065.4259.4767.00
MAERec [17]81.8170.3370.2774.14
E²STR-base81.2669.6669.5173.48
E²STR-ICL83.6476.7774.1078.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.515, + 0.471, + 0.557 + ], + "angle": 0, + "content": "Table 2. Results on cross domain scenarios. Three datasets under two unseen domains are evaluated. All approaches are evaluated in a training-free manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.469, + 0.675 + ], + "angle": 0, + "content": "accuracy over 12 datasets, \\(0.25\\%\\) lower than MAERec [17]. However, by fetching in-context prompts and exploiting incontext information, \\(\\mathrm{E}^2\\)STR-ICL achieves an average word accuracy of \\(91.33\\%\\), which is \\(1.08\\%\\) higher than \\(\\mathrm{E}^2\\)STR-base and \\(0.83\\%\\) higher than MAERec. Please note that this improvement is automatic and training-free." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.676, + 0.47, + 0.782 + ], + "angle": 0, + "content": "Specifically, on the six traditional STR benchmarks (i.e., IIIT, SVT, IC13, IC15, SVTP, and CT80) which have nearly reached saturation in recent years[17], \\(\\mathrm{E}^2\\mathrm{STR}\\) still push the performance limit from \\(97.02\\%\\) to \\(97.74\\%\\), leading to a \\(24\\%\\) error rate decrease. On the 6 larger and harder STR benchmarks (i.e., COCO Text, CTW, TT, HOST, and WOST), \\(\\mathrm{E}^2\\mathrm{STR - ICL}\\) outperforms MAERec by \\(0.94\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.379, + 0.815 + ], + "angle": 0, + "content": "4.2.2 Results on Cross Domain Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.903 + ], + "angle": 0, + "content": "We compare with SOTA methods on cross domain benchmarks. Two novel scenarios are introduced: the industrial scenario (MPSC and EIST) and the handwriting scenario (IAM). In each dataset, only 100 training samples are provided. For \\(\\mathrm{E}^2\\)STR-ICL we simply use the training samples" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.334, + 0.895, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.503, + 0.893, + 0.56 + ], + "angle": 0, + "content": "Figure 5. Comparison with the fine-tuned models. We report the average performance on three cross-domain datasets. Please note that ABINet [10], SATRN [20] and MAERec [17] are fine-tuned with the in-domain data, while our \\(\\mathrm{E}^2\\)STR-ICL is training-free." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.593, + 0.891, + 0.623 + ], + "angle": 0, + "content": "as the in-context pool. We compare the training-free results in Table 2 and the fine-tuning results in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.762 + ], + "angle": 0, + "content": "As we can see, on both industrial and handwriting scenarios our \\(\\mathrm{E}^2\\)STR-ICL reaches SOTA performance. As shown in Table 2, under the training-free constraint \\(\\mathrm{E}^2\\)STR-ICL reaches an average performance of \\(78.17\\%\\), which is \\(4.69\\%\\) higher than \\(\\mathrm{E}^2\\)STR-base and \\(4.03\\%\\) higher than the SOTA method MAERec. Specifically, on EIST and IAM the application of ICL brings a huge improvement of \\(7.11\\%\\) and \\(4.59\\%\\), which demonstrates the extraordinary adaptation ability of \\(\\mathrm{E}^2\\)STR-ICL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We further compare the fine-tuned methods and our \\(\\mathrm{E}^2\\)STR-ICL. We fine-tune ABINet [10], SATRN [20] and MAERec [17] with the same data preserved in the in-context pool. As shown in Figure 5, \\(\\mathrm{E}^2\\)STR-ICL outperforms MAERec by \\(1.16\\%\\) even if the latter is fine-tuned with in-domain data, which is an exciting result given that \\(\\mathrm{E}^2\\)STR-ICL requires no parameter updating. In a word, our \\(\\mathrm{E}^2\\)STR can be rapidly implemented in a training-free manner in various novel scenarios and even achieves better per" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15572" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.169 + ], + "angle": 0, + "content": "
COCOHOSTWordArt
annotation rate10%20%10%20%10%20%
MAERec [17]000000
w/ fine-tuning0.821.671.031.721.342.23
E²STR-base000000
E²STR-ICL10.1212.9212.4313.7626.2232.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.179, + 0.47, + 0.248 + ], + "angle": 0, + "content": "Table 3. Results on hard case rectification. \"Hard Cases\" are test samples misrecognized by both MAERec [17] and our \\(\\mathrm{E}^2\\)STR-base. By providing annotation of a small part of the hard cases, we compare the performance increase in the rest test samples between the fine-tuned MAERec and our \\(\\mathrm{E}^2\\)STR-ICL." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.261, + 0.441, + 0.356 + ], + "angle": 0, + "content": "
Training TaskWord Accuracy
VTTSSSNon-ContextIn-Context
✓69.6926.82
✓✓69.8075.60
✓✓69.6673.09
✓✓✓69.6676.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.367, + 0.468, + 0.409 + ], + "angle": 0, + "content": "Table 4. Ablation of our proposed training strategies, where VT, TS, and SS refer to vanilla STR training, the Transform Strategy, and the Split Strategy. The experiment is performed on EIST." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.435, + 0.377, + 0.449 + ], + "angle": 0, + "content": "formance than the fine-tuned SOTA methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.468, + 0.374, + 0.482 + ], + "angle": 0, + "content": "4.2.3 Results on Hard Case Rectification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.468, + 0.749 + ], + "angle": 0, + "content": "We demonstrate the rectification ability of \\(\\mathrm{E}^2\\mathrm{STR}\\), which can handle hard cases in STR conveniently and effectively, in a training-free manner. Specifically, we define \"hard cases\" as the scene text samples that are wrongly recognized by both \\(\\mathrm{E}^2\\mathrm{STR}\\)-base and the SOTA method MAERec. A small number of hard cases are then annotated, and we study how the model can benefit from the annotated hard cases and decrease the error rate of the rest hard cases. Shown in Table 3, we perform experiments on COCO Text, HOST, and WordArt. As we can see, by annotating \\(10\\%\\) to \\(20\\%\\) of the hard cases, \\(\\mathrm{E}^2\\mathrm{STR}\\)-ICL decreases the error rate of the rest hard cases by up to \\(32\\%\\). This improvement is achieved by simply putting the annotated hard cases into the in-context pool, without any hassle of re-training the model. By comparison, by fine-tuning on the annotated hard cases, MAERec only decreases the error rate by up to \\(2.23\\%\\), completely incomparable to our \\(\\mathrm{E}^2\\mathrm{STR}\\)-ICL." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.241, + 0.771 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Impact of Split-and-Transform Training Strategies. We perform an experiment to show the effectiveness of our proposed Split Strategy and Transform Strategy. Shown in Table 4, the vanilla STR training brings a word accuracy of \\(69.69\\%\\), but the model cannot understand context information, and the performance even severely decreases to \\(26.82\\%\\) when provided with in-context prompts. The application of the Transform Strategy and the Split Strategy in" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.09, + 0.882, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.242, + 0.892, + 0.297 + ], + "angle": 0, + "content": "Figure 6. Comparison between different in-context prompt selection strategies. \"Random Selection\" refers to randomly selecting two samples as in-context prompts from the in-context pool. X-axis is the evaluated benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.446 + ], + "angle": 0, + "content": "the second training stage does not increase the non-context performance (concerning that the synthesized data is typically weaker than the real-world data used in the vanilla training stage), but the model learns to profit from context, and the performance is improved to \\(75.60\\%\\) and \\(73.09\\%\\) respectively when provided with in-context prompts. Finally, the hybrid of the above two strategies further enhances the ICL ability, and the performance reaches \\(76.77\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Impact of Nearest Neighbor In-Context Prompt Selection. We compare the kNN in-context prompt selection strategy to Random Selection, i.e., randomly selecting in-context prompts from the in-context pool. Shown in Figure 6, on all three evaluated datasets, random selection can improve the performance of non-context prediction by a small margin, but is far from comparing with kNN selection. Specifically, on EIST random selection improves the performance of non-context from \\(69.66\\%\\) to \\(70.65\\%\\), while kNN selection reaches \\(76.77\\%\\) word accuracy under the same condition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Impact of In-Context Pool Size. We next study the impact of varying the size of the in-context pool. Shown in Figure 7, we perform experiments on IAM, EIST, and MPSC, by varying the number of samples maintained in the in-context pool. As we can see, in general, the larger in-context pool brings about better performance, and a small number of samples is adequate to bring about huge performance improvement when deploying \\(\\mathrm{E}^2\\)STR-ICL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Impact of the Number of In-Context prompts. We analyze the influence of the number of in-context prompts. Shown in Table 5, the experiment is performed on HOST, ToTal Text, and WordArt. Similar to the in-context pool size, the increase in the number of in-context prompts also generally boosts the performance of \\(\\mathrm{E}^2\\)STR-ICL. Moreover, one to two in-context prompts are adequate for improving the performance by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Computational Complexity. We experimentally compare the inference speed of \\(\\mathrm{E}^2\\mathrm{STR}\\) and MAERec [17]. Shown in Table 6, the inference speed of \\(\\mathrm{E}^2\\mathrm{STR - ICL}\\) is on par" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "15573" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.089, + 0.091, + 0.466, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.259, + 0.473, + 0.302 + ], + "angle": 0, + "content": "Figure 7. The performance change brought by different sizes of the in-context pool. The X-axis is the size of the in-context pool and the Y-axis is the word accuracy results." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.313, + 0.456, + 0.37 + ], + "angle": 0, + "content": "
ICL prompts0124816
HOST73.3074.3474.7574.8374.8374.92
TT93.4694.6894.6894.7894.8894.91
WordArt81.4786.0486.1786.1786.2786.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.379, + 0.471, + 0.409 + ], + "angle": 0, + "content": "Table 5. The performance change brought by the different number of in-context prompts." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.42, + 0.456, + 0.452 + ], + "angle": 0, + "content": "
MAERec [17]E2STR-baseE2STR-ICL
Inference Time (s)0.0920.0710.094
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.461, + 0.47, + 0.49 + ], + "angle": 0, + "content": "Table 6. Comparison of the mean inference time of each test sample. All results are reported under the same hardware environment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.469, + 0.606 + ], + "angle": 0, + "content": "with MAERec. Compared to \\(\\mathrm{E}^2\\mathrm{STR}\\)-base, the in-context prompts of \\(\\mathrm{E}^2\\mathrm{STR}\\)-ICL bring extra consumption, but this leads to a limited inference time increase (i.e., from 0.071 to 0.094). It makes sense since we only maintain the visual tokens in the in-context pool and directly feed the visual tokens of the selected prompts to the language model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.47, + 0.833 + ], + "angle": 0, + "content": "Visualization and Further Analysis. We further study how the test sample learns from context. Shown in Figure 8, we select one context prompt for the test sample, and study the model pays attention to which region of the context image. This is achieved by collecting the attention maps between the language tokens and the image features. As we can see, when the language tokens pay close attention to the misrecognized image region, they also focus on the context image region which has similar patterns. For example, on the last row of Figure 8, \\(\\mathrm{E}^2\\mathrm{STR}\\) misrecognized the test image as \"simplest\" without context. By providing a context prompt \"Display\", one language token focuses on the \"la\" region of both images, which have similar image patterns. Finally, \\(\\mathrm{E}^2\\mathrm{STR}\\) rectified the misrecognized \"e\" to \"a\" with the help of context ground-truth \"la\" of the focused region." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.201, + 0.86 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "There are two limitations in our study. Firstly, there is a very slim chance that \\(\\mathrm{E}^2\\mathrm{STR}\\)-ICL erroneously rectifies pre" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.627, + 0.127 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.129, + 0.583, + 0.14 + ], + "angle": 0, + "content": "Vitoc" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.142, + 0.628, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.192, + 0.626, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.226, + 0.593, + 0.237 + ], + "angle": 0, + "content": "PORTED" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.242, + 0.628, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.28, + 0.593, + 0.291 + ], + "angle": 0, + "content": "simplest" + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.297, + 0.629, + 0.308 + ], + "angle": 0, + "content": "Non-Context Prediction" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.092, + 0.765, + 0.126 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.129, + 0.716, + 0.139 + ], + "angle": 0, + "content": "pal" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.141, + 0.765, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.176, + 0.725, + 0.188 + ], + "angle": 0, + "content": "integral" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.189, + 0.765, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.225, + 0.725, + 0.236 + ], + "angle": 0, + "content": "POWER" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.24, + 0.765, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.279, + 0.723, + 0.291 + ], + "angle": 0, + "content": "Display" + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.297, + 0.756, + 0.309 + ], + "angle": 0, + "content": "In-Context Prompt" + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.092, + 0.888, + 0.126 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.129, + 0.849, + 0.139 + ], + "angle": 0, + "content": "Vital" + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.141, + 0.889, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.177, + 0.864, + 0.188 + ], + "angle": 0, + "content": "accepting" + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.19, + 0.889, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.225, + 0.855, + 0.236 + ], + "angle": 0, + "content": "PORTER" + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.24, + 0.889, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.279, + 0.856, + 0.29 + ], + "angle": 0, + "content": "simplest" + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.297, + 0.888, + 0.308 + ], + "angle": 0, + "content": "In-Context Prediction" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.322, + 0.895, + 0.406 + ], + "angle": 0, + "content": "Figure 8. Cross attention visualization between the language tokens and the vision tokens. Left: Non-context prediction of \\(\\mathrm{E}^2\\mathrm{STR}\\). Error characters are marked in red. Right: In-context prediction of \\(\\mathrm{E}^2\\mathrm{STR - ICL}\\), where only one in-context prompt is selected. We visualize how the language tokens attend to the prompt image and the test image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.434, + 0.893, + 0.48 + ], + "angle": 0, + "content": "dictions due to misleading prompts (please refer to supplementary materials). Additionally, our model still could not recognize characters that are not included in the lexicon." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.496, + 0.619, + 0.512 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.749 + ], + "angle": 0, + "content": "In this paper, we propose \\(\\mathrm{E}^2\\mathrm{STR}\\), an ego-evolving scene text recognizer equipped with in-context learning capabilities. Through our proposed in-context training strategy incorporating context-rich scene text sequences, \\(\\mathrm{E}^2\\mathrm{STR}\\) performs rapid adaptation across diverse scenarios without additional fine-tuning. Extensive experiments demonstrate that \\(\\mathrm{E}^2\\mathrm{STR}\\) not only achieves SOTA performance on common STR benchmarks but also outperforms even the approaches that have been fine-tuned specifically for cross-domain scenarios. The model's ability to easily and effectively handle difficult text cases further underscores its potential as a unified text recognizer. Overall, this research represents a significant step toward efficient and highly adaptive text recognition models well-suited for diverse real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments This work is supported by National Natural Science Foundation of China No.62222602, No.62176092, No.62106075, No.U23A20343. Science and Technology Commission of Shanghai No.21511100700, Shanghai Sailing Program (23YF1410500), Natural Science Foundation of Shanghai (23ZR1420400), Natural Science Foundation of Chongqing, China (CSTB2023NSCQ-JQX0007), (CSTB2023NSCQ-MSX0137), Development Project of Ministry of Industry and Information Technology (ZTZB.23-990-016)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15574" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.284 + ], + "angle": 0, + "content": "[2] Jeonghun Baek, Geewook Kim, Junyeop Lee, Sungrae Park, Dongyoon Han, Sangdoo Yun, Seong Joon Oh, and Hwalsuk Lee. What is wrong with scene text recognition model comparisons? dataset and model analysis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4715-4723, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[3] Darwin Bautista and Rowel Atienza. Scene text recognition with permuted autoregressive sequence models. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.468, + 0.372 + ], + "angle": 0, + "content": "[4] Belval. Generator. https://github.com/Belval/TextRecognitionDataGenerator.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.468, + 0.443 + ], + "angle": 0, + "content": "[5] Fedor Borisyuk, Albert Gordo, and Viswanath Sivakumar. Rosetta: Large scale system for text detection and recognition in images. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 71-79, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.446, + 0.468, + 0.515 + ], + "angle": 0, + "content": "[6] 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, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.518, + 0.47, + 0.586 + ], + "angle": 0, + "content": "[7] Zhanzhan Cheng, Fan Bai, Yunlu Xu, Gang Zheng, Shiliang Pu, and Shuigeng Zhou. Focusing attention: Towards accurate text recognition in natural images. In Proceedings of the IEEE international conference on computer vision, pages 5076-5084, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[8] Chee Kheng Ch'ng and Chee Seng Chan. Total-text: A comprehensive dataset for scene text detection and recognition. In 2017 14th IAPR international conference on document analysis and recognition (ICDAR), pages 935-942. IEEE, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.661, + 0.468, + 0.743 + ], + "angle": 0, + "content": "[9] 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. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[10] Shancheng Fang, Hongtao Xie, Yuxin Wang, Zhendong Mao, and Yongdong Zhang. Read like humans: Autonomous, bidirectional and iterative language modeling for scene text recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7098-7107, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[11] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, pages 369-376, 2006. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[12] Tongkun Guan, Chaochen Gu, Changsheng Lu, Jingzheng Tu, Qi Feng, Kaijie Wu, and Xinping Guan. Industrial scene text detection with refined feature-attentive network. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):6073–6085, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[13] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[14] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. Icl-d3ie: In-context learning with diverse demonstrations updating for document information extraction. arXiv preprint arXiv:2303.05063, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[15] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. Icl-d3ie: In-context learning with diverse demonstrations updating for document information extraction. arXiv preprint arXiv:2303.05063, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[17] Qing Jiang, Jiapeng Wang, Dezhi Peng, Chongyu Liu, and Lianwen Jin. Revisiting scene text recognition: A data perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20543-20554, 2023. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[18] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazan Almazan, and Lluis Pere De Las Heras. Icedar 2013 robust reading competition. In 2013 12th international conference on document analysis and recognition, pages 1484-1493. IEEE, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.67 + ], + "angle": 0, + "content": "[19] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In 2015 13th international conference on document analysis and recognition (ICDAR), pages 1156-1160. IEEE, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.674, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[20] Junyeop Lee, Sungrae Park, Jeonghun Baek, Seong Joon Oh, Seonghyeon Kim, and Hwalsuk Lee. On recognizing texts of arbitrary shapes with 2d self-attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 546–547, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[21] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[22] Hui Li, Peng Wang, Chunhua Shen, and Guyu Zhang. Show, attend and read: A simple and strong baseline for irregular text recognition. In Proceedings of the AAAI conference on artificial intelligence, pages 8610-8617, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Minghui Liao, Jian Zhang, Zhaoyi Wan, Fengming Xie, Jia-jun Liang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Scene text recognition from two-dimensional perspective. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15575" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "ceedings of the AAAI conference on artificial intelligence, pages 8714-8721, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.164 + ], + "angle": 0, + "content": "[24] Wei Liu, Chaofeng Chen, Kwan-Yee K Wong, Zhizhong Su, and Junyu Han. Star-net: a spatial attention residue network for scene text recognition. In BMVC, page 7, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[25] Yuliang Liu, Lianwen Jin, Shuai tao Zhang, Canjie Luo, and Sheng Zhang. Curved scene text detection via transverse and longitudinal sequence connection. Pattern Recognition, 90: 337-345, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.469, + 0.277 + ], + "angle": 0, + "content": "[26] Mengkai Ma, Qiu-Feng Wang, Shan Huang, Shen Huang, Yannis Goulermas, and Kaizhu Huang. Residual attention-based multi-scale script identification in scene text images. Neurocomputing, 421:222-233, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[27] U-V Marti and Horst Bunke. The iam-database: an english sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition, 5: 39-46, 2002. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.39 + ], + "angle": 0, + "content": "[28] Qiang Mei, Qinyou Hu, Chun Yang, Hailin Zheng, and Zhisheng Hu. Port recommendation system for alternative container port destinations using a novel neural language-based algorithm. IEEE Access, 8:199970-199979, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[29] Anand Mishra, Karteek Alahari, and CV Jawahar. Top-down and bottom-up cues for scene text recognition. In 2012 IEEE conference on computer vision and pattern recognition, pages 2687-2694. IEEE, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[30] Imene OUALI, Mohamed BEN HALIMA, and WALI Ali. Augmented reality for scene text recognition, visualization and reading to assist visually impaired people. Procedia Computer Science, 207:158-167, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[31] Trung Quy Phan, Palaiahnakote Shivakumara, Shangxuan Tian, and Chew Lim Tan. Recognizing text with perspective distortion in natural scenes. In Proceedings of the IEEE international conference on computer vision, pages 569-576, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[32] Anhar Risnumawan, Palaiahankote Shivakumara, Chee Seng Chan, and Chew Lim Tan. A robust arbitrary text detection system for natural scene images. Expert Systems with Applications, 41(18):8027-8048, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.688 + ], + "angle": 0, + "content": "[33] Abdul Khader Jilani Saudagar and HabeebVulla Mohammad. Augmented reality mobile application for arabic text extraction, recognition and translation. Journal of Statistics and Management Systems, 21(4):617-629, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[34] Zhenwei Shao, Zhou Yu, Meng Wang, and Jun Yu. Prompting large language models with answer heuristics for knowledge-based visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14974-14983, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[35] Fenfen Sheng, Zhineng Chen, and Bo Xu. Ntrr: A no-recurrence sequence-to-sequence model for scene text recognition. In 2019 International conference on document analysis and recognition (ICDAR), pages 781-786. IEEE, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[36] Baoguang Shi, Mingkun Yang, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Aster: An attentional scene text recognizer with flexible rectification. IEEE transactions on pattern analysis and machine intelligence, 41(9):2035-2048, 2018. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[37] Cunzhao Shi, Chunheng Wang, Baihua Xiao, Song Gao, and Jinlong Hu. End-to-end scene text recognition using tree-structured models. Pattern Recognition, 47(9):2853-2866, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[38] 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. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[39] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie. Coco-text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[40] Zhaoyi Wan, Minghang He, Haoran Chen, Xiang Bai, and Cong Yao. Textscanner: Reading characters in order for robust scene text recognition. In Proceedings of the AAAI conference on artificial intelligence, pages 12120-12127, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[41] Yuxin Wang, Hongtao Xie, Shancheng Fang, Jing Wang, Shenggao Zhu, and Yongdong Zhang. From two to one: A new scene text recognizer with visual language modeling network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14194–14203, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.416, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[42] Xudong Xie, Ling Fu, Zhifei Zhang, Zhaowen Wang, and Xiang Bai. Toward understanding wordart: Corner-guided transformer for scene text recognition. In European Conference on Computer Vision, pages 303-321. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.488, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[43] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. An empirical study of gpt-3 for few-shot knowledge-based vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3081-3089, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.558, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[44] Deli Yu, Xuan Li, Chengquan Zhang, Tao Liu, Junyu Han, Jingtuo Liu, and Errui Ding. Towards accurate scene text recognition with semantic reasoning networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12113-12122, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.629, + 0.892, + 0.697 + ], + "angle": 0, + "content": "[45] Chongsheng Zhang, Yuefeng Tao, Kai Du, Weiping Ding, Bin Wang, Ji Liu, and Wei Wang. Character-level street view text spotting based on deep multisegmentation network for smarter autonomous driving. IEEE Transactions on Artificial Intelligence, 3(2):297-308, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.699, + 0.892, + 0.766 + ], + "angle": 0, + "content": "[46] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.769, + 0.892, + 0.837 + ], + "angle": 0, + "content": "[47] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.837 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15576" + } + ] +] \ No newline at end of file diff --git a/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/bf855ad8-e560-4eb0-9886-16b2ff2266f1_origin.pdf b/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/bf855ad8-e560-4eb0-9886-16b2ff2266f1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0e83b620e1e242f793f76bdc8a9eaca90a8152c1 --- /dev/null +++ b/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/bf855ad8-e560-4eb0-9886-16b2ff2266f1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14c537335ae9a2320ac5e459f782fcce4a27b260b094a39dd0cf8b02bee0f0be +size 610445 diff --git a/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/full.md b/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d513fc4567b4d07a7ce12ea15e0042192b54e38b --- /dev/null +++ b/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/full.md @@ -0,0 +1,353 @@ +# Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer + +Zhen Zhao $^{1,2,*}$ , Jingqun Tang $^{2,\dagger}$ , Chunhui Lin $^{2}$ , Binghong Wu $^{2}$ , Can Huang $^{2}$ , Hao Liu $^{2}$ , Xin Tan $^{1}$ , Zhizhong Zhang $^{1}$ , Yuan Xie $^{1,\dagger}$ $^{1}$ East China Normal University $^{2}$ ByteDance + +{51255901056}@stu.ecnu.edu.cn, {zzzhang,xtan,yxie}@cs.ecnu.edu.cn + +{tangjingqun,linchunhui,wubinghong,liuhao,can.huang}@bytedance.com + +# Abstract + +Scene text recognition (STR) in the wild frequently encounters challenges when coping with domain variations, font diversity, shape deformations, etc. A straightforward solution is performing model fine-tuning tailored to a specific scenario, but it is computationally intensive and requires multiple model copies for various scenarios. Recent studies indicate that large language models (LLMs) can learn from a few demonstration examples in a training-free manner, termed "In-Context Learning" (ICL). Nevertheless, applying LLMs as a text recognizer is unacceptably resource-consuming. Moreover, our pilot experiments on LLMs show that ICL fails in STR, mainly attributed to the insufficient incorporation of contextual information from diverse samples in the training stage. To this end, we introduce $E^2$ STR, a STR model trained with context-rich scene text sequences, where the sequences are generated via our proposed in-context training strategy. $E^2$ STR demonstrates that a regular-sized model is sufficient to achieve effective ICL capabilities in STR. Extensive experiments show that $E^2$ STR exhibits remarkable training-free adaptation in various scenarios and outperforms even the fine-tuned state-of-the-art approaches on public benchmarks. The code is released at https://github.com/bytedance/E2STR. + +# 1. Introduction + +Scene Text Recognition (STR) is a fundamental task in computer vision, with extensive applications in several domains such as autonomous driving [45], augmented reality [30, 33], industrial print recognition [28] and visual understanding [26]. + +Current progress in STR [3, 17, 20, 35] has demonstrated remarkable performance in numerous scenarios. + +However, as shown in Figure 1 (a), STR models are sup + +![](images/2bb54acc35d7fb57ff0359e0d3752e9e0be937b0180d2a18ac947d2f68e0fd1f.jpg) +(a) Diversified scenarios of scene text + +![](images/0332d3b9d72166f36a2f7dac608dcdec76fdcc86fa7080a9ce1debda48c46650.jpg) +(b) Scenario-specific fine-tuning pipeline of current methods + +![](images/933747a14fd1c940284ad7451e6b7d89a3af0d25cd0710451878dda39a6e3859.jpg) +(c) In-context learning pipeline of $\mathbf{E}^2\mathbf{STR}$ +Figure 1. Demonstration of real-world scene text scenarios and the adaptation pipeline. (a) Diversified scenarios of scene text in the real world. (b) The adaptation pipeline of current methods. They typically have to fine-tune upon a trained STR model with the training set, under a specific scenario. (c) The adaptation pipeline of our proposed $\mathrm{E}^2\mathrm{STR}$ . Our method automatically selects in-context prompts and performs training-free adaptation when faced with novel scenarios. Blue characters denote ambiguous scene text that is easily recognized. + +posed to perform robustly over diversified scenarios in the real world, where the scene text is hard to recognize because of domain variation, font diversity, shape deformation, etc. As shown in Figure 1 (b), a straightforward solution involves collecting the corresponding data and then fine-tuning the model for the specific scenario [3, 17, 20]. This process is computationally intensive and requires multiple model copies for diverse scenarios. + +The development of a comprehensive and reliable STR + +![](images/3dba2f422435df1238e2ef183a714acb1715d1b4ae84fd8b39037bde1982db78.jpg) +(b) Context-Rich scene text sequence + +![](images/cf0ef15b006408fe6b66e44f03c601ec822a7c646f5ed00a51c575f73ea054f7.jpg) +(c) ICL results on cross-domain benchmark +Figure 2. Our pilot experiments. (a) The randomly concatenated scene text sequence. (b) Our proposed context-rich scene text sequence. (c) By training an STR model based on the randomly concatenated scene text sequence, we evaluate the model on three cross-domain datasets. + +model that can effectively handle many real-world scenarios remains a significant challenge. + +Fortunately, plenty of studies [1, 6, 21, 38] have shown that Large Language Models (LLMs) can easily adapt without additional training. This adaptation is achieved by leveraging only a handful of input-label pairs as context (prompting information), a phenomenon known as "In-Context Learning" (ICL). The advantages of ICL inspire our interest in implementing it in STR, such that by fetching a few in-context prompts, a single model can be rapidly adapted to various scenarios without fine-tuning. + +However, the equipment of ICL in STR still poses challenges under the existing circumstances. Firstly, it is deemed excessively costly to apply Multi-Modal Large Language Models (M-LLMs) with billions of parameters as a scene text recognizer. And the ICL capabilities in regular-sized models have been barely explored currently. + +Secondly, it is hard to acquire ICL capabilities for a STR model with current training strategies. Previous studies have observed that sending image-text sequences for training would naturally endow ICL for M-LLMs [1, 21, 38], while such a phenomenon is hard to achieve in STR. As shown in Figure 2 (a), we generate sequential training data by randomly concatenating scene text samples. This practice fails as the trained model does not exhibit any performance improvement even when provided with in-domain prompts (Figure 2 (c)). The major cause of this failure is the lack of context in the generated scene text sequences during the training phase. The arbitrary concatenation of scene text does not provide any contextual information (i.e., sample connections) between different samples (Figure 2 (a)). Consequently, the model lacks the ability to effectively use information derived from in-context prompts (Figure 2 (c)), which implies that in-context training is essentially important for the effective implementation of ICL in STR. + +Based on the above analysis, we propose $\mathrm{E}^2\mathrm{STR}$ (Ego-Evolving STR), a paradigm that facilitates adaptation across diverse scenarios in a training-free manner. Specifically, we propose an in-context training strategy, which enables the model to exploit contextual information from the generated context-rich scene text sequences (Figure 2 + +(b)). The context-rich scene text sequences are formed using our ST-strategy, which involves random Splitting and Transformation of scene text, hence generating a set of "sub-samples". The sub-samples are inner-connected in terms of both visual and linguistic aspects. In the inference stage, $\mathrm{E}^2\mathrm{STR}$ fetches in-context prompts based on visual similarities, and utilizes the prompts to assist the recognition, shown in Figure 1 (c). In practice, it is found that with proper training and inference strategies, ICL capabilities can also be observed in regular-sized STR models (hundreds of millions of parameters). + +Finally, the proposed $\mathrm{E}^2\mathrm{STR}$ effectively captures contextual information from the in-context prompts and performs rapid adaptation in various novel scenarios in a training-free manner (Please refer to Section 4.2). On common benchmarks, $\mathrm{E}^2\mathrm{STR}$ achieves SOTA results, with an average improvement of $0.8\%$ over previous methods and $1.1\%$ over itself without ICL. Most importantly, when evaluated on unseen domains, $\mathrm{E}^2\mathrm{STR}$ achieves impressive performance with only a few prompts, even outperforming the fine-tuning results of SOTA methods by $1.2\%$ . Our contributions are summarized below: + +(1) We propose $\mathrm{E}^2\mathrm{STR}$ , a robust STR paradigm that can perform rapid adaptation over diverse scenarios in a training-free manner. +(2) We provide an in-context training strategy for equipping STR models with ICL capabilities, as well as an in-context inference strategy for STR models to leverage contextual information from in-context prompts. +(3) We demonstrate that ICL capabilities can be effectively incorporated into regular-sized STR models via appropriate training and inference strategies. +(4) Extensive experiments show that $\mathrm{E}^2\mathrm{STR}$ exceeds state-of-the-art performance across diverse benchmarks, even surpassing the fine-tuned approaches in unseen domains. + +# 2. Related Work + +# 2.1. Scene Text Recognition + +Recent years have witnessed extensive studies in STR, which can be generally divided into Language-free methods and Language-aware methods. + +Language-free STR. Language-free models directly utilize visual features for prediction, without considering the relationship between the characters. In this branch CTC-based [11] methods [5, 24] play the most prominent part. They typically consist of a CNN for feature extraction and an RNN for sequential feature processing, which are trained end-to-end with the CTC loss [11]. Other methods like [23, 40] focus on treating STR as a character-level segmentation task. The lack of linguistic information limits the application of language-free methods in scenarios with + +occluded or incomplete characters. + +Language-aware STR. Language-aware models leverage linguistic information to assist the recognition, typically utilizing an external language model (LM) [10, 44] or training internal LMs [2, 7, 36]. SRN [44] and ABINet [10] feed visual predictions to an external LM for linguistic refinement. The direct application of an external LM without considering visual features leads to possible erroneous correction. On the other hand, methods like PARSeq [3] and MAERec [17] implicitly train an internal LM in an auto-regressive manner, which have achieved decent performance. In this paper we base our model on the language-aware design, training a transformer-based language decoder inner-connected with the vision encoder. + +# 2.2. Multi-Modal In-Context Learning + +Recent large language models (LLMs) [6, 46] have demonstrated their excellent few-shot adaptation capabilities. By concatenating a few examples with the input as the prompt at reference time, LLMs quickly adapt to novel tasks without parameter updating. This phenomenon introduces a novel learning paradigm termed "In-Context Learning". Meanwhile, unlike LLMs, vision-language models (VLMs) struggle to understand complex multi-modal prompts [47]. A large set of approaches [13-15, 34] have been proposed to empower VLMs with multi-modal in-context learning (M-ICL) capabilities, but they typically utilize vision models (like image caption models) to translate images to text [15, 34, 43], or view the LLM as a scheduler learning to call vision experts based on a few examples [13]. These approaches do not truly establish a VLM with M-ICL capabilities. Recently, several work [1, 21, 38] proposes to train VLMs with sequential multi-modal data, and have achieved great success in prompting VLMs with multi-modal examples. In this paper, we aim to train a scene text recognizer equipped with M-ICL capabilities based on this sequential training paradigm. We demonstrate that the arbitrary concatenation of scene text fails as stated above, which motivates us to generate context-rich scene text sequences. + +# 3. Methodology + +# 3.1. Preliminary of Multi-Modal In-Context Learning + +Multi-modal in-context Learning enables M-LLMs to perform quick adaptation in downstream tasks in a training-free manner, hence eliminating the redundant computation and time expenses of fine-tuning. In this subsection, we introduce how to formulate multi-modal in-context learning for addressing the STR task. + +For a scene text tuple $(\mathbf{x},\mathbf{y})$ where $\pmb{x}$ is the scene image and $\pmb{y}$ is the ground-truth text, the STR task involves generating the label $\pmb{y}$ by maximizing the conditional probabil + +ity under the classic auto-regressive paradigm as follows: $p(\boldsymbol{y}|\boldsymbol{x}) = \prod_{l=1}^{L} p(\boldsymbol{y}_l|\boldsymbol{x}, \boldsymbol{y}_{$ . The training in this phase makes use of the next-token prediction loss: + +$$ +\mathcal {L} = \mathbb {E} _ {(\boldsymbol {x}, \boldsymbol {y}) \sim \mathcal {D}} \left[ - \sum_ {l = 1} ^ {L} \log p (\boldsymbol {y} _ {l} | \boldsymbol {y} _ {< l}, \boldsymbol {x}) \right], \tag {2} +$$ + +where $\mathcal{D}$ is the training set. + +# 3.3.2 In-Context Training + +The objective of the in-context training phase is to equip $\mathrm{E}^2\mathrm{STR}$ with the capability of In-Context Learning. As depicted in the top of Figure 3, the model is trained with context-rich scene text sequences as stated before. In these sequences, we interleave a placeholder $$ in the text for each image. This serves to make the language decoder distinguish between different samples following [1]. In this stage, we propose two strategies to generate context-rich scene text sequences: the Split Strategy and the Transform Strategy (the ST strategy). + +The Split Strategy. As shown in Figure 4 (a), when presented with a training tuple $(\pmb{x},\pmb{y})$ , we split the sample and hence generating a set of "sub-samples". It is evident that the sub-samples exhibit a strong connection to the original training sample. Furthermore, the sub-samples themselves demonstrate interconnectivity as they overlap with one another. Next, we proceed to concatenate the sub-samples with $(\pmb{x},\pmb{y})$ and additional randomly selected samples to form a context-rich sample sequence. + +We randomly shuffle the whole sequence before generating the actual input text (i.e., interleaving the $\langle \mathbf{\Pi} / \mathbf{i}\rangle$ token to the text sequence). + +In practice, to accurately split the training samples, we + +synthesize 600k scene text images based on [4] and record the accurate bounding boxes of every single character. Our subsequent experiments show that the synthesized data does not change $\mathrm{E}^2\mathrm{STR}$ 's non-context text recognition ability, but the Split Strategy based on them equips $\mathrm{E}^2\mathrm{STR}$ with a strong capability of in-context learning. + +The Transform Strategy. As shown in Figure 4 (b), given a training tuple $(\boldsymbol{x}, \boldsymbol{y})$ (whether with character-wise bounding boxes or not), we perform data augmentation (a set of image transformations, e.g., color/direction transformations) on $\boldsymbol{x}$ . In this way, we also generate a set of sub-samples with the same label but different image patterns from the original sample. + +In practice, as depicted in Figure 4 (c), we hybrid the above strategies. The training set is formed by concatenating the synthesized data and the original training data used in the first training phase. For the synthesized data with character-wise bounding boxes, both the Split Strategy and the Transform Strategy are utilized. For the original training data, only the Transform Strategy is implemented. + +Finally, after generating the sample sequence $(X,Y)$ , where $X$ is the image sequence and $Y$ is the text sequence, $X$ is fed into the vision encoder, while $Y$ is processed by the language decoder under the auto-regressive framework. The loss function is formulated as: + +$$ +\mathcal {L} _ {\left(\boldsymbol {X}, \boldsymbol {Y}\right)} = - \sum_ {l = 1} ^ {L} \log p \left(\boldsymbol {Y} _ {l} \mid \boldsymbol {Y} _ {< l}, \boldsymbol {X} _ {\leq l}\right), \tag {3} +$$ + +where $\mathbf{X}_{\leq l}$ is the set of image tokens preceding token $\mathbf{Y}_l$ in the input sequence. + +# 3.4. In-Context Inference + +The In-Context Learning ability is acquired by our $\mathrm{E}^2\mathrm{STR}$ model through the above two-stage training approach. AS shown in the bottom of Figure 3, when presented with a test image $\pmb{x}$ , the framework selects $N$ samples $\{(x_i^c,y_i^c)\}_{i = 1}^N$ from a in-context pool $\mathcal{D}^c$ . The selected samples have the highest visual similarities to $\pmb{x}$ in the latent space. Specifically, we calculate the image embedding $\pmb{I}$ of $\pmb{x}$ by averaging the visual token sequence Encoder $(\pmb{x})$ . The in-context prompts are then formed by choosing $N$ samples from $\mathcal{D}^c$ , where the image embeddings of these samples have the top-N highest cosine similarity with $\pmb{I}$ , i.e., + +$$ +\mathcal {I} = \underset {i \in 1, 2, \dots , | \mathcal {D} ^ {c} |} {\arg \operatorname {T o p N}} \frac {\boldsymbol {I} ^ {\mathrm {T}} \boldsymbol {I} _ {i} ^ {c}}{\| \boldsymbol {I} \| _ {2} \| \boldsymbol {I} _ {i} ^ {c} \| _ {2}}, \tag {4} +$$ + +where $\mathcal{I}$ is the index set of the top-N similar samples in $\mathcal{D}^c$ and $I_{i}^{c}$ is the image embedding of the $i$ -th sample in $\mathcal{D}^c$ . The in-context prompts are then defined as: + +$$ +\boldsymbol {E} = \left\{\left(\boldsymbol {x} _ {i} ^ {c}, \boldsymbol {y} _ {i} ^ {c}\right) | i \in \mathcal {I} \right\}. \tag {5} +$$ + +As shown in the bottom of Figure 3, $\pmb{E}$ is concatenated with the test sample $\pmb{x}$ and our in-context prediction is formulated as $p(\pmb{y}|\pmb{E},\pmb{x})$ . In practice, the in-context pool $\mathcal{D}^c$ retains solely the output tokens generated by the vision encoder, resulting in a highly efficient selection process. Furthermore, because the in-context pool is tiny and we do straight inference without training, the extra consumption is kept to a minimum (Please refer to Section 4.3). + +# 4. Experiment + +# 4.1. Experimental Setup + +Implementation Details. Following MAERec [17], we choose Vision Transformer [9] pre-trained under the MAE [16] framework as the vision encoder. The default language decoder is set as OPT-125M [46]. We use the cosine learning rate scheduler without warm-up and the AdamW optimizer with a weight decay of 0.01. We train our model for 10 epochs with an initial learning rate of 1e-4 during the first training stage, and 5 epochs with an initial learning rate of 5e-6 during the second in-context training stage. The training batch size is 64 for the first stage and 8 for the second stage. During inference for $\mathrm{E}^2\mathrm{STR}$ -ICL, we select two in-context prompts based on the kNN selection strategy. + +Datasets and Metrics. We use the real-world training dataset Union14M-L[17] for the two-stage training. The same training dataset (including the synthesized data) is adopted for all compared methods. $\mathrm{E}^2\mathrm{STR}$ is evaluated under various publicly available benchmarks, including Regular Benchmarks IIIT5k [29], SVT [37], IC13 [18], Irregular Benchmarks IC15 [19], SVTP [31], CUTE80 (CT80) [32], COCO Text (COCO) [39], CTW [25], Total Text (TT) [8], Occluded Benchmarks OST (HOST and WOST) [41] and artistic benchmark WordArt [42]. In cross domain scenarios the evaluated datasets including the metal-surface benchmark MPSC [12] and the handwriting benchmark IAM [27], as well as a more difficult real-world industrial text recognition dataset EIST (Enhanced Industrial Scene Text) collected by us. EIST is collected from the real-world industrial scenario, which contains 200 training samples and 8000 test samples. We use Word Accuracy[17] as the evaluation metric for all compared methods. + +# 4.2. Main Results + +# 4.2.1 Results on Common Benchmarks + +Table 1 presents the performance of $\mathrm{E}^2\mathrm{STR}$ on common benchmarks. We evaluate $\mathrm{E}^2\mathrm{STR}$ on 12 commonly used STR benchmarks and compare with SOTA methods. $\mathrm{E}^2\mathrm{STR}$ -base refers to non-context prediction without prompts. For $\mathrm{E}^2\mathrm{STR}$ -ICL, a tiny in-context pool is maintained by randomly sampling 1000 images from the training data (less than $0.025\%$ of the number of training samples). As we can see, $\mathrm{E}^2\mathrm{STR}$ -base achieves $90.25\%$ average word + +
MethodVenueRegularIrregularOccludedOthersAVG
IIIT3000SVT647IC131015IC152077SVTP645CT80288COCO9896CTW1572TT2201HOST2416WOST2416WordArt1511
ASTER [36]PAMI'1895.0389.4993.7985.4882.0290.2862.2576.5378.6943.3464.6565.5977.26
NRTR [35]ICDAR'1997.4393.8296.0685.1584.0391.3265.9481.7481.8350.8371.5264.0680.31
SAR [22]AAAI'1997.7094.1396.3587.4787.6093.0667.4183.9186.2346.3670.3272.4081.91
SATRN [20]AAAI'2097.8395.8397.4489.4690.8596.1873.0684.6187.9156.7175.6275.7185.10
ABINet [10]CVPR'2197.9095.9896.1691.6690.2393.7571.4683.7286.0156.5475.7575.2584.53
PARSeq* [3]ECCV'2299.1097.8498.1389.2296.9098.61-------
MAERec [17]ICCV'2398.9397.9998.6293.0494.5798.9678.8488.8793.9173.9785.7282.5990.50
\( E^2STR \)-base99.1098.1598.0392.9996.4398.9677.2988.3693.4673.3085.5181.4790.25
\( E^2STR-ICL \)99.2398.6198.7293.8296.7499.3178.3888.9994.6874.7586.5986.1791.33
+ +Table 1. Results on common benchmarks. All methods are trained on the same dataset except for PARSeq.*: PARSeq is trained on its self-collected real-world dataset and we directly quote the results from its original paper. Red and blue values denote the best and the secondary performance. ${\mathrm{E}}^{2}$ STR-base refers to non-context inference. + +
MethodIndustrialHandwritingAVG
MPSC 2941EIST 8000IAM 3000
ASTER [36]63.4848.7652.5054.91
NRTR [35]73.2461.7759.5364.85
SAR [22]73.8558.2656.6362.91
ABINet [10]75.3562.8561.5766.59
SATRN [20]76.1065.4259.4767.00
MAERec [17]81.8170.3370.2774.14
E²STR-base81.2669.6669.5173.48
E²STR-ICL83.6476.7774.1078.17
+ +Table 2. Results on cross domain scenarios. Three datasets under two unseen domains are evaluated. All approaches are evaluated in a training-free manner. + +accuracy over 12 datasets, $0.25\%$ lower than MAERec [17]. However, by fetching in-context prompts and exploiting incontext information, $\mathrm{E}^2$ STR-ICL achieves an average word accuracy of $91.33\%$ , which is $1.08\%$ higher than $\mathrm{E}^2$ STR-base and $0.83\%$ higher than MAERec. Please note that this improvement is automatic and training-free. + +Specifically, on the six traditional STR benchmarks (i.e., IIIT, SVT, IC13, IC15, SVTP, and CT80) which have nearly reached saturation in recent years[17], $\mathrm{E}^2\mathrm{STR}$ still push the performance limit from $97.02\%$ to $97.74\%$ , leading to a $24\%$ error rate decrease. On the 6 larger and harder STR benchmarks (i.e., COCO Text, CTW, TT, HOST, and WOST), $\mathrm{E}^2\mathrm{STR - ICL}$ outperforms MAERec by $0.94\%$ . + +# 4.2.2 Results on Cross Domain Scenarios + +We compare with SOTA methods on cross domain benchmarks. Two novel scenarios are introduced: the industrial scenario (MPSC and EIST) and the handwriting scenario (IAM). In each dataset, only 100 training samples are provided. For $\mathrm{E}^2$ STR-ICL we simply use the training samples + +![](images/01bd4147d21069695247f3d54bd9690b436b9a72d6043d12c65248ee53eaa125.jpg) +Figure 5. Comparison with the fine-tuned models. We report the average performance on three cross-domain datasets. Please note that ABINet [10], SATRN [20] and MAERec [17] are fine-tuned with the in-domain data, while our $\mathrm{E}^2$ STR-ICL is training-free. + +as the in-context pool. We compare the training-free results in Table 2 and the fine-tuning results in Figure 5. + +As we can see, on both industrial and handwriting scenarios our $\mathrm{E}^2$ STR-ICL reaches SOTA performance. As shown in Table 2, under the training-free constraint $\mathrm{E}^2$ STR-ICL reaches an average performance of $78.17\%$ , which is $4.69\%$ higher than $\mathrm{E}^2$ STR-base and $4.03\%$ higher than the SOTA method MAERec. Specifically, on EIST and IAM the application of ICL brings a huge improvement of $7.11\%$ and $4.59\%$ , which demonstrates the extraordinary adaptation ability of $\mathrm{E}^2$ STR-ICL. + +We further compare the fine-tuned methods and our $\mathrm{E}^2$ STR-ICL. We fine-tune ABINet [10], SATRN [20] and MAERec [17] with the same data preserved in the in-context pool. As shown in Figure 5, $\mathrm{E}^2$ STR-ICL outperforms MAERec by $1.16\%$ even if the latter is fine-tuned with in-domain data, which is an exciting result given that $\mathrm{E}^2$ STR-ICL requires no parameter updating. In a word, our $\mathrm{E}^2$ STR can be rapidly implemented in a training-free manner in various novel scenarios and even achieves better per + +
COCOHOSTWordArt
annotation rate10%20%10%20%10%20%
MAERec [17]000000
w/ fine-tuning0.821.671.031.721.342.23
E²STR-base000000
E²STR-ICL10.1212.9212.4313.7626.2232.02
+ +Table 3. Results on hard case rectification. "Hard Cases" are test samples misrecognized by both MAERec [17] and our $\mathrm{E}^2$ STR-base. By providing annotation of a small part of the hard cases, we compare the performance increase in the rest test samples between the fine-tuned MAERec and our $\mathrm{E}^2$ STR-ICL. + +
Training TaskWord Accuracy
VTTSSSNon-ContextIn-Context
✓69.6926.82
✓✓69.8075.60
✓✓69.6673.09
✓✓✓69.6676.77
+ +Table 4. Ablation of our proposed training strategies, where VT, TS, and SS refer to vanilla STR training, the Transform Strategy, and the Split Strategy. The experiment is performed on EIST. + +formance than the fine-tuned SOTA methods. + +# 4.2.3 Results on Hard Case Rectification + +We demonstrate the rectification ability of $\mathrm{E}^2\mathrm{STR}$ , which can handle hard cases in STR conveniently and effectively, in a training-free manner. Specifically, we define "hard cases" as the scene text samples that are wrongly recognized by both $\mathrm{E}^2\mathrm{STR}$ -base and the SOTA method MAERec. A small number of hard cases are then annotated, and we study how the model can benefit from the annotated hard cases and decrease the error rate of the rest hard cases. Shown in Table 3, we perform experiments on COCO Text, HOST, and WordArt. As we can see, by annotating $10\%$ to $20\%$ of the hard cases, $\mathrm{E}^2\mathrm{STR}$ -ICL decreases the error rate of the rest hard cases by up to $32\%$ . This improvement is achieved by simply putting the annotated hard cases into the in-context pool, without any hassle of re-training the model. By comparison, by fine-tuning on the annotated hard cases, MAERec only decreases the error rate by up to $2.23\%$ , completely incomparable to our $\mathrm{E}^2\mathrm{STR}$ -ICL. + +# 4.3. Ablation Studies + +Impact of Split-and-Transform Training Strategies. We perform an experiment to show the effectiveness of our proposed Split Strategy and Transform Strategy. Shown in Table 4, the vanilla STR training brings a word accuracy of $69.69\%$ , but the model cannot understand context information, and the performance even severely decreases to $26.82\%$ when provided with in-context prompts. The application of the Transform Strategy and the Split Strategy in + +![](images/6781f97025bd314b694923578d0f14fad12bdc7a6a8c7bae8cfcc1a141fead86.jpg) +Figure 6. Comparison between different in-context prompt selection strategies. "Random Selection" refers to randomly selecting two samples as in-context prompts from the in-context pool. X-axis is the evaluated benchmarks. + +the second training stage does not increase the non-context performance (concerning that the synthesized data is typically weaker than the real-world data used in the vanilla training stage), but the model learns to profit from context, and the performance is improved to $75.60\%$ and $73.09\%$ respectively when provided with in-context prompts. Finally, the hybrid of the above two strategies further enhances the ICL ability, and the performance reaches $76.77\%$ . + +Impact of Nearest Neighbor In-Context Prompt Selection. We compare the kNN in-context prompt selection strategy to Random Selection, i.e., randomly selecting in-context prompts from the in-context pool. Shown in Figure 6, on all three evaluated datasets, random selection can improve the performance of non-context prediction by a small margin, but is far from comparing with kNN selection. Specifically, on EIST random selection improves the performance of non-context from $69.66\%$ to $70.65\%$ , while kNN selection reaches $76.77\%$ word accuracy under the same condition. + +Impact of In-Context Pool Size. We next study the impact of varying the size of the in-context pool. Shown in Figure 7, we perform experiments on IAM, EIST, and MPSC, by varying the number of samples maintained in the in-context pool. As we can see, in general, the larger in-context pool brings about better performance, and a small number of samples is adequate to bring about huge performance improvement when deploying $\mathrm{E}^2$ STR-ICL. + +Impact of the Number of In-Context prompts. We analyze the influence of the number of in-context prompts. Shown in Table 5, the experiment is performed on HOST, ToTal Text, and WordArt. Similar to the in-context pool size, the increase in the number of in-context prompts also generally boosts the performance of $\mathrm{E}^2$ STR-ICL. Moreover, one to two in-context prompts are adequate for improving the performance by a large margin. + +Computational Complexity. We experimentally compare the inference speed of $\mathrm{E}^2\mathrm{STR}$ and MAERec [17]. Shown in Table 6, the inference speed of $\mathrm{E}^2\mathrm{STR - ICL}$ is on par + +![](images/93d6f6e5a20b02efc97be0831c144ec55ee061278170774389b547806be59f2d.jpg) +Figure 7. The performance change brought by different sizes of the in-context pool. The X-axis is the size of the in-context pool and the Y-axis is the word accuracy results. + +
ICL prompts0124816
HOST73.3074.3474.7574.8374.8374.92
TT93.4694.6894.6894.7894.8894.91
WordArt81.4786.0486.1786.1786.2786.35
+ +Table 5. The performance change brought by the different number of in-context prompts. + +
MAERec [17]E2STR-baseE2STR-ICL
Inference Time (s)0.0920.0710.094
+ +Table 6. Comparison of the mean inference time of each test sample. All results are reported under the same hardware environment. + +with MAERec. Compared to $\mathrm{E}^2\mathrm{STR}$ -base, the in-context prompts of $\mathrm{E}^2\mathrm{STR}$ -ICL bring extra consumption, but this leads to a limited inference time increase (i.e., from 0.071 to 0.094). It makes sense since we only maintain the visual tokens in the in-context pool and directly feed the visual tokens of the selected prompts to the language model. + +Visualization and Further Analysis. We further study how the test sample learns from context. Shown in Figure 8, we select one context prompt for the test sample, and study the model pays attention to which region of the context image. This is achieved by collecting the attention maps between the language tokens and the image features. As we can see, when the language tokens pay close attention to the misrecognized image region, they also focus on the context image region which has similar patterns. For example, on the last row of Figure 8, $\mathrm{E}^2\mathrm{STR}$ misrecognized the test image as "simplest" without context. By providing a context prompt "Display", one language token focuses on the "la" region of both images, which have similar image patterns. Finally, $\mathrm{E}^2\mathrm{STR}$ rectified the misrecognized "e" to "a" with the help of context ground-truth "la" of the focused region. + +# 5. Limitations + +There are two limitations in our study. Firstly, there is a very slim chance that $\mathrm{E}^2\mathrm{STR}$ -ICL erroneously rectifies pre + +![](images/eb73f529238dbad11fb19e80a7ad099dda95c56a0f91d84952d8e3cfa36f5efd.jpg) +Vitoc + +![](images/1d9d82612dc2e15073ed5c84b0a21b97e6a17710f7940946e18c37129fb604d9.jpg) + +![](images/647e6b809f51ed9fdeaa153f5ab50dd8d614ee65af2418719ce8c386618b4c5a.jpg) +PORTED + +![](images/d87c88179af81fa73d266bcbb6bbe3b83d39ed9edf7c45d127984831206842ea.jpg) +simplest +Non-Context Prediction +Figure 8. Cross attention visualization between the language tokens and the vision tokens. Left: Non-context prediction of $\mathrm{E}^2\mathrm{STR}$ . Error characters are marked in red. Right: In-context prediction of $\mathrm{E}^2\mathrm{STR - ICL}$ , where only one in-context prompt is selected. We visualize how the language tokens attend to the prompt image and the test image. + +![](images/21c36fb8edbf9243aa4fc66bd3801cbd202c7520288d34ddb67df5bef3b224bc.jpg) + +![](images/f192a34e78674eda264ef87513a81c6c5eec554e717784f3096c073d12b57fe4.jpg) +pal + +![](images/5aa92e626be07ba7ea68daaa0da7fa03ab4d7b97ac6793aaf4bca22656e40fbb.jpg) +integral +POWER + +![](images/05e2c5bdd61e9db6780bbb6ae809a78fb703830a582f7813493c3c816fbcad7e.jpg) +Display +In-Context Prompt + +![](images/6d2cee8497d2796ae8bf38b286327a2d9027d1cacc034f23d77ff1d545691feb.jpg) + +![](images/d7eec6f5e561f97bc02e9350d0abd870baed38019a7f39d5560abd91aba403d4.jpg) +Vital + +![](images/b71a7df6b4ff1935ea8ad117e3848d53a00ea68b7ef0eec185362c04798717bb.jpg) +accepting +PORTER + +![](images/848adbcc35db8b1f350e1345d21d0b69074e306452934c3907f3608ffbd04d42.jpg) +simplest +In-Context Prediction + +dictions due to misleading prompts (please refer to supplementary materials). Additionally, our model still could not recognize characters that are not included in the lexicon. + +# 6. Conclusion + +In this paper, we propose $\mathrm{E}^2\mathrm{STR}$ , an ego-evolving scene text recognizer equipped with in-context learning capabilities. Through our proposed in-context training strategy incorporating context-rich scene text sequences, $\mathrm{E}^2\mathrm{STR}$ performs rapid adaptation across diverse scenarios without additional fine-tuning. Extensive experiments demonstrate that $\mathrm{E}^2\mathrm{STR}$ not only achieves SOTA performance on common STR benchmarks but also outperforms even the approaches that have been fine-tuned specifically for cross-domain scenarios. The model's ability to easily and effectively handle difficult text cases further underscores its potential as a unified text recognizer. Overall, this research represents a significant step toward efficient and highly adaptive text recognition models well-suited for diverse real-world applications. + +Acknowledgments This work is supported by National Natural Science Foundation of China No.62222602, No.62176092, No.62106075, No.U23A20343. Science and Technology Commission of Shanghai No.21511100700, Shanghai Sailing Program (23YF1410500), Natural Science Foundation of Shanghai (23ZR1420400), Natural Science Foundation of Chongqing, China (CSTB2023NSCQ-JQX0007), (CSTB2023NSCQ-MSX0137), Development Project of Ministry of Industry and Information Technology (ZTZB.23-990-016). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2, 3, 4 +[2] Jeonghun Baek, Geewook Kim, Junyeop Lee, Sungrae Park, Dongyoon Han, Sangdoo Yun, Seong Joon Oh, and Hwalsuk Lee. What is wrong with scene text recognition model comparisons? dataset and model analysis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4715-4723, 2019. 3 +[3] Darwin Bautista and Rowel Atienza. Scene text recognition with permuted autoregressive sequence models. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 1, 3, 6 +[4] Belval. Generator. https://github.com/Belval/TextRecognitionDataGenerator.5 +[5] Fedor Borisyuk, Albert Gordo, and Viswanath Sivakumar. Rosetta: Large scale system for text detection and recognition in images. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 71-79, 2018. 2 +[6] 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, 3 +[7] Zhanzhan Cheng, Fan Bai, Yunlu Xu, Gang Zheng, Shiliang Pu, and Shuigeng Zhou. Focusing attention: Towards accurate text recognition in natural images. In Proceedings of the IEEE international conference on computer vision, pages 5076-5084, 2017. 3 +[8] Chee Kheng Ch'ng and Chee Seng Chan. Total-text: A comprehensive dataset for scene text detection and recognition. In 2017 14th IAPR international conference on document analysis and recognition (ICDAR), pages 935-942. IEEE, 2017. 5 +[9] 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. 5 +[10] Shancheng Fang, Hongtao Xie, Yuxin Wang, Zhendong Mao, and Yongdong Zhang. Read like humans: Autonomous, bidirectional and iterative language modeling for scene text recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7098-7107, 2021. 3, 6 +[11] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, pages 369-376, 2006. 2 + +[12] Tongkun Guan, Chaochen Gu, Changsheng Lu, Jingzheng Tu, Qi Feng, Kaijie Wu, and Xinping Guan. Industrial scene text detection with refined feature-attentive network. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):6073–6085, 2022. 5 +[13] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 3 +[14] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. Icl-d3ie: In-context learning with diverse demonstrations updating for document information extraction. arXiv preprint arXiv:2303.05063, 2023. +[15] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. Icl-d3ie: In-context learning with diverse demonstrations updating for document information extraction. arXiv preprint arXiv:2303.05063, 2023. 3 +[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5 +[17] Qing Jiang, Jiapeng Wang, Dezhi Peng, Chongyu Liu, and Lianwen Jin. Revisiting scene text recognition: A data perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20543-20554, 2023. 1, 3, 5, 6, 7, 8 +[18] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazan Almazan, and Lluis Pere De Las Heras. Icedar 2013 robust reading competition. In 2013 12th international conference on document analysis and recognition, pages 1484-1493. IEEE, 2013. 5 +[19] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In 2015 13th international conference on document analysis and recognition (ICDAR), pages 1156-1160. IEEE, 2015. 5 +[20] Junyeop Lee, Sungrae Park, Jeonghun Baek, Seong Joon Oh, Seonghyeon Kim, and Hwalsuk Lee. On recognizing texts of arbitrary shapes with 2d self-attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 546–547, 2020. 1, 3, 6 +[21] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2, 3 +[22] Hui Li, Peng Wang, Chunhua Shen, and Guyu Zhang. Show, attend and read: A simple and strong baseline for irregular text recognition. In Proceedings of the AAAI conference on artificial intelligence, pages 8610-8617, 2019. 6 +[23] Minghui Liao, Jian Zhang, Zhaoyi Wan, Fengming Xie, Jia-jun Liang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Scene text recognition from two-dimensional perspective. In Pro + +ceedings of the AAAI conference on artificial intelligence, pages 8714-8721, 2019. 2 +[24] Wei Liu, Chaofeng Chen, Kwan-Yee K Wong, Zhizhong Su, and Junyu Han. Star-net: a spatial attention residue network for scene text recognition. In BMVC, page 7, 2016. 2 +[25] Yuliang Liu, Lianwen Jin, Shuai tao Zhang, Canjie Luo, and Sheng Zhang. Curved scene text detection via transverse and longitudinal sequence connection. Pattern Recognition, 90: 337-345, 2019. 5 +[26] Mengkai Ma, Qiu-Feng Wang, Shan Huang, Shen Huang, Yannis Goulermas, and Kaizhu Huang. Residual attention-based multi-scale script identification in scene text images. Neurocomputing, 421:222-233, 2021. 1 +[27] U-V Marti and Horst Bunke. The iam-database: an english sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition, 5: 39-46, 2002. 5 +[28] Qiang Mei, Qinyou Hu, Chun Yang, Hailin Zheng, and Zhisheng Hu. Port recommendation system for alternative container port destinations using a novel neural language-based algorithm. IEEE Access, 8:199970-199979, 2020. 1 +[29] Anand Mishra, Karteek Alahari, and CV Jawahar. Top-down and bottom-up cues for scene text recognition. In 2012 IEEE conference on computer vision and pattern recognition, pages 2687-2694. IEEE, 2012. 5 +[30] Imene OUALI, Mohamed BEN HALIMA, and WALI Ali. Augmented reality for scene text recognition, visualization and reading to assist visually impaired people. Procedia Computer Science, 207:158-167, 2022. 1 +[31] Trung Quy Phan, Palaiahnakote Shivakumara, Shangxuan Tian, and Chew Lim Tan. Recognizing text with perspective distortion in natural scenes. In Proceedings of the IEEE international conference on computer vision, pages 569-576, 2013. 5 +[32] Anhar Risnumawan, Palaiahankote Shivakumara, Chee Seng Chan, and Chew Lim Tan. A robust arbitrary text detection system for natural scene images. Expert Systems with Applications, 41(18):8027-8048, 2014. 5 +[33] Abdul Khader Jilani Saudagar and HabeebVulla Mohammad. Augmented reality mobile application for arabic text extraction, recognition and translation. Journal of Statistics and Management Systems, 21(4):617-629, 2018. 1 +[34] Zhenwei Shao, Zhou Yu, Meng Wang, and Jun Yu. Prompting large language models with answer heuristics for knowledge-based visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14974-14983, 2023. 3 +[35] Fenfen Sheng, Zhineng Chen, and Bo Xu. Ntrr: A no-recurrence sequence-to-sequence model for scene text recognition. In 2019 International conference on document analysis and recognition (ICDAR), pages 781-786. IEEE, 2019. 1, 6 +[36] Baoguang Shi, Mingkun Yang, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Aster: An attentional scene text recognizer with flexible rectification. IEEE transactions on pattern analysis and machine intelligence, 41(9):2035-2048, 2018. 3, 6 + +[37] Cunzhao Shi, Chunheng Wang, Baihua Xiao, Song Gao, and Jinlong Hu. End-to-end scene text recognition using tree-structured models. Pattern Recognition, 47(9):2853-2866, 2014. 5 +[38] 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. 2, 3 +[39] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie. Coco-text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016. 5 +[40] Zhaoyi Wan, Minghang He, Haoran Chen, Xiang Bai, and Cong Yao. Textscanner: Reading characters in order for robust scene text recognition. In Proceedings of the AAAI conference on artificial intelligence, pages 12120-12127, 2020. 2 +[41] Yuxin Wang, Hongtao Xie, Shancheng Fang, Jing Wang, Shenggao Zhu, and Yongdong Zhang. From two to one: A new scene text recognizer with visual language modeling network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14194–14203, 2021. 5 +[42] Xudong Xie, Ling Fu, Zhifei Zhang, Zhaowen Wang, and Xiang Bai. Toward understanding wordart: Corner-guided transformer for scene text recognition. In European Conference on Computer Vision, pages 303-321. Springer, 2022. 5 +[43] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. An empirical study of gpt-3 for few-shot knowledge-based vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3081-3089, 2022. 3 +[44] Deli Yu, Xuan Li, Chengquan Zhang, Tao Liu, Junyu Han, Jingtuo Liu, and Errui Ding. Towards accurate scene text recognition with semantic reasoning networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12113-12122, 2020. 3 +[45] Chongsheng Zhang, Yuefeng Tao, Kai Du, Weiping Ding, Bin Wang, Ji Liu, and Wei Wang. Character-level street view text spotting based on deep multisegmentation network for smarter autonomous driving. IEEE Transactions on Artificial Intelligence, 3(2):297-308, 2021. 1 +[46] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 3, 5 +[47] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 3 \ No newline at end of file diff --git a/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/images.zip b/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b8a96c8e07ab18fbb49ad53c79ae44ea880e45be --- /dev/null +++ b/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:875feedc9909984eebc52c59b2e869fcd47f96a6d9b063e5d84b6e53177a9c48 +size 654317 diff --git a/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/layout.json b/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4e05982b141a8cbc80d269e19036ce1d2692abed --- /dev/null +++ b/2024/Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer/layout.json @@ -0,0 +1,11172 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 547, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 547, + 122 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 547, + 122 + ], + "type": "text", + "content": "Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "spans": [ + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "text", + "content": "Zhen Zhao" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "text", + "content": ", Jingqun Tang" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "text", + "content": ", Chunhui Lin" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "text", + "content": ", Binghong Wu" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "text", + "content": ", Can Huang" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "text", + "content": ", Hao Liu" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "text", + "content": ", Xin Tan" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "text", + "content": ", Zhizhong Zhang" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "text", + "content": ", Yuan Xie" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "text", + "content": " East China Normal University " + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 142, + 482, + 186 + ], + "type": "text", + "content": "ByteDance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 187, + 473, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 187, + 473, + 200 + ], + "spans": [ + { + "bbox": [ + 117, + 187, + 473, + 200 + ], + "type": "text", + "content": "{51255901056}@stu.ecnu.edu.cn, {zzzhang,xtan,yxie}@cs.ecnu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 201, + 473, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 201, + 473, + 213 + ], + "spans": [ + { + "bbox": [ + 117, + 201, + 473, + 213 + ], + "type": "text", + "content": "{tangjingqun,linchunhui,wubinghong,liuhao,can.huang}@bytedance.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 266, + 290, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 266, + 290, + 542 + ], + "spans": [ + { + "bbox": [ + 45, + 266, + 290, + 542 + ], + "type": "text", + "content": "Scene text recognition (STR) in the wild frequently encounters challenges when coping with domain variations, font diversity, shape deformations, etc. A straightforward solution is performing model fine-tuning tailored to a specific scenario, but it is computationally intensive and requires multiple model copies for various scenarios. Recent studies indicate that large language models (LLMs) can learn from a few demonstration examples in a training-free manner, termed \"In-Context Learning\" (ICL). Nevertheless, applying LLMs as a text recognizer is unacceptably resource-consuming. Moreover, our pilot experiments on LLMs show that ICL fails in STR, mainly attributed to the insufficient incorporation of contextual information from diverse samples in the training stage. To this end, we introduce " + }, + { + "bbox": [ + 45, + 266, + 290, + 542 + ], + "type": "inline_equation", + "content": "E^2" + }, + { + "bbox": [ + 45, + 266, + 290, + 542 + ], + "type": "text", + "content": "STR, a STR model trained with context-rich scene text sequences, where the sequences are generated via our proposed in-context training strategy. " + }, + { + "bbox": [ + 45, + 266, + 290, + 542 + ], + "type": "inline_equation", + "content": "E^2" + }, + { + "bbox": [ + 45, + 266, + 290, + 542 + ], + "type": "text", + "content": "STR demonstrates that a regular-sized model is sufficient to achieve effective ICL capabilities in STR. Extensive experiments show that " + }, + { + "bbox": [ + 45, + 266, + 290, + 542 + ], + "type": "inline_equation", + "content": "E^2" + }, + { + "bbox": [ + 45, + 266, + 290, + 542 + ], + "type": "text", + "content": "STR exhibits remarkable training-free adaptation in various scenarios and outperforms even the fine-tuned state-of-the-art approaches on public benchmarks. The code is released at https://github.com/bytedance/E2STR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 566, + 128, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 128, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 128, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 586, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 646 + ], + "type": "text", + "content": "Scene Text Recognition (STR) is a fundamental task in computer vision, with extensive applications in several domains such as autonomous driving [45], augmented reality [30, 33], industrial print recognition [28] and visual understanding [26]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 647, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 670 + ], + "type": "text", + "content": "Current progress in STR [3, 17, 20, 35] has demonstrated remarkable performance in numerous scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 671, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 671, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 59, + 671, + 287, + 685 + ], + "type": "text", + "content": "However, as shown in Figure 1 (a), STR models are sup" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 243, + 544, + 305 + ], + "blocks": [ + { + "bbox": [ + 310, + 243, + 544, + 305 + ], + "lines": [ + { + "bbox": [ + 310, + 243, + 544, + 305 + ], + "spans": [ + { + "bbox": [ + 310, + 243, + 544, + 305 + ], + "type": "image", + "image_path": "2bb54acc35d7fb57ff0359e0d3752e9e0be937b0180d2a18ac947d2f68e0fd1f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 308, + 500, + 318 + ], + "lines": [ + { + "bbox": [ + 354, + 308, + 500, + 318 + ], + "spans": [ + { + "bbox": [ + 354, + 308, + 500, + 318 + ], + "type": "text", + "content": "(a) Diversified scenarios of scene text" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 311, + 321, + 544, + 386 + ], + "blocks": [ + { + "bbox": [ + 311, + 321, + 544, + 386 + ], + "lines": [ + { + "bbox": [ + 311, + 321, + 544, + 386 + ], + "spans": [ + { + "bbox": [ + 311, + 321, + 544, + 386 + ], + "type": "image", + "image_path": "0332d3b9d72166f36a2f7dac608dcdec76fdcc86fa7080a9ce1debda48c46650.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 388, + 546, + 399 + ], + "lines": [ + { + "bbox": [ + 308, + 388, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 546, + 399 + ], + "type": "text", + "content": "(b) Scenario-specific fine-tuning pipeline of current methods" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 311, + 402, + 544, + 460 + ], + "blocks": [ + { + "bbox": [ + 311, + 402, + 544, + 460 + ], + "lines": [ + { + "bbox": [ + 311, + 402, + 544, + 460 + ], + "spans": [ + { + "bbox": [ + 311, + 402, + 544, + 460 + ], + "type": "image", + "image_path": "933747a14fd1c940284ad7451e6b7d89a3af0d25cd0710451878dda39a6e3859.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 462, + 507, + 474 + ], + "lines": [ + { + "bbox": [ + 347, + 462, + 507, + 474 + ], + "spans": [ + { + "bbox": [ + 347, + 462, + 507, + 474 + ], + "type": "text", + "content": "(c) In-context learning pipeline of " + }, + { + "bbox": [ + 347, + 462, + 507, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^2\\mathbf{STR}" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 483, + 547, + 582 + ], + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 582 + ], + "type": "text", + "content": "Figure 1. Demonstration of real-world scene text scenarios and the adaptation pipeline. (a) Diversified scenarios of scene text in the real world. (b) The adaptation pipeline of current methods. They typically have to fine-tune upon a trained STR model with the training set, under a specific scenario. (c) The adaptation pipeline of our proposed " + }, + { + "bbox": [ + 304, + 483, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 483, + 547, + 582 + ], + "type": "text", + "content": ". Our method automatically selects in-context prompts and performs training-free adaptation when faced with novel scenarios. Blue characters denote ambiguous scene text that is easily recognized." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "content": "posed to perform robustly over diversified scenarios in the real world, where the scene text is hard to recognize because of domain variation, font diversity, shape deformation, etc. As shown in Figure 1 (b), a straightforward solution involves collecting the corresponding data and then fine-tuning the model for the specific scenario [3, 17, 20]. This process is computationally intensive and requires multiple model copies for diverse scenarios." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "The development of a comprehensive and reliable STR" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR 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." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 145, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 145, + 703 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 145, + 703 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 61, + 703, + 266, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 266, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 266, + 712 + ], + "type": "text", + "content": "*This work is done when Zhen Zhao is an intern at ByteDance." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15567" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 164, + 126 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 164, + 126 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 164, + 126 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 164, + 126 + ], + "type": "image", + "image_path": "3dba2f422435df1238e2ef183a714acb1715d1b4ae84fd8b39037bde1982db78.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 128, + 148, + 134 + ], + "lines": [ + { + "bbox": [ + 61, + 128, + 148, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 148, + 134 + ], + "type": "text", + "content": "(b) Context-Rich scene text sequence" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 287, + 126 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 287, + 126 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 287, + 126 + ], + "type": "image", + "image_path": "cf0ef15b006408fe6b66e44f03c601ec822a7c646f5ed00a51c575f73ea054f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 128, + 286, + 134 + ], + "lines": [ + { + "bbox": [ + 182, + 128, + 286, + 134 + ], + "spans": [ + { + "bbox": [ + 182, + 128, + 286, + 134 + ], + "type": "text", + "content": "(c) ICL results on cross-domain benchmark" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 144, + 287, + 199 + ], + "lines": [ + { + "bbox": [ + 46, + 144, + 287, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 287, + 199 + ], + "type": "text", + "content": "Figure 2. Our pilot experiments. (a) The randomly concatenated scene text sequence. (b) Our proposed context-rich scene text sequence. (c) By training an STR model based on the randomly concatenated scene text sequence, we evaluate the model on three cross-domain datasets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 221, + 287, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 221, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 287, + 245 + ], + "type": "text", + "content": "model that can effectively handle many real-world scenarios remains a significant challenge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 246, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 287, + 353 + ], + "type": "text", + "content": "Fortunately, plenty of studies [1, 6, 21, 38] have shown that Large Language Models (LLMs) can easily adapt without additional training. This adaptation is achieved by leveraging only a handful of input-label pairs as context (prompting information), a phenomenon known as \"In-Context Learning\" (ICL). The advantages of ICL inspire our interest in implementing it in STR, such that by fetching a few in-context prompts, a single model can be rapidly adapted to various scenarios without fine-tuning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 354, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 425 + ], + "type": "text", + "content": "However, the equipment of ICL in STR still poses challenges under the existing circumstances. Firstly, it is deemed excessively costly to apply Multi-Modal Large Language Models (M-LLMs) with billions of parameters as a scene text recognizer. And the ICL capabilities in regular-sized models have been barely explored currently." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 426, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 640 + ], + "type": "text", + "content": "Secondly, it is hard to acquire ICL capabilities for a STR model with current training strategies. Previous studies have observed that sending image-text sequences for training would naturally endow ICL for M-LLMs [1, 21, 38], while such a phenomenon is hard to achieve in STR. As shown in Figure 2 (a), we generate sequential training data by randomly concatenating scene text samples. This practice fails as the trained model does not exhibit any performance improvement even when provided with in-domain prompts (Figure 2 (c)). The major cause of this failure is the lack of context in the generated scene text sequences during the training phase. The arbitrary concatenation of scene text does not provide any contextual information (i.e., sample connections) between different samples (Figure 2 (a)). Consequently, the model lacks the ability to effectively use information derived from in-context prompts (Figure 2 (c)), which implies that in-context training is essentially important for the effective implementation of ICL in STR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "Based on the above analysis, we propose " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " (Ego-Evolving STR), a paradigm that facilitates adaptation across diverse scenarios in a training-free manner. Specifically, we propose an in-context training strategy, which enables the model to exploit contextual information from the generated context-rich scene text sequences (Figure 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "(b)). The context-rich scene text sequences are formed using our ST-strategy, which involves random Splitting and Transformation of scene text, hence generating a set of \"sub-samples\". The sub-samples are inner-connected in terms of both visual and linguistic aspects. In the inference stage, " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " fetches in-context prompts based on visual similarities, and utilizes the prompts to assist the recognition, shown in Figure 1 (c). In practice, it is found that with proper training and inference strategies, ICL capabilities can also be observed in regular-sized STR models (hundreds of millions of parameters)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "type": "text", + "content": "Finally, the proposed " + }, + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "type": "text", + "content": " effectively captures contextual information from the in-context prompts and performs rapid adaptation in various novel scenarios in a training-free manner (Please refer to Section 4.2). On common benchmarks, " + }, + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "type": "text", + "content": " achieves SOTA results, with an average improvement of " + }, + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "type": "text", + "content": " over previous methods and " + }, + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "type": "text", + "content": " over itself without ICL. Most importantly, when evaluated on unseen domains, " + }, + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "type": "text", + "content": " achieves impressive performance with only a few prompts, even outperforming the fine-tuning results of SOTA methods by " + }, + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 204, + 545, + 335 + ], + "type": "text", + "content": ". Our contributions are summarized below:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 336, + 545, + 505 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 304, + 336, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 372 + ], + "type": "text", + "content": "(1) We propose " + }, + { + "bbox": [ + 304, + 336, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 336, + 545, + 372 + ], + "type": "text", + "content": ", a robust STR paradigm that can perform rapid adaptation over diverse scenarios in a training-free manner." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "content": "(2) We provide an in-context training strategy for equipping STR models with ICL capabilities, as well as an in-context inference strategy for STR models to leverage contextual information from in-context prompts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 421, + 545, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 545, + 457 + ], + "type": "text", + "content": "(3) We demonstrate that ICL capabilities can be effectively incorporated into regular-sized STR models via appropriate training and inference strategies." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 457, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 505 + ], + "type": "text", + "content": "(4) Extensive experiments show that " + }, + { + "bbox": [ + 304, + 457, + 545, + 505 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 457, + 545, + 505 + ], + "type": "text", + "content": " exceeds state-of-the-art performance across diverse benchmarks, even surpassing the fine-tuned approaches in unseen domains." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 518, + 392, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 518, + 392, + 530 + ], + "spans": [ + { + "bbox": [ + 306, + 518, + 392, + 530 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 538, + 437, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 437, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 437, + 551 + ], + "type": "text", + "content": "2.1. Scene Text Recognition" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 557, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 593 + ], + "type": "text", + "content": "Recent years have witnessed extensive studies in STR, which can be generally divided into Language-free methods and Language-aware methods." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Language-free STR. Language-free models directly utilize visual features for prediction, without considering the relationship between the characters. In this branch CTC-based [11] methods [5, 24] play the most prominent part. They typically consist of a CNN for feature extraction and an RNN for sequential feature processing, which are trained end-to-end with the CTC loss [11]. Other methods like [23, 40] focus on treating STR as a character-level segmentation task. The lack of linguistic information limits the application of language-free methods in scenarios with" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15568" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 189, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 189, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 189, + 83 + ], + "type": "text", + "content": "occluded or incomplete characters." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 287, + 240 + ], + "type": "text", + "content": "Language-aware STR. Language-aware models leverage linguistic information to assist the recognition, typically utilizing an external language model (LM) [10, 44] or training internal LMs [2, 7, 36]. SRN [44] and ABINet [10] feed visual predictions to an external LM for linguistic refinement. The direct application of an external LM without considering visual features leads to possible erroneous correction. On the other hand, methods like PARSeq [3] and MAERec [17] implicitly train an internal LM in an auto-regressive manner, which have achieved decent performance. In this paper we base our model on the language-aware design, training a transformer-based language decoder inner-connected with the vision encoder." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 248, + 230, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 230, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 230, + 262 + ], + "type": "text", + "content": "2.2. Multi-Modal In-Context Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 267, + 289, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 289, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 289, + 543 + ], + "type": "text", + "content": "Recent large language models (LLMs) [6, 46] have demonstrated their excellent few-shot adaptation capabilities. By concatenating a few examples with the input as the prompt at reference time, LLMs quickly adapt to novel tasks without parameter updating. This phenomenon introduces a novel learning paradigm termed \"In-Context Learning\". Meanwhile, unlike LLMs, vision-language models (VLMs) struggle to understand complex multi-modal prompts [47]. A large set of approaches [13-15, 34] have been proposed to empower VLMs with multi-modal in-context learning (M-ICL) capabilities, but they typically utilize vision models (like image caption models) to translate images to text [15, 34, 43], or view the LLM as a scheduler learning to call vision experts based on a few examples [13]. These approaches do not truly establish a VLM with M-ICL capabilities. Recently, several work [1, 21, 38] proposes to train VLMs with sequential multi-modal data, and have achieved great success in prompting VLMs with multi-modal examples. In this paper, we aim to train a scene text recognizer equipped with M-ICL capabilities based on this sequential training paradigm. We demonstrate that the arbitrary concatenation of scene text fails as stated above, which motivates us to generate context-rich scene text sequences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 554, + 129, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 129, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 129, + 568 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 287, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 287, + 599 + ], + "type": "text", + "content": "3.1. Preliminary of Multi-Modal In-Context Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "content": "Multi-modal in-context Learning enables M-LLMs to perform quick adaptation in downstream tasks in a training-free manner, hence eliminating the redundant computation and time expenses of fine-tuning. In this subsection, we introduce how to formulate multi-modal in-context learning for addressing the STR task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "For a scene text tuple " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "(\\mathbf{x},\\mathbf{y})" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " is the scene image and " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " is the ground-truth text, the STR task involves generating the label " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " by maximizing the conditional probabil" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "ity under the classic auto-regressive paradigm as follows: " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "p(\\boldsymbol{y}|\\boldsymbol{x}) = \\prod_{l=1}^{L} p(\\boldsymbol{y}_l|\\boldsymbol{x}, \\boldsymbol{y}_{" + }, + { + "bbox": [ + 46, + 596, + 288, + 633 + ], + "type": "text", + "content": ". The training in this phase makes use of the next-token prediction loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 81, + 650, + 287, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 650, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 81, + 650, + 287, + 684 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y}) \\sim \\mathcal {D}} \\left[ - \\sum_ {l = 1} ^ {L} \\log p (\\boldsymbol {y} _ {l} | \\boldsymbol {y} _ {< l}, \\boldsymbol {x}) \\right], \\tag {2}", + "image_path": "7c141a9c8fd0f63b935f29ff0be88b222a483345eb2eeda38d79962502bba264.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 159, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 159, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 159, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 701, + 159, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 701, + 159, + 714 + ], + "type": "text", + "content": " is the training set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 415, + 425, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 425, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 425, + 428 + ], + "type": "text", + "content": "3.3.2 In-Context Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "type": "text", + "content": "The objective of the in-context training phase is to equip " + }, + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "type": "text", + "content": " with the capability of In-Context Learning. As depicted in the top of Figure 3, the model is trained with context-rich scene text sequences as stated before. In these sequences, we interleave a placeholder " + }, + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "type": "text", + "content": " in the text for each image. This serves to make the language decoder distinguish between different samples following [1]. In this stage, we propose two strategies to generate context-rich scene text sequences: the Split Strategy and the Transform Strategy (the ST strategy)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 556, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 664 + ], + "type": "text", + "content": "The Split Strategy. As shown in Figure 4 (a), when presented with a training tuple " + }, + { + "bbox": [ + 304, + 556, + 545, + 664 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\pmb{y})" + }, + { + "bbox": [ + 304, + 556, + 545, + 664 + ], + "type": "text", + "content": ", we split the sample and hence generating a set of \"sub-samples\". It is evident that the sub-samples exhibit a strong connection to the original training sample. Furthermore, the sub-samples themselves demonstrate interconnectivity as they overlap with one another. Next, we proceed to concatenate the sub-samples with " + }, + { + "bbox": [ + 304, + 556, + 545, + 664 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\pmb{y})" + }, + { + "bbox": [ + 304, + 556, + 545, + 664 + ], + "type": "text", + "content": " and additional randomly selected samples to form a context-rich sample sequence." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 700 + ], + "type": "text", + "content": "We randomly shuffle the whole sequence before generating the actual input text (i.e., interleaving the " + }, + { + "bbox": [ + 304, + 665, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{\\Pi} / \\mathbf{i}\\rangle" + }, + { + "bbox": [ + 304, + 665, + 545, + 700 + ], + "type": "text", + "content": " token to the text sequence)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "In practice, to accurately split the training samples, we" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15570" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "synthesize 600k scene text images based on [4] and record the accurate bounding boxes of every single character. Our subsequent experiments show that the synthesized data does not change " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "'s non-context text recognition ability, but the Split Strategy based on them equips " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " with a strong capability of in-context learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 228 + ], + "type": "text", + "content": "The Transform Strategy. As shown in Figure 4 (b), given a training tuple " + }, + { + "bbox": [ + 46, + 144, + 288, + 228 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{x}, \\boldsymbol{y})" + }, + { + "bbox": [ + 46, + 144, + 288, + 228 + ], + "type": "text", + "content": " (whether with character-wise bounding boxes or not), we perform data augmentation (a set of image transformations, e.g., color/direction transformations) on " + }, + { + "bbox": [ + 46, + 144, + 288, + 228 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 46, + 144, + 288, + 228 + ], + "type": "text", + "content": ". In this way, we also generate a set of sub-samples with the same label but different image patterns from the original sample." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 228, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 312 + ], + "type": "text", + "content": "In practice, as depicted in Figure 4 (c), we hybrid the above strategies. The training set is formed by concatenating the synthesized data and the original training data used in the first training phase. For the synthesized data with character-wise bounding boxes, both the Split Strategy and the Transform Strategy are utilized. For the original training data, only the Transform Strategy is implemented." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 312, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 288, + 373 + ], + "type": "text", + "content": "Finally, after generating the sample sequence " + }, + { + "bbox": [ + 47, + 312, + 288, + 373 + ], + "type": "inline_equation", + "content": "(X,Y)" + }, + { + "bbox": [ + 47, + 312, + 288, + 373 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 312, + 288, + 373 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 312, + 288, + 373 + ], + "type": "text", + "content": " is the image sequence and " + }, + { + "bbox": [ + 47, + 312, + 288, + 373 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 312, + 288, + 373 + ], + "type": "text", + "content": " is the text sequence, " + }, + { + "bbox": [ + 47, + 312, + 288, + 373 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 312, + 288, + 373 + ], + "type": "text", + "content": " is fed into the vision encoder, while " + }, + { + "bbox": [ + 47, + 312, + 288, + 373 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 312, + 288, + 373 + ], + "type": "text", + "content": " is processed by the language decoder under the auto-regressive framework. The loss function is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 380, + 287, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 380, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 86, + 380, + 287, + 413 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\left(\\boldsymbol {X}, \\boldsymbol {Y}\\right)} = - \\sum_ {l = 1} ^ {L} \\log p \\left(\\boldsymbol {Y} _ {l} \\mid \\boldsymbol {Y} _ {< l}, \\boldsymbol {X} _ {\\leq l}\\right), \\tag {3}", + "image_path": "64af0be0c0fcae9aeec88c9ecaadfd64c106c32e76cac2cafb4325e2250adaa9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 422, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 422, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\leq l}" + }, + { + "bbox": [ + 46, + 422, + 287, + 448 + ], + "type": "text", + "content": " is the set of image tokens preceding token " + }, + { + "bbox": [ + 46, + 422, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_l" + }, + { + "bbox": [ + 46, + 422, + 287, + 448 + ], + "type": "text", + "content": " in the input sequence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 456, + 168, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 168, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 168, + 467 + ], + "type": "text", + "content": "3.4. In-Context Inference" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "text", + "content": "The In-Context Learning ability is acquired by our " + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "text", + "content": " model through the above two-stage training approach. AS shown in the bottom of Figure 3, when presented with a test image " + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "text", + "content": ", the framework selects " + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "text", + "content": " samples " + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "inline_equation", + "content": "\\{(x_i^c,y_i^c)\\}_{i = 1}^N" + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "text", + "content": " from a in-context pool " + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^c" + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "text", + "content": ". The selected samples have the highest visual similarities to " + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "text", + "content": " in the latent space. Specifically, we calculate the image embedding " + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "text", + "content": " by averaging the visual token sequence Encoder " + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "inline_equation", + "content": "(\\pmb{x})" + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "text", + "content": ". The in-context prompts are then formed by choosing " + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "text", + "content": " samples from " + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^c" + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "text", + "content": ", where the image embeddings of these samples have the top-N highest cosine similarity with " + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 46, + 474, + 287, + 606 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 616, + 287, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 616, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 287, + 644 + ], + "type": "interline_equation", + "content": "\\mathcal {I} = \\underset {i \\in 1, 2, \\dots , | \\mathcal {D} ^ {c} |} {\\arg \\operatorname {T o p N}} \\frac {\\boldsymbol {I} ^ {\\mathrm {T}} \\boldsymbol {I} _ {i} ^ {c}}{\\| \\boldsymbol {I} \\| _ {2} \\| \\boldsymbol {I} _ {i} ^ {c} \\| _ {2}}, \\tag {4}", + "image_path": "ac74670795d7dd9d08f20e5d0859631d1c172c6a94aaa2b88dbdedb97ff78a89.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 654, + 287, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 690 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 654, + 287, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 654, + 287, + 690 + ], + "type": "text", + "content": " is the index set of the top-N similar samples in " + }, + { + "bbox": [ + 46, + 654, + 287, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^c" + }, + { + "bbox": [ + 46, + 654, + 287, + 690 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 654, + 287, + 690 + ], + "type": "inline_equation", + "content": "I_{i}^{c}" + }, + { + "bbox": [ + 46, + 654, + 287, + 690 + ], + "type": "text", + "content": " is the image embedding of the " + }, + { + "bbox": [ + 46, + 654, + 287, + 690 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 654, + 287, + 690 + ], + "type": "text", + "content": "-th sample in " + }, + { + "bbox": [ + 46, + 654, + 287, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^c" + }, + { + "bbox": [ + 46, + 654, + 287, + 690 + ], + "type": "text", + "content": ". The in-context prompts are then defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 119, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 119, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {E} = \\left\\{\\left(\\boldsymbol {x} _ {i} ^ {c}, \\boldsymbol {y} _ {i} ^ {c}\\right) | i \\in \\mathcal {I} \\right\\}. \\tag {5}", + "image_path": "19fa6c2eebdb2a1e07a11afb7f04c5cfa569ea8e0daa57b1c580a8b2dfed014d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "As shown in the bottom of Figure 3, " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " is concatenated with the test sample " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " and our in-context prediction is formulated as " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "p(\\pmb{y}|\\pmb{E},\\pmb{x})" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ". In practice, the in-context pool " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^c" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " retains solely the output tokens generated by the vision encoder, resulting in a highly efficient selection process. Furthermore, because the in-context pool is tiny and we do straight inference without training, the extra consumption is kept to a minimum (Please refer to Section 4.3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 178, + 382, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 178, + 382, + 190 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 382, + 190 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 197, + 422, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 197, + 422, + 210 + ], + "spans": [ + { + "bbox": [ + 305, + 197, + 422, + 210 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 215, + 545, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 215, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 215, + 545, + 359 + ], + "type": "text", + "content": "Implementation Details. Following MAERec [17], we choose Vision Transformer [9] pre-trained under the MAE [16] framework as the vision encoder. The default language decoder is set as OPT-125M [46]. We use the cosine learning rate scheduler without warm-up and the AdamW optimizer with a weight decay of 0.01. We train our model for 10 epochs with an initial learning rate of 1e-4 during the first training stage, and 5 epochs with an initial learning rate of 5e-6 during the second in-context training stage. The training batch size is 64 for the first stage and 8 for the second stage. During inference for " + }, + { + "bbox": [ + 304, + 215, + 545, + 359 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 215, + 545, + 359 + ], + "type": "text", + "content": "-ICL, we select two in-context prompts based on the kNN selection strategy." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 360, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 574 + ], + "type": "text", + "content": "Datasets and Metrics. We use the real-world training dataset Union14M-L[17] for the two-stage training. The same training dataset (including the synthesized data) is adopted for all compared methods. " + }, + { + "bbox": [ + 304, + 360, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 360, + 545, + 574 + ], + "type": "text", + "content": " is evaluated under various publicly available benchmarks, including Regular Benchmarks IIIT5k [29], SVT [37], IC13 [18], Irregular Benchmarks IC15 [19], SVTP [31], CUTE80 (CT80) [32], COCO Text (COCO) [39], CTW [25], Total Text (TT) [8], Occluded Benchmarks OST (HOST and WOST) [41] and artistic benchmark WordArt [42]. In cross domain scenarios the evaluated datasets including the metal-surface benchmark MPSC [12] and the handwriting benchmark IAM [27], as well as a more difficult real-world industrial text recognition dataset EIST (Enhanced Industrial Scene Text) collected by us. EIST is collected from the real-world industrial scenario, which contains 200 training samples and 8000 test samples. We use Word Accuracy[17] as the evaluation metric for all compared methods." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 580, + 391, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 580, + 391, + 592 + ], + "spans": [ + { + "bbox": [ + 305, + 580, + 391, + 592 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 599, + 481, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 481, + 610 + ], + "type": "text", + "content": "4.2.1 Results on Common Benchmarks" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Table 1 presents the performance of " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " on common benchmarks. We evaluate " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " on 12 commonly used STR benchmarks and compare with SOTA methods. " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "-base refers to non-context prediction without prompts. For " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "-ICL, a tiny in-context pool is maintained by randomly sampling 1000 images from the training data (less than " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.025\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " of the number of training samples). As we can see, " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "-base achieves " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "90.25\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " average word" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15571" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 69, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 545, + 205 + ], + "type": "table", + "html": "
MethodVenueRegularIrregularOccludedOthersAVG
IIIT3000SVT647IC131015IC152077SVTP645CT80288COCO9896CTW1572TT2201HOST2416WOST2416WordArt1511
ASTER [36]PAMI'1895.0389.4993.7985.4882.0290.2862.2576.5378.6943.3464.6565.5977.26
NRTR [35]ICDAR'1997.4393.8296.0685.1584.0391.3265.9481.7481.8350.8371.5264.0680.31
SAR [22]AAAI'1997.7094.1396.3587.4787.6093.0667.4183.9186.2346.3670.3272.4081.91
SATRN [20]AAAI'2097.8395.8397.4489.4690.8596.1873.0684.6187.9156.7175.6275.7185.10
ABINet [10]CVPR'2197.9095.9896.1691.6690.2393.7571.4683.7286.0156.5475.7575.2584.53
PARSeq* [3]ECCV'2299.1097.8498.1389.2296.9098.61-------
MAERec [17]ICCV'2398.9397.9998.6293.0494.5798.9678.8488.8793.9173.9785.7282.5990.50
\\( E^2STR \\)-base99.1098.1598.0392.9996.4398.9677.2988.3693.4673.3085.5181.4790.25
\\( E^2STR-ICL \\)99.2398.6198.7293.8296.7499.3178.3888.9994.6874.7586.5986.1791.33
", + "image_path": "2ebaca2d40bfde807c2fb3e072e2e1d5fa23294e2669a32335bbdbba04300329.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 264, + 290, + 399 + ], + "blocks": [ + { + "bbox": [ + 46, + 212, + 546, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 546, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 546, + 246 + ], + "type": "text", + "content": "Table 1. Results on common benchmarks. All methods are trained on the same dataset except for PARSeq.*: PARSeq is trained on its self-collected real-world dataset and we directly quote the results from its original paper. Red and blue values denote the best and the secondary performance. " + }, + { + "bbox": [ + 46, + 212, + 546, + 246 + ], + "type": "inline_equation", + "content": "{\\mathrm{E}}^{2}" + }, + { + "bbox": [ + 46, + 212, + 546, + 246 + ], + "type": "text", + "content": " STR-base refers to non-context inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 264, + 290, + 399 + ], + "lines": [ + { + "bbox": [ + 48, + 264, + 290, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 290, + 399 + ], + "type": "table", + "html": "
MethodIndustrialHandwritingAVG
MPSC 2941EIST 8000IAM 3000
ASTER [36]63.4848.7652.5054.91
NRTR [35]73.2461.7759.5364.85
SAR [22]73.8558.2656.6362.91
ABINet [10]75.3562.8561.5766.59
SATRN [20]76.1065.4259.4767.00
MAERec [17]81.8170.3370.2774.14
E²STR-base81.2669.6669.5173.48
E²STR-ICL83.6476.7774.1078.17
", + "image_path": "44cbb7d7ba427f50a90a8d97ec55d8075b034a4776dcab633ed4ee85b8e86bae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 407, + 288, + 441 + ], + "lines": [ + { + "bbox": [ + 46, + 407, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 288, + 441 + ], + "type": "text", + "content": "Table 2. Results on cross domain scenarios. Three datasets under two unseen domains are evaluated. All approaches are evaluated in a training-free manner." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "type": "text", + "content": "accuracy over 12 datasets, " + }, + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "type": "inline_equation", + "content": "0.25\\%" + }, + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "type": "text", + "content": " lower than MAERec [17]. However, by fetching in-context prompts and exploiting incontext information, " + }, + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "type": "text", + "content": "STR-ICL achieves an average word accuracy of " + }, + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "type": "inline_equation", + "content": "91.33\\%" + }, + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "type": "inline_equation", + "content": "1.08\\%" + }, + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "type": "text", + "content": " higher than " + }, + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "type": "text", + "content": "STR-base and " + }, + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "type": "inline_equation", + "content": "0.83\\%" + }, + { + "bbox": [ + 46, + 463, + 287, + 534 + ], + "type": "text", + "content": " higher than MAERec. Please note that this improvement is automatic and training-free." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "type": "text", + "content": "Specifically, on the six traditional STR benchmarks (i.e., IIIT, SVT, IC13, IC15, SVTP, and CT80) which have nearly reached saturation in recent years[17], " + }, + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "type": "text", + "content": " still push the performance limit from " + }, + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "type": "inline_equation", + "content": "97.02\\%" + }, + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "type": "inline_equation", + "content": "97.74\\%" + }, + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "type": "text", + "content": ", leading to a " + }, + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "type": "text", + "content": " error rate decrease. On the 6 larger and harder STR benchmarks (i.e., COCO Text, CTW, TT, HOST, and WOST), " + }, + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR - ICL}" + }, + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "type": "text", + "content": " outperforms MAERec by " + }, + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "type": "inline_equation", + "content": "0.94\\%" + }, + { + "bbox": [ + 46, + 535, + 287, + 619 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 634, + 231, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 231, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 231, + 645 + ], + "type": "text", + "content": "4.2.2 Results on Cross Domain Scenarios" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "We compare with SOTA methods on cross domain benchmarks. Two novel scenarios are introduced: the industrial scenario (MPSC and EIST) and the handwriting scenario (IAM). In each dataset, only 100 training samples are provided. For " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "STR-ICL we simply use the training samples" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 264, + 547, + 389 + ], + "blocks": [ + { + "bbox": [ + 307, + 264, + 547, + 389 + ], + "lines": [ + { + "bbox": [ + 307, + 264, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 547, + 389 + ], + "type": "image", + "image_path": "01bd4147d21069695247f3d54bd9690b436b9a72d6043d12c65248ee53eaa125.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 398, + 546, + 443 + ], + "lines": [ + { + "bbox": [ + 305, + 398, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 398, + 546, + 443 + ], + "type": "text", + "content": "Figure 5. Comparison with the fine-tuned models. We report the average performance on three cross-domain datasets. Please note that ABINet [10], SATRN [20] and MAERec [17] are fine-tuned with the in-domain data, while our " + }, + { + "bbox": [ + 305, + 398, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 305, + 398, + 546, + 443 + ], + "type": "text", + "content": "STR-ICL is training-free." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 469, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 469, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 305, + 469, + 545, + 493 + ], + "type": "text", + "content": "as the in-context pool. We compare the training-free results in Table 2 and the fine-tuning results in Figure 5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": "As we can see, on both industrial and handwriting scenarios our " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": "STR-ICL reaches SOTA performance. As shown in Table 2, under the training-free constraint " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": "STR-ICL reaches an average performance of " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "78.17\\%" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "4.69\\%" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": " higher than " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": "STR-base and " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "4.03\\%" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": " higher than the SOTA method MAERec. Specifically, on EIST and IAM the application of ICL brings a huge improvement of " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "7.11\\%" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "4.59\\%" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": ", which demonstrates the extraordinary adaptation ability of " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": "STR-ICL." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": "We further compare the fine-tuned methods and our " + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": "STR-ICL. We fine-tune ABINet [10], SATRN [20] and MAERec [17] with the same data preserved in the in-context pool. As shown in Figure 5, " + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": "STR-ICL outperforms MAERec by " + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "inline_equation", + "content": "1.16\\%" + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": " even if the latter is fine-tuned with in-domain data, which is an exciting result given that " + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": "STR-ICL requires no parameter updating. In a word, our " + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": "STR can be rapidly implemented in a training-free manner in various novel scenarios and even achieves better per" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15572" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 133 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 133 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 133 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 133 + ], + "type": "table", + "html": "
COCOHOSTWordArt
annotation rate10%20%10%20%10%20%
MAERec [17]000000
w/ fine-tuning0.821.671.031.721.342.23
E²STR-base000000
E²STR-ICL10.1212.9212.4313.7626.2232.02
", + "image_path": "80c22ec14a1a2a3b520e29614c808b82ca4de267495f3cb6157fa7c7853f25f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 206, + 269, + 281 + ], + "blocks": [ + { + "bbox": [ + 46, + 141, + 287, + 196 + ], + "lines": [ + { + "bbox": [ + 46, + 141, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 46, + 141, + 287, + 196 + ], + "type": "text", + "content": "Table 3. Results on hard case rectification. \"Hard Cases\" are test samples misrecognized by both MAERec [17] and our " + }, + { + "bbox": [ + 46, + 141, + 287, + 196 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 46, + 141, + 287, + 196 + ], + "type": "text", + "content": "STR-base. By providing annotation of a small part of the hard cases, we compare the performance increase in the rest test samples between the fine-tuned MAERec and our " + }, + { + "bbox": [ + 46, + 141, + 287, + 196 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 46, + 141, + 287, + 196 + ], + "type": "text", + "content": "STR-ICL." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 206, + 269, + 281 + ], + "lines": [ + { + "bbox": [ + 65, + 206, + 269, + 281 + ], + "spans": [ + { + "bbox": [ + 65, + 206, + 269, + 281 + ], + "type": "table", + "html": "
Training TaskWord Accuracy
VTTSSSNon-ContextIn-Context
✓69.6926.82
✓✓69.8075.60
✓✓69.6673.09
✓✓✓69.6676.77
", + "image_path": "1d8d6cf8e04cf8ffb53138266ee6dc97b70c6da5d3b4fa301e0d11cd6dc7bba9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 290, + 286, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 290, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 286, + 323 + ], + "type": "text", + "content": "Table 4. Ablation of our proposed training strategies, where VT, TS, and SS refer to vanilla STR training, the Transform Strategy, and the Split Strategy. The experiment is performed on EIST." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 344, + 230, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 344, + 230, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 344, + 230, + 355 + ], + "type": "text", + "content": "formance than the fine-tuned SOTA methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 370, + 228, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 228, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 228, + 381 + ], + "type": "text", + "content": "4.2.3 Results on Hard Case Rectification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "text", + "content": "We demonstrate the rectification ability of " + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "text", + "content": ", which can handle hard cases in STR conveniently and effectively, in a training-free manner. Specifically, we define \"hard cases\" as the scene text samples that are wrongly recognized by both " + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "text", + "content": "-base and the SOTA method MAERec. A small number of hard cases are then annotated, and we study how the model can benefit from the annotated hard cases and decrease the error rate of the rest hard cases. Shown in Table 3, we perform experiments on COCO Text, HOST, and WordArt. As we can see, by annotating " + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "text", + "content": " of the hard cases, " + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "text", + "content": "-ICL decreases the error rate of the rest hard cases by up to " + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "inline_equation", + "content": "32\\%" + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "text", + "content": ". This improvement is achieved by simply putting the annotated hard cases into the in-context pool, without any hassle of re-training the model. By comparison, by fine-tuning on the annotated hard cases, MAERec only decreases the error rate by up to " + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "inline_equation", + "content": "2.23\\%" + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "text", + "content": ", completely incomparable to our " + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 46, + 388, + 286, + 593 + ], + "type": "text", + "content": "-ICL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 599, + 147, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 147, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 147, + 610 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "Impact of Split-and-Transform Training Strategies. We perform an experiment to show the effectiveness of our proposed Split Strategy and Transform Strategy. Shown in Table 4, the vanilla STR training brings a word accuracy of " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "69.69\\%" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": ", but the model cannot understand context information, and the performance even severely decreases to " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "26.82\\%" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " when provided with in-context prompts. The application of the Transform Strategy and the Split Strategy in" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 71, + 539, + 182 + ], + "blocks": [ + { + "bbox": [ + 312, + 71, + 539, + 182 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 539, + 182 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 539, + 182 + ], + "type": "image", + "image_path": "6781f97025bd314b694923578d0f14fad12bdc7a6a8c7bae8cfcc1a141fead86.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 191, + 545, + 235 + ], + "lines": [ + { + "bbox": [ + 305, + 191, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 305, + 191, + 545, + 235 + ], + "type": "text", + "content": "Figure 6. Comparison between different in-context prompt selection strategies. \"Random Selection\" refers to randomly selecting two samples as in-context prompts from the in-context pool. X-axis is the evaluated benchmarks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 257, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 353 + ], + "type": "text", + "content": "the second training stage does not increase the non-context performance (concerning that the synthesized data is typically weaker than the real-world data used in the vanilla training stage), but the model learns to profit from context, and the performance is improved to " + }, + { + "bbox": [ + 304, + 257, + 545, + 353 + ], + "type": "inline_equation", + "content": "75.60\\%" + }, + { + "bbox": [ + 304, + 257, + 545, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 257, + 545, + 353 + ], + "type": "inline_equation", + "content": "73.09\\%" + }, + { + "bbox": [ + 304, + 257, + 545, + 353 + ], + "type": "text", + "content": " respectively when provided with in-context prompts. Finally, the hybrid of the above two strategies further enhances the ICL ability, and the performance reaches " + }, + { + "bbox": [ + 304, + 257, + 545, + 353 + ], + "type": "inline_equation", + "content": "76.77\\%" + }, + { + "bbox": [ + 304, + 257, + 545, + 353 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 354, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 484 + ], + "type": "text", + "content": "Impact of Nearest Neighbor In-Context Prompt Selection. We compare the kNN in-context prompt selection strategy to Random Selection, i.e., randomly selecting in-context prompts from the in-context pool. Shown in Figure 6, on all three evaluated datasets, random selection can improve the performance of non-context prediction by a small margin, but is far from comparing with kNN selection. Specifically, on EIST random selection improves the performance of non-context from " + }, + { + "bbox": [ + 304, + 354, + 545, + 484 + ], + "type": "inline_equation", + "content": "69.66\\%" + }, + { + "bbox": [ + 304, + 354, + 545, + 484 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 354, + 545, + 484 + ], + "type": "inline_equation", + "content": "70.65\\%" + }, + { + "bbox": [ + 304, + 354, + 545, + 484 + ], + "type": "text", + "content": ", while kNN selection reaches " + }, + { + "bbox": [ + 304, + 354, + 545, + 484 + ], + "type": "inline_equation", + "content": "76.77\\%" + }, + { + "bbox": [ + 304, + 354, + 545, + 484 + ], + "type": "text", + "content": " word accuracy under the same condition." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 485, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 581 + ], + "type": "text", + "content": "Impact of In-Context Pool Size. We next study the impact of varying the size of the in-context pool. Shown in Figure 7, we perform experiments on IAM, EIST, and MPSC, by varying the number of samples maintained in the in-context pool. As we can see, in general, the larger in-context pool brings about better performance, and a small number of samples is adequate to bring about huge performance improvement when deploying " + }, + { + "bbox": [ + 304, + 485, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 304, + 485, + 545, + 581 + ], + "type": "text", + "content": "STR-ICL." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": "Impact of the Number of In-Context prompts. We analyze the influence of the number of in-context prompts. Shown in Table 5, the experiment is performed on HOST, ToTal Text, and WordArt. Similar to the in-context pool size, the increase in the number of in-context prompts also generally boosts the performance of " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": "STR-ICL. Moreover, one to two in-context prompts are adequate for improving the performance by a large margin." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Computational Complexity. We experimentally compare the inference speed of " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and MAERec [17]. Shown in Table 6, the inference speed of " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR - ICL}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is on par" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15573" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 72, + 285, + 196 + ], + "blocks": [ + { + "bbox": [ + 54, + 72, + 285, + 196 + ], + "lines": [ + { + "bbox": [ + 54, + 72, + 285, + 196 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 285, + 196 + ], + "type": "image", + "image_path": "93d6f6e5a20b02efc97be0831c144ec55ee061278170774389b547806be59f2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 289, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 239 + ], + "type": "text", + "content": "Figure 7. The performance change brought by different sizes of the in-context pool. The X-axis is the size of the in-context pool and the Y-axis is the word accuracy results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 247, + 279, + 293 + ], + "blocks": [ + { + "bbox": [ + 57, + 247, + 279, + 293 + ], + "lines": [ + { + "bbox": [ + 57, + 247, + 279, + 293 + ], + "spans": [ + { + "bbox": [ + 57, + 247, + 279, + 293 + ], + "type": "table", + "html": "
ICL prompts0124816
HOST73.3074.3474.7574.8374.8374.92
TT93.4694.6894.6894.7894.8894.91
WordArt81.4786.0486.1786.1786.2786.35
", + "image_path": "84b2162522ef1958e2ad0999b2207c5741dd93d60f6a0bfad38a6e7624bc5a27.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 332, + 279, + 357 + ], + "blocks": [ + { + "bbox": [ + 47, + 300, + 288, + 323 + ], + "lines": [ + { + "bbox": [ + 47, + 300, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 288, + 323 + ], + "type": "text", + "content": "Table 5. The performance change brought by the different number of in-context prompts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 332, + 279, + 357 + ], + "lines": [ + { + "bbox": [ + 56, + 332, + 279, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 279, + 357 + ], + "type": "table", + "html": "
MAERec [17]E2STR-baseE2STR-ICL
Inference Time (s)0.0920.0710.094
", + "image_path": "ecde32a1d263add66585695f10a0b1172437394485061c7fef34a8a655bbb6fd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 365, + 287, + 388 + ], + "lines": [ + { + "bbox": [ + 47, + 365, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 287, + 388 + ], + "type": "text", + "content": "Table 6. Comparison of the mean inference time of each test sample. All results are reported under the same hardware environment." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "content": "with MAERec. Compared to " + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "content": "-base, the in-context prompts of " + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "content": "-ICL bring extra consumption, but this leads to a limited inference time increase (i.e., from 0.071 to 0.094). It makes sense since we only maintain the visual tokens in the in-context pool and directly feed the visual tokens of the selected prompts to the language model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 479, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 659 + ], + "type": "text", + "content": "Visualization and Further Analysis. We further study how the test sample learns from context. Shown in Figure 8, we select one context prompt for the test sample, and study the model pays attention to which region of the context image. This is achieved by collecting the attention maps between the language tokens and the image features. As we can see, when the language tokens pay close attention to the misrecognized image region, they also focus on the context image region which has similar patterns. For example, on the last row of Figure 8, " + }, + { + "bbox": [ + 46, + 479, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 46, + 479, + 287, + 659 + ], + "type": "text", + "content": " misrecognized the test image as \"simplest\" without context. By providing a context prompt \"Display\", one language token focuses on the \"la\" region of both images, which have similar image patterns. Finally, " + }, + { + "bbox": [ + 46, + 479, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 46, + 479, + 287, + 659 + ], + "type": "text", + "content": " rectified the misrecognized \"e\" to \"a\" with the help of context ground-truth \"la\" of the focused region." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 669, + 123, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 123, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 123, + 681 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "There are two limitations in our study. Firstly, there is a very slim chance that " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "-ICL erroneously rectifies pre" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 383, + 100 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 383, + 100 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 383, + 100 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 383, + 100 + ], + "type": "image", + "image_path": "eb73f529238dbad11fb19e80a7ad099dda95c56a0f91d84952d8e3cfa36f5efd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 102, + 356, + 110 + ], + "lines": [ + { + "bbox": [ + 335, + 102, + 356, + 110 + ], + "spans": [ + { + "bbox": [ + 335, + 102, + 356, + 110 + ], + "type": "text", + "content": "Vitoc" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 112, + 384, + 150 + ], + "blocks": [ + { + "bbox": [ + 309, + 112, + 384, + 150 + ], + "lines": [ + { + "bbox": [ + 309, + 112, + 384, + 150 + ], + "spans": [ + { + "bbox": [ + 309, + 112, + 384, + 150 + ], + "type": "image", + "image_path": "1d9d82612dc2e15073ed5c84b0a21b97e6a17710f7940946e18c37129fb604d9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 152, + 383, + 177 + ], + "blocks": [ + { + "bbox": [ + 309, + 152, + 383, + 177 + ], + "lines": [ + { + "bbox": [ + 309, + 152, + 383, + 177 + ], + "spans": [ + { + "bbox": [ + 309, + 152, + 383, + 177 + ], + "type": "image", + "image_path": "647e6b809f51ed9fdeaa153f5ab50dd8d614ee65af2418719ce8c386618b4c5a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 178, + 362, + 187 + ], + "lines": [ + { + "bbox": [ + 331, + 178, + 362, + 187 + ], + "spans": [ + { + "bbox": [ + 331, + 178, + 362, + 187 + ], + "type": "text", + "content": "PORTED" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 191, + 384, + 219 + ], + "blocks": [ + { + "bbox": [ + 309, + 191, + 384, + 219 + ], + "lines": [ + { + "bbox": [ + 309, + 191, + 384, + 219 + ], + "spans": [ + { + "bbox": [ + 309, + 191, + 384, + 219 + ], + "type": "image", + "image_path": "d87c88179af81fa73d266bcbb6bbe3b83d39ed9edf7c45d127984831206842ea.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 221, + 362, + 230 + ], + "lines": [ + { + "bbox": [ + 329, + 221, + 362, + 230 + ], + "spans": [ + { + "bbox": [ + 329, + 221, + 362, + 230 + ], + "type": "text", + "content": "simplest" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 235, + 384, + 243 + ], + "lines": [ + { + "bbox": [ + 309, + 235, + 384, + 243 + ], + "spans": [ + { + "bbox": [ + 309, + 235, + 384, + 243 + ], + "type": "text", + "content": "Non-Context Prediction" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 255, + 547, + 321 + ], + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 321 + ], + "type": "text", + "content": "Figure 8. Cross attention visualization between the language tokens and the vision tokens. Left: Non-context prediction of " + }, + { + "bbox": [ + 304, + 255, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 255, + 547, + 321 + ], + "type": "text", + "content": ". Error characters are marked in red. Right: In-context prediction of " + }, + { + "bbox": [ + 304, + 255, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR - ICL}" + }, + { + "bbox": [ + 304, + 255, + 547, + 321 + ], + "type": "text", + "content": ", where only one in-context prompt is selected. We visualize how the language tokens attend to the prompt image and the test image." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 395, + 72, + 468, + 99 + ], + "blocks": [ + { + "bbox": [ + 395, + 72, + 468, + 99 + ], + "lines": [ + { + "bbox": [ + 395, + 72, + 468, + 99 + ], + "spans": [ + { + "bbox": [ + 395, + 72, + 468, + 99 + ], + "type": "image", + "image_path": "21c36fb8edbf9243aa4fc66bd3801cbd202c7520288d34ddb67df5bef3b224bc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 395, + 111, + 468, + 137 + ], + "blocks": [ + { + "bbox": [ + 425, + 102, + 438, + 110 + ], + "lines": [ + { + "bbox": [ + 425, + 102, + 438, + 110 + ], + "spans": [ + { + "bbox": [ + 425, + 102, + 438, + 110 + ], + "type": "text", + "content": "pal" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 395, + 111, + 468, + 137 + ], + "lines": [ + { + "bbox": [ + 395, + 111, + 468, + 137 + ], + "spans": [ + { + "bbox": [ + 395, + 111, + 468, + 137 + ], + "type": "image", + "image_path": "f192a34e78674eda264ef87513a81c6c5eec554e717784f3096c073d12b57fe4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 395, + 149, + 468, + 175 + ], + "blocks": [ + { + "bbox": [ + 413, + 139, + 443, + 148 + ], + "lines": [ + { + "bbox": [ + 413, + 139, + 443, + 148 + ], + "spans": [ + { + "bbox": [ + 413, + 139, + 443, + 148 + ], + "type": "text", + "content": "integral" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 395, + 149, + 468, + 175 + ], + "lines": [ + { + "bbox": [ + 395, + 149, + 468, + 175 + ], + "spans": [ + { + "bbox": [ + 395, + 149, + 468, + 175 + ], + "type": "image", + "image_path": "5aa92e626be07ba7ea68daaa0da7fa03ab4d7b97ac6793aaf4bca22656e40fbb.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 178, + 443, + 186 + ], + "lines": [ + { + "bbox": [ + 413, + 178, + 443, + 186 + ], + "spans": [ + { + "bbox": [ + 413, + 178, + 443, + 186 + ], + "type": "text", + "content": "POWER" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 395, + 190, + 468, + 217 + ], + "blocks": [ + { + "bbox": [ + 395, + 190, + 468, + 217 + ], + "lines": [ + { + "bbox": [ + 395, + 190, + 468, + 217 + ], + "spans": [ + { + "bbox": [ + 395, + 190, + 468, + 217 + ], + "type": "image", + "image_path": "05e2c5bdd61e9db6780bbb6ae809a78fb703830a582f7813493c3c816fbcad7e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 220, + 442, + 230 + ], + "lines": [ + { + "bbox": [ + 413, + 220, + 442, + 230 + ], + "spans": [ + { + "bbox": [ + 413, + 220, + 442, + 230 + ], + "type": "text", + "content": "Display" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 235, + 462, + 244 + ], + "lines": [ + { + "bbox": [ + 400, + 235, + 462, + 244 + ], + "spans": [ + { + "bbox": [ + 400, + 235, + 462, + 244 + ], + "type": "text", + "content": "In-Context Prompt" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 470, + 72, + 543, + 99 + ], + "blocks": [ + { + "bbox": [ + 470, + 72, + 543, + 99 + ], + "lines": [ + { + "bbox": [ + 470, + 72, + 543, + 99 + ], + "spans": [ + { + "bbox": [ + 470, + 72, + 543, + 99 + ], + "type": "image", + "image_path": "6d2cee8497d2796ae8bf38b286327a2d9027d1cacc034f23d77ff1d545691feb.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 470, + 111, + 544, + 137 + ], + "blocks": [ + { + "bbox": [ + 501, + 102, + 519, + 110 + ], + "lines": [ + { + "bbox": [ + 501, + 102, + 519, + 110 + ], + "spans": [ + { + "bbox": [ + 501, + 102, + 519, + 110 + ], + "type": "text", + "content": "Vital" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 111, + 544, + 137 + ], + "lines": [ + { + "bbox": [ + 470, + 111, + 544, + 137 + ], + "spans": [ + { + "bbox": [ + 470, + 111, + 544, + 137 + ], + "type": "image", + "image_path": "d7eec6f5e561f97bc02e9350d0abd870baed38019a7f39d5560abd91aba403d4.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 470, + 150, + 544, + 175 + ], + "blocks": [ + { + "bbox": [ + 492, + 140, + 528, + 148 + ], + "lines": [ + { + "bbox": [ + 492, + 140, + 528, + 148 + ], + "spans": [ + { + "bbox": [ + 492, + 140, + 528, + 148 + ], + "type": "text", + "content": "accepting" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 150, + 544, + 175 + ], + "lines": [ + { + "bbox": [ + 470, + 150, + 544, + 175 + ], + "spans": [ + { + "bbox": [ + 470, + 150, + 544, + 175 + ], + "type": "image", + "image_path": "b71a7df6b4ff1935ea8ad117e3848d53a00ea68b7ef0eec185362c04798717bb.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 178, + 523, + 186 + ], + "lines": [ + { + "bbox": [ + 492, + 178, + 523, + 186 + ], + "spans": [ + { + "bbox": [ + 492, + 178, + 523, + 186 + ], + "type": "text", + "content": "PORTER" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 470, + 190, + 544, + 217 + ], + "blocks": [ + { + "bbox": [ + 470, + 190, + 544, + 217 + ], + "lines": [ + { + "bbox": [ + 470, + 190, + 544, + 217 + ], + "spans": [ + { + "bbox": [ + 470, + 190, + 544, + 217 + ], + "type": "image", + "image_path": "848adbcc35db8b1f350e1345d21d0b69074e306452934c3907f3608ffbd04d42.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 220, + 523, + 229 + ], + "lines": [ + { + "bbox": [ + 492, + 220, + 523, + 229 + ], + "spans": [ + { + "bbox": [ + 492, + 220, + 523, + 229 + ], + "type": "text", + "content": "simplest" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 473, + 235, + 543, + 243 + ], + "lines": [ + { + "bbox": [ + 473, + 235, + 543, + 243 + ], + "spans": [ + { + "bbox": [ + 473, + 235, + 543, + 243 + ], + "type": "text", + "content": "In-Context Prediction" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 343, + 546, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 343, + 546, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 343, + 546, + 380 + ], + "type": "text", + "content": "dictions due to misleading prompts (please refer to supplementary materials). Additionally, our model still could not recognize characters that are not included in the lexicon." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 306, + 392, + 378, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 378, + 405 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 378, + 405 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 413, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 593 + ], + "type": "text", + "content": "In this paper, we propose " + }, + { + "bbox": [ + 304, + 413, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 413, + 545, + 593 + ], + "type": "text", + "content": ", an ego-evolving scene text recognizer equipped with in-context learning capabilities. Through our proposed in-context training strategy incorporating context-rich scene text sequences, " + }, + { + "bbox": [ + 304, + 413, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 413, + 545, + 593 + ], + "type": "text", + "content": " performs rapid adaptation across diverse scenarios without additional fine-tuning. Extensive experiments demonstrate that " + }, + { + "bbox": [ + 304, + 413, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{STR}" + }, + { + "bbox": [ + 304, + 413, + 545, + 593 + ], + "type": "text", + "content": " not only achieves SOTA performance on common STR benchmarks but also outperforms even the approaches that have been fine-tuned specifically for cross-domain scenarios. The model's ability to easily and effectively handle difficult text cases further underscores its potential as a unified text recognizer. Overall, this research represents a significant step toward efficient and highly adaptive text recognition models well-suited for diverse real-world applications." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Acknowledgments This work is supported by National Natural Science Foundation of China No.62222602, No.62176092, No.62106075, No.U23A20343. Science and Technology Commission of Shanghai No.21511100700, Shanghai Sailing Program (23YF1410500), Natural Science Foundation of Shanghai (23ZR1420400), Natural Science Foundation of Chongqing, China (CSTB2023NSCQ-JQX0007), (CSTB2023NSCQ-MSX0137), Development Project of Ministry of Industry and Information Technology (ZTZB.23-990-016)." + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15574" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 224 + ], + "type": "text", + "content": "[2] Jeonghun Baek, Geewook Kim, Junyeop Lee, Sungrae Park, Dongyoon Han, Sangdoo Yun, Seong Joon Oh, and Hwalsuk Lee. What is wrong with scene text recognition model comparisons? dataset and model analysis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4715-4723, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "type": "text", + "content": "[3] Darwin Bautista and Rowel Atienza. Scene text recognition with permuted autoregressive sequence models. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 1, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 272, + 286, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 286, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 286, + 294 + ], + "type": "text", + "content": "[4] Belval. Generator. https://github.com/Belval/TextRecognitionDataGenerator.5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 296, + 286, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 286, + 350 + ], + "type": "text", + "content": "[5] Fedor Borisyuk, Albert Gordo, and Viswanath Sivakumar. Rosetta: Large scale system for text detection and recognition in images. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 71-79, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 353, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 353, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 353, + 286, + 407 + ], + "type": "text", + "content": "[6] 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, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 410, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 410, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 410, + 287, + 464 + ], + "type": "text", + "content": "[7] Zhanzhan Cheng, Fan Bai, Yunlu Xu, Gang Zheng, Shiliang Pu, and Shuigeng Zhou. Focusing attention: Towards accurate text recognition in natural images. In Proceedings of the IEEE international conference on computer vision, pages 5076-5084, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 466, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 286, + 520 + ], + "type": "text", + "content": "[8] Chee Kheng Ch'ng and Chee Seng Chan. Total-text: A comprehensive dataset for scene text detection and recognition. In 2017 14th IAPR international conference on document analysis and recognition (ICDAR), pages 935-942. IEEE, 2017. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 523, + 286, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 523, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 523, + 286, + 588 + ], + "type": "text", + "content": "[9] 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. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 286, + 656 + ], + "type": "text", + "content": "[10] Shancheng Fang, Hongtao Xie, Yuxin Wang, Zhendong Mao, and Yongdong Zhang. Read like humans: Autonomous, bidirectional and iterative language modeling for scene text recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7098-7107, 2021. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "type": "text", + "content": "[11] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, pages 369-376, 2006. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[12] Tongkun Guan, Chaochen Gu, Changsheng Lu, Jingzheng Tu, Qi Feng, Kaijie Wu, and Xinping Guan. Industrial scene text detection with refined feature-attentive network. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):6073–6085, 2022. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[13] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "text", + "content": "[14] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. Icl-d3ie: In-context learning with diverse demonstrations updating for document information extraction. arXiv preprint arXiv:2303.05063, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "text", + "content": "[15] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. Icl-d3ie: In-context learning with diverse demonstrations updating for document information extraction. arXiv preprint arXiv:2303.05063, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "text", + "content": "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 321, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 375 + ], + "type": "text", + "content": "[17] Qing Jiang, Jiapeng Wang, Dezhi Peng, Chongyu Liu, and Lianwen Jin. Revisiting scene text recognition: A data perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20543-20554, 2023. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 377, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 453 + ], + "type": "text", + "content": "[18] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazan Almazan, and Lluis Pere De Las Heras. Icedar 2013 robust reading competition. In 2013 12th international conference on document analysis and recognition, pages 1484-1493. IEEE, 2013. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 455, + 545, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 530 + ], + "type": "text", + "content": "[19] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In 2015 13th international conference on document analysis and recognition (ICDAR), pages 1156-1160. IEEE, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 533, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 588 + ], + "type": "text", + "content": "[20] Junyeop Lee, Sungrae Park, Jeonghun Baek, Seong Joon Oh, Seonghyeon Kim, and Hwalsuk Lee. On recognizing texts of arbitrary shapes with 2d self-attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 546–547, 2020. 1, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "type": "text", + "content": "[21] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[22] Hui Li, Peng Wang, Chunhua Shen, and Guyu Zhang. Show, attend and read: A simple and strong baseline for irregular text recognition. In Proceedings of the AAAI conference on artificial intelligence, pages 8610-8617, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[23] Minghui Liao, Jian Zhang, Zhaoyi Wan, Fengming Xie, Jia-jun Liang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Scene text recognition from two-dimensional perspective. In Pro" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15575" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "ceedings of the AAAI conference on artificial intelligence, pages 8714-8721, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[24] Wei Liu, Chaofeng Chen, Kwan-Yee K Wong, Zhizhong Su, and Junyu Han. Star-net: a spatial attention residue network for scene text recognition. In BMVC, page 7, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 173 + ], + "type": "text", + "content": "[25] Yuliang Liu, Lianwen Jin, Shuai tao Zhang, Canjie Luo, and Sheng Zhang. Curved scene text detection via transverse and longitudinal sequence connection. Pattern Recognition, 90: 337-345, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "type": "text", + "content": "[26] Mengkai Ma, Qiu-Feng Wang, Shan Huang, Shen Huang, Yannis Goulermas, and Kaizhu Huang. Residual attention-based multi-scale script identification in scene text images. Neurocomputing, 421:222-233, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "[27] U-V Marti and Horst Bunke. The iam-database: an english sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition, 5: 39-46, 2002. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[28] Qiang Mei, Qinyou Hu, Chun Yang, Hailin Zheng, and Zhisheng Hu. Port recommendation system for alternative container port destinations using a novel neural language-based algorithm. IEEE Access, 8:199970-199979, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "type": "text", + "content": "[29] Anand Mishra, Karteek Alahari, and CV Jawahar. Top-down and bottom-up cues for scene text recognition. In 2012 IEEE conference on computer vision and pattern recognition, pages 2687-2694. IEEE, 2012. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "text", + "content": "[30] Imene OUALI, Mohamed BEN HALIMA, and WALI Ali. Augmented reality for scene text recognition, visualization and reading to assist visually impaired people. Procedia Computer Science, 207:158-167, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 400, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 287, + 453 + ], + "type": "text", + "content": "[31] Trung Quy Phan, Palaiahnakote Shivakumara, Shangxuan Tian, and Chew Lim Tan. Recognizing text with perspective distortion in natural scenes. In Proceedings of the IEEE international conference on computer vision, pages 569-576, 2013. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "type": "text", + "content": "[32] Anhar Risnumawan, Palaiahankote Shivakumara, Chee Seng Chan, and Chew Lim Tan. A robust arbitrary text detection system for natural scene images. Expert Systems with Applications, 41(18):8027-8048, 2014. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "type": "text", + "content": "[33] Abdul Khader Jilani Saudagar and HabeebVulla Mohammad. Augmented reality mobile application for arabic text extraction, recognition and translation. Journal of Statistics and Management Systems, 21(4):617-629, 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 601 + ], + "type": "text", + "content": "[34] Zhenwei Shao, Zhou Yu, Meng Wang, and Jun Yu. Prompting large language models with answer heuristics for knowledge-based visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14974-14983, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "text", + "content": "[35] Fenfen Sheng, Zhineng Chen, and Bo Xu. Ntrr: A no-recurrence sequence-to-sequence model for scene text recognition. In 2019 International conference on document analysis and recognition (ICDAR), pages 781-786. IEEE, 2019. 1, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[36] Baoguang Shi, Mingkun Yang, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Aster: An attentional scene text recognizer with flexible rectification. IEEE transactions on pattern analysis and machine intelligence, 41(9):2035-2048, 2018. 3, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 662 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[37] Cunzhao Shi, Chunheng Wang, Baihua Xiao, Song Gao, and Jinlong Hu. End-to-end scene text recognition using tree-structured models. Pattern Recognition, 47(9):2853-2866, 2014. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[38] 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. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[39] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie. Coco-text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 260 + ], + "type": "text", + "content": "[40] Zhaoyi Wan, Minghang He, Haoran Chen, Xiang Bai, and Cong Yao. Textscanner: Reading characters in order for robust scene text recognition. In Proceedings of the AAAI conference on artificial intelligence, pages 12120-12127, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 262, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 545, + 327 + ], + "type": "text", + "content": "[41] Yuxin Wang, Hongtao Xie, Shancheng Fang, Jing Wang, Shenggao Zhu, and Yongdong Zhang. From two to one: A new scene text recognizer with visual language modeling network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14194–14203, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 329, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 329, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 329, + 545, + 384 + ], + "type": "text", + "content": "[42] Xudong Xie, Ling Fu, Zhifei Zhang, Zhaowen Wang, and Xiang Bai. Toward understanding wordart: Corner-guided transformer for scene text recognition. In European Conference on Computer Vision, pages 303-321. Springer, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 386, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 386, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 386, + 545, + 441 + ], + "type": "text", + "content": "[43] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. An empirical study of gpt-3 for few-shot knowledge-based vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3081-3089, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 441, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 496 + ], + "type": "text", + "content": "[44] Deli Yu, Xuan Li, Chengquan Zhang, Tao Liu, Junyu Han, Jingtuo Liu, and Errui Ding. Towards accurate scene text recognition with semantic reasoning networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12113-12122, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 498, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 552 + ], + "type": "text", + "content": "[45] Chongsheng Zhang, Yuefeng Tao, Kai Du, Weiping Ding, Bin Wang, Ji Liu, and Wei Wang. Character-level street view text spotting based on deep multisegmentation network for smarter autonomous driving. IEEE Transactions on Artificial Intelligence, 3(2):297-308, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 553, + 545, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 553, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 307, + 553, + 545, + 606 + ], + "type": "text", + "content": "[46] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 609, + 545, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 609, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 307, + 609, + 545, + 662 + ], + "type": "text", + "content": "[47] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15576" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-modal Instruction Tuned LLMs with Fine-grained Visual Perception/1060b587-378f-4876-81ef-a69e77b5ca70_content_list.json b/2024/Multi-modal Instruction Tuned LLMs with Fine-grained Visual Perception/1060b587-378f-4876-81ef-a69e77b5ca70_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1bba468f818deac323ed05bc22bb73798b27f352 --- /dev/null +++ b/2024/Multi-modal Instruction Tuned LLMs with Fine-grained Visual Perception/1060b587-378f-4876-81ef-a69e77b5ca70_content_list.json @@ -0,0 +1,1705 @@ +[ + { + "type": "text", + "text": "Multi-modal Instruction Tuned LLMs with Fine-grained Visual Perception", + "text_level": 1, + "bbox": [ + 107, + 128, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junwen $\\mathrm{He}^{*1,2}$ , Yifan Wang $^{1}$ , Lijun Wang $^{\\dagger 1}$ , Huchuan Lu $^{1}$ , Jun-Yan He $^{2}$ , Jin-Peng Lan $^{2}$ , Bin Luo $^{2}$ , and Xuansong Xie $^{2}$", + "bbox": [ + 197, + 191, + 769, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dalian University of Technology", + "bbox": [ + 346, + 239, + 620, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ DAMO Academy, Alibaba Group", + "bbox": [ + 346, + 260, + 622, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "junwen.he@mail.dlut.edu.cn, {wyfan, ljwang, lhchuan}@dlut.edu.cn {leyuan.hjy, lanjinpeng.ljp, luwu.lb}@alibaba-inc.com, xingtong.xxs@taobao.com", + "bbox": [ + 140, + 276, + 828, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 349, + 313, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal Large Language Model (MLLMs) leverages Large Language Models as a cognitive framework for diverse visual-language tasks. Recent efforts have been made to equip MLLMs with visual perceiving and grounding capabilities. However, there still remains a gap in providing fine-grained pixel-level perceptions and extending interactions beyond text-specific inputs. In this work, we propose AnyRef, a general MLLM model that can generate pixel-wise object perceptions and natural language descriptions from multi-modality references, such as texts, boxes, images, or audio. This innovation empowers users with greater flexibility to engage with the model beyond textual and regional prompts, without modality-specific designs. Through our proposed refocusing mechanism, the generated grounding output is guided to better focus on the referenced object, implicitly incorporating additional pixel-level supervision. This simple modification utilizes attention scores generated during the inference of LLM, eliminating the need for extra computations while exhibiting performance enhancements in both grounding masks and referring expressions. With only publicly available training data, our model achieves state-of-the-art results across multiple benchmarks, including diverse modality referencing segmentation and region-level referring expression generation. Code and models are available at https://github.com/jwh97nn/AnyRef", + "bbox": [ + 75, + 381, + 473, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 792, + 209, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs) have garnered widespread influence across various domains, and advancements have been achieved by augmenting LLMs with visual percep", + "bbox": [ + 75, + 818, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6d499296d1281e16da11b1d354f12ee3ecb9255cd417d9f998dd4b892352937.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 352, + 532, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Can you describe $\\{\\}$ in this image and generate its segmentation mask.", + "bbox": [ + 539, + 356, + 879, + 385 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 555, + 390, + 601, + 402 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 633, + 388, + 673, + 401 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 715, + 388, + 769, + 402 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "